cortex: thesis/cortex.org annotate

annotate thesis/cortex.org @ 471:f14fa9e5b67f

complete first draft of vision.

author	Robert McIntyre <rlm@mit.edu>
date	Fri, 28 Mar 2014 17:31:33 -0400
parents	3401053124b0
children	516a029e0be9

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@470	30 #+end_listing
rlm@470	31
rlm@470	32 #+caption:
rlm@470	33 #+caption:
rlm@470	34 #+caption:
rlm@470	35 #+name: name
rlm@470	36 #+ATTR_LaTeX: :width 10cm
rlm@470	37 [[./images/aurellem-gray.png]]
rlm@470	38
rlm@465	39
rlm@465	40 * COMMENT Empathy and Embodiment as problem solving strategies
rlm@437	41
rlm@437	42 By the end of this thesis, you will have seen a novel approach to
rlm@437	43 interpreting video using embodiment and empathy. You will have also
rlm@437	44 seen one way to efficiently implement empathy for embodied
rlm@447	45 creatures. Finally, you will become familiar with =CORTEX=, a system
rlm@447	46 for designing and simulating creatures with rich senses, which you
rlm@447	47 may choose to use in your own research.
rlm@437	48
rlm@441	49 This is the core vision of my thesis: That one of the important ways
rlm@441	50 in which we understand others is by imagining ourselves in their
rlm@441	51 position and emphatically feeling experiences relative to our own
rlm@441	52 bodies. By understanding events in terms of our own previous
rlm@441	53 corporeal experience, we greatly constrain the possibilities of what
rlm@441	54 would otherwise be an unwieldy exponential search. This extra
rlm@441	55 constraint can be the difference between easily understanding what
rlm@441	56 is happening in a video and being completely lost in a sea of
rlm@441	57 incomprehensible color and movement.
rlm@435	58
rlm@436	59 ** Recognizing actions in video is extremely difficult
rlm@437	60
rlm@447	61 Consider for example the problem of determining what is happening
rlm@447	62 in a video of which this is one frame:
rlm@437	63
rlm@441	64 #+caption: A cat drinking some water. Identifying this action is
rlm@441	65 #+caption: beyond the state of the art for computers.
rlm@441	66 #+ATTR_LaTeX: :width 7cm
rlm@441	67 [[./images/cat-drinking.jpg]]
rlm@441	68
rlm@441	69 It is currently impossible for any computer program to reliably
rlm@447	70 label such a video as ``drinking''. And rightly so -- it is a very
rlm@441	71 hard problem! What features can you describe in terms of low level
rlm@441	72 functions of pixels that can even begin to describe at a high level
rlm@441	73 what is happening here?
rlm@437	74
rlm@447	75 Or suppose that you are building a program that recognizes chairs.
rlm@448	76 How could you ``see'' the chair in figure \ref{hidden-chair}?
rlm@441	77
rlm@441	78 #+caption: The chair in this image is quite obvious to humans, but I
rlm@448	79 #+caption: doubt that any modern computer vision program can find it.
rlm@441	80 #+name: hidden-chair
rlm@441	81 #+ATTR_LaTeX: :width 10cm
rlm@441	82 [[./images/fat-person-sitting-at-desk.jpg]]
rlm@441	83
rlm@441	84 Finally, how is it that you can easily tell the difference between
rlm@441	85 how the girls /muscles/ are working in figure \ref{girl}?
rlm@441	86
rlm@441	87 #+caption: The mysterious ``common sense'' appears here as you are able
rlm@441	88 #+caption: to discern the difference in how the girl's arm muscles
rlm@441	89 #+caption: are activated between the two images.
rlm@441	90 #+name: girl
rlm@448	91 #+ATTR_LaTeX: :width 7cm
rlm@441	92 [[./images/wall-push.png]]
rlm@437	93
rlm@441	94 Each of these examples tells us something about what might be going
rlm@441	95 on in our minds as we easily solve these recognition problems.
rlm@441	96
rlm@441	97 The hidden chairs show us that we are strongly triggered by cues
rlm@447	98 relating to the position of human bodies, and that we can determine
rlm@447	99 the overall physical configuration of a human body even if much of
rlm@447	100 that body is occluded.
rlm@437	101
rlm@441	102 The picture of the girl pushing against the wall tells us that we
rlm@441	103 have common sense knowledge about the kinetics of our own bodies.
rlm@441	104 We know well how our muscles would have to work to maintain us in
rlm@441	105 most positions, and we can easily project this self-knowledge to
rlm@441	106 imagined positions triggered by images of the human body.
rlm@441	107
rlm@441	108 ** =EMPATH= neatly solves recognition problems
rlm@441	109
rlm@441	110 I propose a system that can express the types of recognition
rlm@441	111 problems above in a form amenable to computation. It is split into
rlm@441	112 four parts:
rlm@441	113
rlm@448	114 - Free/Guided Play :: The creature moves around and experiences the
rlm@448	115 world through its unique perspective. Many otherwise
rlm@448	116 complicated actions are easily described in the language of a
rlm@448	117 full suite of body-centered, rich senses. For example,
rlm@448	118 drinking is the feeling of water sliding down your throat, and
rlm@448	119 cooling your insides. It's often accompanied by bringing your
rlm@448	120 hand close to your face, or bringing your face close to water.
rlm@448	121 Sitting down is the feeling of bending your knees, activating
rlm@448	122 your quadriceps, then feeling a surface with your bottom and
rlm@448	123 relaxing your legs. These body-centered action descriptions
rlm@448	124 can be either learned or hard coded.
rlm@448	125 - Posture Imitation :: When trying to interpret a video or image,
rlm@448	126 the creature takes a model of itself and aligns it with
rlm@448	127 whatever it sees. This alignment can even cross species, as
rlm@448	128 when humans try to align themselves with things like ponies,
rlm@448	129 dogs, or other humans with a different body type.
rlm@448	130 - Empathy :: The alignment triggers associations with
rlm@448	131 sensory data from prior experiences. For example, the
rlm@448	132 alignment itself easily maps to proprioceptive data. Any
rlm@448	133 sounds or obvious skin contact in the video can to a lesser
rlm@448	134 extent trigger previous experience. Segments of previous
rlm@448	135 experiences are stitched together to form a coherent and
rlm@448	136 complete sensory portrait of the scene.
rlm@448	137 - Recognition :: With the scene described in terms of first
rlm@448	138 person sensory events, the creature can now run its
rlm@447	139 action-identification programs on this synthesized sensory
rlm@447	140 data, just as it would if it were actually experiencing the
rlm@447	141 scene first-hand. If previous experience has been accurately
rlm@447	142 retrieved, and if it is analogous enough to the scene, then
rlm@447	143 the creature will correctly identify the action in the scene.
rlm@447	144
rlm@441	145 For example, I think humans are able to label the cat video as
rlm@447	146 ``drinking'' because they imagine /themselves/ as the cat, and
rlm@441	147 imagine putting their face up against a stream of water and
rlm@441	148 sticking out their tongue. In that imagined world, they can feel
rlm@441	149 the cool water hitting their tongue, and feel the water entering
rlm@447	150 their body, and are able to recognize that /feeling/ as drinking.
rlm@447	151 So, the label of the action is not really in the pixels of the
rlm@447	152 image, but is found clearly in a simulation inspired by those
rlm@447	153 pixels. An imaginative system, having been trained on drinking and
rlm@447	154 non-drinking examples and learning that the most important
rlm@447	155 component of drinking is the feeling of water sliding down one's
rlm@447	156 throat, would analyze a video of a cat drinking in the following
rlm@447	157 manner:
rlm@441	158
rlm@447	159 1. Create a physical model of the video by putting a ``fuzzy''
rlm@447	160 model of its own body in place of the cat. Possibly also create
rlm@447	161 a simulation of the stream of water.
rlm@441	162
rlm@441	163 2. Play out this simulated scene and generate imagined sensory
rlm@441	164 experience. This will include relevant muscle contractions, a
rlm@441	165 close up view of the stream from the cat's perspective, and most
rlm@441	166 importantly, the imagined feeling of water entering the
rlm@443	167 mouth. The imagined sensory experience can come from a
rlm@441	168 simulation of the event, but can also be pattern-matched from
rlm@441	169 previous, similar embodied experience.
rlm@441	170
rlm@441	171 3. The action is now easily identified as drinking by the sense of
rlm@441	172 taste alone. The other senses (such as the tongue moving in and
rlm@441	173 out) help to give plausibility to the simulated action. Note that
rlm@441	174 the sense of vision, while critical in creating the simulation,
rlm@441	175 is not critical for identifying the action from the simulation.
rlm@441	176
rlm@441	177 For the chair examples, the process is even easier:
rlm@441	178
rlm@441	179 1. Align a model of your body to the person in the image.
rlm@441	180
rlm@441	181 2. Generate proprioceptive sensory data from this alignment.
rlm@437	182
rlm@441	183 3. Use the imagined proprioceptive data as a key to lookup related
rlm@441	184 sensory experience associated with that particular proproceptive
rlm@441	185 feeling.
rlm@437	186
rlm@443	187 4. Retrieve the feeling of your bottom resting on a surface, your
rlm@443	188 knees bent, and your leg muscles relaxed.
rlm@437	189
rlm@441	190 5. This sensory information is consistent with the =sitting?=
rlm@441	191 sensory predicate, so you (and the entity in the image) must be
rlm@441	192 sitting.
rlm@440	193
rlm@441	194 6. There must be a chair-like object since you are sitting.
rlm@440	195
rlm@441	196 Empathy offers yet another alternative to the age-old AI
rlm@441	197 representation question: ``What is a chair?'' --- A chair is the
rlm@441	198 feeling of sitting.
rlm@441	199
rlm@441	200 My program, =EMPATH= uses this empathic problem solving technique
rlm@441	201 to interpret the actions of a simple, worm-like creature.
rlm@437	202
rlm@441	203 #+caption: The worm performs many actions during free play such as
rlm@441	204 #+caption: curling, wiggling, and resting.
rlm@441	205 #+name: worm-intro
rlm@446	206 #+ATTR_LaTeX: :width 15cm
rlm@445	207 [[./images/worm-intro-white.png]]
rlm@437	208
rlm@462	209 #+caption: =EMPATH= recognized and classified each of these
rlm@462	210 #+caption: poses by inferring the complete sensory experience
rlm@462	211 #+caption: from proprioceptive data.
rlm@441	212 #+name: worm-recognition-intro
rlm@446	213 #+ATTR_LaTeX: :width 15cm
rlm@445	214 [[./images/worm-poses.png]]
rlm@441	215
rlm@441	216 One powerful advantage of empathic problem solving is that it
rlm@441	217 factors the action recognition problem into two easier problems. To
rlm@441	218 use empathy, you need an /aligner/, which takes the video and a
rlm@441	219 model of your body, and aligns the model with the video. Then, you
rlm@441	220 need a /recognizer/, which uses the aligned model to interpret the
rlm@441	221 action. The power in this method lies in the fact that you describe
rlm@448	222 all actions form a body-centered viewpoint. You are less tied to
rlm@447	223 the particulars of any visual representation of the actions. If you
rlm@441	224 teach the system what ``running'' is, and you have a good enough
rlm@441	225 aligner, the system will from then on be able to recognize running
rlm@441	226 from any point of view, even strange points of view like above or
rlm@441	227 underneath the runner. This is in contrast to action recognition
rlm@448	228 schemes that try to identify actions using a non-embodied approach.
rlm@448	229 If these systems learn about running as viewed from the side, they
rlm@448	230 will not automatically be able to recognize running from any other
rlm@448	231 viewpoint.
rlm@441	232
rlm@441	233 Another powerful advantage is that using the language of multiple
rlm@441	234 body-centered rich senses to describe body-centerd actions offers a
rlm@441	235 massive boost in descriptive capability. Consider how difficult it
rlm@441	236 would be to compose a set of HOG filters to describe the action of
rlm@447	237 a simple worm-creature ``curling'' so that its head touches its
rlm@447	238 tail, and then behold the simplicity of describing thus action in a
rlm@441	239 language designed for the task (listing \ref{grand-circle-intro}):
rlm@441	240
rlm@446	241 #+caption: Body-centerd actions are best expressed in a body-centered
rlm@446	242 #+caption: language. This code detects when the worm has curled into a
rlm@446	243 #+caption: full circle. Imagine how you would replicate this functionality
rlm@446	244 #+caption: using low-level pixel features such as HOG filters!
rlm@446	245 #+name: grand-circle-intro
rlm@452	246 #+attr_latex: [htpb]
rlm@452	247 #+begin_listing clojure
rlm@446	248 #+begin_src clojure
rlm@446	249 (defn grand-circle?
rlm@446	250 "Does the worm form a majestic circle (one end touching the other)?"
rlm@446	251 [experiences]
rlm@446	252 (and (curled? experiences)
rlm@446	253 (let [worm-touch (:touch (peek experiences))
rlm@446	254 tail-touch (worm-touch 0)
rlm@446	255 head-touch (worm-touch 4)]
rlm@462	256 (and (< 0.2 (contact worm-segment-bottom-tip tail-touch))
rlm@462	257 (< 0.2 (contact worm-segment-top-tip head-touch))))))
rlm@446	258 #+end_src
rlm@446	259 #+end_listing
rlm@446	260
rlm@435	261
rlm@449	262 ** =CORTEX= is a toolkit for building sensate creatures
rlm@435	263
rlm@448	264 I built =CORTEX= to be a general AI research platform for doing
rlm@448	265 experiments involving multiple rich senses and a wide variety and
rlm@448	266 number of creatures. I intend it to be useful as a library for many
rlm@462	267 more projects than just this thesis. =CORTEX= was necessary to meet
rlm@462	268 a need among AI researchers at CSAIL and beyond, which is that
rlm@462	269 people often will invent neat ideas that are best expressed in the
rlm@448	270 language of creatures and senses, but in order to explore those
rlm@448	271 ideas they must first build a platform in which they can create
rlm@448	272 simulated creatures with rich senses! There are many ideas that
rlm@448	273 would be simple to execute (such as =EMPATH=), but attached to them
rlm@448	274 is the multi-month effort to make a good creature simulator. Often,
rlm@448	275 that initial investment of time proves to be too much, and the
rlm@448	276 project must make do with a lesser environment.
rlm@435	277
rlm@448	278 =CORTEX= is well suited as an environment for embodied AI research
rlm@448	279 for three reasons:
rlm@448	280
rlm@448	281 - You can create new creatures using Blender, a popular 3D modeling
rlm@448	282 program. Each sense can be specified using special blender nodes
rlm@448	283 with biologically inspired paramaters. You need not write any
rlm@448	284 code to create a creature, and can use a wide library of
rlm@448	285 pre-existing blender models as a base for your own creatures.
rlm@448	286
rlm@448	287 - =CORTEX= implements a wide variety of senses, including touch,
rlm@448	288 proprioception, vision, hearing, and muscle tension. Complicated
rlm@448	289 senses like touch, and vision involve multiple sensory elements
rlm@448	290 embedded in a 2D surface. You have complete control over the
rlm@448	291 distribution of these sensor elements through the use of simple
rlm@448	292 png image files. In particular, =CORTEX= implements more
rlm@448	293 comprehensive hearing than any other creature simulation system
rlm@448	294 available.
rlm@448	295
rlm@448	296 - =CORTEX= supports any number of creatures and any number of
rlm@448	297 senses. Time in =CORTEX= dialates so that the simulated creatures
rlm@448	298 always precieve a perfectly smooth flow of time, regardless of
rlm@448	299 the actual computational load.
rlm@448	300
rlm@448	301 =CORTEX= is built on top of =jMonkeyEngine3=, which is a video game
rlm@448	302 engine designed to create cross-platform 3D desktop games. =CORTEX=
rlm@448	303 is mainly written in clojure, a dialect of =LISP= that runs on the
rlm@448	304 java virtual machine (JVM). The API for creating and simulating
rlm@449	305 creatures and senses is entirely expressed in clojure, though many
rlm@449	306 senses are implemented at the layer of jMonkeyEngine or below. For
rlm@449	307 example, for the sense of hearing I use a layer of clojure code on
rlm@449	308 top of a layer of java JNI bindings that drive a layer of =C++=
rlm@449	309 code which implements a modified version of =OpenAL= to support
rlm@449	310 multiple listeners. =CORTEX= is the only simulation environment
rlm@449	311 that I know of that can support multiple entities that can each
rlm@449	312 hear the world from their own perspective. Other senses also
rlm@449	313 require a small layer of Java code. =CORTEX= also uses =bullet=, a
rlm@449	314 physics simulator written in =C=.
rlm@448	315
rlm@448	316 #+caption: Here is the worm from above modeled in Blender, a free
rlm@448	317 #+caption: 3D-modeling program. Senses and joints are described
rlm@448	318 #+caption: using special nodes in Blender.
rlm@448	319 #+name: worm-recognition-intro
rlm@448	320 #+ATTR_LaTeX: :width 12cm
rlm@448	321 [[./images/blender-worm.png]]
rlm@448	322
rlm@449	323 Here are some thing I anticipate that =CORTEX= might be used for:
rlm@449	324
rlm@449	325 - exploring new ideas about sensory integration
rlm@449	326 - distributed communication among swarm creatures
rlm@449	327 - self-learning using free exploration,
rlm@449	328 - evolutionary algorithms involving creature construction
rlm@449	329 - exploration of exoitic senses and effectors that are not possible
rlm@449	330 in the real world (such as telekenisis or a semantic sense)
rlm@449	331 - imagination using subworlds
rlm@449	332
rlm@451	333 During one test with =CORTEX=, I created 3,000 creatures each with
rlm@448	334 their own independent senses and ran them all at only 1/80 real
rlm@448	335 time. In another test, I created a detailed model of my own hand,
rlm@448	336 equipped with a realistic distribution of touch (more sensitive at
rlm@448	337 the fingertips), as well as eyes and ears, and it ran at around 1/4
rlm@451	338 real time.
rlm@448	339
rlm@451	340 #+BEGIN_LaTeX
rlm@449	341 \begin{sidewaysfigure}
rlm@449	342 \includegraphics[width=9.5in]{images/full-hand.png}
rlm@451	343 \caption{
rlm@451	344 I modeled my own right hand in Blender and rigged it with all the
rlm@451	345 senses that {\tt CORTEX} supports. My simulated hand has a
rlm@451	346 biologically inspired distribution of touch sensors. The senses are
rlm@451	347 displayed on the right, and the simulation is displayed on the
rlm@451	348 left. Notice that my hand is curling its fingers, that it can see
rlm@451	349 its own finger from the eye in its palm, and that it can feel its
rlm@451	350 own thumb touching its palm.}
rlm@449	351 \end{sidewaysfigure}
rlm@451	352 #+END_LaTeX
rlm@448	353
rlm@437	354 ** Contributions
rlm@435	355
rlm@451	356 - I built =CORTEX=, a comprehensive platform for embodied AI
rlm@451	357 experiments. =CORTEX= supports many features lacking in other
rlm@451	358 systems, such proper simulation of hearing. It is easy to create
rlm@451	359 new =CORTEX= creatures using Blender, a free 3D modeling program.
rlm@449	360
rlm@451	361 - I built =EMPATH=, which uses =CORTEX= to identify the actions of
rlm@451	362 a worm-like creature using a computational model of empathy.
rlm@449	363
rlm@436	364 * Building =CORTEX=
rlm@435	365
rlm@462	366 I intend for =CORTEX= to be used as a general purpose library for
rlm@462	367 building creatures and outfitting them with senses, so that it will
rlm@462	368 be useful for other researchers who want to test out ideas of their
rlm@462	369 own. To this end, wherver I have had to make archetictural choices
rlm@462	370 about =CORTEX=, I have chosen to give as much freedom to the user as
rlm@462	371 possible, so that =CORTEX= may be used for things I have not
rlm@462	372 forseen.
rlm@462	373
rlm@465	374 ** COMMENT Simulation or Reality?
rlm@462	375
rlm@462	376 The most important archetictural decision of all is the choice to
rlm@462	377 use a computer-simulated environemnt in the first place! The world
rlm@462	378 is a vast and rich place, and for now simulations are a very poor
rlm@462	379 reflection of its complexity. It may be that there is a significant
rlm@462	380 qualatative difference between dealing with senses in the real
rlm@468	381 world and dealing with pale facilimilies of them in a simulation.
rlm@468	382 What are the advantages and disadvantages of a simulation vs.
rlm@468	383 reality?
rlm@462	384
rlm@462	385 *** Simulation
rlm@462	386
rlm@462	387 The advantages of virtual reality are that when everything is a
rlm@462	388 simulation, experiments in that simulation are absolutely
rlm@462	389 reproducible. It's also easier to change the character and world
rlm@462	390 to explore new situations and different sensory combinations.
rlm@462	391
rlm@462	392 If the world is to be simulated on a computer, then not only do
rlm@462	393 you have to worry about whether the character's senses are rich
rlm@462	394 enough to learn from the world, but whether the world itself is
rlm@462	395 rendered with enough detail and realism to give enough working
rlm@462	396 material to the character's senses. To name just a few
rlm@462	397 difficulties facing modern physics simulators: destructibility of
rlm@462	398 the environment, simulation of water/other fluids, large areas,
rlm@462	399 nonrigid bodies, lots of objects, smoke. I don't know of any
rlm@462	400 computer simulation that would allow a character to take a rock
rlm@462	401 and grind it into fine dust, then use that dust to make a clay
rlm@462	402 sculpture, at least not without spending years calculating the
rlm@462	403 interactions of every single small grain of dust. Maybe a
rlm@462	404 simulated world with today's limitations doesn't provide enough
rlm@462	405 richness for real intelligence to evolve.
rlm@462	406
rlm@462	407 *** Reality
rlm@462	408
rlm@462	409 The other approach for playing with senses is to hook your
rlm@462	410 software up to real cameras, microphones, robots, etc., and let it
rlm@462	411 loose in the real world. This has the advantage of eliminating
rlm@462	412 concerns about simulating the world at the expense of increasing
rlm@462	413 the complexity of implementing the senses. Instead of just
rlm@462	414 grabbing the current rendered frame for processing, you have to
rlm@462	415 use an actual camera with real lenses and interact with photons to
rlm@462	416 get an image. It is much harder to change the character, which is
rlm@462	417 now partly a physical robot of some sort, since doing so involves
rlm@462	418 changing things around in the real world instead of modifying
rlm@462	419 lines of code. While the real world is very rich and definitely
rlm@462	420 provides enough stimulation for intelligence to develop as
rlm@462	421 evidenced by our own existence, it is also uncontrollable in the
rlm@462	422 sense that a particular situation cannot be recreated perfectly or
rlm@462	423 saved for later use. It is harder to conduct science because it is
rlm@462	424 harder to repeat an experiment. The worst thing about using the
rlm@462	425 real world instead of a simulation is the matter of time. Instead
rlm@462	426 of simulated time you get the constant and unstoppable flow of
rlm@462	427 real time. This severely limits the sorts of software you can use
rlm@462	428 to program the AI because all sense inputs must be handled in real
rlm@462	429 time. Complicated ideas may have to be implemented in hardware or
rlm@462	430 may simply be impossible given the current speed of our
rlm@462	431 processors. Contrast this with a simulation, in which the flow of
rlm@462	432 time in the simulated world can be slowed down to accommodate the
rlm@462	433 limitations of the character's programming. In terms of cost,
rlm@462	434 doing everything in software is far cheaper than building custom
rlm@462	435 real-time hardware. All you need is a laptop and some patience.
rlm@435	436
rlm@465	437 ** COMMENT Because of Time, simulation is perferable to reality
rlm@435	438
rlm@462	439 I envision =CORTEX= being used to support rapid prototyping and
rlm@462	440 iteration of ideas. Even if I could put together a well constructed
rlm@462	441 kit for creating robots, it would still not be enough because of
rlm@462	442 the scourge of real-time processing. Anyone who wants to test their
rlm@462	443 ideas in the real world must always worry about getting their
rlm@465	444 algorithms to run fast enough to process information in real time.
rlm@465	445 The need for real time processing only increases if multiple senses
rlm@465	446 are involved. In the extreme case, even simple algorithms will have
rlm@465	447 to be accelerated by ASIC chips or FPGAs, turning what would
rlm@465	448 otherwise be a few lines of code and a 10x speed penality into a
rlm@465	449 multi-month ordeal. For this reason, =CORTEX= supports
rlm@462	450 /time-dialiation/, which scales back the framerate of the
rlm@465	451 simulation in proportion to the amount of processing each frame.
rlm@465	452 From the perspective of the creatures inside the simulation, time
rlm@465	453 always appears to flow at a constant rate, regardless of how
rlm@462	454 complicated the envorimnent becomes or how many creatures are in
rlm@462	455 the simulation. The cost is that =CORTEX= can sometimes run slower
rlm@462	456 than real time. This can also be an advantage, however ---
rlm@462	457 simulations of very simple creatures in =CORTEX= generally run at
rlm@462	458 40x on my machine!
rlm@462	459
rlm@469	460 ** COMMENT What is a sense?
rlm@468	461
rlm@468	462 If =CORTEX= is to support a wide variety of senses, it would help
rlm@468	463 to have a better understanding of what a ``sense'' actually is!
rlm@468	464 While vision, touch, and hearing all seem like they are quite
rlm@468	465 different things, I was supprised to learn during the course of
rlm@468	466 this thesis that they (and all physical senses) can be expressed as
rlm@468	467 exactly the same mathematical object due to a dimensional argument!
rlm@468	468
rlm@468	469 Human beings are three-dimensional objects, and the nerves that
rlm@468	470 transmit data from our various sense organs to our brain are
rlm@468	471 essentially one-dimensional. This leaves up to two dimensions in
rlm@468	472 which our sensory information may flow. For example, imagine your
rlm@468	473 skin: it is a two-dimensional surface around a three-dimensional
rlm@468	474 object (your body). It has discrete touch sensors embedded at
rlm@468	475 various points, and the density of these sensors corresponds to the
rlm@468	476 sensitivity of that region of skin. Each touch sensor connects to a
rlm@468	477 nerve, all of which eventually are bundled together as they travel
rlm@468	478 up the spinal cord to the brain. Intersect the spinal nerves with a
rlm@468	479 guillotining plane and you will see all of the sensory data of the
rlm@468	480 skin revealed in a roughly circular two-dimensional image which is
rlm@468	481 the cross section of the spinal cord. Points on this image that are
rlm@468	482 close together in this circle represent touch sensors that are
rlm@468	483 /probably/ close together on the skin, although there is of course
rlm@468	484 some cutting and rearrangement that has to be done to transfer the
rlm@468	485 complicated surface of the skin onto a two dimensional image.
rlm@468	486
rlm@468	487 Most human senses consist of many discrete sensors of various
rlm@468	488 properties distributed along a surface at various densities. For
rlm@468	489 skin, it is Pacinian corpuscles, Meissner's corpuscles, Merkel's
rlm@468	490 disks, and Ruffini's endings, which detect pressure and vibration
rlm@468	491 of various intensities. For ears, it is the stereocilia distributed
rlm@468	492 along the basilar membrane inside the cochlea; each one is
rlm@468	493 sensitive to a slightly different frequency of sound. For eyes, it
rlm@468	494 is rods and cones distributed along the surface of the retina. In
rlm@468	495 each case, we can describe the sense with a surface and a
rlm@468	496 distribution of sensors along that surface.
rlm@468	497
rlm@468	498 The neat idea is that every human sense can be effectively
rlm@468	499 described in terms of a surface containing embedded sensors. If the
rlm@468	500 sense had any more dimensions, then there wouldn't be enough room
rlm@468	501 in the spinal chord to transmit the information!
rlm@468	502
rlm@468	503 Therefore, =CORTEX= must support the ability to create objects and
rlm@468	504 then be able to ``paint'' points along their surfaces to describe
rlm@468	505 each sense.
rlm@468	506
rlm@468	507 Fortunately this idea is already a well known computer graphics
rlm@468	508 technique called called /UV-mapping/. The three-dimensional surface
rlm@468	509 of a model is cut and smooshed until it fits on a two-dimensional
rlm@468	510 image. You paint whatever you want on that image, and when the
rlm@468	511 three-dimensional shape is rendered in a game the smooshing and
rlm@468	512 cutting is reversed and the image appears on the three-dimensional
rlm@468	513 object.
rlm@468	514
rlm@468	515 To make a sense, interpret the UV-image as describing the
rlm@468	516 distribution of that senses sensors. To get different types of
rlm@468	517 sensors, you can either use a different color for each type of
rlm@468	518 sensor, or use multiple UV-maps, each labeled with that sensor
rlm@468	519 type. I generally use a white pixel to mean the presence of a
rlm@468	520 sensor and a black pixel to mean the absence of a sensor, and use
rlm@468	521 one UV-map for each sensor-type within a given sense.
rlm@468	522
rlm@468	523 #+CAPTION: The UV-map for an elongated icososphere. The white
rlm@468	524 #+caption: dots each represent a touch sensor. They are dense
rlm@468	525 #+caption: in the regions that describe the tip of the finger,
rlm@468	526 #+caption: and less dense along the dorsal side of the finger
rlm@468	527 #+caption: opposite the tip.
rlm@468	528 #+name: finger-UV
rlm@468	529 #+ATTR_latex: :width 10cm
rlm@468	530 [[./images/finger-UV.png]]
rlm@468	531
rlm@468	532 #+caption: Ventral side of the UV-mapped finger. Notice the
rlm@468	533 #+caption: density of touch sensors at the tip.
rlm@468	534 #+name: finger-side-view
rlm@468	535 #+ATTR_LaTeX: :width 10cm
rlm@468	536 [[./images/finger-1.png]]
rlm@468	537
rlm@465	538 ** COMMENT Video game engines are a great starting point
rlm@462	539
rlm@462	540 I did not need to write my own physics simulation code or shader to
rlm@462	541 build =CORTEX=. Doing so would lead to a system that is impossible
rlm@462	542 for anyone but myself to use anyway. Instead, I use a video game
rlm@462	543 engine as a base and modify it to accomodate the additional needs
rlm@462	544 of =CORTEX=. Video game engines are an ideal starting point to
rlm@462	545 build =CORTEX=, because they are not far from being creature
rlm@463	546 building systems themselves.
rlm@462	547
rlm@462	548 First off, general purpose video game engines come with a physics
rlm@462	549 engine and lighting / sound system. The physics system provides
rlm@462	550 tools that can be co-opted to serve as touch, proprioception, and
rlm@462	551 muscles. Since some games support split screen views, a good video
rlm@462	552 game engine will allow you to efficiently create multiple cameras
rlm@463	553 in the simulated world that can be used as eyes. Video game systems
rlm@463	554 offer integrated asset management for things like textures and
rlm@468	555 creatures models, providing an avenue for defining creatures. They
rlm@468	556 also understand UV-mapping, since this technique is used to apply a
rlm@468	557 texture to a model. Finally, because video game engines support a
rlm@468	558 large number of users, as long as =CORTEX= doesn't stray too far
rlm@468	559 from the base system, other researchers can turn to this community
rlm@468	560 for help when doing their research.
rlm@463	561
rlm@465	562 ** COMMENT =CORTEX= is based on jMonkeyEngine3
rlm@463	563
rlm@463	564 While preparing to build =CORTEX= I studied several video game
rlm@463	565 engines to see which would best serve as a base. The top contenders
rlm@463	566 were:
rlm@463	567
rlm@463	568 - [[http://www.idsoftware.com][Quake II]]/[[http://www.bytonic.de/html/jake2.html][Jake2]] :: The Quake II engine was designed by ID
rlm@463	569 software in 1997. All the source code was released by ID
rlm@463	570 software into the Public Domain several years ago, and as a
rlm@463	571 result it has been ported to many different languages. This
rlm@463	572 engine was famous for its advanced use of realistic shading
rlm@463	573 and had decent and fast physics simulation. The main advantage
rlm@463	574 of the Quake II engine is its simplicity, but I ultimately
rlm@463	575 rejected it because the engine is too tied to the concept of a
rlm@463	576 first-person shooter game. One of the problems I had was that
rlm@463	577 there does not seem to be any easy way to attach multiple
rlm@463	578 cameras to a single character. There are also several physics
rlm@463	579 clipping issues that are corrected in a way that only applies
rlm@463	580 to the main character and do not apply to arbitrary objects.
rlm@463	581
rlm@463	582 - [[http://source.valvesoftware.com/][Source Engine]] :: The Source Engine evolved from the Quake II
rlm@463	583 and Quake I engines and is used by Valve in the Half-Life
rlm@463	584 series of games. The physics simulation in the Source Engine
rlm@463	585 is quite accurate and probably the best out of all the engines
rlm@463	586 I investigated. There is also an extensive community actively
rlm@463	587 working with the engine. However, applications that use the
rlm@463	588 Source Engine must be written in C++, the code is not open, it
rlm@463	589 only runs on Windows, and the tools that come with the SDK to
rlm@463	590 handle models and textures are complicated and awkward to use.
rlm@463	591
rlm@463	592 - [[http://jmonkeyengine.com/][jMonkeyEngine3]] :: jMonkeyEngine3 is a new library for creating
rlm@463	593 games in Java. It uses OpenGL to render to the screen and uses
rlm@463	594 screengraphs to avoid drawing things that do not appear on the
rlm@463	595 screen. It has an active community and several games in the
rlm@463	596 pipeline. The engine was not built to serve any particular
rlm@463	597 game but is instead meant to be used for any 3D game.
rlm@463	598
rlm@463	599 I chose jMonkeyEngine3 because it because it had the most features
rlm@464	600 out of all the free projects I looked at, and because I could then
rlm@463	601 write my code in clojure, an implementation of =LISP= that runs on
rlm@463	602 the JVM.
rlm@435	603
rlm@469	604 ** COMMENT =CORTEX= uses Blender to create creature models
rlm@435	605
rlm@464	606 For the simple worm-like creatures I will use later on in this
rlm@464	607 thesis, I could define a simple API in =CORTEX= that would allow
rlm@464	608 one to create boxes, spheres, etc., and leave that API as the sole
rlm@464	609 way to create creatures. However, for =CORTEX= to truly be useful
rlm@468	610 for other projects, it needs a way to construct complicated
rlm@464	611 creatures. If possible, it would be nice to leverage work that has
rlm@464	612 already been done by the community of 3D modelers, or at least
rlm@464	613 enable people who are talented at moedling but not programming to
rlm@468	614 design =CORTEX= creatures.
rlm@464	615
rlm@464	616 Therefore, I use Blender, a free 3D modeling program, as the main
rlm@464	617 way to create creatures in =CORTEX=. However, the creatures modeled
rlm@464	618 in Blender must also be simple to simulate in jMonkeyEngine3's game
rlm@468	619 engine, and must also be easy to rig with =CORTEX='s senses. I
rlm@468	620 accomplish this with extensive use of Blender's ``empty nodes.''
rlm@464	621
rlm@468	622 Empty nodes have no mass, physical presence, or appearance, but
rlm@468	623 they can hold metadata and have names. I use a tree structure of
rlm@468	624 empty nodes to specify senses in the following manner:
rlm@468	625
rlm@468	626 - Create a single top-level empty node whose name is the name of
rlm@468	627 the sense.
rlm@468	628 - Add empty nodes which each contain meta-data relevant to the
rlm@468	629 sense, including a UV-map describing the number/distribution of
rlm@468	630 sensors if applicable.
rlm@468	631 - Make each empty-node the child of the top-level node.
rlm@468	632
rlm@468	633 #+caption: An example of annoting a creature model with empty
rlm@468	634 #+caption: nodes to describe the layout of senses. There are
rlm@468	635 #+caption: multiple empty nodes which each describe the position
rlm@468	636 #+caption: of muscles, ears, eyes, or joints.
rlm@468	637 #+name: sense-nodes
rlm@468	638 #+ATTR_LaTeX: :width 10cm
rlm@468	639 [[./images/empty-sense-nodes.png]]
rlm@468	640
rlm@469	641 ** COMMENT Bodies are composed of segments connected by joints
rlm@468	642
rlm@468	643 Blender is a general purpose animation tool, which has been used in
rlm@468	644 the past to create high quality movies such as Sintel
rlm@468	645 \cite{sintel}. Though Blender can model and render even complicated
rlm@468	646 things like water, it is crucual to keep models that are meant to
rlm@468	647 be simulated as creatures simple. =Bullet=, which =CORTEX= uses
rlm@468	648 though jMonkeyEngine3, is a rigid-body physics system. This offers
rlm@468	649 a compromise between the expressiveness of a game level and the
rlm@468	650 speed at which it can be simulated, and it means that creatures
rlm@468	651 should be naturally expressed as rigid components held together by
rlm@468	652 joint constraints.
rlm@468	653
rlm@468	654 But humans are more like a squishy bag with wrapped around some
rlm@468	655 hard bones which define the overall shape. When we move, our skin
rlm@468	656 bends and stretches to accomodate the new positions of our bones.
rlm@468	657
rlm@468	658 One way to make bodies composed of rigid pieces connected by joints
rlm@468	659 /seem/ more human-like is to use an /armature/, (or /rigging/)
rlm@468	660 system, which defines a overall ``body mesh'' and defines how the
rlm@468	661 mesh deforms as a function of the position of each ``bone'' which
rlm@468	662 is a standard rigid body. This technique is used extensively to
rlm@468	663 model humans and create realistic animations. It is not a good
rlm@468	664 technique for physical simulation, however because it creates a lie
rlm@468	665 -- the skin is not a physical part of the simulation and does not
rlm@468	666 interact with any objects in the world or itself. Objects will pass
rlm@468	667 right though the skin until they come in contact with the
rlm@468	668 underlying bone, which is a physical object. Whithout simulating
rlm@468	669 the skin, the sense of touch has little meaning, and the creature's
rlm@468	670 own vision will lie to it about the true extent of its body.
rlm@468	671 Simulating the skin as a physical object requires some way to
rlm@468	672 continuously update the physical model of the skin along with the
rlm@468	673 movement of the bones, which is unacceptably slow compared to rigid
rlm@468	674 body simulation.
rlm@468	675
rlm@468	676 Therefore, instead of using the human-like ``deformable bag of
rlm@468	677 bones'' approach, I decided to base my body plans on multiple solid
rlm@468	678 objects that are connected by joints, inspired by the robot =EVE=
rlm@468	679 from the movie WALL-E.
rlm@464	680
rlm@464	681 #+caption: =EVE= from the movie WALL-E. This body plan turns
rlm@464	682 #+caption: out to be much better suited to my purposes than a more
rlm@464	683 #+caption: human-like one.
rlm@465	684 #+ATTR_LaTeX: :width 10cm
rlm@464	685 [[./images/Eve.jpg]]
rlm@464	686
rlm@464	687 =EVE='s body is composed of several rigid components that are held
rlm@464	688 together by invisible joint constraints. This is what I mean by
rlm@464	689 ``eve-like''. The main reason that I use eve-style bodies is for
rlm@464	690 efficiency, and so that there will be correspondence between the
rlm@468	691 AI's semses and the physical presence of its body. Each individual
rlm@464	692 section is simulated by a separate rigid body that corresponds
rlm@464	693 exactly with its visual representation and does not change.
rlm@464	694 Sections are connected by invisible joints that are well supported
rlm@464	695 in jMonkeyEngine3. Bullet, the physics backend for jMonkeyEngine3,
rlm@464	696 can efficiently simulate hundreds of rigid bodies connected by
rlm@468	697 joints. Just because sections are rigid does not mean they have to
rlm@468	698 stay as one piece forever; they can be dynamically replaced with
rlm@468	699 multiple sections to simulate splitting in two. This could be used
rlm@468	700 to simulate retractable claws or =EVE='s hands, which are able to
rlm@468	701 coalesce into one object in the movie.
rlm@465	702
rlm@469	703 *** Solidifying/Connecting a body
rlm@465	704
rlm@469	705 =CORTEX= creates a creature in two steps: first, it traverses the
rlm@469	706 nodes in the blender file and creates physical representations for
rlm@469	707 any of them that have mass defined in their blender meta-data.
rlm@466	708
rlm@466	709 #+caption: Program for iterating through the nodes in a blender file
rlm@466	710 #+caption: and generating physical jMonkeyEngine3 objects with mass
rlm@466	711 #+caption: and a matching physics shape.
rlm@466	712 #+name: name
rlm@466	713 #+begin_listing clojure
rlm@466	714 #+begin_src clojure
rlm@466	715 (defn physical!
rlm@466	716 "Iterate through the nodes in creature and make them real physical
rlm@466	717 objects in the simulation."
rlm@466	718 [#^Node creature]
rlm@466	719 (dorun
rlm@466	720 (map
rlm@466	721 (fn [geom]
rlm@466	722 (let [physics-control
rlm@466	723 (RigidBodyControl.
rlm@466	724 (HullCollisionShape.
rlm@466	725 (.getMesh geom))
rlm@466	726 (if-let [mass (meta-data geom "mass")]
rlm@466	727 (float mass) (float 1)))]
rlm@466	728 (.addControl geom physics-control)))
rlm@466	729 (filter #(isa? (class %) Geometry )
rlm@466	730 (node-seq creature)))))
rlm@466	731 #+end_src
rlm@466	732 #+end_listing
rlm@465	733
rlm@469	734 The next step to making a proper body is to connect those pieces
rlm@469	735 together with joints. jMonkeyEngine has a large array of joints
rlm@469	736 available via =bullet=, such as Point2Point, Cone, Hinge, and a
rlm@469	737 generic Six Degree of Freedom joint, with or without spring
rlm@469	738 restitution.
rlm@465	739
rlm@469	740 Joints are treated a lot like proper senses, in that there is a
rlm@469	741 top-level empty node named ``joints'' whose children each
rlm@469	742 represent a joint.
rlm@466	743
rlm@469	744 #+caption: View of the hand model in Blender showing the main ``joints''
rlm@469	745 #+caption: node (highlighted in yellow) and its children which each
rlm@469	746 #+caption: represent a joint in the hand. Each joint node has metadata
rlm@469	747 #+caption: specifying what sort of joint it is.
rlm@469	748 #+name: blender-hand
rlm@469	749 #+ATTR_LaTeX: :width 10cm
rlm@469	750 [[./images/hand-screenshot1.png]]
rlm@469	751
rlm@469	752
rlm@469	753 =CORTEX='s procedure for binding the creature together with joints
rlm@469	754 is as follows:
rlm@469	755
rlm@469	756 - Find the children of the ``joints'' node.
rlm@469	757 - Determine the two spatials the joint is meant to connect.
rlm@469	758 - Create the joint based on the meta-data of the empty node.
rlm@469	759
rlm@469	760 The higher order function =sense-nodes= from =cortex.sense=
rlm@469	761 simplifies finding the joints based on their parent ``joints''
rlm@469	762 node.
rlm@466	763
rlm@466	764 #+caption: Retrieving the children empty nodes from a single
rlm@466	765 #+caption: named empty node is a common pattern in =CORTEX=
rlm@466	766 #+caption: further instances of this technique for the senses
rlm@466	767 #+caption: will be omitted
rlm@466	768 #+name: get-empty-nodes
rlm@466	769 #+begin_listing clojure
rlm@466	770 #+begin_src clojure
rlm@466	771 (defn sense-nodes
rlm@466	772 "For some senses there is a special empty blender node whose
rlm@466	773 children are considered markers for an instance of that sense. This
rlm@466	774 function generates functions to find those children, given the name
rlm@466	775 of the special parent node."
rlm@466	776 [parent-name]
rlm@466	777 (fn [#^Node creature]
rlm@466	778 (if-let [sense-node (.getChild creature parent-name)]
rlm@466	779 (seq (.getChildren sense-node)) [])))
rlm@466	780
rlm@466	781 (def
rlm@466	782 ^{:doc "Return the children of the creature's \"joints\" node."
rlm@466	783 :arglists '([creature])}
rlm@466	784 joints
rlm@466	785 (sense-nodes "joints"))
rlm@466	786 #+end_src
rlm@466	787 #+end_listing
rlm@466	788
rlm@469	789 To find a joint's targets, =CORTEX= creates a small cube, centered
rlm@469	790 around the empty-node, and grows the cube exponentially until it
rlm@469	791 intersects two physical objects. The objects are ordered according
rlm@469	792 to the joint's rotation, with the first one being the object that
rlm@469	793 has more negative coordinates in the joint's reference frame.
rlm@469	794 Since the objects must be physical, the empty-node itself escapes
rlm@469	795 detection. Because the objects must be physical, =joint-targets=
rlm@469	796 must be called /after/ =physical!= is called.
rlm@464	797
rlm@469	798 #+caption: Program to find the targets of a joint node by
rlm@469	799 #+caption: exponentiallly growth of a search cube.
rlm@469	800 #+name: joint-targets
rlm@469	801 #+begin_listing clojure
rlm@469	802 #+begin_src clojure
rlm@466	803 (defn joint-targets
rlm@466	804 "Return the two closest two objects to the joint object, ordered
rlm@466	805 from bottom to top according to the joint's rotation."
rlm@466	806 [#^Node parts #^Node joint]
rlm@466	807 (loop [radius (float 0.01)]
rlm@466	808 (let [results (CollisionResults.)]
rlm@466	809 (.collideWith
rlm@466	810 parts
rlm@466	811 (BoundingBox. (.getWorldTranslation joint)
rlm@466	812 radius radius radius) results)
rlm@466	813 (let [targets
rlm@466	814 (distinct
rlm@466	815 (map #(.getGeometry %) results))]
rlm@466	816 (if (>= (count targets) 2)
rlm@466	817 (sort-by
rlm@466	818 #(let [joint-ref-frame-position
rlm@466	819 (jme-to-blender
rlm@466	820 (.mult
rlm@466	821 (.inverse (.getWorldRotation joint))
rlm@466	822 (.subtract (.getWorldTranslation %)
rlm@466	823 (.getWorldTranslation joint))))]
rlm@466	824 (.dot (Vector3f. 1 1 1) joint-ref-frame-position))
rlm@466	825 (take 2 targets))
rlm@466	826 (recur (float (* radius 2))))))))
rlm@469	827 #+end_src
rlm@469	828 #+end_listing
rlm@464	829
rlm@469	830 Once =CORTEX= finds all joints and targets, it creates them using
rlm@469	831 a dispatch on the metadata of each joint node.
rlm@466	832
rlm@469	833 #+caption: Program to dispatch on blender metadata and create joints
rlm@469	834 #+caption: sutiable for physical simulation.
rlm@469	835 #+name: joint-dispatch
rlm@469	836 #+begin_listing clojure
rlm@469	837 #+begin_src clojure
rlm@466	838 (defmulti joint-dispatch
rlm@466	839 "Translate blender pseudo-joints into real JME joints."
rlm@466	840 (fn [constraints & _]
rlm@466	841 (:type constraints)))
rlm@466	842
rlm@466	843 (defmethod joint-dispatch :point
rlm@466	844 [constraints control-a control-b pivot-a pivot-b rotation]
rlm@466	845 (doto (SixDofJoint. control-a control-b pivot-a pivot-b false)
rlm@466	846 (.setLinearLowerLimit Vector3f/ZERO)
rlm@466	847 (.setLinearUpperLimit Vector3f/ZERO)))
rlm@466	848
rlm@466	849 (defmethod joint-dispatch :hinge
rlm@466	850 [constraints control-a control-b pivot-a pivot-b rotation]
rlm@466	851 (let [axis (if-let [axis (:axis constraints)] axis Vector3f/UNIT_X)
rlm@466	852 [limit-1 limit-2] (:limit constraints)
rlm@466	853 hinge-axis (.mult rotation (blender-to-jme axis))]
rlm@466	854 (doto (HingeJoint. control-a control-b pivot-a pivot-b
rlm@466	855 hinge-axis hinge-axis)
rlm@466	856 (.setLimit limit-1 limit-2))))
rlm@466	857
rlm@466	858 (defmethod joint-dispatch :cone
rlm@466	859 [constraints control-a control-b pivot-a pivot-b rotation]
rlm@466	860 (let [limit-xz (:limit-xz constraints)
rlm@466	861 limit-xy (:limit-xy constraints)
rlm@466	862 twist (:twist constraints)]
rlm@466	863 (doto (ConeJoint. control-a control-b pivot-a pivot-b
rlm@466	864 rotation rotation)
rlm@466	865 (.setLimit (float limit-xz) (float limit-xy)
rlm@466	866 (float twist)))))
rlm@469	867 #+end_src
rlm@469	868 #+end_listing
rlm@466	869
rlm@469	870 All that is left for joints it to combine the above pieces into a
rlm@469	871 something that can operate on the collection of nodes that a
rlm@469	872 blender file represents.
rlm@466	873
rlm@469	874 #+caption: Program to completely create a joint given information
rlm@469	875 #+caption: from a blender file.
rlm@469	876 #+name: connect
rlm@469	877 #+begin_listing clojure
rlm@466	878 #+begin_src clojure
rlm@466	879 (defn connect
rlm@466	880 "Create a joint between 'obj-a and 'obj-b at the location of
rlm@466	881 'joint. The type of joint is determined by the metadata on 'joint.
rlm@466	882
rlm@466	883 Here are some examples:
rlm@466	884 {:type :point}
rlm@466	885 {:type :hinge :limit [0 (/ Math/PI 2)] :axis (Vector3f. 0 1 0)}
rlm@466	886 (:axis defaults to (Vector3f. 1 0 0) if not provided for hinge joints)
rlm@466	887
rlm@466	888 {:type :cone :limit-xz 0]
rlm@466	889 :limit-xy 0]
rlm@466	890 :twist 0]} (use XZY rotation mode in blender!)"
rlm@466	891 [#^Node obj-a #^Node obj-b #^Node joint]
rlm@466	892 (let [control-a (.getControl obj-a RigidBodyControl)
rlm@466	893 control-b (.getControl obj-b RigidBodyControl)
rlm@466	894 joint-center (.getWorldTranslation joint)
rlm@466	895 joint-rotation (.toRotationMatrix (.getWorldRotation joint))
rlm@466	896 pivot-a (world-to-local obj-a joint-center)
rlm@466	897 pivot-b (world-to-local obj-b joint-center)]
rlm@466	898 (if-let
rlm@466	899 [constraints (map-vals eval (read-string (meta-data joint "joint")))]
rlm@466	900 ;; A side-effect of creating a joint registers
rlm@466	901 ;; it with both physics objects which in turn
rlm@466	902 ;; will register the joint with the physics system
rlm@466	903 ;; when the simulation is started.
rlm@466	904 (joint-dispatch constraints
rlm@466	905 control-a control-b
rlm@466	906 pivot-a pivot-b
rlm@466	907 joint-rotation))))
rlm@469	908 #+end_src
rlm@469	909 #+end_listing
rlm@466	910
rlm@469	911 In general, whenever =CORTEX= exposes a sense (or in this case
rlm@469	912 physicality), it provides a function of the type =sense!=, which
rlm@469	913 takes in a collection of nodes and augments it to support that
rlm@469	914 sense. The function returns any controlls necessary to use that
rlm@469	915 sense. In this case =body!= cerates a physical body and returns no
rlm@469	916 control functions.
rlm@466	917
rlm@469	918 #+caption: Program to give joints to a creature.
rlm@469	919 #+name: name
rlm@469	920 #+begin_listing clojure
rlm@469	921 #+begin_src clojure
rlm@466	922 (defn joints!
rlm@466	923 "Connect the solid parts of the creature with physical joints. The
rlm@466	924 joints are taken from the \"joints\" node in the creature."
rlm@466	925 [#^Node creature]
rlm@466	926 (dorun
rlm@466	927 (map
rlm@466	928 (fn [joint]
rlm@466	929 (let [[obj-a obj-b] (joint-targets creature joint)]
rlm@466	930 (connect obj-a obj-b joint)))
rlm@466	931 (joints creature))))
rlm@466	932 (defn body!
rlm@466	933 "Endow the creature with a physical body connected with joints. The
rlm@466	934 particulars of the joints and the masses of each body part are
rlm@466	935 determined in blender."
rlm@466	936 [#^Node creature]
rlm@466	937 (physical! creature)
rlm@466	938 (joints! creature))
rlm@469	939 #+end_src
rlm@469	940 #+end_listing
rlm@466	941
rlm@469	942 All of the code you have just seen amounts to only 130 lines, yet
rlm@469	943 because it builds on top of Blender and jMonkeyEngine3, those few
rlm@469	944 lines pack quite a punch!
rlm@466	945
rlm@469	946 The hand from figure \ref{blender-hand}, which was modeled after
rlm@469	947 my own right hand, can now be given joints and simulated as a
rlm@469	948 creature.
rlm@466	949
rlm@469	950 #+caption: With the ability to create physical creatures from blender,
rlm@469	951 #+caption: =CORTEX= gets one step closer to becomming a full creature
rlm@469	952 #+caption: simulation environment.
rlm@469	953 #+name: name
rlm@469	954 #+ATTR_LaTeX: :width 15cm
rlm@469	955 [[./images/physical-hand.png]]
rlm@468	956
rlm@436	957 ** Eyes reuse standard video game components
rlm@436	958
rlm@470	959 Vision is one of the most important senses for humans, so I need to
rlm@470	960 build a simulated sense of vision for my AI. I will do this with
rlm@470	961 simulated eyes. Each eye can be independently moved and should see
rlm@470	962 its own version of the world depending on where it is.
rlm@470	963
rlm@470	964 Making these simulated eyes a reality is simple because
rlm@470	965 jMonkeyEngine already contains extensive support for multiple views
rlm@470	966 of the same 3D simulated world. The reason jMonkeyEngine has this
rlm@470	967 support is because the support is necessary to create games with
rlm@470	968 split-screen views. Multiple views are also used to create
rlm@470	969 efficient pseudo-reflections by rendering the scene from a certain
rlm@470	970 perspective and then projecting it back onto a surface in the 3D
rlm@470	971 world.
rlm@470	972
rlm@470	973 #+caption: jMonkeyEngine supports multiple views to enable
rlm@470	974 #+caption: split-screen games, like GoldenEye, which was one of
rlm@470	975 #+caption: the first games to use split-screen views.
rlm@470	976 #+name: name
rlm@470	977 #+ATTR_LaTeX: :width 10cm
rlm@470	978 [[./images/goldeneye-4-player.png]]
rlm@470	979
rlm@470	980 *** A Brief Description of jMonkeyEngine's Rendering Pipeline
rlm@470	981
rlm@470	982 jMonkeyEngine allows you to create a =ViewPort=, which represents a
rlm@470	983 view of the simulated world. You can create as many of these as you
rlm@470	984 want. Every frame, the =RenderManager= iterates through each
rlm@470	985 =ViewPort=, rendering the scene in the GPU. For each =ViewPort= there
rlm@470	986 is a =FrameBuffer= which represents the rendered image in the GPU.
rlm@470	987
rlm@470	988 #+caption: =ViewPorts= are cameras in the world. During each frame,
rlm@470	989 #+caption: the =RenderManager= records a snapshot of what each view
rlm@470	990 #+caption: is currently seeing; these snapshots are =FrameBuffer= objects.
rlm@470	991 #+name: name
rlm@470	992 #+ATTR_LaTeX: :width 10cm
rlm@470	993 [[../images/diagram_rendermanager2.png]]
rlm@470	994
rlm@470	995 Each =ViewPort= can have any number of attached =SceneProcessor=
rlm@470	996 objects, which are called every time a new frame is rendered. A
rlm@470	997 =SceneProcessor= receives its =ViewPort's= =FrameBuffer= and can do
rlm@470	998 whatever it wants to the data. Often this consists of invoking GPU
rlm@470	999 specific operations on the rendered image. The =SceneProcessor= can
rlm@470	1000 also copy the GPU image data to RAM and process it with the CPU.
rlm@470	1001
rlm@470	1002 *** Appropriating Views for Vision
rlm@470	1003
rlm@470	1004 Each eye in the simulated creature needs its own =ViewPort= so
rlm@470	1005 that it can see the world from its own perspective. To this
rlm@470	1006 =ViewPort=, I add a =SceneProcessor= that feeds the visual data to
rlm@470	1007 any arbitrary continuation function for further processing. That
rlm@470	1008 continuation function may perform both CPU and GPU operations on
rlm@470	1009 the data. To make this easy for the continuation function, the
rlm@470	1010 =SceneProcessor= maintains appropriately sized buffers in RAM to
rlm@470	1011 hold the data. It does not do any copying from the GPU to the CPU
rlm@470	1012 itself because it is a slow operation.
rlm@470	1013
rlm@470	1014 #+caption: Function to make the rendered secne in jMonkeyEngine
rlm@470	1015 #+caption: available for further processing.
rlm@470	1016 #+name: pipeline-1
rlm@470	1017 #+begin_listing clojure
rlm@470	1018 #+begin_src clojure
rlm@470	1019 (defn vision-pipeline
rlm@470	1020 "Create a SceneProcessor object which wraps a vision processing
rlm@470	1021 continuation function. The continuation is a function that takes
rlm@470	1022 [#^Renderer r #^FrameBuffer fb #^ByteBuffer b #^BufferedImage bi],
rlm@470	1023 each of which has already been appropriately sized."
rlm@470	1024 [continuation]
rlm@470	1025 (let [byte-buffer (atom nil)
rlm@470	1026 renderer (atom nil)
rlm@470	1027 image (atom nil)]
rlm@470	1028 (proxy [SceneProcessor] []
rlm@470	1029 (initialize
rlm@470	1030 [renderManager viewPort]
rlm@470	1031 (let [cam (.getCamera viewPort)
rlm@470	1032 width (.getWidth cam)
rlm@470	1033 height (.getHeight cam)]
rlm@470	1034 (reset! renderer (.getRenderer renderManager))
rlm@470	1035 (reset! byte-buffer
rlm@470	1036 (BufferUtils/createByteBuffer
rlm@470	1037 (* width height 4)))
rlm@470	1038 (reset! image (BufferedImage.
rlm@470	1039 width height
rlm@470	1040 BufferedImage/TYPE_4BYTE_ABGR))))
rlm@470	1041 (isInitialized [] (not (nil? @byte-buffer)))
rlm@470	1042 (reshape [_ _ _])
rlm@470	1043 (preFrame [_])
rlm@470	1044 (postQueue [_])
rlm@470	1045 (postFrame
rlm@470	1046 [#^FrameBuffer fb]
rlm@470	1047 (.clear @byte-buffer)
rlm@470	1048 (continuation @renderer fb @byte-buffer @image))
rlm@470	1049 (cleanup []))))
rlm@470	1050 #+end_src
rlm@470	1051 #+end_listing
rlm@470	1052
rlm@470	1053 The continuation function given to =vision-pipeline= above will be
rlm@470	1054 given a =Renderer= and three containers for image data. The
rlm@470	1055 =FrameBuffer= references the GPU image data, but the pixel data
rlm@470	1056 can not be used directly on the CPU. The =ByteBuffer= and
rlm@470	1057 =BufferedImage= are initially "empty" but are sized to hold the
rlm@470	1058 data in the =FrameBuffer=. I call transferring the GPU image data
rlm@470	1059 to the CPU structures "mixing" the image data.
rlm@470	1060
rlm@470	1061 *** Optical sensor arrays are described with images and referenced with metadata
rlm@470	1062
rlm@470	1063 The vision pipeline described above handles the flow of rendered
rlm@470	1064 images. Now, =CORTEX= needs simulated eyes to serve as the source
rlm@470	1065 of these images.
rlm@470	1066
rlm@470	1067 An eye is described in blender in the same way as a joint. They
rlm@470	1068 are zero dimensional empty objects with no geometry whose local
rlm@470	1069 coordinate system determines the orientation of the resulting eye.
rlm@470	1070 All eyes are children of a parent node named "eyes" just as all
rlm@470	1071 joints have a parent named "joints". An eye binds to the nearest
rlm@470	1072 physical object with =bind-sense=.
rlm@470	1073
rlm@470	1074 #+caption: Here, the camera is created based on metadata on the
rlm@470	1075 #+caption: eye-node and attached to the nearest physical object
rlm@470	1076 #+caption: with =bind-sense=
rlm@470	1077 #+name: add-eye
rlm@470	1078 #+begin_listing clojure
rlm@470	1079 (defn add-eye!
rlm@470	1080 "Create a Camera centered on the current position of 'eye which
rlm@470	1081 follows the closest physical node in 'creature. The camera will
rlm@470	1082 point in the X direction and use the Z vector as up as determined
rlm@470	1083 by the rotation of these vectors in blender coordinate space. Use
rlm@470	1084 XZY rotation for the node in blender."
rlm@470	1085 [#^Node creature #^Spatial eye]
rlm@470	1086 (let [target (closest-node creature eye)
rlm@470	1087 [cam-width cam-height]
rlm@470	1088 ;;[640 480] ;; graphics card on laptop doesn't support
rlm@470	1089 ;; arbitray dimensions.
rlm@470	1090 (eye-dimensions eye)
rlm@470	1091 cam (Camera. cam-width cam-height)
rlm@470	1092 rot (.getWorldRotation eye)]
rlm@470	1093 (.setLocation cam (.getWorldTranslation eye))
rlm@470	1094 (.lookAtDirection
rlm@470	1095 cam ; this part is not a mistake and
rlm@470	1096 (.mult rot Vector3f/UNIT_X) ; is consistent with using Z in
rlm@470	1097 (.mult rot Vector3f/UNIT_Y)) ; blender as the UP vector.
rlm@470	1098 (.setFrustumPerspective
rlm@470	1099 cam (float 45)
rlm@470	1100 (float (/ (.getWidth cam) (.getHeight cam)))
rlm@470	1101 (float 1)
rlm@470	1102 (float 1000))
rlm@470	1103 (bind-sense target cam) cam))
rlm@470	1104 #+end_listing
rlm@470	1105
rlm@470	1106 *** Simulated Retina
rlm@470	1107
rlm@470	1108 An eye is a surface (the retina) which contains many discrete
rlm@470	1109 sensors to detect light. These sensors can have different
rlm@470	1110 light-sensing properties. In humans, each discrete sensor is
rlm@470	1111 sensitive to red, blue, green, or gray. These different types of
rlm@470	1112 sensors can have different spatial distributions along the retina.
rlm@470	1113 In humans, there is a fovea in the center of the retina which has
rlm@470	1114 a very high density of color sensors, and a blind spot which has
rlm@470	1115 no sensors at all. Sensor density decreases in proportion to
rlm@470	1116 distance from the fovea.
rlm@470	1117
rlm@470	1118 I want to be able to model any retinal configuration, so my
rlm@470	1119 eye-nodes in blender contain metadata pointing to images that
rlm@470	1120 describe the precise position of the individual sensors using
rlm@470	1121 white pixels. The meta-data also describes the precise sensitivity
rlm@470	1122 to light that the sensors described in the image have. An eye can
rlm@470	1123 contain any number of these images. For example, the metadata for
rlm@470	1124 an eye might look like this:
rlm@470	1125
rlm@470	1126 #+begin_src clojure
rlm@470	1127 {0xFF0000 "Models/test-creature/retina-small.png"}
rlm@470	1128 #+end_src
rlm@470	1129
rlm@470	1130 #+caption: An example retinal profile image. White pixels are
rlm@470	1131 #+caption: photo-sensitive elements. The distribution of white
rlm@470	1132 #+caption: pixels is denser in the middle and falls off at the
rlm@470	1133 #+caption: edges and is inspired by the human retina.
rlm@470	1134 #+name: retina
rlm@470	1135 #+ATTR_LaTeX: :width 10cm
rlm@470	1136 [[./images/retina-small.png]]
rlm@470	1137
rlm@470	1138 Together, the number 0xFF0000 and the image image above describe
rlm@470	1139 the placement of red-sensitive sensory elements.
rlm@470	1140
rlm@470	1141 Meta-data to very crudely approximate a human eye might be
rlm@470	1142 something like this:
rlm@470	1143
rlm@470	1144 #+begin_src clojure
rlm@470	1145 (let [retinal-profile "Models/test-creature/retina-small.png"]
rlm@470	1146 {0xFF0000 retinal-profile
rlm@470	1147 0x00FF00 retinal-profile
rlm@470	1148 0x0000FF retinal-profile
rlm@470	1149 0xFFFFFF retinal-profile})
rlm@470	1150 #+end_src
rlm@470	1151
rlm@470	1152 The numbers that serve as keys in the map determine a sensor's
rlm@470	1153 relative sensitivity to the channels red, green, and blue. These
rlm@470	1154 sensitivity values are packed into an integer in the order
rlm@470	1155 =\|_\|R\|G\|B\|= in 8-bit fields. The RGB values of a pixel in the
rlm@470	1156 image are added together with these sensitivities as linear
rlm@470	1157 weights. Therefore, 0xFF0000 means sensitive to red only while
rlm@470	1158 0xFFFFFF means sensitive to all colors equally (gray).
rlm@470	1159
rlm@470	1160 #+caption: This is the core of vision in =CORTEX=. A given eye node
rlm@470	1161 #+caption: is converted into a function that returns visual
rlm@470	1162 #+caption: information from the simulation.
rlm@471	1163 #+name: vision-kernel
rlm@470	1164 #+begin_listing clojure
rlm@470	1165 (defn vision-kernel
rlm@470	1166 "Returns a list of functions, each of which will return a color
rlm@470	1167 channel's worth of visual information when called inside a running
rlm@470	1168 simulation."
rlm@470	1169 [#^Node creature #^Spatial eye & {skip :skip :or {skip 0}}]
rlm@470	1170 (let [retinal-map (retina-sensor-profile eye)
rlm@470	1171 camera (add-eye! creature eye)
rlm@470	1172 vision-image
rlm@470	1173 (atom
rlm@470	1174 (BufferedImage. (.getWidth camera)
rlm@470	1175 (.getHeight camera)
rlm@470	1176 BufferedImage/TYPE_BYTE_BINARY))
rlm@470	1177 register-eye!
rlm@470	1178 (runonce
rlm@470	1179 (fn [world]
rlm@470	1180 (add-camera!
rlm@470	1181 world camera
rlm@470	1182 (let [counter (atom 0)]
rlm@470	1183 (fn [r fb bb bi]
rlm@470	1184 (if (zero? (rem (swap! counter inc) (inc skip)))
rlm@470	1185 (reset! vision-image
rlm@470	1186 (BufferedImage! r fb bb bi))))))))]
rlm@470	1187 (vec
rlm@470	1188 (map
rlm@470	1189 (fn [[key image]]
rlm@470	1190 (let [whites (white-coordinates image)
rlm@470	1191 topology (vec (collapse whites))
rlm@470	1192 sensitivity (sensitivity-presets key key)]
rlm@470	1193 (attached-viewport.
rlm@470	1194 (fn [world]
rlm@470	1195 (register-eye! world)
rlm@470	1196 (vector
rlm@470	1197 topology
rlm@470	1198 (vec
rlm@470	1199 (for [[x y] whites]
rlm@470	1200 (pixel-sense
rlm@470	1201 sensitivity
rlm@470	1202 (.getRGB @vision-image x y))))))
rlm@470	1203 register-eye!)))
rlm@470	1204 retinal-map))))
rlm@470	1205 #+end_listing
rlm@470	1206
rlm@470	1207 Note that since each of the functions generated by =vision-kernel=
rlm@470	1208 shares the same =register-eye!= function, the eye will be
rlm@470	1209 registered only once the first time any of the functions from the
rlm@470	1210 list returned by =vision-kernel= is called. Each of the functions
rlm@470	1211 returned by =vision-kernel= also allows access to the =Viewport=
rlm@470	1212 through which it receives images.
rlm@470	1213
rlm@470	1214 All the hard work has been done; all that remains is to apply
rlm@470	1215 =vision-kernel= to each eye in the creature and gather the results
rlm@470	1216 into one list of functions.
rlm@470	1217
rlm@470	1218
rlm@470	1219 #+caption: With =vision!=, =CORTEX= is already a fine simulation
rlm@470	1220 #+caption: environment for experimenting with different types of
rlm@470	1221 #+caption: eyes.
rlm@470	1222 #+name: vision!
rlm@470	1223 #+begin_listing clojure
rlm@470	1224 (defn vision!
rlm@470	1225 "Returns a list of functions, each of which returns visual sensory
rlm@470	1226 data when called inside a running simulation."
rlm@470	1227 [#^Node creature & {skip :skip :or {skip 0}}]
rlm@470	1228 (reduce
rlm@470	1229 concat
rlm@470	1230 (for [eye (eyes creature)]
rlm@470	1231 (vision-kernel creature eye))))
rlm@470	1232 #+end_listing
rlm@470	1233
rlm@471	1234 #+caption: Simulated vision with a test creature and the
rlm@471	1235 #+caption: human-like eye approximation. Notice how each channel
rlm@471	1236 #+caption: of the eye responds differently to the differently
rlm@471	1237 #+caption: colored balls.
rlm@471	1238 #+name: worm-vision-test.
rlm@471	1239 #+ATTR_LaTeX: :width 13cm
rlm@471	1240 [[./images/worm-vision.png]]
rlm@470	1241
rlm@471	1242 The vision code is not much more complicated than the body code,
rlm@471	1243 and enables multiple further paths for simulated vision. For
rlm@471	1244 example, it is quite easy to create bifocal vision -- you just
rlm@471	1245 make two eyes next to each other in blender! It is also possible
rlm@471	1246 to encode vision transforms in the retinal files. For example, the
rlm@471	1247 human like retina file in figure \ref{retina} approximates a
rlm@471	1248 log-polar transform.
rlm@470	1249
rlm@471	1250 This vision code has already been absorbed by the jMonkeyEngine
rlm@471	1251 community and is now (in modified form) part of a system for
rlm@471	1252 capturing in-game video to a file.
rlm@470	1253
rlm@436	1254 ** Hearing is hard; =CORTEX= does it right
rlm@436	1255
rlm@436	1256 ** Touch uses hundreds of hair-like elements
rlm@436	1257
rlm@440	1258 ** Proprioception is the sense that makes everything ``real''
rlm@436	1259
rlm@436	1260 ** Muscles are both effectors and sensors
rlm@436	1261
rlm@436	1262 ** =CORTEX= brings complex creatures to life!
rlm@436	1263
rlm@436	1264 ** =CORTEX= enables many possiblities for further research
rlm@435	1265
rlm@465	1266 * COMMENT Empathy in a simulated worm
rlm@435	1267
rlm@449	1268 Here I develop a computational model of empathy, using =CORTEX= as a
rlm@449	1269 base. Empathy in this context is the ability to observe another
rlm@449	1270 creature and infer what sorts of sensations that creature is
rlm@449	1271 feeling. My empathy algorithm involves multiple phases. First is
rlm@449	1272 free-play, where the creature moves around and gains sensory
rlm@449	1273 experience. From this experience I construct a representation of the
rlm@449	1274 creature's sensory state space, which I call \Phi-space. Using
rlm@449	1275 \Phi-space, I construct an efficient function which takes the
rlm@449	1276 limited data that comes from observing another creature and enriches
rlm@449	1277 it full compliment of imagined sensory data. I can then use the
rlm@449	1278 imagined sensory data to recognize what the observed creature is
rlm@449	1279 doing and feeling, using straightforward embodied action predicates.
rlm@449	1280 This is all demonstrated with using a simple worm-like creature, and
rlm@449	1281 recognizing worm-actions based on limited data.
rlm@449	1282
rlm@449	1283 #+caption: Here is the worm with which we will be working.
rlm@449	1284 #+caption: It is composed of 5 segments. Each segment has a
rlm@449	1285 #+caption: pair of extensor and flexor muscles. Each of the
rlm@449	1286 #+caption: worm's four joints is a hinge joint which allows
rlm@451	1287 #+caption: about 30 degrees of rotation to either side. Each segment
rlm@449	1288 #+caption: of the worm is touch-capable and has a uniform
rlm@449	1289 #+caption: distribution of touch sensors on each of its faces.
rlm@449	1290 #+caption: Each joint has a proprioceptive sense to detect
rlm@449	1291 #+caption: relative positions. The worm segments are all the
rlm@449	1292 #+caption: same except for the first one, which has a much
rlm@449	1293 #+caption: higher weight than the others to allow for easy
rlm@449	1294 #+caption: manual motor control.
rlm@449	1295 #+name: basic-worm-view
rlm@449	1296 #+ATTR_LaTeX: :width 10cm
rlm@449	1297 [[./images/basic-worm-view.png]]
rlm@449	1298
rlm@449	1299 #+caption: Program for reading a worm from a blender file and
rlm@449	1300 #+caption: outfitting it with the senses of proprioception,
rlm@449	1301 #+caption: touch, and the ability to move, as specified in the
rlm@449	1302 #+caption: blender file.
rlm@449	1303 #+name: get-worm
rlm@449	1304 #+begin_listing clojure
rlm@449	1305 #+begin_src clojure
rlm@449	1306 (defn worm []
rlm@449	1307 (let [model (load-blender-model "Models/worm/worm.blend")]
rlm@449	1308 {:body (doto model (body!))
rlm@449	1309 :touch (touch! model)
rlm@449	1310 :proprioception (proprioception! model)
rlm@449	1311 :muscles (movement! model)}))
rlm@449	1312 #+end_src
rlm@449	1313 #+end_listing
rlm@452	1314
rlm@436	1315 ** Embodiment factors action recognition into managable parts
rlm@435	1316
rlm@449	1317 Using empathy, I divide the problem of action recognition into a
rlm@449	1318 recognition process expressed in the language of a full compliment
rlm@449	1319 of senses, and an imaganitive process that generates full sensory
rlm@449	1320 data from partial sensory data. Splitting the action recognition
rlm@449	1321 problem in this manner greatly reduces the total amount of work to
rlm@449	1322 recognize actions: The imaganitive process is mostly just matching
rlm@449	1323 previous experience, and the recognition process gets to use all
rlm@449	1324 the senses to directly describe any action.
rlm@449	1325
rlm@436	1326 ** Action recognition is easy with a full gamut of senses
rlm@435	1327
rlm@449	1328 Embodied representations using multiple senses such as touch,
rlm@449	1329 proprioception, and muscle tension turns out be be exceedingly
rlm@449	1330 efficient at describing body-centered actions. It is the ``right
rlm@449	1331 language for the job''. For example, it takes only around 5 lines
rlm@449	1332 of LISP code to describe the action of ``curling'' using embodied
rlm@451	1333 primitives. It takes about 10 lines to describe the seemingly
rlm@449	1334 complicated action of wiggling.
rlm@449	1335
rlm@449	1336 The following action predicates each take a stream of sensory
rlm@449	1337 experience, observe however much of it they desire, and decide
rlm@449	1338 whether the worm is doing the action they describe. =curled?=
rlm@449	1339 relies on proprioception, =resting?= relies on touch, =wiggling?=
rlm@449	1340 relies on a fourier analysis of muscle contraction, and
rlm@449	1341 =grand-circle?= relies on touch and reuses =curled?= as a gaurd.
rlm@449	1342
rlm@449	1343 #+caption: Program for detecting whether the worm is curled. This is the
rlm@449	1344 #+caption: simplest action predicate, because it only uses the last frame
rlm@449	1345 #+caption: of sensory experience, and only uses proprioceptive data. Even
rlm@449	1346 #+caption: this simple predicate, however, is automatically frame
rlm@449	1347 #+caption: independent and ignores vermopomorphic differences such as
rlm@449	1348 #+caption: worm textures and colors.
rlm@449	1349 #+name: curled
rlm@452	1350 #+attr_latex: [htpb]
rlm@452	1351 #+begin_listing clojure
rlm@449	1352 #+begin_src clojure
rlm@449	1353 (defn curled?
rlm@449	1354 "Is the worm curled up?"
rlm@449	1355 [experiences]
rlm@449	1356 (every?
rlm@449	1357 (fn [[_ _ bend]]
rlm@449	1358 (> (Math/sin bend) 0.64))
rlm@449	1359 (:proprioception (peek experiences))))
rlm@449	1360 #+end_src
rlm@449	1361 #+end_listing
rlm@449	1362
rlm@449	1363 #+caption: Program for summarizing the touch information in a patch
rlm@449	1364 #+caption: of skin.
rlm@449	1365 #+name: touch-summary
rlm@452	1366 #+attr_latex: [htpb]
rlm@452	1367
rlm@452	1368 #+begin_listing clojure
rlm@449	1369 #+begin_src clojure
rlm@449	1370 (defn contact
rlm@449	1371 "Determine how much contact a particular worm segment has with
rlm@449	1372 other objects. Returns a value between 0 and 1, where 1 is full
rlm@449	1373 contact and 0 is no contact."
rlm@449	1374 [touch-region [coords contact :as touch]]
rlm@449	1375 (-> (zipmap coords contact)
rlm@449	1376 (select-keys touch-region)
rlm@449	1377 (vals)
rlm@449	1378 (#(map first %))
rlm@449	1379 (average)
rlm@449	1380 (* 10)
rlm@449	1381 (- 1)
rlm@449	1382 (Math/abs)))
rlm@449	1383 #+end_src
rlm@449	1384 #+end_listing
rlm@449	1385
rlm@449	1386
rlm@449	1387 #+caption: Program for detecting whether the worm is at rest. This program
rlm@449	1388 #+caption: uses a summary of the tactile information from the underbelly
rlm@449	1389 #+caption: of the worm, and is only true if every segment is touching the
rlm@449	1390 #+caption: floor. Note that this function contains no references to
rlm@449	1391 #+caption: proprioction at all.
rlm@449	1392 #+name: resting
rlm@452	1393 #+attr_latex: [htpb]
rlm@452	1394 #+begin_listing clojure
rlm@449	1395 #+begin_src clojure
rlm@449	1396 (def worm-segment-bottom (rect-region [8 15] [14 22]))
rlm@449	1397
rlm@449	1398 (defn resting?
rlm@449	1399 "Is the worm resting on the ground?"
rlm@449	1400 [experiences]
rlm@449	1401 (every?
rlm@449	1402 (fn [touch-data]
rlm@449	1403 (< 0.9 (contact worm-segment-bottom touch-data)))
rlm@449	1404 (:touch (peek experiences))))
rlm@449	1405 #+end_src
rlm@449	1406 #+end_listing
rlm@449	1407
rlm@449	1408 #+caption: Program for detecting whether the worm is curled up into a
rlm@449	1409 #+caption: full circle. Here the embodied approach begins to shine, as
rlm@449	1410 #+caption: I am able to both use a previous action predicate (=curled?=)
rlm@449	1411 #+caption: as well as the direct tactile experience of the head and tail.
rlm@449	1412 #+name: grand-circle
rlm@452	1413 #+attr_latex: [htpb]
rlm@452	1414 #+begin_listing clojure
rlm@449	1415 #+begin_src clojure
rlm@449	1416 (def worm-segment-bottom-tip (rect-region [15 15] [22 22]))
rlm@449	1417
rlm@449	1418 (def worm-segment-top-tip (rect-region [0 15] [7 22]))
rlm@449	1419
rlm@449	1420 (defn grand-circle?
rlm@449	1421 "Does the worm form a majestic circle (one end touching the other)?"
rlm@449	1422 [experiences]
rlm@449	1423 (and (curled? experiences)
rlm@449	1424 (let [worm-touch (:touch (peek experiences))
rlm@449	1425 tail-touch (worm-touch 0)
rlm@449	1426 head-touch (worm-touch 4)]
rlm@449	1427 (and (< 0.55 (contact worm-segment-bottom-tip tail-touch))
rlm@449	1428 (< 0.55 (contact worm-segment-top-tip head-touch))))))
rlm@449	1429 #+end_src
rlm@449	1430 #+end_listing
rlm@449	1431
rlm@449	1432
rlm@449	1433 #+caption: Program for detecting whether the worm has been wiggling for
rlm@449	1434 #+caption: the last few frames. It uses a fourier analysis of the muscle
rlm@449	1435 #+caption: contractions of the worm's tail to determine wiggling. This is
rlm@449	1436 #+caption: signigicant because there is no particular frame that clearly
rlm@449	1437 #+caption: indicates that the worm is wiggling --- only when multiple frames
rlm@449	1438 #+caption: are analyzed together is the wiggling revealed. Defining
rlm@449	1439 #+caption: wiggling this way also gives the worm an opportunity to learn
rlm@449	1440 #+caption: and recognize ``frustrated wiggling'', where the worm tries to
rlm@449	1441 #+caption: wiggle but can't. Frustrated wiggling is very visually different
rlm@449	1442 #+caption: from actual wiggling, but this definition gives it to us for free.
rlm@449	1443 #+name: wiggling
rlm@452	1444 #+attr_latex: [htpb]
rlm@452	1445 #+begin_listing clojure
rlm@449	1446 #+begin_src clojure
rlm@449	1447 (defn fft [nums]
rlm@449	1448 (map
rlm@449	1449 #(.getReal %)
rlm@449	1450 (.transform
rlm@449	1451 (FastFourierTransformer. DftNormalization/STANDARD)
rlm@449	1452 (double-array nums) TransformType/FORWARD)))
rlm@449	1453
rlm@449	1454 (def indexed (partial map-indexed vector))
rlm@449	1455
rlm@449	1456 (defn max-indexed [s]
rlm@449	1457 (first (sort-by (comp - second) (indexed s))))
rlm@449	1458
rlm@449	1459 (defn wiggling?
rlm@449	1460 "Is the worm wiggling?"
rlm@449	1461 [experiences]
rlm@449	1462 (let [analysis-interval 0x40]
rlm@449	1463 (when (> (count experiences) analysis-interval)
rlm@449	1464 (let [a-flex 3
rlm@449	1465 a-ex 2
rlm@449	1466 muscle-activity
rlm@449	1467 (map :muscle (vector:last-n experiences analysis-interval))
rlm@449	1468 base-activity
rlm@449	1469 (map #(- (% a-flex) (% a-ex)) muscle-activity)]
rlm@449	1470 (= 2
rlm@449	1471 (first
rlm@449	1472 (max-indexed
rlm@449	1473 (map #(Math/abs %)
rlm@449	1474 (take 20 (fft base-activity))))))))))
rlm@449	1475 #+end_src
rlm@449	1476 #+end_listing
rlm@449	1477
rlm@449	1478 With these action predicates, I can now recognize the actions of
rlm@449	1479 the worm while it is moving under my control and I have access to
rlm@449	1480 all the worm's senses.
rlm@449	1481
rlm@449	1482 #+caption: Use the action predicates defined earlier to report on
rlm@449	1483 #+caption: what the worm is doing while in simulation.
rlm@449	1484 #+name: report-worm-activity
rlm@452	1485 #+attr_latex: [htpb]
rlm@452	1486 #+begin_listing clojure
rlm@449	1487 #+begin_src clojure
rlm@449	1488 (defn debug-experience
rlm@449	1489 [experiences text]
rlm@449	1490 (cond
rlm@449	1491 (grand-circle? experiences) (.setText text "Grand Circle")
rlm@449	1492 (curled? experiences) (.setText text "Curled")
rlm@449	1493 (wiggling? experiences) (.setText text "Wiggling")
rlm@449	1494 (resting? experiences) (.setText text "Resting")))
rlm@449	1495 #+end_src
rlm@449	1496 #+end_listing
rlm@449	1497
rlm@449	1498 #+caption: Using =debug-experience=, the body-centered predicates
rlm@449	1499 #+caption: work together to classify the behaviour of the worm.
rlm@451	1500 #+caption: the predicates are operating with access to the worm's
rlm@451	1501 #+caption: full sensory data.
rlm@449	1502 #+name: basic-worm-view
rlm@449	1503 #+ATTR_LaTeX: :width 10cm
rlm@449	1504 [[./images/worm-identify-init.png]]
rlm@449	1505
rlm@449	1506 These action predicates satisfy the recognition requirement of an
rlm@451	1507 empathic recognition system. There is power in the simplicity of
rlm@451	1508 the action predicates. They describe their actions without getting
rlm@451	1509 confused in visual details of the worm. Each one is frame
rlm@451	1510 independent, but more than that, they are each indepent of
rlm@449	1511 irrelevant visual details of the worm and the environment. They
rlm@449	1512 will work regardless of whether the worm is a different color or
rlm@451	1513 hevaily textured, or if the environment has strange lighting.
rlm@449	1514
rlm@449	1515 The trick now is to make the action predicates work even when the
rlm@449	1516 sensory data on which they depend is absent. If I can do that, then
rlm@449	1517 I will have gained much,
rlm@435	1518
rlm@436	1519 ** \Phi-space describes the worm's experiences
rlm@449	1520
rlm@449	1521 As a first step towards building empathy, I need to gather all of
rlm@449	1522 the worm's experiences during free play. I use a simple vector to
rlm@449	1523 store all the experiences.
rlm@449	1524
rlm@449	1525 Each element of the experience vector exists in the vast space of
rlm@449	1526 all possible worm-experiences. Most of this vast space is actually
rlm@449	1527 unreachable due to physical constraints of the worm's body. For
rlm@449	1528 example, the worm's segments are connected by hinge joints that put
rlm@451	1529 a practical limit on the worm's range of motions without limiting
rlm@451	1530 its degrees of freedom. Some groupings of senses are impossible;
rlm@451	1531 the worm can not be bent into a circle so that its ends are
rlm@451	1532 touching and at the same time not also experience the sensation of
rlm@451	1533 touching itself.
rlm@449	1534
rlm@451	1535 As the worm moves around during free play and its experience vector
rlm@451	1536 grows larger, the vector begins to define a subspace which is all
rlm@451	1537 the sensations the worm can practicaly experience during normal
rlm@451	1538 operation. I call this subspace \Phi-space, short for
rlm@451	1539 physical-space. The experience vector defines a path through
rlm@451	1540 \Phi-space. This path has interesting properties that all derive
rlm@451	1541 from physical embodiment. The proprioceptive components are
rlm@451	1542 completely smooth, because in order for the worm to move from one
rlm@451	1543 position to another, it must pass through the intermediate
rlm@451	1544 positions. The path invariably forms loops as actions are repeated.
rlm@451	1545 Finally and most importantly, proprioception actually gives very
rlm@451	1546 strong inference about the other senses. For example, when the worm
rlm@451	1547 is flat, you can infer that it is touching the ground and that its
rlm@451	1548 muscles are not active, because if the muscles were active, the
rlm@451	1549 worm would be moving and would not be perfectly flat. In order to
rlm@451	1550 stay flat, the worm has to be touching the ground, or it would
rlm@451	1551 again be moving out of the flat position due to gravity. If the
rlm@451	1552 worm is positioned in such a way that it interacts with itself,
rlm@451	1553 then it is very likely to be feeling the same tactile feelings as
rlm@451	1554 the last time it was in that position, because it has the same body
rlm@451	1555 as then. If you observe multiple frames of proprioceptive data,
rlm@451	1556 then you can become increasingly confident about the exact
rlm@451	1557 activations of the worm's muscles, because it generally takes a
rlm@451	1558 unique combination of muscle contractions to transform the worm's
rlm@451	1559 body along a specific path through \Phi-space.
rlm@449	1560
rlm@449	1561 There is a simple way of taking \Phi-space and the total ordering
rlm@449	1562 provided by an experience vector and reliably infering the rest of
rlm@449	1563 the senses.
rlm@435	1564
rlm@436	1565 ** Empathy is the process of tracing though \Phi-space
rlm@449	1566
rlm@450	1567 Here is the core of a basic empathy algorithm, starting with an
rlm@451	1568 experience vector:
rlm@451	1569
rlm@451	1570 First, group the experiences into tiered proprioceptive bins. I use
rlm@451	1571 powers of 10 and 3 bins, and the smallest bin has an approximate
rlm@451	1572 size of 0.001 radians in all proprioceptive dimensions.
rlm@450	1573
rlm@450	1574 Then, given a sequence of proprioceptive input, generate a set of
rlm@451	1575 matching experience records for each input, using the tiered
rlm@451	1576 proprioceptive bins.
rlm@449	1577
rlm@450	1578 Finally, to infer sensory data, select the longest consective chain
rlm@451	1579 of experiences. Conecutive experience means that the experiences
rlm@451	1580 appear next to each other in the experience vector.
rlm@449	1581
rlm@450	1582 This algorithm has three advantages:
rlm@450	1583
rlm@450	1584 1. It's simple
rlm@450	1585
rlm@451	1586 3. It's very fast -- retrieving possible interpretations takes
rlm@451	1587 constant time. Tracing through chains of interpretations takes
rlm@451	1588 time proportional to the average number of experiences in a
rlm@451	1589 proprioceptive bin. Redundant experiences in \Phi-space can be
rlm@451	1590 merged to save computation.
rlm@450	1591
rlm@450	1592 2. It protects from wrong interpretations of transient ambiguous
rlm@451	1593 proprioceptive data. For example, if the worm is flat for just
rlm@450	1594 an instant, this flattness will not be interpreted as implying
rlm@450	1595 that the worm has its muscles relaxed, since the flattness is
rlm@450	1596 part of a longer chain which includes a distinct pattern of
rlm@451	1597 muscle activation. Markov chains or other memoryless statistical
rlm@451	1598 models that operate on individual frames may very well make this
rlm@451	1599 mistake.
rlm@450	1600
rlm@450	1601 #+caption: Program to convert an experience vector into a
rlm@450	1602 #+caption: proprioceptively binned lookup function.
rlm@450	1603 #+name: bin
rlm@452	1604 #+attr_latex: [htpb]
rlm@452	1605 #+begin_listing clojure
rlm@450	1606 #+begin_src clojure
rlm@449	1607 (defn bin [digits]
rlm@449	1608 (fn [angles]
rlm@449	1609 (->> angles
rlm@449	1610 (flatten)
rlm@449	1611 (map (juxt #(Math/sin %) #(Math/cos %)))
rlm@449	1612 (flatten)
rlm@449	1613 (mapv #(Math/round (* % (Math/pow 10 (dec digits))))))))
rlm@449	1614
rlm@449	1615 (defn gen-phi-scan
rlm@450	1616 "Nearest-neighbors with binning. Only returns a result if
rlm@450	1617 the propriceptive data is within 10% of a previously recorded
rlm@450	1618 result in all dimensions."
rlm@450	1619 [phi-space]
rlm@449	1620 (let [bin-keys (map bin [3 2 1])
rlm@449	1621 bin-maps
rlm@449	1622 (map (fn [bin-key]
rlm@449	1623 (group-by
rlm@449	1624 (comp bin-key :proprioception phi-space)
rlm@449	1625 (range (count phi-space)))) bin-keys)
rlm@449	1626 lookups (map (fn [bin-key bin-map]
rlm@450	1627 (fn [proprio] (bin-map (bin-key proprio))))
rlm@450	1628 bin-keys bin-maps)]
rlm@449	1629 (fn lookup [proprio-data]
rlm@449	1630 (set (some #(% proprio-data) lookups)))))
rlm@450	1631 #+end_src
rlm@450	1632 #+end_listing
rlm@449	1633
rlm@451	1634 #+caption: =longest-thread= finds the longest path of consecutive
rlm@451	1635 #+caption: experiences to explain proprioceptive worm data.
rlm@451	1636 #+name: phi-space-history-scan
rlm@451	1637 #+ATTR_LaTeX: :width 10cm
rlm@451	1638 [[./images/aurellem-gray.png]]
rlm@451	1639
rlm@451	1640 =longest-thread= infers sensory data by stitching together pieces
rlm@451	1641 from previous experience. It prefers longer chains of previous
rlm@451	1642 experience to shorter ones. For example, during training the worm
rlm@451	1643 might rest on the ground for one second before it performs its
rlm@451	1644 excercises. If during recognition the worm rests on the ground for
rlm@451	1645 five seconds, =longest-thread= will accomodate this five second
rlm@451	1646 rest period by looping the one second rest chain five times.
rlm@451	1647
rlm@451	1648 =longest-thread= takes time proportinal to the average number of
rlm@451	1649 entries in a proprioceptive bin, because for each element in the
rlm@451	1650 starting bin it performes a series of set lookups in the preceeding
rlm@451	1651 bins. If the total history is limited, then this is only a constant
rlm@451	1652 multiple times the number of entries in the starting bin. This
rlm@451	1653 analysis also applies even if the action requires multiple longest
rlm@451	1654 chains -- it's still the average number of entries in a
rlm@451	1655 proprioceptive bin times the desired chain length. Because
rlm@451	1656 =longest-thread= is so efficient and simple, I can interpret
rlm@451	1657 worm-actions in real time.
rlm@449	1658
rlm@450	1659 #+caption: Program to calculate empathy by tracing though \Phi-space
rlm@450	1660 #+caption: and finding the longest (ie. most coherent) interpretation
rlm@450	1661 #+caption: of the data.
rlm@450	1662 #+name: longest-thread
rlm@452	1663 #+attr_latex: [htpb]
rlm@452	1664 #+begin_listing clojure
rlm@450	1665 #+begin_src clojure
rlm@449	1666 (defn longest-thread
rlm@449	1667 "Find the longest thread from phi-index-sets. The index sets should
rlm@449	1668 be ordered from most recent to least recent."
rlm@449	1669 [phi-index-sets]
rlm@449	1670 (loop [result '()
rlm@449	1671 [thread-bases & remaining :as phi-index-sets] phi-index-sets]
rlm@449	1672 (if (empty? phi-index-sets)
rlm@449	1673 (vec result)
rlm@449	1674 (let [threads
rlm@449	1675 (for [thread-base thread-bases]
rlm@449	1676 (loop [thread (list thread-base)
rlm@449	1677 remaining remaining]
rlm@449	1678 (let [next-index (dec (first thread))]
rlm@449	1679 (cond (empty? remaining) thread
rlm@449	1680 (contains? (first remaining) next-index)
rlm@449	1681 (recur
rlm@449	1682 (cons next-index thread) (rest remaining))
rlm@449	1683 :else thread))))
rlm@449	1684 longest-thread
rlm@449	1685 (reduce (fn [thread-a thread-b]
rlm@449	1686 (if (> (count thread-a) (count thread-b))
rlm@449	1687 thread-a thread-b))
rlm@449	1688 '(nil)
rlm@449	1689 threads)]
rlm@449	1690 (recur (concat longest-thread result)
rlm@449	1691 (drop (count longest-thread) phi-index-sets))))))
rlm@450	1692 #+end_src
rlm@450	1693 #+end_listing
rlm@450	1694
rlm@451	1695 There is one final piece, which is to replace missing sensory data
rlm@451	1696 with a best-guess estimate. While I could fill in missing data by
rlm@451	1697 using a gradient over the closest known sensory data points,
rlm@451	1698 averages can be misleading. It is certainly possible to create an
rlm@451	1699 impossible sensory state by averaging two possible sensory states.
rlm@451	1700 Therefore, I simply replicate the most recent sensory experience to
rlm@451	1701 fill in the gaps.
rlm@449	1702
rlm@449	1703 #+caption: Fill in blanks in sensory experience by replicating the most
rlm@449	1704 #+caption: recent experience.
rlm@449	1705 #+name: infer-nils
rlm@452	1706 #+attr_latex: [htpb]
rlm@452	1707 #+begin_listing clojure
rlm@449	1708 #+begin_src clojure
rlm@449	1709 (defn infer-nils
rlm@449	1710 "Replace nils with the next available non-nil element in the
rlm@449	1711 sequence, or barring that, 0."
rlm@449	1712 [s]
rlm@449	1713 (loop [i (dec (count s))
rlm@449	1714 v (transient s)]
rlm@449	1715 (if (zero? i) (persistent! v)
rlm@449	1716 (if-let [cur (v i)]
rlm@449	1717 (if (get v (dec i) 0)
rlm@449	1718 (recur (dec i) v)
rlm@449	1719 (recur (dec i) (assoc! v (dec i) cur)))
rlm@449	1720 (recur i (assoc! v i 0))))))
rlm@449	1721 #+end_src
rlm@449	1722 #+end_listing
rlm@435	1723
rlm@441	1724 ** Efficient action recognition with =EMPATH=
rlm@451	1725
rlm@451	1726 To use =EMPATH= with the worm, I first need to gather a set of
rlm@451	1727 experiences from the worm that includes the actions I want to
rlm@452	1728 recognize. The =generate-phi-space= program (listing
rlm@451	1729 \ref{generate-phi-space} runs the worm through a series of
rlm@451	1730 exercices and gatheres those experiences into a vector. The
rlm@451	1731 =do-all-the-things= program is a routine expressed in a simple
rlm@452	1732 muscle contraction script language for automated worm control. It
rlm@452	1733 causes the worm to rest, curl, and wiggle over about 700 frames
rlm@452	1734 (approx. 11 seconds).
rlm@425	1735
rlm@451	1736 #+caption: Program to gather the worm's experiences into a vector for
rlm@451	1737 #+caption: further processing. The =motor-control-program= line uses
rlm@451	1738 #+caption: a motor control script that causes the worm to execute a series
rlm@451	1739 #+caption: of ``exercices'' that include all the action predicates.
rlm@451	1740 #+name: generate-phi-space
rlm@452	1741 #+attr_latex: [htpb]
rlm@452	1742 #+begin_listing clojure
rlm@451	1743 #+begin_src clojure
rlm@451	1744 (def do-all-the-things
rlm@451	1745 (concat
rlm@451	1746 curl-script
rlm@451	1747 [[300 :d-ex 40]
rlm@451	1748 [320 :d-ex 0]]
rlm@451	1749 (shift-script 280 (take 16 wiggle-script))))
rlm@451	1750
rlm@451	1751 (defn generate-phi-space []
rlm@451	1752 (let [experiences (atom [])]
rlm@451	1753 (run-world
rlm@451	1754 (apply-map
rlm@451	1755 worm-world
rlm@451	1756 (merge
rlm@451	1757 (worm-world-defaults)
rlm@451	1758 {:end-frame 700
rlm@451	1759 :motor-control
rlm@451	1760 (motor-control-program worm-muscle-labels do-all-the-things)
rlm@451	1761 :experiences experiences})))
rlm@451	1762 @experiences))
rlm@451	1763 #+end_src
rlm@451	1764 #+end_listing
rlm@451	1765
rlm@451	1766 #+caption: Use longest thread and a phi-space generated from a short
rlm@451	1767 #+caption: exercise routine to interpret actions during free play.
rlm@451	1768 #+name: empathy-debug
rlm@452	1769 #+attr_latex: [htpb]
rlm@452	1770 #+begin_listing clojure
rlm@451	1771 #+begin_src clojure
rlm@451	1772 (defn init []
rlm@451	1773 (def phi-space (generate-phi-space))
rlm@451	1774 (def phi-scan (gen-phi-scan phi-space)))
rlm@451	1775
rlm@451	1776 (defn empathy-demonstration []
rlm@451	1777 (let [proprio (atom ())]
rlm@451	1778 (fn
rlm@451	1779 [experiences text]
rlm@451	1780 (let [phi-indices (phi-scan (:proprioception (peek experiences)))]
rlm@451	1781 (swap! proprio (partial cons phi-indices))
rlm@451	1782 (let [exp-thread (longest-thread (take 300 @proprio))
rlm@451	1783 empathy (mapv phi-space (infer-nils exp-thread))]
rlm@451	1784 (println-repl (vector:last-n exp-thread 22))
rlm@451	1785 (cond
rlm@451	1786 (grand-circle? empathy) (.setText text "Grand Circle")
rlm@451	1787 (curled? empathy) (.setText text "Curled")
rlm@451	1788 (wiggling? empathy) (.setText text "Wiggling")
rlm@451	1789 (resting? empathy) (.setText text "Resting")
rlm@451	1790 :else (.setText text "Unknown")))))))
rlm@451	1791
rlm@451	1792 (defn empathy-experiment [record]
rlm@451	1793 (.start (worm-world :experience-watch (debug-experience-phi)
rlm@451	1794 :record record :worm worm*)))
rlm@451	1795 #+end_src
rlm@451	1796 #+end_listing
rlm@451	1797
rlm@451	1798 The result of running =empathy-experiment= is that the system is
rlm@451	1799 generally able to interpret worm actions using the action-predicates
rlm@451	1800 on simulated sensory data just as well as with actual data. Figure
rlm@451	1801 \ref{empathy-debug-image} was generated using =empathy-experiment=:
rlm@451	1802
rlm@451	1803 #+caption: From only proprioceptive data, =EMPATH= was able to infer
rlm@451	1804 #+caption: the complete sensory experience and classify four poses
rlm@451	1805 #+caption: (The last panel shows a composite image of \emph{wriggling},
rlm@451	1806 #+caption: a dynamic pose.)
rlm@451	1807 #+name: empathy-debug-image
rlm@451	1808 #+ATTR_LaTeX: :width 10cm :placement [H]
rlm@451	1809 [[./images/empathy-1.png]]
rlm@451	1810
rlm@451	1811 One way to measure the performance of =EMPATH= is to compare the
rlm@451	1812 sutiability of the imagined sense experience to trigger the same
rlm@451	1813 action predicates as the real sensory experience.
rlm@451	1814
rlm@451	1815 #+caption: Determine how closely empathy approximates actual
rlm@451	1816 #+caption: sensory data.
rlm@451	1817 #+name: test-empathy-accuracy
rlm@452	1818 #+attr_latex: [htpb]
rlm@452	1819 #+begin_listing clojure
rlm@451	1820 #+begin_src clojure
rlm@451	1821 (def worm-action-label
rlm@451	1822 (juxt grand-circle? curled? wiggling?))
rlm@451	1823
rlm@451	1824 (defn compare-empathy-with-baseline [matches]
rlm@451	1825 (let [proprio (atom ())]
rlm@451	1826 (fn
rlm@451	1827 [experiences text]
rlm@451	1828 (let [phi-indices (phi-scan (:proprioception (peek experiences)))]
rlm@451	1829 (swap! proprio (partial cons phi-indices))
rlm@451	1830 (let [exp-thread (longest-thread (take 300 @proprio))
rlm@451	1831 empathy (mapv phi-space (infer-nils exp-thread))
rlm@451	1832 experience-matches-empathy
rlm@451	1833 (= (worm-action-label experiences)
rlm@451	1834 (worm-action-label empathy))]
rlm@451	1835 (println-repl experience-matches-empathy)
rlm@451	1836 (swap! matches #(conj % experience-matches-empathy)))))))
rlm@451	1837
rlm@451	1838 (defn accuracy [v]
rlm@451	1839 (float (/ (count (filter true? v)) (count v))))
rlm@451	1840
rlm@451	1841 (defn test-empathy-accuracy []
rlm@451	1842 (let [res (atom [])]
rlm@451	1843 (run-world
rlm@451	1844 (worm-world :experience-watch
rlm@451	1845 (compare-empathy-with-baseline res)
rlm@451	1846 :worm worm*))
rlm@451	1847 (accuracy @res)))
rlm@451	1848 #+end_src
rlm@451	1849 #+end_listing
rlm@451	1850
rlm@451	1851 Running =test-empathy-accuracy= using the very short exercise
rlm@451	1852 program defined in listing \ref{generate-phi-space}, and then doing
rlm@451	1853 a similar pattern of activity manually yeilds an accuracy of around
rlm@451	1854 73%. This is based on very limited worm experience. By training the
rlm@451	1855 worm for longer, the accuracy dramatically improves.
rlm@451	1856
rlm@451	1857 #+caption: Program to generate \Phi-space using manual training.
rlm@451	1858 #+name: manual-phi-space
rlm@452	1859 #+attr_latex: [htpb]
rlm@451	1860 #+begin_listing clojure
rlm@451	1861 #+begin_src clojure
rlm@451	1862 (defn init-interactive []
rlm@451	1863 (def phi-space
rlm@451	1864 (let [experiences (atom [])]
rlm@451	1865 (run-world
rlm@451	1866 (apply-map
rlm@451	1867 worm-world
rlm@451	1868 (merge
rlm@451	1869 (worm-world-defaults)
rlm@451	1870 {:experiences experiences})))
rlm@451	1871 @experiences))
rlm@451	1872 (def phi-scan (gen-phi-scan phi-space)))
rlm@451	1873 #+end_src
rlm@451	1874 #+end_listing
rlm@451	1875
rlm@451	1876 After about 1 minute of manual training, I was able to achieve 95%
rlm@451	1877 accuracy on manual testing of the worm using =init-interactive= and
rlm@452	1878 =test-empathy-accuracy=. The majority of errors are near the
rlm@452	1879 boundaries of transitioning from one type of action to another.
rlm@452	1880 During these transitions the exact label for the action is more open
rlm@452	1881 to interpretation, and dissaggrement between empathy and experience
rlm@452	1882 is more excusable.
rlm@450	1883
rlm@449	1884 ** Digression: bootstrapping touch using free exploration
rlm@449	1885
rlm@452	1886 In the previous section I showed how to compute actions in terms of
rlm@452	1887 body-centered predicates which relied averate touch activation of
rlm@452	1888 pre-defined regions of the worm's skin. What if, instead of recieving
rlm@452	1889 touch pre-grouped into the six faces of each worm segment, the true
rlm@452	1890 topology of the worm's skin was unknown? This is more similiar to how
rlm@452	1891 a nerve fiber bundle might be arranged. While two fibers that are
rlm@452	1892 close in a nerve bundle /might/ correspond to two touch sensors that
rlm@452	1893 are close together on the skin, the process of taking a complicated
rlm@452	1894 surface and forcing it into essentially a circle requires some cuts
rlm@452	1895 and rerragenments.
rlm@452	1896
rlm@452	1897 In this section I show how to automatically learn the skin-topology of
rlm@452	1898 a worm segment by free exploration. As the worm rolls around on the
rlm@452	1899 floor, large sections of its surface get activated. If the worm has
rlm@452	1900 stopped moving, then whatever region of skin that is touching the
rlm@452	1901 floor is probably an important region, and should be recorded.
rlm@452	1902
rlm@452	1903 #+caption: Program to detect whether the worm is in a resting state
rlm@452	1904 #+caption: with one face touching the floor.
rlm@452	1905 #+name: pure-touch
rlm@452	1906 #+begin_listing clojure
rlm@452	1907 #+begin_src clojure
rlm@452	1908 (def full-contact [(float 0.0) (float 0.1)])
rlm@452	1909
rlm@452	1910 (defn pure-touch?
rlm@452	1911 "This is worm specific code to determine if a large region of touch
rlm@452	1912 sensors is either all on or all off."
rlm@452	1913 [[coords touch :as touch-data]]
rlm@452	1914 (= (set (map first touch)) (set full-contact)))
rlm@452	1915 #+end_src
rlm@452	1916 #+end_listing
rlm@452	1917
rlm@452	1918 After collecting these important regions, there will many nearly
rlm@452	1919 similiar touch regions. While for some purposes the subtle
rlm@452	1920 differences between these regions will be important, for my
rlm@452	1921 purposes I colapse them into mostly non-overlapping sets using
rlm@452	1922 =remove-similiar= in listing \ref{remove-similiar}
rlm@452	1923
rlm@452	1924 #+caption: Program to take a lits of set of points and ``collapse them''
rlm@452	1925 #+caption: so that the remaining sets in the list are siginificantly
rlm@452	1926 #+caption: different from each other. Prefer smaller sets to larger ones.
rlm@452	1927 #+name: remove-similiar
rlm@452	1928 #+begin_listing clojure
rlm@452	1929 #+begin_src clojure
rlm@452	1930 (defn remove-similar
rlm@452	1931 [coll]
rlm@452	1932 (loop [result () coll (sort-by (comp - count) coll)]
rlm@452	1933 (if (empty? coll) result
rlm@452	1934 (let [[x & xs] coll
rlm@452	1935 c (count x)]
rlm@452	1936 (if (some
rlm@452	1937 (fn [other-set]
rlm@452	1938 (let [oc (count other-set)]
rlm@452	1939 (< (- (count (union other-set x)) c) (* oc 0.1))))
rlm@452	1940 xs)
rlm@452	1941 (recur result xs)
rlm@452	1942 (recur (cons x result) xs))))))
rlm@452	1943 #+end_src
rlm@452	1944 #+end_listing
rlm@452	1945
rlm@452	1946 Actually running this simulation is easy given =CORTEX='s facilities.
rlm@452	1947
rlm@452	1948 #+caption: Collect experiences while the worm moves around. Filter the touch
rlm@452	1949 #+caption: sensations by stable ones, collapse similiar ones together,
rlm@452	1950 #+caption: and report the regions learned.
rlm@452	1951 #+name: learn-touch
rlm@452	1952 #+begin_listing clojure
rlm@452	1953 #+begin_src clojure
rlm@452	1954 (defn learn-touch-regions []
rlm@452	1955 (let [experiences (atom [])
rlm@452	1956 world (apply-map
rlm@452	1957 worm-world
rlm@452	1958 (assoc (worm-segment-defaults)
rlm@452	1959 :experiences experiences))]
rlm@452	1960 (run-world world)
rlm@452	1961 (->>
rlm@452	1962 @experiences
rlm@452	1963 (drop 175)
rlm@452	1964 ;; access the single segment's touch data
rlm@452	1965 (map (comp first :touch))
rlm@452	1966 ;; only deal with "pure" touch data to determine surfaces
rlm@452	1967 (filter pure-touch?)
rlm@452	1968 ;; associate coordinates with touch values
rlm@452	1969 (map (partial apply zipmap))
rlm@452	1970 ;; select those regions where contact is being made
rlm@452	1971 (map (partial group-by second))
rlm@452	1972 (map #(get % full-contact))
rlm@452	1973 (map (partial map first))
rlm@452	1974 ;; remove redundant/subset regions
rlm@452	1975 (map set)
rlm@452	1976 remove-similar)))
rlm@452	1977
rlm@452	1978 (defn learn-and-view-touch-regions []
rlm@452	1979 (map view-touch-region
rlm@452	1980 (learn-touch-regions)))
rlm@452	1981 #+end_src
rlm@452	1982 #+end_listing
rlm@452	1983
rlm@452	1984 The only thing remining to define is the particular motion the worm
rlm@452	1985 must take. I accomplish this with a simple motor control program.
rlm@452	1986
rlm@452	1987 #+caption: Motor control program for making the worm roll on the ground.
rlm@452	1988 #+caption: This could also be replaced with random motion.
rlm@452	1989 #+name: worm-roll
rlm@452	1990 #+begin_listing clojure
rlm@452	1991 #+begin_src clojure
rlm@452	1992 (defn touch-kinesthetics []
rlm@452	1993 [[170 :lift-1 40]
rlm@452	1994 [190 :lift-1 19]
rlm@452	1995 [206 :lift-1 0]
rlm@452	1996
rlm@452	1997 [400 :lift-2 40]
rlm@452	1998 [410 :lift-2 0]
rlm@452	1999
rlm@452	2000 [570 :lift-2 40]
rlm@452	2001 [590 :lift-2 21]
rlm@452	2002 [606 :lift-2 0]
rlm@452	2003
rlm@452	2004 [800 :lift-1 30]
rlm@452	2005 [809 :lift-1 0]
rlm@452	2006
rlm@452	2007 [900 :roll-2 40]
rlm@452	2008 [905 :roll-2 20]
rlm@452	2009 [910 :roll-2 0]
rlm@452	2010
rlm@452	2011 [1000 :roll-2 40]
rlm@452	2012 [1005 :roll-2 20]
rlm@452	2013 [1010 :roll-2 0]
rlm@452	2014
rlm@452	2015 [1100 :roll-2 40]
rlm@452	2016 [1105 :roll-2 20]
rlm@452	2017 [1110 :roll-2 0]
rlm@452	2018 ])
rlm@452	2019 #+end_src
rlm@452	2020 #+end_listing
rlm@452	2021
rlm@452	2022
rlm@452	2023 #+caption: The small worm rolls around on the floor, driven
rlm@452	2024 #+caption: by the motor control program in listing \ref{worm-roll}.
rlm@452	2025 #+name: worm-roll
rlm@452	2026 #+ATTR_LaTeX: :width 12cm
rlm@452	2027 [[./images/worm-roll.png]]
rlm@452	2028
rlm@452	2029
rlm@452	2030 #+caption: After completing its adventures, the worm now knows
rlm@452	2031 #+caption: how its touch sensors are arranged along its skin. These
rlm@452	2032 #+caption: are the regions that were deemed important by
rlm@452	2033 #+caption: =learn-touch-regions=. Note that the worm has discovered
rlm@452	2034 #+caption: that it has six sides.
rlm@452	2035 #+name: worm-touch-map
rlm@452	2036 #+ATTR_LaTeX: :width 12cm
rlm@452	2037 [[./images/touch-learn.png]]
rlm@452	2038
rlm@452	2039 While simple, =learn-touch-regions= exploits regularities in both
rlm@452	2040 the worm's physiology and the worm's environment to correctly
rlm@452	2041 deduce that the worm has six sides. Note that =learn-touch-regions=
rlm@452	2042 would work just as well even if the worm's touch sense data were
rlm@452	2043 completely scrambled. The cross shape is just for convienence. This
rlm@452	2044 example justifies the use of pre-defined touch regions in =EMPATH=.
rlm@452	2045
rlm@465	2046 * COMMENT Contributions
rlm@454	2047
rlm@461	2048 In this thesis you have seen the =CORTEX= system, a complete
rlm@461	2049 environment for creating simulated creatures. You have seen how to
rlm@461	2050 implement five senses including touch, proprioception, hearing,
rlm@461	2051 vision, and muscle tension. You have seen how to create new creatues
rlm@461	2052 using blender, a 3D modeling tool. I hope that =CORTEX= will be
rlm@461	2053 useful in further research projects. To this end I have included the
rlm@461	2054 full source to =CORTEX= along with a large suite of tests and
rlm@461	2055 examples. I have also created a user guide for =CORTEX= which is
rlm@461	2056 inculded in an appendix to this thesis.
rlm@447	2057
rlm@461	2058 You have also seen how I used =CORTEX= as a platform to attach the
rlm@461	2059 /action recognition/ problem, which is the problem of recognizing
rlm@461	2060 actions in video. You saw a simple system called =EMPATH= which
rlm@461	2061 ientifies actions by first describing actions in a body-centerd,
rlm@461	2062 rich sense language, then infering a full range of sensory
rlm@461	2063 experience from limited data using previous experience gained from
rlm@461	2064 free play.
rlm@447	2065
rlm@461	2066 As a minor digression, you also saw how I used =CORTEX= to enable a
rlm@461	2067 tiny worm to discover the topology of its skin simply by rolling on
rlm@461	2068 the ground.
rlm@461	2069
rlm@461	2070 In conclusion, the main contributions of this thesis are:
rlm@461	2071
rlm@461	2072 - =CORTEX=, a system for creating simulated creatures with rich
rlm@461	2073 senses.
rlm@461	2074 - =EMPATH=, a program for recognizing actions by imagining sensory
rlm@461	2075 experience.
rlm@447	2076
rlm@447	2077 # An anatomical joke:
rlm@447	2078 # - Training
rlm@447	2079 # - Skeletal imitation
rlm@447	2080 # - Sensory fleshing-out
rlm@447	2081 # - Classification

Mercurial > cortex

annotate thesis/cortex.org @ 471:f14fa9e5b67f