cortex: thesis/cortex.org annotate

annotate thesis/cortex.org @ 482:074eadc919fe

author	Robert McIntyre <rlm@mit.edu>
date	Fri, 28 Mar 2014 23:33:22 -0400
parents	6e68720e1c13
children	3046d963ec1a

rev	line source
rlm@425	1 #+title: =CORTEX=
rlm@425	2 #+author: Robert McIntyre
rlm@425	3 #+email: rlm@mit.edu
rlm@425	4 #+description: Using embodied AI to facilitate Artificial Imagination.
rlm@425	5 #+keywords: AI, clojure, embodiment
rlm@451	6 #+LaTeX_CLASS_OPTIONS: [nofloat]
rlm@422	7
rlm@465	8 * COMMENT templates
rlm@470	9 #+caption:
rlm@470	10 #+caption:
rlm@470	11 #+caption:
rlm@470	12 #+caption:
rlm@470	13 #+name: name
rlm@470	14 #+begin_listing clojure
rlm@479	15 #+BEGIN_SRC clojure
rlm@479	16 #+END_SRC
rlm@470	17 #+end_listing
rlm@465	18
rlm@470	19 #+caption:
rlm@470	20 #+caption:
rlm@470	21 #+caption:
rlm@470	22 #+name: name
rlm@470	23 #+ATTR_LaTeX: :width 10cm
rlm@470	24 [[./images/aurellem-gray.png]]
rlm@470	25
rlm@470	26 #+caption:
rlm@470	27 #+caption:
rlm@470	28 #+caption:
rlm@470	29 #+caption:
rlm@470	30 #+name: name
rlm@470	31 #+begin_listing clojure
rlm@475	32 #+BEGIN_SRC clojure
rlm@475	33 #+END_SRC
rlm@470	34 #+end_listing
rlm@470	35
rlm@470	36 #+caption:
rlm@470	37 #+caption:
rlm@470	38 #+caption:
rlm@470	39 #+name: name
rlm@470	40 #+ATTR_LaTeX: :width 10cm
rlm@470	41 [[./images/aurellem-gray.png]]
rlm@470	42
rlm@465	43
rlm@465	44 * COMMENT Empathy and Embodiment as problem solving strategies
rlm@437	45
rlm@437	46 By the end of this thesis, you will have seen a novel approach to
rlm@437	47 interpreting video using embodiment and empathy. You will have also
rlm@437	48 seen one way to efficiently implement empathy for embodied
rlm@447	49 creatures. Finally, you will become familiar with =CORTEX=, a system
rlm@447	50 for designing and simulating creatures with rich senses, which you
rlm@447	51 may choose to use in your own research.
rlm@437	52
rlm@441	53 This is the core vision of my thesis: That one of the important ways
rlm@441	54 in which we understand others is by imagining ourselves in their
rlm@441	55 position and emphatically feeling experiences relative to our own
rlm@441	56 bodies. By understanding events in terms of our own previous
rlm@441	57 corporeal experience, we greatly constrain the possibilities of what
rlm@441	58 would otherwise be an unwieldy exponential search. This extra
rlm@441	59 constraint can be the difference between easily understanding what
rlm@441	60 is happening in a video and being completely lost in a sea of
rlm@441	61 incomprehensible color and movement.
rlm@435	62
rlm@436	63 ** Recognizing actions in video is extremely difficult
rlm@437	64
rlm@447	65 Consider for example the problem of determining what is happening
rlm@447	66 in a video of which this is one frame:
rlm@437	67
rlm@441	68 #+caption: A cat drinking some water. Identifying this action is
rlm@441	69 #+caption: beyond the state of the art for computers.
rlm@441	70 #+ATTR_LaTeX: :width 7cm
rlm@441	71 [[./images/cat-drinking.jpg]]
rlm@441	72
rlm@441	73 It is currently impossible for any computer program to reliably
rlm@447	74 label such a video as ``drinking''. And rightly so -- it is a very
rlm@441	75 hard problem! What features can you describe in terms of low level
rlm@441	76 functions of pixels that can even begin to describe at a high level
rlm@441	77 what is happening here?
rlm@437	78
rlm@447	79 Or suppose that you are building a program that recognizes chairs.
rlm@448	80 How could you ``see'' the chair in figure \ref{hidden-chair}?
rlm@441	81
rlm@441	82 #+caption: The chair in this image is quite obvious to humans, but I
rlm@448	83 #+caption: doubt that any modern computer vision program can find it.
rlm@441	84 #+name: hidden-chair
rlm@441	85 #+ATTR_LaTeX: :width 10cm
rlm@441	86 [[./images/fat-person-sitting-at-desk.jpg]]
rlm@441	87
rlm@441	88 Finally, how is it that you can easily tell the difference between
rlm@441	89 how the girls /muscles/ are working in figure \ref{girl}?
rlm@441	90
rlm@441	91 #+caption: The mysterious ``common sense'' appears here as you are able
rlm@441	92 #+caption: to discern the difference in how the girl's arm muscles
rlm@441	93 #+caption: are activated between the two images.
rlm@441	94 #+name: girl
rlm@448	95 #+ATTR_LaTeX: :width 7cm
rlm@441	96 [[./images/wall-push.png]]
rlm@437	97
rlm@441	98 Each of these examples tells us something about what might be going
rlm@441	99 on in our minds as we easily solve these recognition problems.
rlm@441	100
rlm@441	101 The hidden chairs show us that we are strongly triggered by cues
rlm@447	102 relating to the position of human bodies, and that we can determine
rlm@447	103 the overall physical configuration of a human body even if much of
rlm@447	104 that body is occluded.
rlm@437	105
rlm@441	106 The picture of the girl pushing against the wall tells us that we
rlm@441	107 have common sense knowledge about the kinetics of our own bodies.
rlm@441	108 We know well how our muscles would have to work to maintain us in
rlm@441	109 most positions, and we can easily project this self-knowledge to
rlm@441	110 imagined positions triggered by images of the human body.
rlm@441	111
rlm@441	112 ** =EMPATH= neatly solves recognition problems
rlm@441	113
rlm@441	114 I propose a system that can express the types of recognition
rlm@441	115 problems above in a form amenable to computation. It is split into
rlm@441	116 four parts:
rlm@441	117
rlm@448	118 - Free/Guided Play :: The creature moves around and experiences the
rlm@448	119 world through its unique perspective. Many otherwise
rlm@448	120 complicated actions are easily described in the language of a
rlm@448	121 full suite of body-centered, rich senses. For example,
rlm@448	122 drinking is the feeling of water sliding down your throat, and
rlm@448	123 cooling your insides. It's often accompanied by bringing your
rlm@448	124 hand close to your face, or bringing your face close to water.
rlm@448	125 Sitting down is the feeling of bending your knees, activating
rlm@448	126 your quadriceps, then feeling a surface with your bottom and
rlm@448	127 relaxing your legs. These body-centered action descriptions
rlm@448	128 can be either learned or hard coded.
rlm@448	129 - Posture Imitation :: When trying to interpret a video or image,
rlm@448	130 the creature takes a model of itself and aligns it with
rlm@448	131 whatever it sees. This alignment can even cross species, as
rlm@448	132 when humans try to align themselves with things like ponies,
rlm@448	133 dogs, or other humans with a different body type.
rlm@448	134 - Empathy :: The alignment triggers associations with
rlm@448	135 sensory data from prior experiences. For example, the
rlm@448	136 alignment itself easily maps to proprioceptive data. Any
rlm@448	137 sounds or obvious skin contact in the video can to a lesser
rlm@448	138 extent trigger previous experience. Segments of previous
rlm@448	139 experiences are stitched together to form a coherent and
rlm@448	140 complete sensory portrait of the scene.
rlm@448	141 - Recognition :: With the scene described in terms of first
rlm@448	142 person sensory events, the creature can now run its
rlm@447	143 action-identification programs on this synthesized sensory
rlm@447	144 data, just as it would if it were actually experiencing the
rlm@447	145 scene first-hand. If previous experience has been accurately
rlm@447	146 retrieved, and if it is analogous enough to the scene, then
rlm@447	147 the creature will correctly identify the action in the scene.
rlm@447	148
rlm@441	149 For example, I think humans are able to label the cat video as
rlm@447	150 ``drinking'' because they imagine /themselves/ as the cat, and
rlm@441	151 imagine putting their face up against a stream of water and
rlm@441	152 sticking out their tongue. In that imagined world, they can feel
rlm@441	153 the cool water hitting their tongue, and feel the water entering
rlm@447	154 their body, and are able to recognize that /feeling/ as drinking.
rlm@447	155 So, the label of the action is not really in the pixels of the
rlm@447	156 image, but is found clearly in a simulation inspired by those
rlm@447	157 pixels. An imaginative system, having been trained on drinking and
rlm@447	158 non-drinking examples and learning that the most important
rlm@447	159 component of drinking is the feeling of water sliding down one's
rlm@447	160 throat, would analyze a video of a cat drinking in the following
rlm@447	161 manner:
rlm@441	162
rlm@447	163 1. Create a physical model of the video by putting a ``fuzzy''
rlm@447	164 model of its own body in place of the cat. Possibly also create
rlm@447	165 a simulation of the stream of water.
rlm@441	166
rlm@441	167 2. Play out this simulated scene and generate imagined sensory
rlm@441	168 experience. This will include relevant muscle contractions, a
rlm@441	169 close up view of the stream from the cat's perspective, and most
rlm@441	170 importantly, the imagined feeling of water entering the
rlm@443	171 mouth. The imagined sensory experience can come from a
rlm@441	172 simulation of the event, but can also be pattern-matched from
rlm@441	173 previous, similar embodied experience.
rlm@441	174
rlm@441	175 3. The action is now easily identified as drinking by the sense of
rlm@441	176 taste alone. The other senses (such as the tongue moving in and
rlm@441	177 out) help to give plausibility to the simulated action. Note that
rlm@441	178 the sense of vision, while critical in creating the simulation,
rlm@441	179 is not critical for identifying the action from the simulation.
rlm@441	180
rlm@441	181 For the chair examples, the process is even easier:
rlm@441	182
rlm@441	183 1. Align a model of your body to the person in the image.
rlm@441	184
rlm@441	185 2. Generate proprioceptive sensory data from this alignment.
rlm@437	186
rlm@441	187 3. Use the imagined proprioceptive data as a key to lookup related
rlm@441	188 sensory experience associated with that particular proproceptive
rlm@441	189 feeling.
rlm@437	190
rlm@443	191 4. Retrieve the feeling of your bottom resting on a surface, your
rlm@443	192 knees bent, and your leg muscles relaxed.
rlm@437	193
rlm@441	194 5. This sensory information is consistent with the =sitting?=
rlm@441	195 sensory predicate, so you (and the entity in the image) must be
rlm@441	196 sitting.
rlm@440	197
rlm@441	198 6. There must be a chair-like object since you are sitting.
rlm@440	199
rlm@441	200 Empathy offers yet another alternative to the age-old AI
rlm@441	201 representation question: ``What is a chair?'' --- A chair is the
rlm@441	202 feeling of sitting.
rlm@441	203
rlm@441	204 My program, =EMPATH= uses this empathic problem solving technique
rlm@441	205 to interpret the actions of a simple, worm-like creature.
rlm@437	206
rlm@441	207 #+caption: The worm performs many actions during free play such as
rlm@441	208 #+caption: curling, wiggling, and resting.
rlm@441	209 #+name: worm-intro
rlm@446	210 #+ATTR_LaTeX: :width 15cm
rlm@445	211 [[./images/worm-intro-white.png]]
rlm@437	212
rlm@462	213 #+caption: =EMPATH= recognized and classified each of these
rlm@462	214 #+caption: poses by inferring the complete sensory experience
rlm@462	215 #+caption: from proprioceptive data.
rlm@441	216 #+name: worm-recognition-intro
rlm@446	217 #+ATTR_LaTeX: :width 15cm
rlm@445	218 [[./images/worm-poses.png]]
rlm@441	219
rlm@441	220 One powerful advantage of empathic problem solving is that it
rlm@441	221 factors the action recognition problem into two easier problems. To
rlm@441	222 use empathy, you need an /aligner/, which takes the video and a
rlm@441	223 model of your body, and aligns the model with the video. Then, you
rlm@441	224 need a /recognizer/, which uses the aligned model to interpret the
rlm@441	225 action. The power in this method lies in the fact that you describe
rlm@448	226 all actions form a body-centered viewpoint. You are less tied to
rlm@447	227 the particulars of any visual representation of the actions. If you
rlm@441	228 teach the system what ``running'' is, and you have a good enough
rlm@441	229 aligner, the system will from then on be able to recognize running
rlm@441	230 from any point of view, even strange points of view like above or
rlm@441	231 underneath the runner. This is in contrast to action recognition
rlm@448	232 schemes that try to identify actions using a non-embodied approach.
rlm@448	233 If these systems learn about running as viewed from the side, they
rlm@448	234 will not automatically be able to recognize running from any other
rlm@448	235 viewpoint.
rlm@441	236
rlm@441	237 Another powerful advantage is that using the language of multiple
rlm@441	238 body-centered rich senses to describe body-centerd actions offers a
rlm@441	239 massive boost in descriptive capability. Consider how difficult it
rlm@441	240 would be to compose a set of HOG filters to describe the action of
rlm@447	241 a simple worm-creature ``curling'' so that its head touches its
rlm@447	242 tail, and then behold the simplicity of describing thus action in a
rlm@441	243 language designed for the task (listing \ref{grand-circle-intro}):
rlm@441	244
rlm@446	245 #+caption: Body-centerd actions are best expressed in a body-centered
rlm@446	246 #+caption: language. This code detects when the worm has curled into a
rlm@446	247 #+caption: full circle. Imagine how you would replicate this functionality
rlm@446	248 #+caption: using low-level pixel features such as HOG filters!
rlm@446	249 #+name: grand-circle-intro
rlm@452	250 #+attr_latex: [htpb]
rlm@452	251 #+begin_listing clojure
rlm@446	252 #+begin_src clojure
rlm@446	253 (defn grand-circle?
rlm@446	254 "Does the worm form a majestic circle (one end touching the other)?"
rlm@446	255 [experiences]
rlm@446	256 (and (curled? experiences)
rlm@446	257 (let [worm-touch (:touch (peek experiences))
rlm@446	258 tail-touch (worm-touch 0)
rlm@446	259 head-touch (worm-touch 4)]
rlm@462	260 (and (< 0.2 (contact worm-segment-bottom-tip tail-touch))
rlm@462	261 (< 0.2 (contact worm-segment-top-tip head-touch))))))
rlm@446	262 #+end_src
rlm@446	263 #+end_listing
rlm@446	264
rlm@449	265 ** =CORTEX= is a toolkit for building sensate creatures
rlm@435	266
rlm@448	267 I built =CORTEX= to be a general AI research platform for doing
rlm@448	268 experiments involving multiple rich senses and a wide variety and
rlm@448	269 number of creatures. I intend it to be useful as a library for many
rlm@462	270 more projects than just this thesis. =CORTEX= was necessary to meet
rlm@462	271 a need among AI researchers at CSAIL and beyond, which is that
rlm@462	272 people often will invent neat ideas that are best expressed in the
rlm@448	273 language of creatures and senses, but in order to explore those
rlm@448	274 ideas they must first build a platform in which they can create
rlm@448	275 simulated creatures with rich senses! There are many ideas that
rlm@448	276 would be simple to execute (such as =EMPATH=), but attached to them
rlm@448	277 is the multi-month effort to make a good creature simulator. Often,
rlm@448	278 that initial investment of time proves to be too much, and the
rlm@448	279 project must make do with a lesser environment.
rlm@435	280
rlm@448	281 =CORTEX= is well suited as an environment for embodied AI research
rlm@448	282 for three reasons:
rlm@448	283
rlm@448	284 - You can create new creatures using Blender, a popular 3D modeling
rlm@448	285 program. Each sense can be specified using special blender nodes
rlm@448	286 with biologically inspired paramaters. You need not write any
rlm@448	287 code to create a creature, and can use a wide library of
rlm@448	288 pre-existing blender models as a base for your own creatures.
rlm@448	289
rlm@448	290 - =CORTEX= implements a wide variety of senses, including touch,
rlm@448	291 proprioception, vision, hearing, and muscle tension. Complicated
rlm@448	292 senses like touch, and vision involve multiple sensory elements
rlm@448	293 embedded in a 2D surface. You have complete control over the
rlm@448	294 distribution of these sensor elements through the use of simple
rlm@448	295 png image files. In particular, =CORTEX= implements more
rlm@448	296 comprehensive hearing than any other creature simulation system
rlm@448	297 available.
rlm@448	298
rlm@448	299 - =CORTEX= supports any number of creatures and any number of
rlm@448	300 senses. Time in =CORTEX= dialates so that the simulated creatures
rlm@448	301 always precieve a perfectly smooth flow of time, regardless of
rlm@448	302 the actual computational load.
rlm@448	303
rlm@448	304 =CORTEX= is built on top of =jMonkeyEngine3=, which is a video game
rlm@448	305 engine designed to create cross-platform 3D desktop games. =CORTEX=
rlm@448	306 is mainly written in clojure, a dialect of =LISP= that runs on the
rlm@448	307 java virtual machine (JVM). The API for creating and simulating
rlm@449	308 creatures and senses is entirely expressed in clojure, though many
rlm@449	309 senses are implemented at the layer of jMonkeyEngine or below. For
rlm@449	310 example, for the sense of hearing I use a layer of clojure code on
rlm@449	311 top of a layer of java JNI bindings that drive a layer of =C++=
rlm@449	312 code which implements a modified version of =OpenAL= to support
rlm@449	313 multiple listeners. =CORTEX= is the only simulation environment
rlm@449	314 that I know of that can support multiple entities that can each
rlm@449	315 hear the world from their own perspective. Other senses also
rlm@449	316 require a small layer of Java code. =CORTEX= also uses =bullet=, a
rlm@449	317 physics simulator written in =C=.
rlm@448	318
rlm@448	319 #+caption: Here is the worm from above modeled in Blender, a free
rlm@448	320 #+caption: 3D-modeling program. Senses and joints are described
rlm@448	321 #+caption: using special nodes in Blender.
rlm@448	322 #+name: worm-recognition-intro
rlm@448	323 #+ATTR_LaTeX: :width 12cm
rlm@448	324 [[./images/blender-worm.png]]
rlm@448	325
rlm@449	326 Here are some thing I anticipate that =CORTEX= might be used for:
rlm@449	327
rlm@449	328 - exploring new ideas about sensory integration
rlm@449	329 - distributed communication among swarm creatures
rlm@449	330 - self-learning using free exploration,
rlm@449	331 - evolutionary algorithms involving creature construction
rlm@449	332 - exploration of exoitic senses and effectors that are not possible
rlm@449	333 in the real world (such as telekenisis or a semantic sense)
rlm@449	334 - imagination using subworlds
rlm@449	335
rlm@451	336 During one test with =CORTEX=, I created 3,000 creatures each with
rlm@448	337 their own independent senses and ran them all at only 1/80 real
rlm@448	338 time. In another test, I created a detailed model of my own hand,
rlm@448	339 equipped with a realistic distribution of touch (more sensitive at
rlm@448	340 the fingertips), as well as eyes and ears, and it ran at around 1/4
rlm@451	341 real time.
rlm@448	342
rlm@451	343 #+BEGIN_LaTeX
rlm@449	344 \begin{sidewaysfigure}
rlm@449	345 \includegraphics[width=9.5in]{images/full-hand.png}
rlm@451	346 \caption{
rlm@451	347 I modeled my own right hand in Blender and rigged it with all the
rlm@451	348 senses that {\tt CORTEX} supports. My simulated hand has a
rlm@451	349 biologically inspired distribution of touch sensors. The senses are
rlm@451	350 displayed on the right, and the simulation is displayed on the
rlm@451	351 left. Notice that my hand is curling its fingers, that it can see
rlm@451	352 its own finger from the eye in its palm, and that it can feel its
rlm@451	353 own thumb touching its palm.}
rlm@449	354 \end{sidewaysfigure}
rlm@451	355 #+END_LaTeX
rlm@448	356
rlm@437	357 ** Contributions
rlm@435	358
rlm@451	359 - I built =CORTEX=, a comprehensive platform for embodied AI
rlm@451	360 experiments. =CORTEX= supports many features lacking in other
rlm@451	361 systems, such proper simulation of hearing. It is easy to create
rlm@451	362 new =CORTEX= creatures using Blender, a free 3D modeling program.
rlm@449	363
rlm@451	364 - I built =EMPATH=, which uses =CORTEX= to identify the actions of
rlm@451	365 a worm-like creature using a computational model of empathy.
rlm@449	366
rlm@436	367 * Building =CORTEX=
rlm@435	368
rlm@462	369 I intend for =CORTEX= to be used as a general purpose library for
rlm@462	370 building creatures and outfitting them with senses, so that it will
rlm@462	371 be useful for other researchers who want to test out ideas of their
rlm@462	372 own. To this end, wherver I have had to make archetictural choices
rlm@462	373 about =CORTEX=, I have chosen to give as much freedom to the user as
rlm@462	374 possible, so that =CORTEX= may be used for things I have not
rlm@462	375 forseen.
rlm@462	376
rlm@465	377 ** COMMENT Simulation or Reality?
rlm@462	378
rlm@462	379 The most important archetictural decision of all is the choice to
rlm@462	380 use a computer-simulated environemnt in the first place! The world
rlm@462	381 is a vast and rich place, and for now simulations are a very poor
rlm@462	382 reflection of its complexity. It may be that there is a significant
rlm@462	383 qualatative difference between dealing with senses in the real
rlm@468	384 world and dealing with pale facilimilies of them in a simulation.
rlm@468	385 What are the advantages and disadvantages of a simulation vs.
rlm@468	386 reality?
rlm@462	387
rlm@462	388 *** Simulation
rlm@462	389
rlm@462	390 The advantages of virtual reality are that when everything is a
rlm@462	391 simulation, experiments in that simulation are absolutely
rlm@462	392 reproducible. It's also easier to change the character and world
rlm@462	393 to explore new situations and different sensory combinations.
rlm@462	394
rlm@462	395 If the world is to be simulated on a computer, then not only do
rlm@462	396 you have to worry about whether the character's senses are rich
rlm@462	397 enough to learn from the world, but whether the world itself is
rlm@462	398 rendered with enough detail and realism to give enough working
rlm@462	399 material to the character's senses. To name just a few
rlm@462	400 difficulties facing modern physics simulators: destructibility of
rlm@462	401 the environment, simulation of water/other fluids, large areas,
rlm@462	402 nonrigid bodies, lots of objects, smoke. I don't know of any
rlm@462	403 computer simulation that would allow a character to take a rock
rlm@462	404 and grind it into fine dust, then use that dust to make a clay
rlm@462	405 sculpture, at least not without spending years calculating the
rlm@462	406 interactions of every single small grain of dust. Maybe a
rlm@462	407 simulated world with today's limitations doesn't provide enough
rlm@462	408 richness for real intelligence to evolve.
rlm@462	409
rlm@462	410 *** Reality
rlm@462	411
rlm@462	412 The other approach for playing with senses is to hook your
rlm@462	413 software up to real cameras, microphones, robots, etc., and let it
rlm@462	414 loose in the real world. This has the advantage of eliminating
rlm@462	415 concerns about simulating the world at the expense of increasing
rlm@462	416 the complexity of implementing the senses. Instead of just
rlm@462	417 grabbing the current rendered frame for processing, you have to
rlm@462	418 use an actual camera with real lenses and interact with photons to
rlm@462	419 get an image. It is much harder to change the character, which is
rlm@462	420 now partly a physical robot of some sort, since doing so involves
rlm@462	421 changing things around in the real world instead of modifying
rlm@462	422 lines of code. While the real world is very rich and definitely
rlm@462	423 provides enough stimulation for intelligence to develop as
rlm@462	424 evidenced by our own existence, it is also uncontrollable in the
rlm@462	425 sense that a particular situation cannot be recreated perfectly or
rlm@462	426 saved for later use. It is harder to conduct science because it is
rlm@462	427 harder to repeat an experiment. The worst thing about using the
rlm@462	428 real world instead of a simulation is the matter of time. Instead
rlm@462	429 of simulated time you get the constant and unstoppable flow of
rlm@462	430 real time. This severely limits the sorts of software you can use
rlm@462	431 to program the AI because all sense inputs must be handled in real
rlm@462	432 time. Complicated ideas may have to be implemented in hardware or
rlm@462	433 may simply be impossible given the current speed of our
rlm@462	434 processors. Contrast this with a simulation, in which the flow of
rlm@462	435 time in the simulated world can be slowed down to accommodate the
rlm@462	436 limitations of the character's programming. In terms of cost,
rlm@462	437 doing everything in software is far cheaper than building custom
rlm@462	438 real-time hardware. All you need is a laptop and some patience.
rlm@435	439
rlm@465	440 ** COMMENT Because of Time, simulation is perferable to reality
rlm@435	441
rlm@462	442 I envision =CORTEX= being used to support rapid prototyping and
rlm@462	443 iteration of ideas. Even if I could put together a well constructed
rlm@462	444 kit for creating robots, it would still not be enough because of
rlm@462	445 the scourge of real-time processing. Anyone who wants to test their
rlm@462	446 ideas in the real world must always worry about getting their
rlm@465	447 algorithms to run fast enough to process information in real time.
rlm@465	448 The need for real time processing only increases if multiple senses
rlm@465	449 are involved. In the extreme case, even simple algorithms will have
rlm@465	450 to be accelerated by ASIC chips or FPGAs, turning what would
rlm@465	451 otherwise be a few lines of code and a 10x speed penality into a
rlm@465	452 multi-month ordeal. For this reason, =CORTEX= supports
rlm@462	453 /time-dialiation/, which scales back the framerate of the
rlm@465	454 simulation in proportion to the amount of processing each frame.
rlm@465	455 From the perspective of the creatures inside the simulation, time
rlm@465	456 always appears to flow at a constant rate, regardless of how
rlm@462	457 complicated the envorimnent becomes or how many creatures are in
rlm@462	458 the simulation. The cost is that =CORTEX= can sometimes run slower
rlm@462	459 than real time. This can also be an advantage, however ---
rlm@462	460 simulations of very simple creatures in =CORTEX= generally run at
rlm@462	461 40x on my machine!
rlm@462	462
rlm@469	463 ** COMMENT What is a sense?
rlm@468	464
rlm@468	465 If =CORTEX= is to support a wide variety of senses, it would help
rlm@468	466 to have a better understanding of what a ``sense'' actually is!
rlm@468	467 While vision, touch, and hearing all seem like they are quite
rlm@468	468 different things, I was supprised to learn during the course of
rlm@468	469 this thesis that they (and all physical senses) can be expressed as
rlm@468	470 exactly the same mathematical object due to a dimensional argument!
rlm@468	471
rlm@468	472 Human beings are three-dimensional objects, and the nerves that
rlm@468	473 transmit data from our various sense organs to our brain are
rlm@468	474 essentially one-dimensional. This leaves up to two dimensions in
rlm@468	475 which our sensory information may flow. For example, imagine your
rlm@468	476 skin: it is a two-dimensional surface around a three-dimensional
rlm@468	477 object (your body). It has discrete touch sensors embedded at
rlm@468	478 various points, and the density of these sensors corresponds to the
rlm@468	479 sensitivity of that region of skin. Each touch sensor connects to a
rlm@468	480 nerve, all of which eventually are bundled together as they travel
rlm@468	481 up the spinal cord to the brain. Intersect the spinal nerves with a
rlm@468	482 guillotining plane and you will see all of the sensory data of the
rlm@468	483 skin revealed in a roughly circular two-dimensional image which is
rlm@468	484 the cross section of the spinal cord. Points on this image that are
rlm@468	485 close together in this circle represent touch sensors that are
rlm@468	486 /probably/ close together on the skin, although there is of course
rlm@468	487 some cutting and rearrangement that has to be done to transfer the
rlm@468	488 complicated surface of the skin onto a two dimensional image.
rlm@468	489
rlm@468	490 Most human senses consist of many discrete sensors of various
rlm@468	491 properties distributed along a surface at various densities. For
rlm@468	492 skin, it is Pacinian corpuscles, Meissner's corpuscles, Merkel's
rlm@468	493 disks, and Ruffini's endings, which detect pressure and vibration
rlm@468	494 of various intensities. For ears, it is the stereocilia distributed
rlm@468	495 along the basilar membrane inside the cochlea; each one is
rlm@468	496 sensitive to a slightly different frequency of sound. For eyes, it
rlm@468	497 is rods and cones distributed along the surface of the retina. In
rlm@468	498 each case, we can describe the sense with a surface and a
rlm@468	499 distribution of sensors along that surface.
rlm@468	500
rlm@468	501 The neat idea is that every human sense can be effectively
rlm@468	502 described in terms of a surface containing embedded sensors. If the
rlm@468	503 sense had any more dimensions, then there wouldn't be enough room
rlm@468	504 in the spinal chord to transmit the information!
rlm@468	505
rlm@468	506 Therefore, =CORTEX= must support the ability to create objects and
rlm@468	507 then be able to ``paint'' points along their surfaces to describe
rlm@468	508 each sense.
rlm@468	509
rlm@468	510 Fortunately this idea is already a well known computer graphics
rlm@468	511 technique called called /UV-mapping/. The three-dimensional surface
rlm@468	512 of a model is cut and smooshed until it fits on a two-dimensional
rlm@468	513 image. You paint whatever you want on that image, and when the
rlm@468	514 three-dimensional shape is rendered in a game the smooshing and
rlm@468	515 cutting is reversed and the image appears on the three-dimensional
rlm@468	516 object.
rlm@468	517
rlm@468	518 To make a sense, interpret the UV-image as describing the
rlm@468	519 distribution of that senses sensors. To get different types of
rlm@468	520 sensors, you can either use a different color for each type of
rlm@468	521 sensor, or use multiple UV-maps, each labeled with that sensor
rlm@468	522 type. I generally use a white pixel to mean the presence of a
rlm@468	523 sensor and a black pixel to mean the absence of a sensor, and use
rlm@468	524 one UV-map for each sensor-type within a given sense.
rlm@468	525
rlm@468	526 #+CAPTION: The UV-map for an elongated icososphere. The white
rlm@468	527 #+caption: dots each represent a touch sensor. They are dense
rlm@468	528 #+caption: in the regions that describe the tip of the finger,
rlm@468	529 #+caption: and less dense along the dorsal side of the finger
rlm@468	530 #+caption: opposite the tip.
rlm@468	531 #+name: finger-UV
rlm@468	532 #+ATTR_latex: :width 10cm
rlm@468	533 [[./images/finger-UV.png]]
rlm@468	534
rlm@468	535 #+caption: Ventral side of the UV-mapped finger. Notice the
rlm@468	536 #+caption: density of touch sensors at the tip.
rlm@468	537 #+name: finger-side-view
rlm@468	538 #+ATTR_LaTeX: :width 10cm
rlm@468	539 [[./images/finger-1.png]]
rlm@468	540
rlm@465	541 ** COMMENT Video game engines are a great starting point
rlm@462	542
rlm@462	543 I did not need to write my own physics simulation code or shader to
rlm@462	544 build =CORTEX=. Doing so would lead to a system that is impossible
rlm@462	545 for anyone but myself to use anyway. Instead, I use a video game
rlm@462	546 engine as a base and modify it to accomodate the additional needs
rlm@462	547 of =CORTEX=. Video game engines are an ideal starting point to
rlm@462	548 build =CORTEX=, because they are not far from being creature
rlm@463	549 building systems themselves.
rlm@462	550
rlm@462	551 First off, general purpose video game engines come with a physics
rlm@462	552 engine and lighting / sound system. The physics system provides
rlm@462	553 tools that can be co-opted to serve as touch, proprioception, and
rlm@462	554 muscles. Since some games support split screen views, a good video
rlm@462	555 game engine will allow you to efficiently create multiple cameras
rlm@463	556 in the simulated world that can be used as eyes. Video game systems
rlm@463	557 offer integrated asset management for things like textures and
rlm@468	558 creatures models, providing an avenue for defining creatures. They
rlm@468	559 also understand UV-mapping, since this technique is used to apply a
rlm@468	560 texture to a model. Finally, because video game engines support a
rlm@468	561 large number of users, as long as =CORTEX= doesn't stray too far
rlm@468	562 from the base system, other researchers can turn to this community
rlm@468	563 for help when doing their research.
rlm@463	564
rlm@465	565 ** COMMENT =CORTEX= is based on jMonkeyEngine3
rlm@463	566
rlm@463	567 While preparing to build =CORTEX= I studied several video game
rlm@463	568 engines to see which would best serve as a base. The top contenders
rlm@463	569 were:
rlm@463	570
rlm@463	571 - [[http://www.idsoftware.com][Quake II]]/[[http://www.bytonic.de/html/jake2.html][Jake2]] :: The Quake II engine was designed by ID
rlm@463	572 software in 1997. All the source code was released by ID
rlm@463	573 software into the Public Domain several years ago, and as a
rlm@463	574 result it has been ported to many different languages. This
rlm@463	575 engine was famous for its advanced use of realistic shading
rlm@463	576 and had decent and fast physics simulation. The main advantage
rlm@463	577 of the Quake II engine is its simplicity, but I ultimately
rlm@463	578 rejected it because the engine is too tied to the concept of a
rlm@463	579 first-person shooter game. One of the problems I had was that
rlm@463	580 there does not seem to be any easy way to attach multiple
rlm@463	581 cameras to a single character. There are also several physics
rlm@463	582 clipping issues that are corrected in a way that only applies
rlm@463	583 to the main character and do not apply to arbitrary objects.
rlm@463	584
rlm@463	585 - [[http://source.valvesoftware.com/][Source Engine]] :: The Source Engine evolved from the Quake II
rlm@463	586 and Quake I engines and is used by Valve in the Half-Life
rlm@463	587 series of games. The physics simulation in the Source Engine
rlm@463	588 is quite accurate and probably the best out of all the engines
rlm@463	589 I investigated. There is also an extensive community actively
rlm@463	590 working with the engine. However, applications that use the
rlm@463	591 Source Engine must be written in C++, the code is not open, it
rlm@463	592 only runs on Windows, and the tools that come with the SDK to
rlm@463	593 handle models and textures are complicated and awkward to use.
rlm@463	594
rlm@463	595 - [[http://jmonkeyengine.com/][jMonkeyEngine3]] :: jMonkeyEngine3 is a new library for creating
rlm@463	596 games in Java. It uses OpenGL to render to the screen and uses
rlm@463	597 screengraphs to avoid drawing things that do not appear on the
rlm@463	598 screen. It has an active community and several games in the
rlm@463	599 pipeline. The engine was not built to serve any particular
rlm@463	600 game but is instead meant to be used for any 3D game.
rlm@463	601
rlm@463	602 I chose jMonkeyEngine3 because it because it had the most features
rlm@464	603 out of all the free projects I looked at, and because I could then
rlm@463	604 write my code in clojure, an implementation of =LISP= that runs on
rlm@463	605 the JVM.
rlm@435	606
rlm@469	607 ** COMMENT =CORTEX= uses Blender to create creature models
rlm@435	608
rlm@464	609 For the simple worm-like creatures I will use later on in this
rlm@464	610 thesis, I could define a simple API in =CORTEX= that would allow
rlm@464	611 one to create boxes, spheres, etc., and leave that API as the sole
rlm@464	612 way to create creatures. However, for =CORTEX= to truly be useful
rlm@468	613 for other projects, it needs a way to construct complicated
rlm@464	614 creatures. If possible, it would be nice to leverage work that has
rlm@464	615 already been done by the community of 3D modelers, or at least
rlm@464	616 enable people who are talented at moedling but not programming to
rlm@468	617 design =CORTEX= creatures.
rlm@464	618
rlm@464	619 Therefore, I use Blender, a free 3D modeling program, as the main
rlm@464	620 way to create creatures in =CORTEX=. However, the creatures modeled
rlm@464	621 in Blender must also be simple to simulate in jMonkeyEngine3's game
rlm@468	622 engine, and must also be easy to rig with =CORTEX='s senses. I
rlm@468	623 accomplish this with extensive use of Blender's ``empty nodes.''
rlm@464	624
rlm@468	625 Empty nodes have no mass, physical presence, or appearance, but
rlm@468	626 they can hold metadata and have names. I use a tree structure of
rlm@468	627 empty nodes to specify senses in the following manner:
rlm@468	628
rlm@468	629 - Create a single top-level empty node whose name is the name of
rlm@468	630 the sense.
rlm@468	631 - Add empty nodes which each contain meta-data relevant to the
rlm@468	632 sense, including a UV-map describing the number/distribution of
rlm@468	633 sensors if applicable.
rlm@468	634 - Make each empty-node the child of the top-level node.
rlm@468	635
rlm@468	636 #+caption: An example of annoting a creature model with empty
rlm@468	637 #+caption: nodes to describe the layout of senses. There are
rlm@468	638 #+caption: multiple empty nodes which each describe the position
rlm@468	639 #+caption: of muscles, ears, eyes, or joints.
rlm@468	640 #+name: sense-nodes
rlm@468	641 #+ATTR_LaTeX: :width 10cm
rlm@468	642 [[./images/empty-sense-nodes.png]]
rlm@468	643
rlm@469	644 ** COMMENT Bodies are composed of segments connected by joints
rlm@468	645
rlm@468	646 Blender is a general purpose animation tool, which has been used in
rlm@468	647 the past to create high quality movies such as Sintel
rlm@468	648 \cite{sintel}. Though Blender can model and render even complicated
rlm@468	649 things like water, it is crucual to keep models that are meant to
rlm@468	650 be simulated as creatures simple. =Bullet=, which =CORTEX= uses
rlm@468	651 though jMonkeyEngine3, is a rigid-body physics system. This offers
rlm@468	652 a compromise between the expressiveness of a game level and the
rlm@468	653 speed at which it can be simulated, and it means that creatures
rlm@468	654 should be naturally expressed as rigid components held together by
rlm@468	655 joint constraints.
rlm@468	656
rlm@468	657 But humans are more like a squishy bag with wrapped around some
rlm@468	658 hard bones which define the overall shape. When we move, our skin
rlm@468	659 bends and stretches to accomodate the new positions of our bones.
rlm@468	660
rlm@468	661 One way to make bodies composed of rigid pieces connected by joints
rlm@468	662 /seem/ more human-like is to use an /armature/, (or /rigging/)
rlm@468	663 system, which defines a overall ``body mesh'' and defines how the
rlm@468	664 mesh deforms as a function of the position of each ``bone'' which
rlm@468	665 is a standard rigid body. This technique is used extensively to
rlm@468	666 model humans and create realistic animations. It is not a good
rlm@468	667 technique for physical simulation, however because it creates a lie
rlm@468	668 -- the skin is not a physical part of the simulation and does not
rlm@468	669 interact with any objects in the world or itself. Objects will pass
rlm@468	670 right though the skin until they come in contact with the
rlm@468	671 underlying bone, which is a physical object. Whithout simulating
rlm@468	672 the skin, the sense of touch has little meaning, and the creature's
rlm@468	673 own vision will lie to it about the true extent of its body.
rlm@468	674 Simulating the skin as a physical object requires some way to
rlm@468	675 continuously update the physical model of the skin along with the
rlm@468	676 movement of the bones, which is unacceptably slow compared to rigid
rlm@468	677 body simulation.
rlm@468	678
rlm@468	679 Therefore, instead of using the human-like ``deformable bag of
rlm@468	680 bones'' approach, I decided to base my body plans on multiple solid
rlm@468	681 objects that are connected by joints, inspired by the robot =EVE=
rlm@468	682 from the movie WALL-E.
rlm@464	683
rlm@464	684 #+caption: =EVE= from the movie WALL-E. This body plan turns
rlm@464	685 #+caption: out to be much better suited to my purposes than a more
rlm@464	686 #+caption: human-like one.
rlm@465	687 #+ATTR_LaTeX: :width 10cm
rlm@464	688 [[./images/Eve.jpg]]
rlm@464	689
rlm@464	690 =EVE='s body is composed of several rigid components that are held
rlm@464	691 together by invisible joint constraints. This is what I mean by
rlm@464	692 ``eve-like''. The main reason that I use eve-style bodies is for
rlm@464	693 efficiency, and so that there will be correspondence between the
rlm@468	694 AI's semses and the physical presence of its body. Each individual
rlm@464	695 section is simulated by a separate rigid body that corresponds
rlm@464	696 exactly with its visual representation and does not change.
rlm@464	697 Sections are connected by invisible joints that are well supported
rlm@464	698 in jMonkeyEngine3. Bullet, the physics backend for jMonkeyEngine3,
rlm@464	699 can efficiently simulate hundreds of rigid bodies connected by
rlm@468	700 joints. Just because sections are rigid does not mean they have to
rlm@468	701 stay as one piece forever; they can be dynamically replaced with
rlm@468	702 multiple sections to simulate splitting in two. This could be used
rlm@468	703 to simulate retractable claws or =EVE='s hands, which are able to
rlm@468	704 coalesce into one object in the movie.
rlm@465	705
rlm@469	706 *** Solidifying/Connecting a body
rlm@465	707
rlm@469	708 =CORTEX= creates a creature in two steps: first, it traverses the
rlm@469	709 nodes in the blender file and creates physical representations for
rlm@469	710 any of them that have mass defined in their blender meta-data.
rlm@466	711
rlm@466	712 #+caption: Program for iterating through the nodes in a blender file
rlm@466	713 #+caption: and generating physical jMonkeyEngine3 objects with mass
rlm@466	714 #+caption: and a matching physics shape.
rlm@466	715 #+name: name
rlm@466	716 #+begin_listing clojure
rlm@466	717 #+begin_src clojure
rlm@466	718 (defn physical!
rlm@466	719 "Iterate through the nodes in creature and make them real physical
rlm@466	720 objects in the simulation."
rlm@466	721 [#^Node creature]
rlm@466	722 (dorun
rlm@466	723 (map
rlm@466	724 (fn [geom]
rlm@466	725 (let [physics-control
rlm@466	726 (RigidBodyControl.
rlm@466	727 (HullCollisionShape.
rlm@466	728 (.getMesh geom))
rlm@466	729 (if-let [mass (meta-data geom "mass")]
rlm@466	730 (float mass) (float 1)))]
rlm@466	731 (.addControl geom physics-control)))
rlm@466	732 (filter #(isa? (class %) Geometry )
rlm@466	733 (node-seq creature)))))
rlm@466	734 #+end_src
rlm@466	735 #+end_listing
rlm@465	736
rlm@469	737 The next step to making a proper body is to connect those pieces
rlm@469	738 together with joints. jMonkeyEngine has a large array of joints
rlm@469	739 available via =bullet=, such as Point2Point, Cone, Hinge, and a
rlm@469	740 generic Six Degree of Freedom joint, with or without spring
rlm@469	741 restitution.
rlm@465	742
rlm@469	743 Joints are treated a lot like proper senses, in that there is a
rlm@469	744 top-level empty node named ``joints'' whose children each
rlm@469	745 represent a joint.
rlm@466	746
rlm@469	747 #+caption: View of the hand model in Blender showing the main ``joints''
rlm@469	748 #+caption: node (highlighted in yellow) and its children which each
rlm@469	749 #+caption: represent a joint in the hand. Each joint node has metadata
rlm@469	750 #+caption: specifying what sort of joint it is.
rlm@469	751 #+name: blender-hand
rlm@469	752 #+ATTR_LaTeX: :width 10cm
rlm@469	753 [[./images/hand-screenshot1.png]]
rlm@469	754
rlm@469	755
rlm@469	756 =CORTEX='s procedure for binding the creature together with joints
rlm@469	757 is as follows:
rlm@469	758
rlm@469	759 - Find the children of the ``joints'' node.
rlm@469	760 - Determine the two spatials the joint is meant to connect.
rlm@469	761 - Create the joint based on the meta-data of the empty node.
rlm@469	762
rlm@469	763 The higher order function =sense-nodes= from =cortex.sense=
rlm@469	764 simplifies finding the joints based on their parent ``joints''
rlm@469	765 node.
rlm@466	766
rlm@466	767 #+caption: Retrieving the children empty nodes from a single
rlm@466	768 #+caption: named empty node is a common pattern in =CORTEX=
rlm@466	769 #+caption: further instances of this technique for the senses
rlm@466	770 #+caption: will be omitted
rlm@466	771 #+name: get-empty-nodes
rlm@466	772 #+begin_listing clojure
rlm@466	773 #+begin_src clojure
rlm@466	774 (defn sense-nodes
rlm@466	775 "For some senses there is a special empty blender node whose
rlm@466	776 children are considered markers for an instance of that sense. This
rlm@466	777 function generates functions to find those children, given the name
rlm@466	778 of the special parent node."
rlm@466	779 [parent-name]
rlm@466	780 (fn [#^Node creature]
rlm@466	781 (if-let [sense-node (.getChild creature parent-name)]
rlm@466	782 (seq (.getChildren sense-node)) [])))
rlm@466	783
rlm@466	784 (def
rlm@466	785 ^{:doc "Return the children of the creature's \"joints\" node."
rlm@466	786 :arglists '([creature])}
rlm@466	787 joints
rlm@466	788 (sense-nodes "joints"))
rlm@466	789 #+end_src
rlm@466	790 #+end_listing
rlm@466	791
rlm@469	792 To find a joint's targets, =CORTEX= creates a small cube, centered
rlm@469	793 around the empty-node, and grows the cube exponentially until it
rlm@469	794 intersects two physical objects. The objects are ordered according
rlm@469	795 to the joint's rotation, with the first one being the object that
rlm@469	796 has more negative coordinates in the joint's reference frame.
rlm@469	797 Since the objects must be physical, the empty-node itself escapes
rlm@469	798 detection. Because the objects must be physical, =joint-targets=
rlm@469	799 must be called /after/ =physical!= is called.
rlm@464	800
rlm@469	801 #+caption: Program to find the targets of a joint node by
rlm@469	802 #+caption: exponentiallly growth of a search cube.
rlm@469	803 #+name: joint-targets
rlm@469	804 #+begin_listing clojure
rlm@469	805 #+begin_src clojure
rlm@466	806 (defn joint-targets
rlm@466	807 "Return the two closest two objects to the joint object, ordered
rlm@466	808 from bottom to top according to the joint's rotation."
rlm@466	809 [#^Node parts #^Node joint]
rlm@466	810 (loop [radius (float 0.01)]
rlm@466	811 (let [results (CollisionResults.)]
rlm@466	812 (.collideWith
rlm@466	813 parts
rlm@466	814 (BoundingBox. (.getWorldTranslation joint)
rlm@466	815 radius radius radius) results)
rlm@466	816 (let [targets
rlm@466	817 (distinct
rlm@466	818 (map #(.getGeometry %) results))]
rlm@466	819 (if (>= (count targets) 2)
rlm@466	820 (sort-by
rlm@466	821 #(let [joint-ref-frame-position
rlm@466	822 (jme-to-blender
rlm@466	823 (.mult
rlm@466	824 (.inverse (.getWorldRotation joint))
rlm@466	825 (.subtract (.getWorldTranslation %)
rlm@466	826 (.getWorldTranslation joint))))]
rlm@466	827 (.dot (Vector3f. 1 1 1) joint-ref-frame-position))
rlm@466	828 (take 2 targets))
rlm@466	829 (recur (float (* radius 2))))))))
rlm@469	830 #+end_src
rlm@469	831 #+end_listing
rlm@464	832
rlm@469	833 Once =CORTEX= finds all joints and targets, it creates them using
rlm@469	834 a dispatch on the metadata of each joint node.
rlm@466	835
rlm@469	836 #+caption: Program to dispatch on blender metadata and create joints
rlm@469	837 #+caption: sutiable for physical simulation.
rlm@469	838 #+name: joint-dispatch
rlm@469	839 #+begin_listing clojure
rlm@469	840 #+begin_src clojure
rlm@466	841 (defmulti joint-dispatch
rlm@466	842 "Translate blender pseudo-joints into real JME joints."
rlm@466	843 (fn [constraints & _]
rlm@466	844 (:type constraints)))
rlm@466	845
rlm@466	846 (defmethod joint-dispatch :point
rlm@466	847 [constraints control-a control-b pivot-a pivot-b rotation]
rlm@466	848 (doto (SixDofJoint. control-a control-b pivot-a pivot-b false)
rlm@466	849 (.setLinearLowerLimit Vector3f/ZERO)
rlm@466	850 (.setLinearUpperLimit Vector3f/ZERO)))
rlm@466	851
rlm@466	852 (defmethod joint-dispatch :hinge
rlm@466	853 [constraints control-a control-b pivot-a pivot-b rotation]
rlm@466	854 (let [axis (if-let [axis (:axis constraints)] axis Vector3f/UNIT_X)
rlm@466	855 [limit-1 limit-2] (:limit constraints)
rlm@466	856 hinge-axis (.mult rotation (blender-to-jme axis))]
rlm@466	857 (doto (HingeJoint. control-a control-b pivot-a pivot-b
rlm@466	858 hinge-axis hinge-axis)
rlm@466	859 (.setLimit limit-1 limit-2))))
rlm@466	860
rlm@466	861 (defmethod joint-dispatch :cone
rlm@466	862 [constraints control-a control-b pivot-a pivot-b rotation]
rlm@466	863 (let [limit-xz (:limit-xz constraints)
rlm@466	864 limit-xy (:limit-xy constraints)
rlm@466	865 twist (:twist constraints)]
rlm@466	866 (doto (ConeJoint. control-a control-b pivot-a pivot-b
rlm@466	867 rotation rotation)
rlm@466	868 (.setLimit (float limit-xz) (float limit-xy)
rlm@466	869 (float twist)))))
rlm@469	870 #+end_src
rlm@469	871 #+end_listing
rlm@466	872
rlm@469	873 All that is left for joints it to combine the above pieces into a
rlm@469	874 something that can operate on the collection of nodes that a
rlm@469	875 blender file represents.
rlm@466	876
rlm@469	877 #+caption: Program to completely create a joint given information
rlm@469	878 #+caption: from a blender file.
rlm@469	879 #+name: connect
rlm@469	880 #+begin_listing clojure
rlm@466	881 #+begin_src clojure
rlm@466	882 (defn connect
rlm@466	883 "Create a joint between 'obj-a and 'obj-b at the location of
rlm@466	884 'joint. The type of joint is determined by the metadata on 'joint.
rlm@466	885
rlm@466	886 Here are some examples:
rlm@466	887 {:type :point}
rlm@466	888 {:type :hinge :limit [0 (/ Math/PI 2)] :axis (Vector3f. 0 1 0)}
rlm@466	889 (:axis defaults to (Vector3f. 1 0 0) if not provided for hinge joints)
rlm@466	890
rlm@466	891 {:type :cone :limit-xz 0]
rlm@466	892 :limit-xy 0]
rlm@466	893 :twist 0]} (use XZY rotation mode in blender!)"
rlm@466	894 [#^Node obj-a #^Node obj-b #^Node joint]
rlm@466	895 (let [control-a (.getControl obj-a RigidBodyControl)
rlm@466	896 control-b (.getControl obj-b RigidBodyControl)
rlm@466	897 joint-center (.getWorldTranslation joint)
rlm@466	898 joint-rotation (.toRotationMatrix (.getWorldRotation joint))
rlm@466	899 pivot-a (world-to-local obj-a joint-center)
rlm@466	900 pivot-b (world-to-local obj-b joint-center)]
rlm@466	901 (if-let
rlm@466	902 [constraints (map-vals eval (read-string (meta-data joint "joint")))]
rlm@466	903 ;; A side-effect of creating a joint registers
rlm@466	904 ;; it with both physics objects which in turn
rlm@466	905 ;; will register the joint with the physics system
rlm@466	906 ;; when the simulation is started.
rlm@466	907 (joint-dispatch constraints
rlm@466	908 control-a control-b
rlm@466	909 pivot-a pivot-b
rlm@466	910 joint-rotation))))
rlm@469	911 #+end_src
rlm@469	912 #+end_listing
rlm@466	913
rlm@469	914 In general, whenever =CORTEX= exposes a sense (or in this case
rlm@469	915 physicality), it provides a function of the type =sense!=, which
rlm@469	916 takes in a collection of nodes and augments it to support that
rlm@469	917 sense. The function returns any controlls necessary to use that
rlm@469	918 sense. In this case =body!= cerates a physical body and returns no
rlm@469	919 control functions.
rlm@466	920
rlm@469	921 #+caption: Program to give joints to a creature.
rlm@469	922 #+name: name
rlm@469	923 #+begin_listing clojure
rlm@469	924 #+begin_src clojure
rlm@466	925 (defn joints!
rlm@466	926 "Connect the solid parts of the creature with physical joints. The
rlm@466	927 joints are taken from the \"joints\" node in the creature."
rlm@466	928 [#^Node creature]
rlm@466	929 (dorun
rlm@466	930 (map
rlm@466	931 (fn [joint]
rlm@466	932 (let [[obj-a obj-b] (joint-targets creature joint)]
rlm@466	933 (connect obj-a obj-b joint)))
rlm@466	934 (joints creature))))
rlm@466	935 (defn body!
rlm@466	936 "Endow the creature with a physical body connected with joints. The
rlm@466	937 particulars of the joints and the masses of each body part are
rlm@466	938 determined in blender."
rlm@466	939 [#^Node creature]
rlm@466	940 (physical! creature)
rlm@466	941 (joints! creature))
rlm@469	942 #+end_src
rlm@469	943 #+end_listing
rlm@466	944
rlm@469	945 All of the code you have just seen amounts to only 130 lines, yet
rlm@469	946 because it builds on top of Blender and jMonkeyEngine3, those few
rlm@469	947 lines pack quite a punch!
rlm@466	948
rlm@469	949 The hand from figure \ref{blender-hand}, which was modeled after
rlm@469	950 my own right hand, can now be given joints and simulated as a
rlm@469	951 creature.
rlm@466	952
rlm@469	953 #+caption: With the ability to create physical creatures from blender,
rlm@469	954 #+caption: =CORTEX= gets one step closer to becomming a full creature
rlm@469	955 #+caption: simulation environment.
rlm@469	956 #+name: name
rlm@469	957 #+ATTR_LaTeX: :width 15cm
rlm@469	958 [[./images/physical-hand.png]]
rlm@468	959
rlm@472	960 ** COMMENT Eyes reuse standard video game components
rlm@436	961
rlm@470	962 Vision is one of the most important senses for humans, so I need to
rlm@470	963 build a simulated sense of vision for my AI. I will do this with
rlm@470	964 simulated eyes. Each eye can be independently moved and should see
rlm@470	965 its own version of the world depending on where it is.
rlm@470	966
rlm@470	967 Making these simulated eyes a reality is simple because
rlm@470	968 jMonkeyEngine already contains extensive support for multiple views
rlm@470	969 of the same 3D simulated world. The reason jMonkeyEngine has this
rlm@470	970 support is because the support is necessary to create games with
rlm@470	971 split-screen views. Multiple views are also used to create
rlm@470	972 efficient pseudo-reflections by rendering the scene from a certain
rlm@470	973 perspective and then projecting it back onto a surface in the 3D
rlm@470	974 world.
rlm@470	975
rlm@470	976 #+caption: jMonkeyEngine supports multiple views to enable
rlm@470	977 #+caption: split-screen games, like GoldenEye, which was one of
rlm@470	978 #+caption: the first games to use split-screen views.
rlm@470	979 #+name: name
rlm@470	980 #+ATTR_LaTeX: :width 10cm
rlm@470	981 [[./images/goldeneye-4-player.png]]
rlm@470	982
rlm@470	983 *** A Brief Description of jMonkeyEngine's Rendering Pipeline
rlm@470	984
rlm@470	985 jMonkeyEngine allows you to create a =ViewPort=, which represents a
rlm@470	986 view of the simulated world. You can create as many of these as you
rlm@470	987 want. Every frame, the =RenderManager= iterates through each
rlm@470	988 =ViewPort=, rendering the scene in the GPU. For each =ViewPort= there
rlm@470	989 is a =FrameBuffer= which represents the rendered image in the GPU.
rlm@470	990
rlm@470	991 #+caption: =ViewPorts= are cameras in the world. During each frame,
rlm@470	992 #+caption: the =RenderManager= records a snapshot of what each view
rlm@470	993 #+caption: is currently seeing; these snapshots are =FrameBuffer= objects.
rlm@470	994 #+name: name
rlm@470	995 #+ATTR_LaTeX: :width 10cm
rlm@470	996 [[../images/diagram_rendermanager2.png]]
rlm@470	997
rlm@470	998 Each =ViewPort= can have any number of attached =SceneProcessor=
rlm@470	999 objects, which are called every time a new frame is rendered. A
rlm@470	1000 =SceneProcessor= receives its =ViewPort's= =FrameBuffer= and can do
rlm@470	1001 whatever it wants to the data. Often this consists of invoking GPU
rlm@470	1002 specific operations on the rendered image. The =SceneProcessor= can
rlm@470	1003 also copy the GPU image data to RAM and process it with the CPU.
rlm@470	1004
rlm@470	1005 *** Appropriating Views for Vision
rlm@470	1006
rlm@470	1007 Each eye in the simulated creature needs its own =ViewPort= so
rlm@470	1008 that it can see the world from its own perspective. To this
rlm@470	1009 =ViewPort=, I add a =SceneProcessor= that feeds the visual data to
rlm@470	1010 any arbitrary continuation function for further processing. That
rlm@470	1011 continuation function may perform both CPU and GPU operations on
rlm@470	1012 the data. To make this easy for the continuation function, the
rlm@470	1013 =SceneProcessor= maintains appropriately sized buffers in RAM to
rlm@470	1014 hold the data. It does not do any copying from the GPU to the CPU
rlm@470	1015 itself because it is a slow operation.
rlm@470	1016
rlm@470	1017 #+caption: Function to make the rendered secne in jMonkeyEngine
rlm@470	1018 #+caption: available for further processing.
rlm@470	1019 #+name: pipeline-1
rlm@470	1020 #+begin_listing clojure
rlm@470	1021 #+begin_src clojure
rlm@470	1022 (defn vision-pipeline
rlm@470	1023 "Create a SceneProcessor object which wraps a vision processing
rlm@470	1024 continuation function. The continuation is a function that takes
rlm@470	1025 [#^Renderer r #^FrameBuffer fb #^ByteBuffer b #^BufferedImage bi],
rlm@470	1026 each of which has already been appropriately sized."
rlm@470	1027 [continuation]
rlm@470	1028 (let [byte-buffer (atom nil)
rlm@470	1029 renderer (atom nil)
rlm@470	1030 image (atom nil)]
rlm@470	1031 (proxy [SceneProcessor] []
rlm@470	1032 (initialize
rlm@470	1033 [renderManager viewPort]
rlm@470	1034 (let [cam (.getCamera viewPort)
rlm@470	1035 width (.getWidth cam)
rlm@470	1036 height (.getHeight cam)]
rlm@470	1037 (reset! renderer (.getRenderer renderManager))
rlm@470	1038 (reset! byte-buffer
rlm@470	1039 (BufferUtils/createByteBuffer
rlm@470	1040 (* width height 4)))
rlm@470	1041 (reset! image (BufferedImage.
rlm@470	1042 width height
rlm@470	1043 BufferedImage/TYPE_4BYTE_ABGR))))
rlm@470	1044 (isInitialized [] (not (nil? @byte-buffer)))
rlm@470	1045 (reshape [_ _ _])
rlm@470	1046 (preFrame [_])
rlm@470	1047 (postQueue [_])
rlm@470	1048 (postFrame
rlm@470	1049 [#^FrameBuffer fb]
rlm@470	1050 (.clear @byte-buffer)
rlm@470	1051 (continuation @renderer fb @byte-buffer @image))
rlm@470	1052 (cleanup []))))
rlm@470	1053 #+end_src
rlm@470	1054 #+end_listing
rlm@470	1055
rlm@470	1056 The continuation function given to =vision-pipeline= above will be
rlm@470	1057 given a =Renderer= and three containers for image data. The
rlm@470	1058 =FrameBuffer= references the GPU image data, but the pixel data
rlm@470	1059 can not be used directly on the CPU. The =ByteBuffer= and
rlm@470	1060 =BufferedImage= are initially "empty" but are sized to hold the
rlm@470	1061 data in the =FrameBuffer=. I call transferring the GPU image data
rlm@470	1062 to the CPU structures "mixing" the image data.
rlm@470	1063
rlm@470	1064 *** Optical sensor arrays are described with images and referenced with metadata
rlm@470	1065
rlm@470	1066 The vision pipeline described above handles the flow of rendered
rlm@470	1067 images. Now, =CORTEX= needs simulated eyes to serve as the source
rlm@470	1068 of these images.
rlm@470	1069
rlm@470	1070 An eye is described in blender in the same way as a joint. They
rlm@470	1071 are zero dimensional empty objects with no geometry whose local
rlm@470	1072 coordinate system determines the orientation of the resulting eye.
rlm@470	1073 All eyes are children of a parent node named "eyes" just as all
rlm@470	1074 joints have a parent named "joints". An eye binds to the nearest
rlm@470	1075 physical object with =bind-sense=.
rlm@470	1076
rlm@470	1077 #+caption: Here, the camera is created based on metadata on the
rlm@470	1078 #+caption: eye-node and attached to the nearest physical object
rlm@470	1079 #+caption: with =bind-sense=
rlm@470	1080 #+name: add-eye
rlm@470	1081 #+begin_listing clojure
rlm@470	1082 (defn add-eye!
rlm@470	1083 "Create a Camera centered on the current position of 'eye which
rlm@470	1084 follows the closest physical node in 'creature. The camera will
rlm@470	1085 point in the X direction and use the Z vector as up as determined
rlm@470	1086 by the rotation of these vectors in blender coordinate space. Use
rlm@470	1087 XZY rotation for the node in blender."
rlm@470	1088 [#^Node creature #^Spatial eye]
rlm@470	1089 (let [target (closest-node creature eye)
rlm@470	1090 [cam-width cam-height]
rlm@470	1091 ;;[640 480] ;; graphics card on laptop doesn't support
rlm@470	1092 ;; arbitray dimensions.
rlm@470	1093 (eye-dimensions eye)
rlm@470	1094 cam (Camera. cam-width cam-height)
rlm@470	1095 rot (.getWorldRotation eye)]
rlm@470	1096 (.setLocation cam (.getWorldTranslation eye))
rlm@470	1097 (.lookAtDirection
rlm@470	1098 cam ; this part is not a mistake and
rlm@470	1099 (.mult rot Vector3f/UNIT_X) ; is consistent with using Z in
rlm@470	1100 (.mult rot Vector3f/UNIT_Y)) ; blender as the UP vector.
rlm@470	1101 (.setFrustumPerspective
rlm@470	1102 cam (float 45)
rlm@470	1103 (float (/ (.getWidth cam) (.getHeight cam)))
rlm@470	1104 (float 1)
rlm@470	1105 (float 1000))
rlm@470	1106 (bind-sense target cam) cam))
rlm@470	1107 #+end_listing
rlm@470	1108
rlm@470	1109 *** Simulated Retina
rlm@470	1110
rlm@470	1111 An eye is a surface (the retina) which contains many discrete
rlm@470	1112 sensors to detect light. These sensors can have different
rlm@470	1113 light-sensing properties. In humans, each discrete sensor is
rlm@470	1114 sensitive to red, blue, green, or gray. These different types of
rlm@470	1115 sensors can have different spatial distributions along the retina.
rlm@470	1116 In humans, there is a fovea in the center of the retina which has
rlm@470	1117 a very high density of color sensors, and a blind spot which has
rlm@470	1118 no sensors at all. Sensor density decreases in proportion to
rlm@470	1119 distance from the fovea.
rlm@470	1120
rlm@470	1121 I want to be able to model any retinal configuration, so my
rlm@470	1122 eye-nodes in blender contain metadata pointing to images that
rlm@470	1123 describe the precise position of the individual sensors using
rlm@470	1124 white pixels. The meta-data also describes the precise sensitivity
rlm@470	1125 to light that the sensors described in the image have. An eye can
rlm@470	1126 contain any number of these images. For example, the metadata for
rlm@470	1127 an eye might look like this:
rlm@470	1128
rlm@470	1129 #+begin_src clojure
rlm@470	1130 {0xFF0000 "Models/test-creature/retina-small.png"}
rlm@470	1131 #+end_src
rlm@470	1132
rlm@470	1133 #+caption: An example retinal profile image. White pixels are
rlm@470	1134 #+caption: photo-sensitive elements. The distribution of white
rlm@470	1135 #+caption: pixels is denser in the middle and falls off at the
rlm@470	1136 #+caption: edges and is inspired by the human retina.
rlm@470	1137 #+name: retina
rlm@470	1138 #+ATTR_LaTeX: :width 10cm
rlm@470	1139 [[./images/retina-small.png]]
rlm@470	1140
rlm@470	1141 Together, the number 0xFF0000 and the image image above describe
rlm@470	1142 the placement of red-sensitive sensory elements.
rlm@470	1143
rlm@470	1144 Meta-data to very crudely approximate a human eye might be
rlm@470	1145 something like this:
rlm@470	1146
rlm@470	1147 #+begin_src clojure
rlm@470	1148 (let [retinal-profile "Models/test-creature/retina-small.png"]
rlm@470	1149 {0xFF0000 retinal-profile
rlm@470	1150 0x00FF00 retinal-profile
rlm@470	1151 0x0000FF retinal-profile
rlm@470	1152 0xFFFFFF retinal-profile})
rlm@470	1153 #+end_src
rlm@470	1154
rlm@470	1155 The numbers that serve as keys in the map determine a sensor's
rlm@470	1156 relative sensitivity to the channels red, green, and blue. These
rlm@470	1157 sensitivity values are packed into an integer in the order
rlm@470	1158 =\|_\|R\|G\|B\|= in 8-bit fields. The RGB values of a pixel in the
rlm@470	1159 image are added together with these sensitivities as linear
rlm@470	1160 weights. Therefore, 0xFF0000 means sensitive to red only while
rlm@470	1161 0xFFFFFF means sensitive to all colors equally (gray).
rlm@470	1162
rlm@470	1163 #+caption: This is the core of vision in =CORTEX=. A given eye node
rlm@470	1164 #+caption: is converted into a function that returns visual
rlm@470	1165 #+caption: information from the simulation.
rlm@471	1166 #+name: vision-kernel
rlm@470	1167 #+begin_listing clojure
rlm@470	1168 (defn vision-kernel
rlm@470	1169 "Returns a list of functions, each of which will return a color
rlm@470	1170 channel's worth of visual information when called inside a running
rlm@470	1171 simulation."
rlm@470	1172 [#^Node creature #^Spatial eye & {skip :skip :or {skip 0}}]
rlm@470	1173 (let [retinal-map (retina-sensor-profile eye)
rlm@470	1174 camera (add-eye! creature eye)
rlm@470	1175 vision-image
rlm@470	1176 (atom
rlm@470	1177 (BufferedImage. (.getWidth camera)
rlm@470	1178 (.getHeight camera)
rlm@470	1179 BufferedImage/TYPE_BYTE_BINARY))
rlm@470	1180 register-eye!
rlm@470	1181 (runonce
rlm@470	1182 (fn [world]
rlm@470	1183 (add-camera!
rlm@470	1184 world camera
rlm@470	1185 (let [counter (atom 0)]
rlm@470	1186 (fn [r fb bb bi]
rlm@470	1187 (if (zero? (rem (swap! counter inc) (inc skip)))
rlm@470	1188 (reset! vision-image
rlm@470	1189 (BufferedImage! r fb bb bi))))))))]
rlm@470	1190 (vec
rlm@470	1191 (map
rlm@470	1192 (fn [[key image]]
rlm@470	1193 (let [whites (white-coordinates image)
rlm@470	1194 topology (vec (collapse whites))
rlm@470	1195 sensitivity (sensitivity-presets key key)]
rlm@470	1196 (attached-viewport.
rlm@470	1197 (fn [world]
rlm@470	1198 (register-eye! world)
rlm@470	1199 (vector
rlm@470	1200 topology
rlm@470	1201 (vec
rlm@470	1202 (for [[x y] whites]
rlm@470	1203 (pixel-sense
rlm@470	1204 sensitivity
rlm@470	1205 (.getRGB @vision-image x y))))))
rlm@470	1206 register-eye!)))
rlm@470	1207 retinal-map))))
rlm@470	1208 #+end_listing
rlm@470	1209
rlm@470	1210 Note that since each of the functions generated by =vision-kernel=
rlm@470	1211 shares the same =register-eye!= function, the eye will be
rlm@470	1212 registered only once the first time any of the functions from the
rlm@470	1213 list returned by =vision-kernel= is called. Each of the functions
rlm@470	1214 returned by =vision-kernel= also allows access to the =Viewport=
rlm@470	1215 through which it receives images.
rlm@470	1216
rlm@470	1217 All the hard work has been done; all that remains is to apply
rlm@470	1218 =vision-kernel= to each eye in the creature and gather the results
rlm@470	1219 into one list of functions.
rlm@470	1220
rlm@470	1221
rlm@470	1222 #+caption: With =vision!=, =CORTEX= is already a fine simulation
rlm@470	1223 #+caption: environment for experimenting with different types of
rlm@470	1224 #+caption: eyes.
rlm@470	1225 #+name: vision!
rlm@470	1226 #+begin_listing clojure
rlm@470	1227 (defn vision!
rlm@470	1228 "Returns a list of functions, each of which returns visual sensory
rlm@470	1229 data when called inside a running simulation."
rlm@470	1230 [#^Node creature & {skip :skip :or {skip 0}}]
rlm@470	1231 (reduce
rlm@470	1232 concat
rlm@470	1233 (for [eye (eyes creature)]
rlm@470	1234 (vision-kernel creature eye))))
rlm@470	1235 #+end_listing
rlm@470	1236
rlm@471	1237 #+caption: Simulated vision with a test creature and the
rlm@471	1238 #+caption: human-like eye approximation. Notice how each channel
rlm@471	1239 #+caption: of the eye responds differently to the differently
rlm@471	1240 #+caption: colored balls.
rlm@471	1241 #+name: worm-vision-test.
rlm@471	1242 #+ATTR_LaTeX: :width 13cm
rlm@471	1243 [[./images/worm-vision.png]]
rlm@470	1244
rlm@471	1245 The vision code is not much more complicated than the body code,
rlm@471	1246 and enables multiple further paths for simulated vision. For
rlm@471	1247 example, it is quite easy to create bifocal vision -- you just
rlm@471	1248 make two eyes next to each other in blender! It is also possible
rlm@471	1249 to encode vision transforms in the retinal files. For example, the
rlm@471	1250 human like retina file in figure \ref{retina} approximates a
rlm@471	1251 log-polar transform.
rlm@470	1252
rlm@471	1253 This vision code has already been absorbed by the jMonkeyEngine
rlm@471	1254 community and is now (in modified form) part of a system for
rlm@471	1255 capturing in-game video to a file.
rlm@470	1256
rlm@473	1257 ** COMMENT Hearing is hard; =CORTEX= does it right
rlm@473	1258
rlm@472	1259 At the end of this section I will have simulated ears that work the
rlm@472	1260 same way as the simulated eyes in the last section. I will be able to
rlm@472	1261 place any number of ear-nodes in a blender file, and they will bind to
rlm@472	1262 the closest physical object and follow it as it moves around. Each ear
rlm@472	1263 will provide access to the sound data it picks up between every frame.
rlm@472	1264
rlm@472	1265 Hearing is one of the more difficult senses to simulate, because there
rlm@472	1266 is less support for obtaining the actual sound data that is processed
rlm@472	1267 by jMonkeyEngine3. There is no "split-screen" support for rendering
rlm@472	1268 sound from different points of view, and there is no way to directly
rlm@472	1269 access the rendered sound data.
rlm@472	1270
rlm@472	1271 =CORTEX='s hearing is unique because it does not have any
rlm@472	1272 limitations compared to other simulation environments. As far as I
rlm@472	1273 know, there is no other system that supports multiple listerers,
rlm@472	1274 and the sound demo at the end of this section is the first time
rlm@472	1275 it's been done in a video game environment.
rlm@472	1276
rlm@472	1277 *** Brief Description of jMonkeyEngine's Sound System
rlm@472	1278
rlm@472	1279 jMonkeyEngine's sound system works as follows:
rlm@472	1280
rlm@472	1281 - jMonkeyEngine uses the =AppSettings= for the particular
rlm@472	1282 application to determine what sort of =AudioRenderer= should be
rlm@472	1283 used.
rlm@472	1284 - Although some support is provided for multiple AudioRendering
rlm@472	1285 backends, jMonkeyEngine at the time of this writing will either
rlm@472	1286 pick no =AudioRenderer= at all, or the =LwjglAudioRenderer=.
rlm@472	1287 - jMonkeyEngine tries to figure out what sort of system you're
rlm@472	1288 running and extracts the appropriate native libraries.
rlm@472	1289 - The =LwjglAudioRenderer= uses the [[http://lwjgl.org/][=LWJGL=]] (LightWeight Java Game
rlm@472	1290 Library) bindings to interface with a C library called [[http://kcat.strangesoft.net/openal.html][=OpenAL=]]
rlm@472	1291 - =OpenAL= renders the 3D sound and feeds the rendered sound
rlm@472	1292 directly to any of various sound output devices with which it
rlm@472	1293 knows how to communicate.
rlm@472	1294
rlm@472	1295 A consequence of this is that there's no way to access the actual
rlm@472	1296 sound data produced by =OpenAL=. Even worse, =OpenAL= only supports
rlm@472	1297 one /listener/ (it renders sound data from only one perspective),
rlm@472	1298 which normally isn't a problem for games, but becomes a problem
rlm@472	1299 when trying to make multiple AI creatures that can each hear the
rlm@472	1300 world from a different perspective.
rlm@472	1301
rlm@472	1302 To make many AI creatures in jMonkeyEngine that can each hear the
rlm@472	1303 world from their own perspective, or to make a single creature with
rlm@472	1304 many ears, it is necessary to go all the way back to =OpenAL= and
rlm@472	1305 implement support for simulated hearing there.
rlm@472	1306
rlm@472	1307 *** Extending =OpenAl=
rlm@472	1308
rlm@472	1309 Extending =OpenAL= to support multiple listeners requires 500
rlm@472	1310 lines of =C= code and is too hairy to mention here. Instead, I
rlm@472	1311 will show a small amount of extension code and go over the high
rlm@472	1312 level stragety. Full source is of course available with the
rlm@472	1313 =CORTEX= distribution if you're interested.
rlm@472	1314
rlm@472	1315 =OpenAL= goes to great lengths to support many different systems,
rlm@472	1316 all with different sound capabilities and interfaces. It
rlm@472	1317 accomplishes this difficult task by providing code for many
rlm@472	1318 different sound backends in pseudo-objects called /Devices/.
rlm@472	1319 There's a device for the Linux Open Sound System and the Advanced
rlm@472	1320 Linux Sound Architecture, there's one for Direct Sound on Windows,
rlm@472	1321 and there's even one for Solaris. =OpenAL= solves the problem of
rlm@472	1322 platform independence by providing all these Devices.
rlm@472	1323
rlm@472	1324 Wrapper libraries such as LWJGL are free to examine the system on
rlm@472	1325 which they are running and then select an appropriate device for
rlm@472	1326 that system.
rlm@472	1327
rlm@472	1328 There are also a few "special" devices that don't interface with
rlm@472	1329 any particular system. These include the Null Device, which
rlm@472	1330 doesn't do anything, and the Wave Device, which writes whatever
rlm@472	1331 sound it receives to a file, if everything has been set up
rlm@472	1332 correctly when configuring =OpenAL=.
rlm@472	1333
rlm@472	1334 Actual mixing (doppler shift and distance.environment-based
rlm@472	1335 attenuation) of the sound data happens in the Devices, and they
rlm@472	1336 are the only point in the sound rendering process where this data
rlm@472	1337 is available.
rlm@472	1338
rlm@472	1339 Therefore, in order to support multiple listeners, and get the
rlm@472	1340 sound data in a form that the AIs can use, it is necessary to
rlm@472	1341 create a new Device which supports this feature.
rlm@472	1342
rlm@472	1343 Adding a device to OpenAL is rather tricky -- there are five
rlm@472	1344 separate files in the =OpenAL= source tree that must be modified
rlm@472	1345 to do so. I named my device the "Multiple Audio Send" Device, or
rlm@472	1346 =Send= Device for short, since it sends audio data back to the
rlm@472	1347 calling application like an Aux-Send cable on a mixing board.
rlm@472	1348
rlm@472	1349 The main idea behind the Send device is to take advantage of the
rlm@472	1350 fact that LWJGL only manages one /context/ when using OpenAL. A
rlm@472	1351 /context/ is like a container that holds samples and keeps track
rlm@472	1352 of where the listener is. In order to support multiple listeners,
rlm@472	1353 the Send device identifies the LWJGL context as the master
rlm@472	1354 context, and creates any number of slave contexts to represent
rlm@472	1355 additional listeners. Every time the device renders sound, it
rlm@472	1356 synchronizes every source from the master LWJGL context to the
rlm@472	1357 slave contexts. Then, it renders each context separately, using a
rlm@472	1358 different listener for each one. The rendered sound is made
rlm@472	1359 available via JNI to jMonkeyEngine.
rlm@472	1360
rlm@472	1361 Switching between contexts is not the normal operation of a
rlm@472	1362 Device, and one of the problems with doing so is that a Device
rlm@472	1363 normally keeps around a few pieces of state such as the
rlm@472	1364 =ClickRemoval= array above which will become corrupted if the
rlm@472	1365 contexts are not rendered in parallel. The solution is to create a
rlm@472	1366 copy of this normally global device state for each context, and
rlm@472	1367 copy it back and forth into and out of the actual device state
rlm@472	1368 whenever a context is rendered.
rlm@472	1369
rlm@472	1370 The core of the =Send= device is the =syncSources= function, which
rlm@472	1371 does the job of copying all relevant data from one context to
rlm@472	1372 another.
rlm@472	1373
rlm@472	1374 #+caption: Program for extending =OpenAL= to support multiple
rlm@472	1375 #+caption: listeners via context copying/switching.
rlm@472	1376 #+name: sync-openal-sources
rlm@472	1377 #+begin_listing C
rlm@472	1378 void syncSources(ALsource masterSource, ALsource slaveSource,
rlm@472	1379 ALCcontext masterCtx, ALCcontext slaveCtx){
rlm@472	1380 ALuint master = masterSource->source;
rlm@472	1381 ALuint slave = slaveSource->source;
rlm@472	1382 ALCcontext *current = alcGetCurrentContext();
rlm@472	1383
rlm@472	1384 syncSourcef(master,slave,masterCtx,slaveCtx,AL_PITCH);
rlm@472	1385 syncSourcef(master,slave,masterCtx,slaveCtx,AL_GAIN);
rlm@472	1386 syncSourcef(master,slave,masterCtx,slaveCtx,AL_MAX_DISTANCE);
rlm@472	1387 syncSourcef(master,slave,masterCtx,slaveCtx,AL_ROLLOFF_FACTOR);
rlm@472	1388 syncSourcef(master,slave,masterCtx,slaveCtx,AL_REFERENCE_DISTANCE);
rlm@472	1389 syncSourcef(master,slave,masterCtx,slaveCtx,AL_MIN_GAIN);
rlm@472	1390 syncSourcef(master,slave,masterCtx,slaveCtx,AL_MAX_GAIN);
rlm@472	1391 syncSourcef(master,slave,masterCtx,slaveCtx,AL_CONE_OUTER_GAIN);
rlm@472	1392 syncSourcef(master,slave,masterCtx,slaveCtx,AL_CONE_INNER_ANGLE);
rlm@472	1393 syncSourcef(master,slave,masterCtx,slaveCtx,AL_CONE_OUTER_ANGLE);
rlm@472	1394 syncSourcef(master,slave,masterCtx,slaveCtx,AL_SEC_OFFSET);
rlm@472	1395 syncSourcef(master,slave,masterCtx,slaveCtx,AL_SAMPLE_OFFSET);
rlm@472	1396 syncSourcef(master,slave,masterCtx,slaveCtx,AL_BYTE_OFFSET);
rlm@472	1397
rlm@472	1398 syncSource3f(master,slave,masterCtx,slaveCtx,AL_POSITION);
rlm@472	1399 syncSource3f(master,slave,masterCtx,slaveCtx,AL_VELOCITY);
rlm@472	1400 syncSource3f(master,slave,masterCtx,slaveCtx,AL_DIRECTION);
rlm@472	1401
rlm@472	1402 syncSourcei(master,slave,masterCtx,slaveCtx,AL_SOURCE_RELATIVE);
rlm@472	1403 syncSourcei(master,slave,masterCtx,slaveCtx,AL_LOOPING);
rlm@472	1404
rlm@472	1405 alcMakeContextCurrent(masterCtx);
rlm@472	1406 ALint source_type;
rlm@472	1407 alGetSourcei(master, AL_SOURCE_TYPE, &source_type);
rlm@472	1408
rlm@472	1409 // Only static sources are currently synchronized!
rlm@472	1410 if (AL_STATIC == source_type){
rlm@472	1411 ALint master_buffer;
rlm@472	1412 ALint slave_buffer;
rlm@472	1413 alGetSourcei(master, AL_BUFFER, &master_buffer);
rlm@472	1414 alcMakeContextCurrent(slaveCtx);
rlm@472	1415 alGetSourcei(slave, AL_BUFFER, &slave_buffer);
rlm@472	1416 if (master_buffer != slave_buffer){
rlm@472	1417 alSourcei(slave, AL_BUFFER, master_buffer);
rlm@472	1418 }
rlm@472	1419 }
rlm@472	1420
rlm@472	1421 // Synchronize the state of the two sources.
rlm@472	1422 alcMakeContextCurrent(masterCtx);
rlm@472	1423 ALint masterState;
rlm@472	1424 ALint slaveState;
rlm@472	1425
rlm@472	1426 alGetSourcei(master, AL_SOURCE_STATE, &masterState);
rlm@472	1427 alcMakeContextCurrent(slaveCtx);
rlm@472	1428 alGetSourcei(slave, AL_SOURCE_STATE, &slaveState);
rlm@472	1429
rlm@472	1430 if (masterState != slaveState){
rlm@472	1431 switch (masterState){
rlm@472	1432 case AL_INITIAL : alSourceRewind(slave); break;
rlm@472	1433 case AL_PLAYING : alSourcePlay(slave); break;
rlm@472	1434 case AL_PAUSED : alSourcePause(slave); break;
rlm@472	1435 case AL_STOPPED : alSourceStop(slave); break;
rlm@472	1436 }
rlm@472	1437 }
rlm@472	1438 // Restore whatever context was previously active.
rlm@472	1439 alcMakeContextCurrent(current);
rlm@472	1440 }
rlm@472	1441 #+end_listing
rlm@472	1442
rlm@472	1443 With this special context-switching device, and some ugly JNI
rlm@472	1444 bindings that are not worth mentioning, =CORTEX= gains the ability
rlm@472	1445 to access multiple sound streams from =OpenAL=.
rlm@472	1446
rlm@472	1447 #+caption: Program to create an ear from a blender empty node. The ear
rlm@472	1448 #+caption: follows around the nearest physical object and passes
rlm@472	1449 #+caption: all sensory data to a continuation function.
rlm@472	1450 #+name: add-ear
rlm@472	1451 #+begin_listing clojure
rlm@472	1452 (defn add-ear!
rlm@472	1453 "Create a Listener centered on the current position of 'ear
rlm@472	1454 which follows the closest physical node in 'creature and
rlm@472	1455 sends sound data to 'continuation."
rlm@472	1456 [#^Application world #^Node creature #^Spatial ear continuation]
rlm@472	1457 (let [target (closest-node creature ear)
rlm@472	1458 lis (Listener.)
rlm@472	1459 audio-renderer (.getAudioRenderer world)
rlm@472	1460 sp (hearing-pipeline continuation)]
rlm@472	1461 (.setLocation lis (.getWorldTranslation ear))
rlm@472	1462 (.setRotation lis (.getWorldRotation ear))
rlm@472	1463 (bind-sense target lis)
rlm@472	1464 (update-listener-velocity! target lis)
rlm@472	1465 (.addListener audio-renderer lis)
rlm@472	1466 (.registerSoundProcessor audio-renderer lis sp)))
rlm@472	1467 #+end_listing
rlm@472	1468
rlm@472	1469
rlm@472	1470 The =Send= device, unlike most of the other devices in =OpenAL=,
rlm@472	1471 does not render sound unless asked. This enables the system to
rlm@472	1472 slow down or speed up depending on the needs of the AIs who are
rlm@472	1473 using it to listen. If the device tried to render samples in
rlm@472	1474 real-time, a complicated AI whose mind takes 100 seconds of
rlm@472	1475 computer time to simulate 1 second of AI-time would miss almost
rlm@472	1476 all of the sound in its environment!
rlm@472	1477
rlm@472	1478 #+caption: Program to enable arbitrary hearing in =CORTEX=
rlm@472	1479 #+name: hearing
rlm@472	1480 #+begin_listing clojure
rlm@472	1481 (defn hearing-kernel
rlm@472	1482 "Returns a function which returns auditory sensory data when called
rlm@472	1483 inside a running simulation."
rlm@472	1484 [#^Node creature #^Spatial ear]
rlm@472	1485 (let [hearing-data (atom [])
rlm@472	1486 register-listener!
rlm@472	1487 (runonce
rlm@472	1488 (fn [#^Application world]
rlm@472	1489 (add-ear!
rlm@472	1490 world creature ear
rlm@472	1491 (comp #(reset! hearing-data %)
rlm@472	1492 byteBuffer->pulse-vector))))]
rlm@472	1493 (fn [#^Application world]
rlm@472	1494 (register-listener! world)
rlm@472	1495 (let [data @hearing-data
rlm@472	1496 topology
rlm@472	1497 (vec (map #(vector % 0) (range 0 (count data))))]
rlm@472	1498 [topology data]))))
rlm@472	1499
rlm@472	1500 (defn hearing!
rlm@472	1501 "Endow the creature in a particular world with the sense of
rlm@472	1502 hearing. Will return a sequence of functions, one for each ear,
rlm@472	1503 which when called will return the auditory data from that ear."
rlm@472	1504 [#^Node creature]
rlm@472	1505 (for [ear (ears creature)]
rlm@472	1506 (hearing-kernel creature ear)))
rlm@472	1507 #+end_listing
rlm@472	1508
rlm@472	1509 Armed with these functions, =CORTEX= is able to test possibly the
rlm@472	1510 first ever instance of multiple listeners in a video game engine
rlm@472	1511 based simulation!
rlm@472	1512
rlm@472	1513 #+caption: Here a simple creature responds to sound by changing
rlm@472	1514 #+caption: its color from gray to green when the total volume
rlm@472	1515 #+caption: goes over a threshold.
rlm@472	1516 #+name: sound-test
rlm@472	1517 #+begin_listing java
rlm@472	1518 /**
rlm@472	1519 * Respond to sound! This is the brain of an AI entity that
rlm@472	1520 * hears its surroundings and reacts to them.
rlm@472	1521 */
rlm@472	1522 public void process(ByteBuffer audioSamples,
rlm@472	1523 int numSamples, AudioFormat format) {
rlm@472	1524 audioSamples.clear();
rlm@472	1525 byte[] data = new byte[numSamples];
rlm@472	1526 float[] out = new float[numSamples];
rlm@472	1527 audioSamples.get(data);
rlm@472	1528 FloatSampleTools.
rlm@472	1529 byte2floatInterleaved
rlm@472	1530 (data, 0, out, 0, numSamples/format.getFrameSize(), format);
rlm@472	1531
rlm@472	1532 float max = Float.NEGATIVE_INFINITY;
rlm@472	1533 for (float f : out){if (f > max) max = f;}
rlm@472	1534 audioSamples.clear();
rlm@472	1535
rlm@472	1536 if (max > 0.1){
rlm@472	1537 entity.getMaterial().setColor("Color", ColorRGBA.Green);
rlm@472	1538 }
rlm@472	1539 else {
rlm@472	1540 entity.getMaterial().setColor("Color", ColorRGBA.Gray);
rlm@472	1541 }
rlm@472	1542 #+end_listing
rlm@472	1543
rlm@472	1544 #+caption: First ever simulation of multiple listerners in =CORTEX=.
rlm@472	1545 #+caption: Each cube is a creature which processes sound data with
rlm@472	1546 #+caption: the =process= function from listing \ref{sound-test}.
rlm@472	1547 #+caption: the ball is constantally emiting a pure tone of
rlm@472	1548 #+caption: constant volume. As it approaches the cubes, they each
rlm@472	1549 #+caption: change color in response to the sound.
rlm@472	1550 #+name: sound-cubes.
rlm@472	1551 #+ATTR_LaTeX: :width 10cm
rlm@472	1552 [[./images/aurellem-gray.png]]
rlm@472	1553
rlm@472	1554 This system of hearing has also been co-opted by the
rlm@472	1555 jMonkeyEngine3 community and is used to record audio for demo
rlm@472	1556 videos.
rlm@472	1557
rlm@477	1558 ** COMMENT Touch uses hundreds of hair-like elements
rlm@436	1559
rlm@474	1560 Touch is critical to navigation and spatial reasoning and as such I
rlm@474	1561 need a simulated version of it to give to my AI creatures.
rlm@474	1562
rlm@474	1563 Human skin has a wide array of touch sensors, each of which
rlm@474	1564 specialize in detecting different vibrational modes and pressures.
rlm@474	1565 These sensors can integrate a vast expanse of skin (i.e. your
rlm@474	1566 entire palm), or a tiny patch of skin at the tip of your finger.
rlm@474	1567 The hairs of the skin help detect objects before they even come
rlm@474	1568 into contact with the skin proper.
rlm@474	1569
rlm@474	1570 However, touch in my simulated world can not exactly correspond to
rlm@474	1571 human touch because my creatures are made out of completely rigid
rlm@474	1572 segments that don't deform like human skin.
rlm@474	1573
rlm@474	1574 Instead of measuring deformation or vibration, I surround each
rlm@474	1575 rigid part with a plenitude of hair-like objects (/feelers/) which
rlm@474	1576 do not interact with the physical world. Physical objects can pass
rlm@474	1577 through them with no effect. The feelers are able to tell when
rlm@474	1578 other objects pass through them, and they constantly report how
rlm@474	1579 much of their extent is covered. So even though the creature's body
rlm@474	1580 parts do not deform, the feelers create a margin around those body
rlm@474	1581 parts which achieves a sense of touch which is a hybrid between a
rlm@474	1582 human's sense of deformation and sense from hairs.
rlm@474	1583
rlm@474	1584 Implementing touch in jMonkeyEngine follows a different technical
rlm@474	1585 route than vision and hearing. Those two senses piggybacked off
rlm@474	1586 jMonkeyEngine's 3D audio and video rendering subsystems. To
rlm@474	1587 simulate touch, I use jMonkeyEngine's physics system to execute
rlm@474	1588 many small collision detections, one for each feeler. The placement
rlm@474	1589 of the feelers is determined by a UV-mapped image which shows where
rlm@474	1590 each feeler should be on the 3D surface of the body.
rlm@474	1591
rlm@477	1592 *** Defining Touch Meta-Data in Blender
rlm@474	1593
rlm@474	1594 Each geometry can have a single UV map which describes the
rlm@474	1595 position of the feelers which will constitute its sense of touch.
rlm@474	1596 This image path is stored under the ``touch'' key. The image itself
rlm@474	1597 is black and white, with black meaning a feeler length of 0 (no
rlm@474	1598 feeler is present) and white meaning a feeler length of =scale=,
rlm@474	1599 which is a float stored under the key "scale".
rlm@474	1600
rlm@475	1601 #+caption: Touch does not use empty nodes, to store metadata,
rlm@475	1602 #+caption: because the metadata of each solid part of a
rlm@475	1603 #+caption: creature's body is sufficient.
rlm@475	1604 #+name: touch-meta-data
rlm@475	1605 #+begin_listing clojure
rlm@477	1606 #+BEGIN_SRC clojure
rlm@474	1607 (defn tactile-sensor-profile
rlm@474	1608 "Return the touch-sensor distribution image in BufferedImage format,
rlm@474	1609 or nil if it does not exist."
rlm@474	1610 [#^Geometry obj]
rlm@474	1611 (if-let [image-path (meta-data obj "touch")]
rlm@474	1612 (load-image image-path)))
rlm@474	1613
rlm@474	1614 (defn tactile-scale
rlm@474	1615 "Return the length of each feeler. Default scale is 0.01
rlm@474	1616 jMonkeyEngine units."
rlm@474	1617 [#^Geometry obj]
rlm@474	1618 (if-let [scale (meta-data obj "scale")]
rlm@474	1619 scale 0.1))
rlm@477	1620 #+END_SRC
rlm@475	1621 #+end_listing
rlm@474	1622
rlm@475	1623 Here is an example of a UV-map which specifies the position of
rlm@475	1624 touch sensors along the surface of the upper segment of a fingertip.
rlm@474	1625
rlm@475	1626 #+caption: This is the tactile-sensor-profile for the upper segment
rlm@475	1627 #+caption: of a fingertip. It defines regions of high touch sensitivity
rlm@475	1628 #+caption: (where there are many white pixels) and regions of low
rlm@475	1629 #+caption: sensitivity (where white pixels are sparse).
rlm@475	1630 #+name: fimgertip-UV
rlm@477	1631 #+ATTR_LaTeX: :width 13cm
rlm@477	1632 [[./images/finger-UV.png]]
rlm@474	1633
rlm@477	1634 *** Implementation Summary
rlm@474	1635
rlm@474	1636 To simulate touch there are three conceptual steps. For each solid
rlm@474	1637 object in the creature, you first have to get UV image and scale
rlm@474	1638 parameter which define the position and length of the feelers.
rlm@474	1639 Then, you use the triangles which comprise the mesh and the UV
rlm@474	1640 data stored in the mesh to determine the world-space position and
rlm@474	1641 orientation of each feeler. Then once every frame, update these
rlm@474	1642 positions and orientations to match the current position and
rlm@474	1643 orientation of the object, and use physics collision detection to
rlm@474	1644 gather tactile data.
rlm@474	1645
rlm@474	1646 Extracting the meta-data has already been described. The third
rlm@474	1647 step, physics collision detection, is handled in =touch-kernel=.
rlm@474	1648 Translating the positions and orientations of the feelers from the
rlm@474	1649 UV-map to world-space is itself a three-step process.
rlm@474	1650
rlm@475	1651 - Find the triangles which make up the mesh in pixel-space and in
rlm@475	1652 world-space. (=triangles= =pixel-triangles=).
rlm@474	1653
rlm@475	1654 - Find the coordinates of each feeler in world-space. These are
rlm@475	1655 the origins of the feelers. (=feeler-origins=).
rlm@474	1656
rlm@475	1657 - Calculate the normals of the triangles in world space, and add
rlm@475	1658 them to each of the origins of the feelers. These are the
rlm@475	1659 normalized coordinates of the tips of the feelers.
rlm@475	1660 (=feeler-tips=).
rlm@474	1661
rlm@477	1662 *** Triangle Math
rlm@474	1663
rlm@475	1664 The rigid objects which make up a creature have an underlying
rlm@475	1665 =Geometry=, which is a =Mesh= plus a =Material= and other
rlm@475	1666 important data involved with displaying the object.
rlm@475	1667
rlm@475	1668 A =Mesh= is composed of =Triangles=, and each =Triangle= has three
rlm@475	1669 vertices which have coordinates in world space and UV space.
rlm@475	1670
rlm@475	1671 Here, =triangles= gets all the world-space triangles which
rlm@475	1672 comprise a mesh, while =pixel-triangles= gets those same triangles
rlm@475	1673 expressed in pixel coordinates (which are UV coordinates scaled to
rlm@475	1674 fit the height and width of the UV image).
rlm@474	1675
rlm@475	1676 #+caption: Programs to extract triangles from a geometry and get
rlm@475	1677 #+caption: their verticies in both world and UV-coordinates.
rlm@475	1678 #+name: get-triangles
rlm@475	1679 #+begin_listing clojure
rlm@477	1680 #+BEGIN_SRC clojure
rlm@474	1681 (defn triangle
rlm@474	1682 "Get the triangle specified by triangle-index from the mesh."
rlm@474	1683 [#^Geometry geo triangle-index]
rlm@474	1684 (triangle-seq
rlm@474	1685 (let [scratch (Triangle.)]
rlm@474	1686 (.getTriangle (.getMesh geo) triangle-index scratch) scratch)))
rlm@474	1687
rlm@474	1688 (defn triangles
rlm@474	1689 "Return a sequence of all the Triangles which comprise a given
rlm@474	1690 Geometry."
rlm@474	1691 [#^Geometry geo]
rlm@474	1692 (map (partial triangle geo) (range (.getTriangleCount (.getMesh geo)))))
rlm@474	1693
rlm@474	1694 (defn triangle-vertex-indices
rlm@474	1695 "Get the triangle vertex indices of a given triangle from a given
rlm@474	1696 mesh."
rlm@474	1697 [#^Mesh mesh triangle-index]
rlm@474	1698 (let [indices (int-array 3)]
rlm@474	1699 (.getTriangle mesh triangle-index indices)
rlm@474	1700 (vec indices)))
rlm@474	1701
rlm@475	1702 (defn vertex-UV-coord
rlm@474	1703 "Get the UV-coordinates of the vertex named by vertex-index"
rlm@474	1704 [#^Mesh mesh vertex-index]
rlm@474	1705 (let [UV-buffer
rlm@474	1706 (.getData
rlm@474	1707 (.getBuffer
rlm@474	1708 mesh
rlm@474	1709 VertexBuffer$Type/TexCoord))]
rlm@474	1710 [(.get UV-buffer (* vertex-index 2))
rlm@474	1711 (.get UV-buffer (+ 1 (* vertex-index 2)))]))
rlm@474	1712
rlm@474	1713 (defn pixel-triangle [#^Geometry geo image index]
rlm@474	1714 (let [mesh (.getMesh geo)
rlm@474	1715 width (.getWidth image)
rlm@474	1716 height (.getHeight image)]
rlm@474	1717 (vec (map (fn [[u v]] (vector (* width u) (* height v)))
rlm@474	1718 (map (partial vertex-UV-coord mesh)
rlm@474	1719 (triangle-vertex-indices mesh index))))))
rlm@474	1720
rlm@474	1721 (defn pixel-triangles
rlm@474	1722 "The pixel-space triangles of the Geometry, in the same order as
rlm@474	1723 (triangles geo)"
rlm@474	1724 [#^Geometry geo image]
rlm@474	1725 (let [height (.getHeight image)
rlm@474	1726 width (.getWidth image)]
rlm@474	1727 (map (partial pixel-triangle geo image)
rlm@474	1728 (range (.getTriangleCount (.getMesh geo))))))
rlm@477	1729 #+END_SRC
rlm@475	1730 #+end_listing
rlm@475	1731
rlm@474	1732 *** The Affine Transform from one Triangle to Another
rlm@474	1733
rlm@475	1734 =pixel-triangles= gives us the mesh triangles expressed in pixel
rlm@475	1735 coordinates and =triangles= gives us the mesh triangles expressed
rlm@475	1736 in world coordinates. The tactile-sensor-profile gives the
rlm@475	1737 position of each feeler in pixel-space. In order to convert
rlm@475	1738 pixel-space coordinates into world-space coordinates we need
rlm@475	1739 something that takes coordinates on the surface of one triangle
rlm@475	1740 and gives the corresponding coordinates on the surface of another
rlm@475	1741 triangle.
rlm@475	1742
rlm@475	1743 Triangles are [[http://mathworld.wolfram.com/AffineTransformation.html ][affine]], which means any triangle can be transformed
rlm@475	1744 into any other by a combination of translation, scaling, and
rlm@475	1745 rotation. The affine transformation from one triangle to another
rlm@475	1746 is readily computable if the triangle is expressed in terms of a
rlm@475	1747 $4x4$ matrix.
rlm@476	1748
rlm@476	1749 #+BEGIN_LaTeX
rlm@476	1750 $$
rlm@475	1751 \begin{bmatrix}
rlm@475	1752 x_1 & x_2 & x_3 & n_x \\
rlm@475	1753 y_1 & y_2 & y_3 & n_y \\
rlm@475	1754 z_1 & z_2 & z_3 & n_z \\
rlm@475	1755 1 & 1 & 1 & 1
rlm@475	1756 \end{bmatrix}
rlm@476	1757 $$
rlm@476	1758 #+END_LaTeX
rlm@475	1759
rlm@475	1760 Here, the first three columns of the matrix are the vertices of
rlm@475	1761 the triangle. The last column is the right-handed unit normal of
rlm@475	1762 the triangle.
rlm@475	1763
rlm@476	1764 With two triangles $T_{1}$ and $T_{2}$ each expressed as a
rlm@476	1765 matrix like above, the affine transform from $T_{1}$ to $T_{2}$
rlm@476	1766 is $T_{2}T_{1}^{-1}$.
rlm@475	1767
rlm@475	1768 The clojure code below recapitulates the formulas above, using
rlm@475	1769 jMonkeyEngine's =Matrix4f= objects, which can describe any affine
rlm@475	1770 transformation.
rlm@474	1771
rlm@475	1772 #+caption: Program to interpert triangles as affine transforms.
rlm@475	1773 #+name: triangle-affine
rlm@475	1774 #+begin_listing clojure
rlm@475	1775 #+BEGIN_SRC clojure
rlm@474	1776 (defn triangle->matrix4f
rlm@474	1777 "Converts the triangle into a 4x4 matrix: The first three columns
rlm@474	1778 contain the vertices of the triangle; the last contains the unit
rlm@474	1779 normal of the triangle. The bottom row is filled with 1s."
rlm@474	1780 [#^Triangle t]
rlm@474	1781 (let [mat (Matrix4f.)
rlm@474	1782 [vert-1 vert-2 vert-3]
rlm@474	1783 (mapv #(.get t %) (range 3))
rlm@474	1784 unit-normal (do (.calculateNormal t)(.getNormal t))
rlm@474	1785 vertices [vert-1 vert-2 vert-3 unit-normal]]
rlm@474	1786 (dorun
rlm@474	1787 (for [row (range 4) col (range 3)]
rlm@474	1788 (do
rlm@474	1789 (.set mat col row (.get (vertices row) col))
rlm@474	1790 (.set mat 3 row 1)))) mat))
rlm@474	1791
rlm@474	1792 (defn triangles->affine-transform
rlm@474	1793 "Returns the affine transformation that converts each vertex in the
rlm@474	1794 first triangle into the corresponding vertex in the second
rlm@474	1795 triangle."
rlm@474	1796 [#^Triangle tri-1 #^Triangle tri-2]
rlm@474	1797 (.mult
rlm@474	1798 (triangle->matrix4f tri-2)
rlm@474	1799 (.invert (triangle->matrix4f tri-1))))
rlm@475	1800 #+END_SRC
rlm@475	1801 #+end_listing
rlm@474	1802
rlm@477	1803 *** Triangle Boundaries
rlm@474	1804
rlm@474	1805 For efficiency's sake I will divide the tactile-profile image into
rlm@474	1806 small squares which inscribe each pixel-triangle, then extract the
rlm@474	1807 points which lie inside the triangle and map them to 3D-space using
rlm@474	1808 =triangle-transform= above. To do this I need a function,
rlm@474	1809 =convex-bounds= which finds the smallest box which inscribes a 2D
rlm@474	1810 triangle.
rlm@474	1811
rlm@474	1812 =inside-triangle?= determines whether a point is inside a triangle
rlm@474	1813 in 2D pixel-space.
rlm@474	1814
rlm@475	1815 #+caption: Program to efficiently determine point includion
rlm@475	1816 #+caption: in a triangle.
rlm@475	1817 #+name: in-triangle
rlm@475	1818 #+begin_listing clojure
rlm@475	1819 #+BEGIN_SRC clojure
rlm@474	1820 (defn convex-bounds
rlm@474	1821 "Returns the smallest square containing the given vertices, as a
rlm@474	1822 vector of integers [left top width height]."
rlm@474	1823 [verts]
rlm@474	1824 (let [xs (map first verts)
rlm@474	1825 ys (map second verts)
rlm@474	1826 x0 (Math/floor (apply min xs))
rlm@474	1827 y0 (Math/floor (apply min ys))
rlm@474	1828 x1 (Math/ceil (apply max xs))
rlm@474	1829 y1 (Math/ceil (apply max ys))]
rlm@474	1830 [x0 y0 (- x1 x0) (- y1 y0)]))
rlm@474	1831
rlm@474	1832 (defn same-side?
rlm@474	1833 "Given the points p1 and p2 and the reference point ref, is point p
rlm@474	1834 on the same side of the line that goes through p1 and p2 as ref is?"
rlm@474	1835 [p1 p2 ref p]
rlm@474	1836 (<=
rlm@474	1837 0
rlm@474	1838 (.dot
rlm@474	1839 (.cross (.subtract p2 p1) (.subtract p p1))
rlm@474	1840 (.cross (.subtract p2 p1) (.subtract ref p1)))))
rlm@474	1841
rlm@474	1842 (defn inside-triangle?
rlm@474	1843 "Is the point inside the triangle?"
rlm@474	1844 {:author "Dylan Holmes"}
rlm@474	1845 [#^Triangle tri #^Vector3f p]
rlm@474	1846 (let [[vert-1 vert-2 vert-3] [(.get1 tri) (.get2 tri) (.get3 tri)]]
rlm@474	1847 (and
rlm@474	1848 (same-side? vert-1 vert-2 vert-3 p)
rlm@474	1849 (same-side? vert-2 vert-3 vert-1 p)
rlm@474	1850 (same-side? vert-3 vert-1 vert-2 p))))
rlm@475	1851 #+END_SRC
rlm@475	1852 #+end_listing
rlm@474	1853
rlm@477	1854 *** Feeler Coordinates
rlm@474	1855
rlm@475	1856 The triangle-related functions above make short work of
rlm@475	1857 calculating the positions and orientations of each feeler in
rlm@475	1858 world-space.
rlm@474	1859
rlm@475	1860 #+caption: Program to get the coordinates of ``feelers '' in
rlm@475	1861 #+caption: both world and UV-coordinates.
rlm@475	1862 #+name: feeler-coordinates
rlm@475	1863 #+begin_listing clojure
rlm@475	1864 #+BEGIN_SRC clojure
rlm@474	1865 (defn feeler-pixel-coords
rlm@474	1866 "Returns the coordinates of the feelers in pixel space in lists, one
rlm@474	1867 list for each triangle, ordered in the same way as (triangles) and
rlm@474	1868 (pixel-triangles)."
rlm@474	1869 [#^Geometry geo image]
rlm@474	1870 (map
rlm@474	1871 (fn [pixel-triangle]
rlm@474	1872 (filter
rlm@474	1873 (fn [coord]
rlm@474	1874 (inside-triangle? (->triangle pixel-triangle)
rlm@474	1875 (->vector3f coord)))
rlm@474	1876 (white-coordinates image (convex-bounds pixel-triangle))))
rlm@474	1877 (pixel-triangles geo image)))
rlm@474	1878
rlm@474	1879 (defn feeler-world-coords
rlm@474	1880 "Returns the coordinates of the feelers in world space in lists, one
rlm@474	1881 list for each triangle, ordered in the same way as (triangles) and
rlm@474	1882 (pixel-triangles)."
rlm@474	1883 [#^Geometry geo image]
rlm@474	1884 (let [transforms
rlm@474	1885 (map #(triangles->affine-transform
rlm@474	1886 (->triangle %1) (->triangle %2))
rlm@474	1887 (pixel-triangles geo image)
rlm@474	1888 (triangles geo))]
rlm@474	1889 (map (fn [transform coords]
rlm@474	1890 (map #(.mult transform (->vector3f %)) coords))
rlm@474	1891 transforms (feeler-pixel-coords geo image))))
rlm@475	1892 #+END_SRC
rlm@475	1893 #+end_listing
rlm@474	1894
rlm@475	1895 #+caption: Program to get the position of the base and tip of
rlm@475	1896 #+caption: each ``feeler''
rlm@475	1897 #+name: feeler-tips
rlm@475	1898 #+begin_listing clojure
rlm@475	1899 #+BEGIN_SRC clojure
rlm@474	1900 (defn feeler-origins
rlm@474	1901 "The world space coordinates of the root of each feeler."
rlm@474	1902 [#^Geometry geo image]
rlm@474	1903 (reduce concat (feeler-world-coords geo image)))
rlm@474	1904
rlm@474	1905 (defn feeler-tips
rlm@474	1906 "The world space coordinates of the tip of each feeler."
rlm@474	1907 [#^Geometry geo image]
rlm@474	1908 (let [world-coords (feeler-world-coords geo image)
rlm@474	1909 normals
rlm@474	1910 (map
rlm@474	1911 (fn [triangle]
rlm@474	1912 (.calculateNormal triangle)
rlm@474	1913 (.clone (.getNormal triangle)))
rlm@474	1914 (map ->triangle (triangles geo)))]
rlm@474	1915
rlm@474	1916 (mapcat (fn [origins normal]
rlm@474	1917 (map #(.add % normal) origins))
rlm@474	1918 world-coords normals)))
rlm@474	1919
rlm@474	1920 (defn touch-topology
rlm@474	1921 [#^Geometry geo image]
rlm@474	1922 (collapse (reduce concat (feeler-pixel-coords geo image))))
rlm@475	1923 #+END_SRC
rlm@475	1924 #+end_listing
rlm@474	1925
rlm@477	1926 *** Simulated Touch
rlm@474	1927
rlm@475	1928 Now that the functions to construct feelers are complete,
rlm@475	1929 =touch-kernel= generates functions to be called from within a
rlm@475	1930 simulation that perform the necessary physics collisions to
rlm@475	1931 collect tactile data, and =touch!= recursively applies it to every
rlm@475	1932 node in the creature.
rlm@474	1933
rlm@475	1934 #+caption: Efficient program to transform a ray from
rlm@475	1935 #+caption: one position to another.
rlm@475	1936 #+name: set-ray
rlm@475	1937 #+begin_listing clojure
rlm@475	1938 #+BEGIN_SRC clojure
rlm@474	1939 (defn set-ray [#^Ray ray #^Matrix4f transform
rlm@474	1940 #^Vector3f origin #^Vector3f tip]
rlm@474	1941 ;; Doing everything locally reduces garbage collection by enough to
rlm@474	1942 ;; be worth it.
rlm@474	1943 (.mult transform origin (.getOrigin ray))
rlm@474	1944 (.mult transform tip (.getDirection ray))
rlm@474	1945 (.subtractLocal (.getDirection ray) (.getOrigin ray))
rlm@474	1946 (.normalizeLocal (.getDirection ray)))
rlm@475	1947 #+END_SRC
rlm@475	1948 #+end_listing
rlm@474	1949
rlm@475	1950 #+caption: This is the core of touch in =CORTEX= each feeler
rlm@475	1951 #+caption: follows the object it is bound to, reporting any
rlm@475	1952 #+caption: collisions that may happen.
rlm@475	1953 #+name: touch-kernel
rlm@475	1954 #+begin_listing clojure
rlm@475	1955 #+BEGIN_SRC clojure
rlm@474	1956 (defn touch-kernel
rlm@474	1957 "Constructs a function which will return tactile sensory data from
rlm@474	1958 'geo when called from inside a running simulation"
rlm@474	1959 [#^Geometry geo]
rlm@474	1960 (if-let
rlm@474	1961 [profile (tactile-sensor-profile geo)]
rlm@474	1962 (let [ray-reference-origins (feeler-origins geo profile)
rlm@474	1963 ray-reference-tips (feeler-tips geo profile)
rlm@474	1964 ray-length (tactile-scale geo)
rlm@474	1965 current-rays (map (fn [_] (Ray.)) ray-reference-origins)
rlm@474	1966 topology (touch-topology geo profile)
rlm@474	1967 correction (float (* ray-length -0.2))]
rlm@474	1968 ;; slight tolerance for very close collisions.
rlm@474	1969 (dorun
rlm@474	1970 (map (fn [origin tip]
rlm@474	1971 (.addLocal origin (.mult (.subtract tip origin)
rlm@474	1972 correction)))
rlm@474	1973 ray-reference-origins ray-reference-tips))
rlm@474	1974 (dorun (map #(.setLimit % ray-length) current-rays))
rlm@474	1975 (fn [node]
rlm@474	1976 (let [transform (.getWorldMatrix geo)]
rlm@474	1977 (dorun
rlm@474	1978 (map (fn [ray ref-origin ref-tip]
rlm@474	1979 (set-ray ray transform ref-origin ref-tip))
rlm@474	1980 current-rays ray-reference-origins
rlm@474	1981 ray-reference-tips))
rlm@474	1982 (vector
rlm@474	1983 topology
rlm@474	1984 (vec
rlm@474	1985 (for [ray current-rays]
rlm@474	1986 (do
rlm@474	1987 (let [results (CollisionResults.)]
rlm@474	1988 (.collideWith node ray results)
rlm@474	1989 (let [touch-objects
rlm@474	1990 (filter #(not (= geo (.getGeometry %)))
rlm@474	1991 results)
rlm@474	1992 limit (.getLimit ray)]
rlm@474	1993 [(if (empty? touch-objects)
rlm@474	1994 limit
rlm@474	1995 (let [response
rlm@474	1996 (apply min (map #(.getDistance %)
rlm@474	1997 touch-objects))]
rlm@474	1998 (FastMath/clamp
rlm@474	1999 (float
rlm@474	2000 (if (> response limit) (float 0.0)
rlm@474	2001 (+ response correction)))
rlm@474	2002 (float 0.0)
rlm@474	2003 limit)))
rlm@474	2004 limit])))))))))))
rlm@475	2005 #+END_SRC
rlm@475	2006 #+end_listing
rlm@474	2007
rlm@475	2008 Armed with the =touch!= function, =CORTEX= becomes capable of
rlm@475	2009 giving creatures a sense of touch. A simple test is to create a
rlm@475	2010 cube that is outfitted with a uniform distrubition of touch
rlm@475	2011 sensors. It can feel the ground and any balls that it touches.
rlm@475	2012
rlm@475	2013 #+caption: =CORTEX= interface for creating touch in a simulated
rlm@475	2014 #+caption: creature.
rlm@475	2015 #+name: touch
rlm@475	2016 #+begin_listing clojure
rlm@475	2017 #+BEGIN_SRC clojure
rlm@474	2018 (defn touch!
rlm@474	2019 "Endow the creature with the sense of touch. Returns a sequence of
rlm@474	2020 functions, one for each body part with a tactile-sensor-profile,
rlm@474	2021 each of which when called returns sensory data for that body part."
rlm@474	2022 [#^Node creature]
rlm@474	2023 (filter
rlm@474	2024 (comp not nil?)
rlm@474	2025 (map touch-kernel
rlm@474	2026 (filter #(isa? (class %) Geometry)
rlm@474	2027 (node-seq creature)))))
rlm@475	2028 #+END_SRC
rlm@475	2029 #+end_listing
rlm@475	2030
rlm@475	2031 The tactile-sensor-profile image for the touch cube is a simple
rlm@475	2032 cross with a unifom distribution of touch sensors:
rlm@474	2033
rlm@475	2034 #+caption: The touch profile for the touch-cube. Each pure white
rlm@475	2035 #+caption: pixel defines a touch sensitive feeler.
rlm@475	2036 #+name: touch-cube-uv-map
rlm@475	2037 #+ATTR_LaTeX: :width 10cm
rlm@475	2038 [[./images/touch-profile.png]]
rlm@474	2039
rlm@475	2040 #+caption: The touch cube reacts to canonballs. The black, red,
rlm@475	2041 #+caption: and white cross on the right is a visual display of
rlm@475	2042 #+caption: the creature's touch. White means that it is feeling
rlm@475	2043 #+caption: something strongly, black is not feeling anything,
rlm@475	2044 #+caption: and gray is in-between. The cube can feel both the
rlm@475	2045 #+caption: floor and the ball. Notice that when the ball causes
rlm@475	2046 #+caption: the cube to tip, that the bottom face can still feel
rlm@475	2047 #+caption: part of the ground.
rlm@475	2048 #+name: touch-cube-uv-map
rlm@475	2049 #+ATTR_LaTeX: :width 15cm
rlm@475	2050 [[./images/touch-cube.png]]
rlm@474	2051
rlm@480	2052 ** COMMENT Proprioception is the sense that makes everything ``real''
rlm@436	2053
rlm@479	2054 Close your eyes, and touch your nose with your right index finger.
rlm@479	2055 How did you do it? You could not see your hand, and neither your
rlm@479	2056 hand nor your nose could use the sense of touch to guide the path
rlm@479	2057 of your hand. There are no sound cues, and Taste and Smell
rlm@479	2058 certainly don't provide any help. You know where your hand is
rlm@479	2059 without your other senses because of Proprioception.
rlm@479	2060
rlm@479	2061 Humans can sometimes loose this sense through viral infections or
rlm@479	2062 damage to the spinal cord or brain, and when they do, they loose
rlm@479	2063 the ability to control their own bodies without looking directly at
rlm@479	2064 the parts they want to move. In [[http://en.wikipedia.org/wiki/The_Man_Who_Mistook_His_Wife_for_a_Hat][The Man Who Mistook His Wife for a
rlm@479	2065 Hat]], a woman named Christina looses this sense and has to learn how
rlm@479	2066 to move by carefully watching her arms and legs. She describes
rlm@479	2067 proprioception as the "eyes of the body, the way the body sees
rlm@479	2068 itself".
rlm@479	2069
rlm@479	2070 Proprioception in humans is mediated by [[http://en.wikipedia.org/wiki/Articular_capsule][joint capsules]], [[http://en.wikipedia.org/wiki/Muscle_spindle][muscle
rlm@479	2071 spindles]], and the [[http://en.wikipedia.org/wiki/Golgi_tendon_organ][Golgi tendon organs]]. These measure the relative
rlm@479	2072 positions of each body part by monitoring muscle strain and length.
rlm@479	2073
rlm@479	2074 It's clear that this is a vital sense for fluid, graceful movement.
rlm@479	2075 It's also particularly easy to implement in jMonkeyEngine.
rlm@479	2076
rlm@479	2077 My simulated proprioception calculates the relative angles of each
rlm@479	2078 joint from the rest position defined in the blender file. This
rlm@479	2079 simulates the muscle-spindles and joint capsules. I will deal with
rlm@479	2080 Golgi tendon organs, which calculate muscle strain, in the next
rlm@479	2081 section.
rlm@479	2082
rlm@479	2083 *** Helper functions
rlm@479	2084
rlm@479	2085 =absolute-angle= calculates the angle between two vectors,
rlm@479	2086 relative to a third axis vector. This angle is the number of
rlm@479	2087 radians you have to move counterclockwise around the axis vector
rlm@479	2088 to get from the first to the second vector. It is not commutative
rlm@479	2089 like a normal dot-product angle is.
rlm@479	2090
rlm@479	2091 The purpose of these functions is to build a system of angle
rlm@479	2092 measurement that is biologically plausable.
rlm@479	2093
rlm@479	2094 #+caption: Program to measure angles along a vector
rlm@479	2095 #+name: helpers
rlm@479	2096 #+begin_listing clojure
rlm@479	2097 #+BEGIN_SRC clojure
rlm@479	2098 (defn right-handed?
rlm@479	2099 "true iff the three vectors form a right handed coordinate
rlm@479	2100 system. The three vectors do not have to be normalized or
rlm@479	2101 orthogonal."
rlm@479	2102 [vec1 vec2 vec3]
rlm@479	2103 (pos? (.dot (.cross vec1 vec2) vec3)))
rlm@479	2104
rlm@479	2105 (defn absolute-angle
rlm@479	2106 "The angle between 'vec1 and 'vec2 around 'axis. In the range
rlm@479	2107 [0 (* 2 Math/PI)]."
rlm@479	2108 [vec1 vec2 axis]
rlm@479	2109 (let [angle (.angleBetween vec1 vec2)]
rlm@479	2110 (if (right-handed? vec1 vec2 axis)
rlm@479	2111 angle (- (* 2 Math/PI) angle))))
rlm@479	2112 #+END_SRC
rlm@479	2113 #+end_listing
rlm@479	2114
rlm@479	2115 *** Proprioception Kernel
rlm@479	2116
rlm@479	2117 Given a joint, =proprioception-kernel= produces a function that
rlm@479	2118 calculates the Euler angles between the the objects the joint
rlm@479	2119 connects. The only tricky part here is making the angles relative
rlm@479	2120 to the joint's initial ``straightness''.
rlm@479	2121
rlm@479	2122 #+caption: Program to return biologially reasonable proprioceptive
rlm@479	2123 #+caption: data for each joint.
rlm@479	2124 #+name: proprioception
rlm@479	2125 #+begin_listing clojure
rlm@479	2126 #+BEGIN_SRC clojure
rlm@479	2127 (defn proprioception-kernel
rlm@479	2128 "Returns a function which returns proprioceptive sensory data when
rlm@479	2129 called inside a running simulation."
rlm@479	2130 [#^Node parts #^Node joint]
rlm@479	2131 (let [[obj-a obj-b] (joint-targets parts joint)
rlm@479	2132 joint-rot (.getWorldRotation joint)
rlm@479	2133 x0 (.mult joint-rot Vector3f/UNIT_X)
rlm@479	2134 y0 (.mult joint-rot Vector3f/UNIT_Y)
rlm@479	2135 z0 (.mult joint-rot Vector3f/UNIT_Z)]
rlm@479	2136 (fn []
rlm@479	2137 (let [rot-a (.clone (.getWorldRotation obj-a))
rlm@479	2138 rot-b (.clone (.getWorldRotation obj-b))
rlm@479	2139 x (.mult rot-a x0)
rlm@479	2140 y (.mult rot-a y0)
rlm@479	2141 z (.mult rot-a z0)
rlm@479	2142
rlm@479	2143 X (.mult rot-b x0)
rlm@479	2144 Y (.mult rot-b y0)
rlm@479	2145 Z (.mult rot-b z0)
rlm@479	2146 heading (Math/atan2 (.dot X z) (.dot X x))
rlm@479	2147 pitch (Math/atan2 (.dot X y) (.dot X x))
rlm@479	2148
rlm@479	2149 ;; rotate x-vector back to origin
rlm@479	2150 reverse
rlm@479	2151 (doto (Quaternion.)
rlm@479	2152 (.fromAngleAxis
rlm@479	2153 (.angleBetween X x)
rlm@479	2154 (let [cross (.normalize (.cross X x))]
rlm@479	2155 (if (= 0 (.length cross)) y cross))))
rlm@479	2156 roll (absolute-angle (.mult reverse Y) y x)]
rlm@479	2157 [heading pitch roll]))))
rlm@479	2158
rlm@479	2159 (defn proprioception!
rlm@479	2160 "Endow the creature with the sense of proprioception. Returns a
rlm@479	2161 sequence of functions, one for each child of the \"joints\" node in
rlm@479	2162 the creature, which each report proprioceptive information about
rlm@479	2163 that joint."
rlm@479	2164 [#^Node creature]
rlm@479	2165 ;; extract the body's joints
rlm@479	2166 (let [senses (map (partial proprioception-kernel creature)
rlm@479	2167 (joints creature))]
rlm@479	2168 (fn []
rlm@479	2169 (map #(%) senses))))
rlm@479	2170 #+END_SRC
rlm@479	2171 #+end_listing
rlm@479	2172
rlm@479	2173 =proprioception!= maps =proprioception-kernel= across all the
rlm@479	2174 joints of the creature. It uses the same list of joints that
rlm@479	2175 =joints= uses. Proprioception is the easiest sense to implement in
rlm@479	2176 =CORTEX=, and it will play a crucial role when efficiently
rlm@479	2177 implementing empathy.
rlm@479	2178
rlm@479	2179 #+caption: In the upper right corner, the three proprioceptive
rlm@479	2180 #+caption: angle measurements are displayed. Red is yaw, Green is
rlm@479	2181 #+caption: pitch, and White is roll.
rlm@479	2182 #+name: proprio
rlm@479	2183 #+ATTR_LaTeX: :width 11cm
rlm@479	2184 [[./images/proprio.png]]
rlm@479	2185
rlm@481	2186 ** COMMENT Muscles are both effectors and sensors
rlm@481	2187
rlm@481	2188 Surprisingly enough, terrestrial creatures only move by using
rlm@481	2189 torque applied about their joints. There's not a single straight
rlm@481	2190 line of force in the human body at all! (A straight line of force
rlm@481	2191 would correspond to some sort of jet or rocket propulsion.)
rlm@481	2192
rlm@481	2193 In humans, muscles are composed of muscle fibers which can contract
rlm@481	2194 to exert force. The muscle fibers which compose a muscle are
rlm@481	2195 partitioned into discrete groups which are each controlled by a
rlm@481	2196 single alpha motor neuron. A single alpha motor neuron might
rlm@481	2197 control as little as three or as many as one thousand muscle
rlm@481	2198 fibers. When the alpha motor neuron is engaged by the spinal cord,
rlm@481	2199 it activates all of the muscle fibers to which it is attached. The
rlm@481	2200 spinal cord generally engages the alpha motor neurons which control
rlm@481	2201 few muscle fibers before the motor neurons which control many
rlm@481	2202 muscle fibers. This recruitment strategy allows for precise
rlm@481	2203 movements at low strength. The collection of all motor neurons that
rlm@481	2204 control a muscle is called the motor pool. The brain essentially
rlm@481	2205 says "activate 30% of the motor pool" and the spinal cord recruits
rlm@481	2206 motor neurons until 30% are activated. Since the distribution of
rlm@481	2207 power among motor neurons is unequal and recruitment goes from
rlm@481	2208 weakest to strongest, the first 30% of the motor pool might be 5%
rlm@481	2209 of the strength of the muscle.
rlm@481	2210
rlm@481	2211 My simulated muscles follow a similar design: Each muscle is
rlm@481	2212 defined by a 1-D array of numbers (the "motor pool"). Each entry in
rlm@481	2213 the array represents a motor neuron which controls a number of
rlm@481	2214 muscle fibers equal to the value of the entry. Each muscle has a
rlm@481	2215 scalar strength factor which determines the total force the muscle
rlm@481	2216 can exert when all motor neurons are activated. The effector
rlm@481	2217 function for a muscle takes a number to index into the motor pool,
rlm@481	2218 and then "activates" all the motor neurons whose index is lower or
rlm@481	2219 equal to the number. Each motor-neuron will apply force in
rlm@481	2220 proportion to its value in the array. Lower values cause less
rlm@481	2221 force. The lower values can be put at the "beginning" of the 1-D
rlm@481	2222 array to simulate the layout of actual human muscles, which are
rlm@481	2223 capable of more precise movements when exerting less force. Or, the
rlm@481	2224 motor pool can simulate more exotic recruitment strategies which do
rlm@481	2225 not correspond to human muscles.
rlm@481	2226
rlm@481	2227 This 1D array is defined in an image file for ease of
rlm@481	2228 creation/visualization. Here is an example muscle profile image.
rlm@481	2229
rlm@481	2230 #+caption: A muscle profile image that describes the strengths
rlm@481	2231 #+caption: of each motor neuron in a muscle. White is weakest
rlm@481	2232 #+caption: and dark red is strongest. This particular pattern
rlm@481	2233 #+caption: has weaker motor neurons at the beginning, just
rlm@481	2234 #+caption: like human muscle.
rlm@481	2235 #+name: muscle-recruit
rlm@481	2236 #+ATTR_LaTeX: :width 7cm
rlm@481	2237 [[./images/basic-muscle.png]]
rlm@481	2238
rlm@481	2239 *** Muscle meta-data
rlm@481	2240
rlm@481	2241 #+caption: Program to deal with loading muscle data from a blender
rlm@481	2242 #+caption: file's metadata.
rlm@481	2243 #+name: motor-pool
rlm@481	2244 #+begin_listing clojure
rlm@481	2245 #+BEGIN_SRC clojure
rlm@481	2246 (defn muscle-profile-image
rlm@481	2247 "Get the muscle-profile image from the node's blender meta-data."
rlm@481	2248 [#^Node muscle]
rlm@481	2249 (if-let [image (meta-data muscle "muscle")]
rlm@481	2250 (load-image image)))
rlm@481	2251
rlm@481	2252 (defn muscle-strength
rlm@481	2253 "Return the strength of this muscle, or 1 if it is not defined."
rlm@481	2254 [#^Node muscle]
rlm@481	2255 (if-let [strength (meta-data muscle "strength")]
rlm@481	2256 strength 1))
rlm@481	2257
rlm@481	2258 (defn motor-pool
rlm@481	2259 "Return a vector where each entry is the strength of the \"motor
rlm@481	2260 neuron\" at that part in the muscle."
rlm@481	2261 [#^Node muscle]
rlm@481	2262 (let [profile (muscle-profile-image muscle)]
rlm@481	2263 (vec
rlm@481	2264 (let [width (.getWidth profile)]
rlm@481	2265 (for [x (range width)]
rlm@481	2266 (- 255
rlm@481	2267 (bit-and
rlm@481	2268 0x0000FF
rlm@481	2269 (.getRGB profile x 0))))))))
rlm@481	2270 #+END_SRC
rlm@481	2271 #+end_listing
rlm@481	2272
rlm@481	2273 Of note here is =motor-pool= which interprets the muscle-profile
rlm@481	2274 image in a way that allows me to use gradients between white and
rlm@481	2275 red, instead of shades of gray as I've been using for all the
rlm@481	2276 other senses. This is purely an aesthetic touch.
rlm@481	2277
rlm@481	2278 *** Creating muscles
rlm@481	2279
rlm@481	2280 #+caption: This is the core movement functoion in =CORTEX=, which
rlm@481	2281 #+caption: implements muscles that report on their activation.
rlm@481	2282 #+name: muscle-kernel
rlm@481	2283 #+begin_listing clojure
rlm@481	2284 #+BEGIN_SRC clojure
rlm@481	2285 (defn movement-kernel
rlm@481	2286 "Returns a function which when called with a integer value inside a
rlm@481	2287 running simulation will cause movement in the creature according
rlm@481	2288 to the muscle's position and strength profile. Each function
rlm@481	2289 returns the amount of force applied / max force."
rlm@481	2290 [#^Node creature #^Node muscle]
rlm@481	2291 (let [target (closest-node creature muscle)
rlm@481	2292 axis
rlm@481	2293 (.mult (.getWorldRotation muscle) Vector3f/UNIT_Y)
rlm@481	2294 strength (muscle-strength muscle)
rlm@481	2295
rlm@481	2296 pool (motor-pool muscle)
rlm@481	2297 pool-integral (reductions + pool)
rlm@481	2298 forces
rlm@481	2299 (vec (map #(float (* strength (/ % (last pool-integral))))
rlm@481	2300 pool-integral))
rlm@481	2301 control (.getControl target RigidBodyControl)]
rlm@481	2302 ;;(println-repl (.getName target) axis)
rlm@481	2303 (fn [n]
rlm@481	2304 (let [pool-index (max 0 (min n (dec (count pool))))
rlm@481	2305 force (forces pool-index)]
rlm@481	2306 (.applyTorque control (.mult axis force))
rlm@481	2307 (float (/ force strength))))))
rlm@481	2308
rlm@481	2309 (defn movement!
rlm@481	2310 "Endow the creature with the power of movement. Returns a sequence
rlm@481	2311 of functions, each of which accept an integer value and will
rlm@481	2312 activate their corresponding muscle."
rlm@481	2313 [#^Node creature]
rlm@481	2314 (for [muscle (muscles creature)]
rlm@481	2315 (movement-kernel creature muscle)))
rlm@481	2316 #+END_SRC
rlm@481	2317 #+end_listing
rlm@481	2318
rlm@481	2319
rlm@481	2320 =movement-kernel= creates a function that will move the nearest
rlm@481	2321 physical object to the muscle node. The muscle exerts a rotational
rlm@481	2322 force dependent on it's orientation to the object in the blender
rlm@481	2323 file. The function returned by =movement-kernel= is also a sense
rlm@481	2324 function: it returns the percent of the total muscle strength that
rlm@481	2325 is currently being employed. This is analogous to muscle tension
rlm@481	2326 in humans and completes the sense of proprioception begun in the
rlm@481	2327 last section.
rlm@481	2328
rlm@481	2329
rlm@481	2330
rlm@436	2331 ** =CORTEX= brings complex creatures to life!
rlm@436	2332
rlm@436	2333 ** =CORTEX= enables many possiblities for further research
rlm@474	2334
rlm@465	2335 * COMMENT Empathy in a simulated worm
rlm@435	2336
rlm@449	2337 Here I develop a computational model of empathy, using =CORTEX= as a
rlm@449	2338 base. Empathy in this context is the ability to observe another
rlm@449	2339 creature and infer what sorts of sensations that creature is
rlm@449	2340 feeling. My empathy algorithm involves multiple phases. First is
rlm@449	2341 free-play, where the creature moves around and gains sensory
rlm@449	2342 experience. From this experience I construct a representation of the
rlm@449	2343 creature's sensory state space, which I call \Phi-space. Using
rlm@449	2344 \Phi-space, I construct an efficient function which takes the
rlm@449	2345 limited data that comes from observing another creature and enriches
rlm@449	2346 it full compliment of imagined sensory data. I can then use the
rlm@449	2347 imagined sensory data to recognize what the observed creature is
rlm@449	2348 doing and feeling, using straightforward embodied action predicates.
rlm@449	2349 This is all demonstrated with using a simple worm-like creature, and
rlm@449	2350 recognizing worm-actions based on limited data.
rlm@449	2351
rlm@449	2352 #+caption: Here is the worm with which we will be working.
rlm@449	2353 #+caption: It is composed of 5 segments. Each segment has a
rlm@449	2354 #+caption: pair of extensor and flexor muscles. Each of the
rlm@449	2355 #+caption: worm's four joints is a hinge joint which allows
rlm@451	2356 #+caption: about 30 degrees of rotation to either side. Each segment
rlm@449	2357 #+caption: of the worm is touch-capable and has a uniform
rlm@449	2358 #+caption: distribution of touch sensors on each of its faces.
rlm@449	2359 #+caption: Each joint has a proprioceptive sense to detect
rlm@449	2360 #+caption: relative positions. The worm segments are all the
rlm@449	2361 #+caption: same except for the first one, which has a much
rlm@449	2362 #+caption: higher weight than the others to allow for easy
rlm@449	2363 #+caption: manual motor control.
rlm@449	2364 #+name: basic-worm-view
rlm@449	2365 #+ATTR_LaTeX: :width 10cm
rlm@449	2366 [[./images/basic-worm-view.png]]
rlm@449	2367
rlm@449	2368 #+caption: Program for reading a worm from a blender file and
rlm@449	2369 #+caption: outfitting it with the senses of proprioception,
rlm@449	2370 #+caption: touch, and the ability to move, as specified in the
rlm@449	2371 #+caption: blender file.
rlm@449	2372 #+name: get-worm
rlm@449	2373 #+begin_listing clojure
rlm@449	2374 #+begin_src clojure
rlm@449	2375 (defn worm []
rlm@449	2376 (let [model (load-blender-model "Models/worm/worm.blend")]
rlm@449	2377 {:body (doto model (body!))
rlm@449	2378 :touch (touch! model)
rlm@449	2379 :proprioception (proprioception! model)
rlm@449	2380 :muscles (movement! model)}))
rlm@449	2381 #+end_src
rlm@449	2382 #+end_listing
rlm@452	2383
rlm@436	2384 ** Embodiment factors action recognition into managable parts
rlm@435	2385
rlm@449	2386 Using empathy, I divide the problem of action recognition into a
rlm@449	2387 recognition process expressed in the language of a full compliment
rlm@449	2388 of senses, and an imaganitive process that generates full sensory
rlm@449	2389 data from partial sensory data. Splitting the action recognition
rlm@449	2390 problem in this manner greatly reduces the total amount of work to
rlm@449	2391 recognize actions: The imaganitive process is mostly just matching
rlm@449	2392 previous experience, and the recognition process gets to use all
rlm@449	2393 the senses to directly describe any action.
rlm@449	2394
rlm@436	2395 ** Action recognition is easy with a full gamut of senses
rlm@435	2396
rlm@449	2397 Embodied representations using multiple senses such as touch,
rlm@449	2398 proprioception, and muscle tension turns out be be exceedingly
rlm@449	2399 efficient at describing body-centered actions. It is the ``right
rlm@449	2400 language for the job''. For example, it takes only around 5 lines
rlm@449	2401 of LISP code to describe the action of ``curling'' using embodied
rlm@451	2402 primitives. It takes about 10 lines to describe the seemingly
rlm@449	2403 complicated action of wiggling.
rlm@449	2404
rlm@449	2405 The following action predicates each take a stream of sensory
rlm@449	2406 experience, observe however much of it they desire, and decide
rlm@449	2407 whether the worm is doing the action they describe. =curled?=
rlm@449	2408 relies on proprioception, =resting?= relies on touch, =wiggling?=
rlm@449	2409 relies on a fourier analysis of muscle contraction, and
rlm@449	2410 =grand-circle?= relies on touch and reuses =curled?= as a gaurd.
rlm@449	2411
rlm@449	2412 #+caption: Program for detecting whether the worm is curled. This is the
rlm@449	2413 #+caption: simplest action predicate, because it only uses the last frame
rlm@449	2414 #+caption: of sensory experience, and only uses proprioceptive data. Even
rlm@449	2415 #+caption: this simple predicate, however, is automatically frame
rlm@449	2416 #+caption: independent and ignores vermopomorphic differences such as
rlm@449	2417 #+caption: worm textures and colors.
rlm@449	2418 #+name: curled
rlm@452	2419 #+attr_latex: [htpb]
rlm@452	2420 #+begin_listing clojure
rlm@449	2421 #+begin_src clojure
rlm@449	2422 (defn curled?
rlm@449	2423 "Is the worm curled up?"
rlm@449	2424 [experiences]
rlm@449	2425 (every?
rlm@449	2426 (fn [[_ _ bend]]
rlm@449	2427 (> (Math/sin bend) 0.64))
rlm@449	2428 (:proprioception (peek experiences))))
rlm@449	2429 #+end_src
rlm@449	2430 #+end_listing
rlm@449	2431
rlm@449	2432 #+caption: Program for summarizing the touch information in a patch
rlm@449	2433 #+caption: of skin.
rlm@449	2434 #+name: touch-summary
rlm@452	2435 #+attr_latex: [htpb]
rlm@452	2436
rlm@452	2437 #+begin_listing clojure
rlm@449	2438 #+begin_src clojure
rlm@449	2439 (defn contact
rlm@449	2440 "Determine how much contact a particular worm segment has with
rlm@449	2441 other objects. Returns a value between 0 and 1, where 1 is full
rlm@449	2442 contact and 0 is no contact."
rlm@449	2443 [touch-region [coords contact :as touch]]
rlm@449	2444 (-> (zipmap coords contact)
rlm@449	2445 (select-keys touch-region)
rlm@449	2446 (vals)
rlm@449	2447 (#(map first %))
rlm@449	2448 (average)
rlm@449	2449 (* 10)
rlm@449	2450 (- 1)
rlm@449	2451 (Math/abs)))
rlm@449	2452 #+end_src
rlm@449	2453 #+end_listing
rlm@449	2454
rlm@449	2455
rlm@449	2456 #+caption: Program for detecting whether the worm is at rest. This program
rlm@449	2457 #+caption: uses a summary of the tactile information from the underbelly
rlm@449	2458 #+caption: of the worm, and is only true if every segment is touching the
rlm@449	2459 #+caption: floor. Note that this function contains no references to
rlm@449	2460 #+caption: proprioction at all.
rlm@449	2461 #+name: resting
rlm@452	2462 #+attr_latex: [htpb]
rlm@452	2463 #+begin_listing clojure
rlm@449	2464 #+begin_src clojure
rlm@449	2465 (def worm-segment-bottom (rect-region [8 15] [14 22]))
rlm@449	2466
rlm@449	2467 (defn resting?
rlm@449	2468 "Is the worm resting on the ground?"
rlm@449	2469 [experiences]
rlm@449	2470 (every?
rlm@449	2471 (fn [touch-data]
rlm@449	2472 (< 0.9 (contact worm-segment-bottom touch-data)))
rlm@449	2473 (:touch (peek experiences))))
rlm@449	2474 #+end_src
rlm@449	2475 #+end_listing
rlm@449	2476
rlm@449	2477 #+caption: Program for detecting whether the worm is curled up into a
rlm@449	2478 #+caption: full circle. Here the embodied approach begins to shine, as
rlm@449	2479 #+caption: I am able to both use a previous action predicate (=curled?=)
rlm@449	2480 #+caption: as well as the direct tactile experience of the head and tail.
rlm@449	2481 #+name: grand-circle
rlm@452	2482 #+attr_latex: [htpb]
rlm@452	2483 #+begin_listing clojure
rlm@449	2484 #+begin_src clojure
rlm@449	2485 (def worm-segment-bottom-tip (rect-region [15 15] [22 22]))
rlm@449	2486
rlm@449	2487 (def worm-segment-top-tip (rect-region [0 15] [7 22]))
rlm@449	2488
rlm@449	2489 (defn grand-circle?
rlm@449	2490 "Does the worm form a majestic circle (one end touching the other)?"
rlm@449	2491 [experiences]
rlm@449	2492 (and (curled? experiences)
rlm@449	2493 (let [worm-touch (:touch (peek experiences))
rlm@449	2494 tail-touch (worm-touch 0)
rlm@449	2495 head-touch (worm-touch 4)]
rlm@449	2496 (and (< 0.55 (contact worm-segment-bottom-tip tail-touch))
rlm@449	2497 (< 0.55 (contact worm-segment-top-tip head-touch))))))
rlm@449	2498 #+end_src
rlm@449	2499 #+end_listing
rlm@449	2500
rlm@449	2501
rlm@449	2502 #+caption: Program for detecting whether the worm has been wiggling for
rlm@449	2503 #+caption: the last few frames. It uses a fourier analysis of the muscle
rlm@449	2504 #+caption: contractions of the worm's tail to determine wiggling. This is
rlm@449	2505 #+caption: signigicant because there is no particular frame that clearly
rlm@449	2506 #+caption: indicates that the worm is wiggling --- only when multiple frames
rlm@449	2507 #+caption: are analyzed together is the wiggling revealed. Defining
rlm@449	2508 #+caption: wiggling this way also gives the worm an opportunity to learn
rlm@449	2509 #+caption: and recognize ``frustrated wiggling'', where the worm tries to
rlm@449	2510 #+caption: wiggle but can't. Frustrated wiggling is very visually different
rlm@449	2511 #+caption: from actual wiggling, but this definition gives it to us for free.
rlm@449	2512 #+name: wiggling
rlm@452	2513 #+attr_latex: [htpb]
rlm@452	2514 #+begin_listing clojure
rlm@449	2515 #+begin_src clojure
rlm@449	2516 (defn fft [nums]
rlm@449	2517 (map
rlm@449	2518 #(.getReal %)
rlm@449	2519 (.transform
rlm@449	2520 (FastFourierTransformer. DftNormalization/STANDARD)
rlm@449	2521 (double-array nums) TransformType/FORWARD)))
rlm@449	2522
rlm@449	2523 (def indexed (partial map-indexed vector))
rlm@449	2524
rlm@449	2525 (defn max-indexed [s]
rlm@449	2526 (first (sort-by (comp - second) (indexed s))))
rlm@449	2527
rlm@449	2528 (defn wiggling?
rlm@449	2529 "Is the worm wiggling?"
rlm@449	2530 [experiences]
rlm@449	2531 (let [analysis-interval 0x40]
rlm@449	2532 (when (> (count experiences) analysis-interval)
rlm@449	2533 (let [a-flex 3
rlm@449	2534 a-ex 2
rlm@449	2535 muscle-activity
rlm@449	2536 (map :muscle (vector:last-n experiences analysis-interval))
rlm@449	2537 base-activity
rlm@449	2538 (map #(- (% a-flex) (% a-ex)) muscle-activity)]
rlm@449	2539 (= 2
rlm@449	2540 (first
rlm@449	2541 (max-indexed
rlm@449	2542 (map #(Math/abs %)
rlm@449	2543 (take 20 (fft base-activity))))))))))
rlm@449	2544 #+end_src
rlm@449	2545 #+end_listing
rlm@449	2546
rlm@449	2547 With these action predicates, I can now recognize the actions of
rlm@449	2548 the worm while it is moving under my control and I have access to
rlm@449	2549 all the worm's senses.
rlm@449	2550
rlm@449	2551 #+caption: Use the action predicates defined earlier to report on
rlm@449	2552 #+caption: what the worm is doing while in simulation.
rlm@449	2553 #+name: report-worm-activity
rlm@452	2554 #+attr_latex: [htpb]
rlm@452	2555 #+begin_listing clojure
rlm@449	2556 #+begin_src clojure
rlm@449	2557 (defn debug-experience
rlm@449	2558 [experiences text]
rlm@449	2559 (cond
rlm@449	2560 (grand-circle? experiences) (.setText text "Grand Circle")
rlm@449	2561 (curled? experiences) (.setText text "Curled")
rlm@449	2562 (wiggling? experiences) (.setText text "Wiggling")
rlm@449	2563 (resting? experiences) (.setText text "Resting")))
rlm@449	2564 #+end_src
rlm@449	2565 #+end_listing
rlm@449	2566
rlm@449	2567 #+caption: Using =debug-experience=, the body-centered predicates
rlm@449	2568 #+caption: work together to classify the behaviour of the worm.
rlm@451	2569 #+caption: the predicates are operating with access to the worm's
rlm@451	2570 #+caption: full sensory data.
rlm@449	2571 #+name: basic-worm-view
rlm@449	2572 #+ATTR_LaTeX: :width 10cm
rlm@449	2573 [[./images/worm-identify-init.png]]
rlm@449	2574
rlm@449	2575 These action predicates satisfy the recognition requirement of an
rlm@451	2576 empathic recognition system. There is power in the simplicity of
rlm@451	2577 the action predicates. They describe their actions without getting
rlm@451	2578 confused in visual details of the worm. Each one is frame
rlm@451	2579 independent, but more than that, they are each indepent of
rlm@449	2580 irrelevant visual details of the worm and the environment. They
rlm@449	2581 will work regardless of whether the worm is a different color or
rlm@451	2582 hevaily textured, or if the environment has strange lighting.
rlm@449	2583
rlm@449	2584 The trick now is to make the action predicates work even when the
rlm@449	2585 sensory data on which they depend is absent. If I can do that, then
rlm@449	2586 I will have gained much,
rlm@435	2587
rlm@436	2588 ** \Phi-space describes the worm's experiences
rlm@449	2589
rlm@449	2590 As a first step towards building empathy, I need to gather all of
rlm@449	2591 the worm's experiences during free play. I use a simple vector to
rlm@449	2592 store all the experiences.
rlm@449	2593
rlm@449	2594 Each element of the experience vector exists in the vast space of
rlm@449	2595 all possible worm-experiences. Most of this vast space is actually
rlm@449	2596 unreachable due to physical constraints of the worm's body. For
rlm@449	2597 example, the worm's segments are connected by hinge joints that put
rlm@451	2598 a practical limit on the worm's range of motions without limiting
rlm@451	2599 its degrees of freedom. Some groupings of senses are impossible;
rlm@451	2600 the worm can not be bent into a circle so that its ends are
rlm@451	2601 touching and at the same time not also experience the sensation of
rlm@451	2602 touching itself.
rlm@449	2603
rlm@451	2604 As the worm moves around during free play and its experience vector
rlm@451	2605 grows larger, the vector begins to define a subspace which is all
rlm@451	2606 the sensations the worm can practicaly experience during normal
rlm@451	2607 operation. I call this subspace \Phi-space, short for
rlm@451	2608 physical-space. The experience vector defines a path through
rlm@451	2609 \Phi-space. This path has interesting properties that all derive
rlm@451	2610 from physical embodiment. The proprioceptive components are
rlm@451	2611 completely smooth, because in order for the worm to move from one
rlm@451	2612 position to another, it must pass through the intermediate
rlm@451	2613 positions. The path invariably forms loops as actions are repeated.
rlm@451	2614 Finally and most importantly, proprioception actually gives very
rlm@451	2615 strong inference about the other senses. For example, when the worm
rlm@451	2616 is flat, you can infer that it is touching the ground and that its
rlm@451	2617 muscles are not active, because if the muscles were active, the
rlm@451	2618 worm would be moving and would not be perfectly flat. In order to
rlm@451	2619 stay flat, the worm has to be touching the ground, or it would
rlm@451	2620 again be moving out of the flat position due to gravity. If the
rlm@451	2621 worm is positioned in such a way that it interacts with itself,
rlm@451	2622 then it is very likely to be feeling the same tactile feelings as
rlm@451	2623 the last time it was in that position, because it has the same body
rlm@451	2624 as then. If you observe multiple frames of proprioceptive data,
rlm@451	2625 then you can become increasingly confident about the exact
rlm@451	2626 activations of the worm's muscles, because it generally takes a
rlm@451	2627 unique combination of muscle contractions to transform the worm's
rlm@451	2628 body along a specific path through \Phi-space.
rlm@449	2629
rlm@449	2630 There is a simple way of taking \Phi-space and the total ordering
rlm@449	2631 provided by an experience vector and reliably infering the rest of
rlm@449	2632 the senses.
rlm@435	2633
rlm@436	2634 ** Empathy is the process of tracing though \Phi-space
rlm@449	2635
rlm@450	2636 Here is the core of a basic empathy algorithm, starting with an
rlm@451	2637 experience vector:
rlm@451	2638
rlm@451	2639 First, group the experiences into tiered proprioceptive bins. I use
rlm@451	2640 powers of 10 and 3 bins, and the smallest bin has an approximate
rlm@451	2641 size of 0.001 radians in all proprioceptive dimensions.
rlm@450	2642
rlm@450	2643 Then, given a sequence of proprioceptive input, generate a set of
rlm@451	2644 matching experience records for each input, using the tiered
rlm@451	2645 proprioceptive bins.
rlm@449	2646
rlm@450	2647 Finally, to infer sensory data, select the longest consective chain
rlm@451	2648 of experiences. Conecutive experience means that the experiences
rlm@451	2649 appear next to each other in the experience vector.
rlm@449	2650
rlm@450	2651 This algorithm has three advantages:
rlm@450	2652
rlm@450	2653 1. It's simple
rlm@450	2654
rlm@451	2655 3. It's very fast -- retrieving possible interpretations takes
rlm@451	2656 constant time. Tracing through chains of interpretations takes
rlm@451	2657 time proportional to the average number of experiences in a
rlm@451	2658 proprioceptive bin. Redundant experiences in \Phi-space can be
rlm@451	2659 merged to save computation.
rlm@450	2660
rlm@450	2661 2. It protects from wrong interpretations of transient ambiguous
rlm@451	2662 proprioceptive data. For example, if the worm is flat for just
rlm@450	2663 an instant, this flattness will not be interpreted as implying
rlm@450	2664 that the worm has its muscles relaxed, since the flattness is
rlm@450	2665 part of a longer chain which includes a distinct pattern of
rlm@451	2666 muscle activation. Markov chains or other memoryless statistical
rlm@451	2667 models that operate on individual frames may very well make this
rlm@451	2668 mistake.
rlm@450	2669
rlm@450	2670 #+caption: Program to convert an experience vector into a
rlm@450	2671 #+caption: proprioceptively binned lookup function.
rlm@450	2672 #+name: bin
rlm@452	2673 #+attr_latex: [htpb]
rlm@452	2674 #+begin_listing clojure
rlm@450	2675 #+begin_src clojure
rlm@449	2676 (defn bin [digits]
rlm@449	2677 (fn [angles]
rlm@449	2678 (->> angles
rlm@449	2679 (flatten)
rlm@449	2680 (map (juxt #(Math/sin %) #(Math/cos %)))
rlm@449	2681 (flatten)
rlm@449	2682 (mapv #(Math/round (* % (Math/pow 10 (dec digits))))))))
rlm@449	2683
rlm@449	2684 (defn gen-phi-scan
rlm@450	2685 "Nearest-neighbors with binning. Only returns a result if
rlm@450	2686 the propriceptive data is within 10% of a previously recorded
rlm@450	2687 result in all dimensions."
rlm@450	2688 [phi-space]
rlm@449	2689 (let [bin-keys (map bin [3 2 1])
rlm@449	2690 bin-maps
rlm@449	2691 (map (fn [bin-key]
rlm@449	2692 (group-by
rlm@449	2693 (comp bin-key :proprioception phi-space)
rlm@449	2694 (range (count phi-space)))) bin-keys)
rlm@449	2695 lookups (map (fn [bin-key bin-map]
rlm@450	2696 (fn [proprio] (bin-map (bin-key proprio))))
rlm@450	2697 bin-keys bin-maps)]
rlm@449	2698 (fn lookup [proprio-data]
rlm@449	2699 (set (some #(% proprio-data) lookups)))))
rlm@450	2700 #+end_src
rlm@450	2701 #+end_listing
rlm@449	2702
rlm@451	2703 #+caption: =longest-thread= finds the longest path of consecutive
rlm@451	2704 #+caption: experiences to explain proprioceptive worm data.
rlm@451	2705 #+name: phi-space-history-scan
rlm@451	2706 #+ATTR_LaTeX: :width 10cm
rlm@451	2707 [[./images/aurellem-gray.png]]
rlm@451	2708
rlm@451	2709 =longest-thread= infers sensory data by stitching together pieces
rlm@451	2710 from previous experience. It prefers longer chains of previous
rlm@451	2711 experience to shorter ones. For example, during training the worm
rlm@451	2712 might rest on the ground for one second before it performs its
rlm@451	2713 excercises. If during recognition the worm rests on the ground for
rlm@451	2714 five seconds, =longest-thread= will accomodate this five second
rlm@451	2715 rest period by looping the one second rest chain five times.
rlm@451	2716
rlm@451	2717 =longest-thread= takes time proportinal to the average number of
rlm@451	2718 entries in a proprioceptive bin, because for each element in the
rlm@451	2719 starting bin it performes a series of set lookups in the preceeding
rlm@451	2720 bins. If the total history is limited, then this is only a constant
rlm@451	2721 multiple times the number of entries in the starting bin. This
rlm@451	2722 analysis also applies even if the action requires multiple longest
rlm@451	2723 chains -- it's still the average number of entries in a
rlm@451	2724 proprioceptive bin times the desired chain length. Because
rlm@451	2725 =longest-thread= is so efficient and simple, I can interpret
rlm@451	2726 worm-actions in real time.
rlm@449	2727
rlm@450	2728 #+caption: Program to calculate empathy by tracing though \Phi-space
rlm@450	2729 #+caption: and finding the longest (ie. most coherent) interpretation
rlm@450	2730 #+caption: of the data.
rlm@450	2731 #+name: longest-thread
rlm@452	2732 #+attr_latex: [htpb]
rlm@452	2733 #+begin_listing clojure
rlm@450	2734 #+begin_src clojure
rlm@449	2735 (defn longest-thread
rlm@449	2736 "Find the longest thread from phi-index-sets. The index sets should
rlm@449	2737 be ordered from most recent to least recent."
rlm@449	2738 [phi-index-sets]
rlm@449	2739 (loop [result '()
rlm@449	2740 [thread-bases & remaining :as phi-index-sets] phi-index-sets]
rlm@449	2741 (if (empty? phi-index-sets)
rlm@449	2742 (vec result)
rlm@449	2743 (let [threads
rlm@449	2744 (for [thread-base thread-bases]
rlm@449	2745 (loop [thread (list thread-base)
rlm@449	2746 remaining remaining]
rlm@449	2747 (let [next-index (dec (first thread))]
rlm@449	2748 (cond (empty? remaining) thread
rlm@449	2749 (contains? (first remaining) next-index)
rlm@449	2750 (recur
rlm@449	2751 (cons next-index thread) (rest remaining))
rlm@449	2752 :else thread))))
rlm@449	2753 longest-thread
rlm@449	2754 (reduce (fn [thread-a thread-b]
rlm@449	2755 (if (> (count thread-a) (count thread-b))
rlm@449	2756 thread-a thread-b))
rlm@449	2757 '(nil)
rlm@449	2758 threads)]
rlm@449	2759 (recur (concat longest-thread result)
rlm@449	2760 (drop (count longest-thread) phi-index-sets))))))
rlm@450	2761 #+end_src
rlm@450	2762 #+end_listing
rlm@450	2763
rlm@451	2764 There is one final piece, which is to replace missing sensory data
rlm@451	2765 with a best-guess estimate. While I could fill in missing data by
rlm@451	2766 using a gradient over the closest known sensory data points,
rlm@451	2767 averages can be misleading. It is certainly possible to create an
rlm@451	2768 impossible sensory state by averaging two possible sensory states.
rlm@451	2769 Therefore, I simply replicate the most recent sensory experience to
rlm@451	2770 fill in the gaps.
rlm@449	2771
rlm@449	2772 #+caption: Fill in blanks in sensory experience by replicating the most
rlm@449	2773 #+caption: recent experience.
rlm@449	2774 #+name: infer-nils
rlm@452	2775 #+attr_latex: [htpb]
rlm@452	2776 #+begin_listing clojure
rlm@449	2777 #+begin_src clojure
rlm@449	2778 (defn infer-nils
rlm@449	2779 "Replace nils with the next available non-nil element in the
rlm@449	2780 sequence, or barring that, 0."
rlm@449	2781 [s]
rlm@449	2782 (loop [i (dec (count s))
rlm@449	2783 v (transient s)]
rlm@449	2784 (if (zero? i) (persistent! v)
rlm@449	2785 (if-let [cur (v i)]
rlm@449	2786 (if (get v (dec i) 0)
rlm@449	2787 (recur (dec i) v)
rlm@449	2788 (recur (dec i) (assoc! v (dec i) cur)))
rlm@449	2789 (recur i (assoc! v i 0))))))
rlm@449	2790 #+end_src
rlm@449	2791 #+end_listing
rlm@435	2792
rlm@441	2793 ** Efficient action recognition with =EMPATH=
rlm@451	2794
rlm@451	2795 To use =EMPATH= with the worm, I first need to gather a set of
rlm@451	2796 experiences from the worm that includes the actions I want to
rlm@452	2797 recognize. The =generate-phi-space= program (listing
rlm@451	2798 \ref{generate-phi-space} runs the worm through a series of
rlm@451	2799 exercices and gatheres those experiences into a vector. The
rlm@451	2800 =do-all-the-things= program is a routine expressed in a simple
rlm@452	2801 muscle contraction script language for automated worm control. It
rlm@452	2802 causes the worm to rest, curl, and wiggle over about 700 frames
rlm@452	2803 (approx. 11 seconds).
rlm@425	2804
rlm@451	2805 #+caption: Program to gather the worm's experiences into a vector for
rlm@451	2806 #+caption: further processing. The =motor-control-program= line uses
rlm@451	2807 #+caption: a motor control script that causes the worm to execute a series
rlm@451	2808 #+caption: of ``exercices'' that include all the action predicates.
rlm@451	2809 #+name: generate-phi-space
rlm@452	2810 #+attr_latex: [htpb]
rlm@452	2811 #+begin_listing clojure
rlm@451	2812 #+begin_src clojure
rlm@451	2813 (def do-all-the-things
rlm@451	2814 (concat
rlm@451	2815 curl-script
rlm@451	2816 [[300 :d-ex 40]
rlm@451	2817 [320 :d-ex 0]]
rlm@451	2818 (shift-script 280 (take 16 wiggle-script))))
rlm@451	2819
rlm@451	2820 (defn generate-phi-space []
rlm@451	2821 (let [experiences (atom [])]
rlm@451	2822 (run-world
rlm@451	2823 (apply-map
rlm@451	2824 worm-world
rlm@451	2825 (merge
rlm@451	2826 (worm-world-defaults)
rlm@451	2827 {:end-frame 700
rlm@451	2828 :motor-control
rlm@451	2829 (motor-control-program worm-muscle-labels do-all-the-things)
rlm@451	2830 :experiences experiences})))
rlm@451	2831 @experiences))
rlm@451	2832 #+end_src
rlm@451	2833 #+end_listing
rlm@451	2834
rlm@451	2835 #+caption: Use longest thread and a phi-space generated from a short
rlm@451	2836 #+caption: exercise routine to interpret actions during free play.
rlm@451	2837 #+name: empathy-debug
rlm@452	2838 #+attr_latex: [htpb]
rlm@452	2839 #+begin_listing clojure
rlm@451	2840 #+begin_src clojure
rlm@451	2841 (defn init []
rlm@451	2842 (def phi-space (generate-phi-space))
rlm@451	2843 (def phi-scan (gen-phi-scan phi-space)))
rlm@451	2844
rlm@451	2845 (defn empathy-demonstration []
rlm@451	2846 (let [proprio (atom ())]
rlm@451	2847 (fn
rlm@451	2848 [experiences text]
rlm@451	2849 (let [phi-indices (phi-scan (:proprioception (peek experiences)))]
rlm@451	2850 (swap! proprio (partial cons phi-indices))
rlm@451	2851 (let [exp-thread (longest-thread (take 300 @proprio))
rlm@451	2852 empathy (mapv phi-space (infer-nils exp-thread))]
rlm@451	2853 (println-repl (vector:last-n exp-thread 22))
rlm@451	2854 (cond
rlm@451	2855 (grand-circle? empathy) (.setText text "Grand Circle")
rlm@451	2856 (curled? empathy) (.setText text "Curled")
rlm@451	2857 (wiggling? empathy) (.setText text "Wiggling")
rlm@451	2858 (resting? empathy) (.setText text "Resting")
rlm@451	2859 :else (.setText text "Unknown")))))))
rlm@451	2860
rlm@451	2861 (defn empathy-experiment [record]
rlm@451	2862 (.start (worm-world :experience-watch (debug-experience-phi)
rlm@451	2863 :record record :worm worm*)))
rlm@451	2864 #+end_src
rlm@451	2865 #+end_listing
rlm@451	2866
rlm@451	2867 The result of running =empathy-experiment= is that the system is
rlm@451	2868 generally able to interpret worm actions using the action-predicates
rlm@451	2869 on simulated sensory data just as well as with actual data. Figure
rlm@451	2870 \ref{empathy-debug-image} was generated using =empathy-experiment=:
rlm@451	2871
rlm@451	2872 #+caption: From only proprioceptive data, =EMPATH= was able to infer
rlm@451	2873 #+caption: the complete sensory experience and classify four poses
rlm@451	2874 #+caption: (The last panel shows a composite image of \emph{wriggling},
rlm@451	2875 #+caption: a dynamic pose.)
rlm@451	2876 #+name: empathy-debug-image
rlm@451	2877 #+ATTR_LaTeX: :width 10cm :placement [H]
rlm@451	2878 [[./images/empathy-1.png]]
rlm@451	2879
rlm@451	2880 One way to measure the performance of =EMPATH= is to compare the
rlm@451	2881 sutiability of the imagined sense experience to trigger the same
rlm@451	2882 action predicates as the real sensory experience.
rlm@451	2883
rlm@451	2884 #+caption: Determine how closely empathy approximates actual
rlm@451	2885 #+caption: sensory data.
rlm@451	2886 #+name: test-empathy-accuracy
rlm@452	2887 #+attr_latex: [htpb]
rlm@452	2888 #+begin_listing clojure
rlm@451	2889 #+begin_src clojure
rlm@451	2890 (def worm-action-label
rlm@451	2891 (juxt grand-circle? curled? wiggling?))
rlm@451	2892
rlm@451	2893 (defn compare-empathy-with-baseline [matches]
rlm@451	2894 (let [proprio (atom ())]
rlm@451	2895 (fn
rlm@451	2896 [experiences text]
rlm@451	2897 (let [phi-indices (phi-scan (:proprioception (peek experiences)))]
rlm@451	2898 (swap! proprio (partial cons phi-indices))
rlm@451	2899 (let [exp-thread (longest-thread (take 300 @proprio))
rlm@451	2900 empathy (mapv phi-space (infer-nils exp-thread))
rlm@451	2901 experience-matches-empathy
rlm@451	2902 (= (worm-action-label experiences)
rlm@451	2903 (worm-action-label empathy))]
rlm@451	2904 (println-repl experience-matches-empathy)
rlm@451	2905 (swap! matches #(conj % experience-matches-empathy)))))))
rlm@451	2906
rlm@451	2907 (defn accuracy [v]
rlm@451	2908 (float (/ (count (filter true? v)) (count v))))
rlm@451	2909
rlm@451	2910 (defn test-empathy-accuracy []
rlm@451	2911 (let [res (atom [])]
rlm@451	2912 (run-world
rlm@451	2913 (worm-world :experience-watch
rlm@451	2914 (compare-empathy-with-baseline res)
rlm@451	2915 :worm worm*))
rlm@451	2916 (accuracy @res)))
rlm@451	2917 #+end_src
rlm@451	2918 #+end_listing
rlm@451	2919
rlm@451	2920 Running =test-empathy-accuracy= using the very short exercise
rlm@451	2921 program defined in listing \ref{generate-phi-space}, and then doing
rlm@451	2922 a similar pattern of activity manually yeilds an accuracy of around
rlm@451	2923 73%. This is based on very limited worm experience. By training the
rlm@451	2924 worm for longer, the accuracy dramatically improves.
rlm@451	2925
rlm@451	2926 #+caption: Program to generate \Phi-space using manual training.
rlm@451	2927 #+name: manual-phi-space
rlm@452	2928 #+attr_latex: [htpb]
rlm@451	2929 #+begin_listing clojure
rlm@451	2930 #+begin_src clojure
rlm@451	2931 (defn init-interactive []
rlm@451	2932 (def phi-space
rlm@451	2933 (let [experiences (atom [])]
rlm@451	2934 (run-world
rlm@451	2935 (apply-map
rlm@451	2936 worm-world
rlm@451	2937 (merge
rlm@451	2938 (worm-world-defaults)
rlm@451	2939 {:experiences experiences})))
rlm@451	2940 @experiences))
rlm@451	2941 (def phi-scan (gen-phi-scan phi-space)))
rlm@451	2942 #+end_src
rlm@451	2943 #+end_listing
rlm@451	2944
rlm@451	2945 After about 1 minute of manual training, I was able to achieve 95%
rlm@451	2946 accuracy on manual testing of the worm using =init-interactive= and
rlm@452	2947 =test-empathy-accuracy=. The majority of errors are near the
rlm@452	2948 boundaries of transitioning from one type of action to another.
rlm@452	2949 During these transitions the exact label for the action is more open
rlm@452	2950 to interpretation, and dissaggrement between empathy and experience
rlm@452	2951 is more excusable.
rlm@450	2952
rlm@449	2953 ** Digression: bootstrapping touch using free exploration
rlm@449	2954
rlm@452	2955 In the previous section I showed how to compute actions in terms of
rlm@452	2956 body-centered predicates which relied averate touch activation of
rlm@452	2957 pre-defined regions of the worm's skin. What if, instead of recieving
rlm@452	2958 touch pre-grouped into the six faces of each worm segment, the true
rlm@452	2959 topology of the worm's skin was unknown? This is more similiar to how
rlm@452	2960 a nerve fiber bundle might be arranged. While two fibers that are
rlm@452	2961 close in a nerve bundle /might/ correspond to two touch sensors that
rlm@452	2962 are close together on the skin, the process of taking a complicated
rlm@452	2963 surface and forcing it into essentially a circle requires some cuts
rlm@452	2964 and rerragenments.
rlm@452	2965
rlm@452	2966 In this section I show how to automatically learn the skin-topology of
rlm@452	2967 a worm segment by free exploration. As the worm rolls around on the
rlm@452	2968 floor, large sections of its surface get activated. If the worm has
rlm@452	2969 stopped moving, then whatever region of skin that is touching the
rlm@452	2970 floor is probably an important region, and should be recorded.
rlm@452	2971
rlm@452	2972 #+caption: Program to detect whether the worm is in a resting state
rlm@452	2973 #+caption: with one face touching the floor.
rlm@452	2974 #+name: pure-touch
rlm@452	2975 #+begin_listing clojure
rlm@452	2976 #+begin_src clojure
rlm@452	2977 (def full-contact [(float 0.0) (float 0.1)])
rlm@452	2978
rlm@452	2979 (defn pure-touch?
rlm@452	2980 "This is worm specific code to determine if a large region of touch
rlm@452	2981 sensors is either all on or all off."
rlm@452	2982 [[coords touch :as touch-data]]
rlm@452	2983 (= (set (map first touch)) (set full-contact)))
rlm@452	2984 #+end_src
rlm@452	2985 #+end_listing
rlm@452	2986
rlm@452	2987 After collecting these important regions, there will many nearly
rlm@452	2988 similiar touch regions. While for some purposes the subtle
rlm@452	2989 differences between these regions will be important, for my
rlm@452	2990 purposes I colapse them into mostly non-overlapping sets using
rlm@452	2991 =remove-similiar= in listing \ref{remove-similiar}
rlm@452	2992
rlm@452	2993 #+caption: Program to take a lits of set of points and ``collapse them''
rlm@452	2994 #+caption: so that the remaining sets in the list are siginificantly
rlm@452	2995 #+caption: different from each other. Prefer smaller sets to larger ones.
rlm@452	2996 #+name: remove-similiar
rlm@452	2997 #+begin_listing clojure
rlm@452	2998 #+begin_src clojure
rlm@452	2999 (defn remove-similar
rlm@452	3000 [coll]
rlm@452	3001 (loop [result () coll (sort-by (comp - count) coll)]
rlm@452	3002 (if (empty? coll) result
rlm@452	3003 (let [[x & xs] coll
rlm@452	3004 c (count x)]
rlm@452	3005 (if (some
rlm@452	3006 (fn [other-set]
rlm@452	3007 (let [oc (count other-set)]
rlm@452	3008 (< (- (count (union other-set x)) c) (* oc 0.1))))
rlm@452	3009 xs)
rlm@452	3010 (recur result xs)
rlm@452	3011 (recur (cons x result) xs))))))
rlm@452	3012 #+end_src
rlm@452	3013 #+end_listing
rlm@452	3014
rlm@452	3015 Actually running this simulation is easy given =CORTEX='s facilities.
rlm@452	3016
rlm@452	3017 #+caption: Collect experiences while the worm moves around. Filter the touch
rlm@452	3018 #+caption: sensations by stable ones, collapse similiar ones together,
rlm@452	3019 #+caption: and report the regions learned.
rlm@452	3020 #+name: learn-touch
rlm@452	3021 #+begin_listing clojure
rlm@452	3022 #+begin_src clojure
rlm@452	3023 (defn learn-touch-regions []
rlm@452	3024 (let [experiences (atom [])
rlm@452	3025 world (apply-map
rlm@452	3026 worm-world
rlm@452	3027 (assoc (worm-segment-defaults)
rlm@452	3028 :experiences experiences))]
rlm@452	3029 (run-world world)
rlm@452	3030 (->>
rlm@452	3031 @experiences
rlm@452	3032 (drop 175)
rlm@452	3033 ;; access the single segment's touch data
rlm@452	3034 (map (comp first :touch))
rlm@452	3035 ;; only deal with "pure" touch data to determine surfaces
rlm@452	3036 (filter pure-touch?)
rlm@452	3037 ;; associate coordinates with touch values
rlm@452	3038 (map (partial apply zipmap))
rlm@452	3039 ;; select those regions where contact is being made
rlm@452	3040 (map (partial group-by second))
rlm@452	3041 (map #(get % full-contact))
rlm@452	3042 (map (partial map first))
rlm@452	3043 ;; remove redundant/subset regions
rlm@452	3044 (map set)
rlm@452	3045 remove-similar)))
rlm@452	3046
rlm@452	3047 (defn learn-and-view-touch-regions []
rlm@452	3048 (map view-touch-region
rlm@452	3049 (learn-touch-regions)))
rlm@452	3050 #+end_src
rlm@452	3051 #+end_listing
rlm@452	3052
rlm@452	3053 The only thing remining to define is the particular motion the worm
rlm@452	3054 must take. I accomplish this with a simple motor control program.
rlm@452	3055
rlm@452	3056 #+caption: Motor control program for making the worm roll on the ground.
rlm@452	3057 #+caption: This could also be replaced with random motion.
rlm@452	3058 #+name: worm-roll
rlm@452	3059 #+begin_listing clojure
rlm@452	3060 #+begin_src clojure
rlm@452	3061 (defn touch-kinesthetics []
rlm@452	3062 [[170 :lift-1 40]
rlm@452	3063 [190 :lift-1 19]
rlm@452	3064 [206 :lift-1 0]
rlm@452	3065
rlm@452	3066 [400 :lift-2 40]
rlm@452	3067 [410 :lift-2 0]
rlm@452	3068
rlm@452	3069 [570 :lift-2 40]
rlm@452	3070 [590 :lift-2 21]
rlm@452	3071 [606 :lift-2 0]
rlm@452	3072
rlm@452	3073 [800 :lift-1 30]
rlm@452	3074 [809 :lift-1 0]
rlm@452	3075
rlm@452	3076 [900 :roll-2 40]
rlm@452	3077 [905 :roll-2 20]
rlm@452	3078 [910 :roll-2 0]
rlm@452	3079
rlm@452	3080 [1000 :roll-2 40]
rlm@452	3081 [1005 :roll-2 20]
rlm@452	3082 [1010 :roll-2 0]
rlm@452	3083
rlm@452	3084 [1100 :roll-2 40]
rlm@452	3085 [1105 :roll-2 20]
rlm@452	3086 [1110 :roll-2 0]
rlm@452	3087 ])
rlm@452	3088 #+end_src
rlm@452	3089 #+end_listing
rlm@452	3090
rlm@452	3091
rlm@452	3092 #+caption: The small worm rolls around on the floor, driven
rlm@452	3093 #+caption: by the motor control program in listing \ref{worm-roll}.
rlm@452	3094 #+name: worm-roll
rlm@452	3095 #+ATTR_LaTeX: :width 12cm
rlm@452	3096 [[./images/worm-roll.png]]
rlm@452	3097
rlm@452	3098
rlm@452	3099 #+caption: After completing its adventures, the worm now knows
rlm@452	3100 #+caption: how its touch sensors are arranged along its skin. These
rlm@452	3101 #+caption: are the regions that were deemed important by
rlm@452	3102 #+caption: =learn-touch-regions=. Note that the worm has discovered
rlm@452	3103 #+caption: that it has six sides.
rlm@452	3104 #+name: worm-touch-map
rlm@452	3105 #+ATTR_LaTeX: :width 12cm
rlm@452	3106 [[./images/touch-learn.png]]
rlm@452	3107
rlm@452	3108 While simple, =learn-touch-regions= exploits regularities in both
rlm@452	3109 the worm's physiology and the worm's environment to correctly
rlm@452	3110 deduce that the worm has six sides. Note that =learn-touch-regions=
rlm@452	3111 would work just as well even if the worm's touch sense data were
rlm@452	3112 completely scrambled. The cross shape is just for convienence. This
rlm@452	3113 example justifies the use of pre-defined touch regions in =EMPATH=.
rlm@452	3114
rlm@465	3115 * COMMENT Contributions
rlm@454	3116
rlm@461	3117 In this thesis you have seen the =CORTEX= system, a complete
rlm@461	3118 environment for creating simulated creatures. You have seen how to
rlm@461	3119 implement five senses including touch, proprioception, hearing,
rlm@461	3120 vision, and muscle tension. You have seen how to create new creatues
rlm@461	3121 using blender, a 3D modeling tool. I hope that =CORTEX= will be
rlm@461	3122 useful in further research projects. To this end I have included the
rlm@461	3123 full source to =CORTEX= along with a large suite of tests and
rlm@461	3124 examples. I have also created a user guide for =CORTEX= which is
rlm@461	3125 inculded in an appendix to this thesis.
rlm@447	3126
rlm@461	3127 You have also seen how I used =CORTEX= as a platform to attach the
rlm@461	3128 /action recognition/ problem, which is the problem of recognizing
rlm@461	3129 actions in video. You saw a simple system called =EMPATH= which
rlm@461	3130 ientifies actions by first describing actions in a body-centerd,
rlm@461	3131 rich sense language, then infering a full range of sensory
rlm@461	3132 experience from limited data using previous experience gained from
rlm@461	3133 free play.
rlm@447	3134
rlm@461	3135 As a minor digression, you also saw how I used =CORTEX= to enable a
rlm@461	3136 tiny worm to discover the topology of its skin simply by rolling on
rlm@461	3137 the ground.
rlm@461	3138
rlm@461	3139 In conclusion, the main contributions of this thesis are:
rlm@461	3140
rlm@461	3141 - =CORTEX=, a system for creating simulated creatures with rich
rlm@461	3142 senses.
rlm@461	3143 - =EMPATH=, a program for recognizing actions by imagining sensory
rlm@461	3144 experience.
rlm@447	3145
rlm@447	3146 # An anatomical joke:
rlm@447	3147 # - Training
rlm@447	3148 # - Skeletal imitation
rlm@447	3149 # - Sensory fleshing-out
rlm@447	3150 # - Classification

Mercurial > cortex

annotate thesis/cortex.org @ 482:074eadc919fe