cortex: thesis/cortex.org annotate

annotate thesis/cortex.org @ 477:ba54df21fc7c

complete first draft of touch.

author	Robert McIntyre <rlm@mit.edu>
date	Fri, 28 Mar 2014 22:51:14 -0400
parents	5a15611fbb9f
children	a5480a06d5fe

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

Mercurial > cortex

annotate thesis/cortex.org @ 477:ba54df21fc7c