cortex: thesis/cortex.org annotate

annotate thesis/cortex.org @ 469:ae10f35022ba

completed first draft of body section.

author	Robert McIntyre <rlm@mit.edu>
date	Fri, 28 Mar 2014 16:34:35 -0400
parents	258078f78b33
children	3401053124b0

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@465	9 #+caption:
rlm@465	10 #+caption:
rlm@465	11 #+caption:
rlm@465	12 #+caption:
rlm@465	13 #+name: name
rlm@465	14 #+begin_listing clojure
rlm@465	15 #+begin_src clojure
rlm@465	16 #+end_src
rlm@465	17 #+end_listing
rlm@465	18
rlm@465	19 #+caption:
rlm@465	20 #+caption:
rlm@465	21 #+caption:
rlm@465	22 #+name: name
rlm@465	23 #+ATTR_LaTeX: :width 10cm
rlm@466	24 [[./images/aurellem-gray.png]]
rlm@465	25
rlm@465	26 * COMMENT Empathy and Embodiment as problem solving strategies
rlm@437	27
rlm@437	28 By the end of this thesis, you will have seen a novel approach to
rlm@437	29 interpreting video using embodiment and empathy. You will have also
rlm@437	30 seen one way to efficiently implement empathy for embodied
rlm@447	31 creatures. Finally, you will become familiar with =CORTEX=, a system
rlm@447	32 for designing and simulating creatures with rich senses, which you
rlm@447	33 may choose to use in your own research.
rlm@437	34
rlm@441	35 This is the core vision of my thesis: That one of the important ways
rlm@441	36 in which we understand others is by imagining ourselves in their
rlm@441	37 position and emphatically feeling experiences relative to our own
rlm@441	38 bodies. By understanding events in terms of our own previous
rlm@441	39 corporeal experience, we greatly constrain the possibilities of what
rlm@441	40 would otherwise be an unwieldy exponential search. This extra
rlm@441	41 constraint can be the difference between easily understanding what
rlm@441	42 is happening in a video and being completely lost in a sea of
rlm@441	43 incomprehensible color and movement.
rlm@435	44
rlm@436	45 ** Recognizing actions in video is extremely difficult
rlm@437	46
rlm@447	47 Consider for example the problem of determining what is happening
rlm@447	48 in a video of which this is one frame:
rlm@437	49
rlm@441	50 #+caption: A cat drinking some water. Identifying this action is
rlm@441	51 #+caption: beyond the state of the art for computers.
rlm@441	52 #+ATTR_LaTeX: :width 7cm
rlm@441	53 [[./images/cat-drinking.jpg]]
rlm@441	54
rlm@441	55 It is currently impossible for any computer program to reliably
rlm@447	56 label such a video as ``drinking''. And rightly so -- it is a very
rlm@441	57 hard problem! What features can you describe in terms of low level
rlm@441	58 functions of pixels that can even begin to describe at a high level
rlm@441	59 what is happening here?
rlm@437	60
rlm@447	61 Or suppose that you are building a program that recognizes chairs.
rlm@448	62 How could you ``see'' the chair in figure \ref{hidden-chair}?
rlm@441	63
rlm@441	64 #+caption: The chair in this image is quite obvious to humans, but I
rlm@448	65 #+caption: doubt that any modern computer vision program can find it.
rlm@441	66 #+name: hidden-chair
rlm@441	67 #+ATTR_LaTeX: :width 10cm
rlm@441	68 [[./images/fat-person-sitting-at-desk.jpg]]
rlm@441	69
rlm@441	70 Finally, how is it that you can easily tell the difference between
rlm@441	71 how the girls /muscles/ are working in figure \ref{girl}?
rlm@441	72
rlm@441	73 #+caption: The mysterious ``common sense'' appears here as you are able
rlm@441	74 #+caption: to discern the difference in how the girl's arm muscles
rlm@441	75 #+caption: are activated between the two images.
rlm@441	76 #+name: girl
rlm@448	77 #+ATTR_LaTeX: :width 7cm
rlm@441	78 [[./images/wall-push.png]]
rlm@437	79
rlm@441	80 Each of these examples tells us something about what might be going
rlm@441	81 on in our minds as we easily solve these recognition problems.
rlm@441	82
rlm@441	83 The hidden chairs show us that we are strongly triggered by cues
rlm@447	84 relating to the position of human bodies, and that we can determine
rlm@447	85 the overall physical configuration of a human body even if much of
rlm@447	86 that body is occluded.
rlm@437	87
rlm@441	88 The picture of the girl pushing against the wall tells us that we
rlm@441	89 have common sense knowledge about the kinetics of our own bodies.
rlm@441	90 We know well how our muscles would have to work to maintain us in
rlm@441	91 most positions, and we can easily project this self-knowledge to
rlm@441	92 imagined positions triggered by images of the human body.
rlm@441	93
rlm@441	94 ** =EMPATH= neatly solves recognition problems
rlm@441	95
rlm@441	96 I propose a system that can express the types of recognition
rlm@441	97 problems above in a form amenable to computation. It is split into
rlm@441	98 four parts:
rlm@441	99
rlm@448	100 - Free/Guided Play :: The creature moves around and experiences the
rlm@448	101 world through its unique perspective. Many otherwise
rlm@448	102 complicated actions are easily described in the language of a
rlm@448	103 full suite of body-centered, rich senses. For example,
rlm@448	104 drinking is the feeling of water sliding down your throat, and
rlm@448	105 cooling your insides. It's often accompanied by bringing your
rlm@448	106 hand close to your face, or bringing your face close to water.
rlm@448	107 Sitting down is the feeling of bending your knees, activating
rlm@448	108 your quadriceps, then feeling a surface with your bottom and
rlm@448	109 relaxing your legs. These body-centered action descriptions
rlm@448	110 can be either learned or hard coded.
rlm@448	111 - Posture Imitation :: When trying to interpret a video or image,
rlm@448	112 the creature takes a model of itself and aligns it with
rlm@448	113 whatever it sees. This alignment can even cross species, as
rlm@448	114 when humans try to align themselves with things like ponies,
rlm@448	115 dogs, or other humans with a different body type.
rlm@448	116 - Empathy :: The alignment triggers associations with
rlm@448	117 sensory data from prior experiences. For example, the
rlm@448	118 alignment itself easily maps to proprioceptive data. Any
rlm@448	119 sounds or obvious skin contact in the video can to a lesser
rlm@448	120 extent trigger previous experience. Segments of previous
rlm@448	121 experiences are stitched together to form a coherent and
rlm@448	122 complete sensory portrait of the scene.
rlm@448	123 - Recognition :: With the scene described in terms of first
rlm@448	124 person sensory events, the creature can now run its
rlm@447	125 action-identification programs on this synthesized sensory
rlm@447	126 data, just as it would if it were actually experiencing the
rlm@447	127 scene first-hand. If previous experience has been accurately
rlm@447	128 retrieved, and if it is analogous enough to the scene, then
rlm@447	129 the creature will correctly identify the action in the scene.
rlm@447	130
rlm@441	131 For example, I think humans are able to label the cat video as
rlm@447	132 ``drinking'' because they imagine /themselves/ as the cat, and
rlm@441	133 imagine putting their face up against a stream of water and
rlm@441	134 sticking out their tongue. In that imagined world, they can feel
rlm@441	135 the cool water hitting their tongue, and feel the water entering
rlm@447	136 their body, and are able to recognize that /feeling/ as drinking.
rlm@447	137 So, the label of the action is not really in the pixels of the
rlm@447	138 image, but is found clearly in a simulation inspired by those
rlm@447	139 pixels. An imaginative system, having been trained on drinking and
rlm@447	140 non-drinking examples and learning that the most important
rlm@447	141 component of drinking is the feeling of water sliding down one's
rlm@447	142 throat, would analyze a video of a cat drinking in the following
rlm@447	143 manner:
rlm@441	144
rlm@447	145 1. Create a physical model of the video by putting a ``fuzzy''
rlm@447	146 model of its own body in place of the cat. Possibly also create
rlm@447	147 a simulation of the stream of water.
rlm@441	148
rlm@441	149 2. Play out this simulated scene and generate imagined sensory
rlm@441	150 experience. This will include relevant muscle contractions, a
rlm@441	151 close up view of the stream from the cat's perspective, and most
rlm@441	152 importantly, the imagined feeling of water entering the
rlm@443	153 mouth. The imagined sensory experience can come from a
rlm@441	154 simulation of the event, but can also be pattern-matched from
rlm@441	155 previous, similar embodied experience.
rlm@441	156
rlm@441	157 3. The action is now easily identified as drinking by the sense of
rlm@441	158 taste alone. The other senses (such as the tongue moving in and
rlm@441	159 out) help to give plausibility to the simulated action. Note that
rlm@441	160 the sense of vision, while critical in creating the simulation,
rlm@441	161 is not critical for identifying the action from the simulation.
rlm@441	162
rlm@441	163 For the chair examples, the process is even easier:
rlm@441	164
rlm@441	165 1. Align a model of your body to the person in the image.
rlm@441	166
rlm@441	167 2. Generate proprioceptive sensory data from this alignment.
rlm@437	168
rlm@441	169 3. Use the imagined proprioceptive data as a key to lookup related
rlm@441	170 sensory experience associated with that particular proproceptive
rlm@441	171 feeling.
rlm@437	172
rlm@443	173 4. Retrieve the feeling of your bottom resting on a surface, your
rlm@443	174 knees bent, and your leg muscles relaxed.
rlm@437	175
rlm@441	176 5. This sensory information is consistent with the =sitting?=
rlm@441	177 sensory predicate, so you (and the entity in the image) must be
rlm@441	178 sitting.
rlm@440	179
rlm@441	180 6. There must be a chair-like object since you are sitting.
rlm@440	181
rlm@441	182 Empathy offers yet another alternative to the age-old AI
rlm@441	183 representation question: ``What is a chair?'' --- A chair is the
rlm@441	184 feeling of sitting.
rlm@441	185
rlm@441	186 My program, =EMPATH= uses this empathic problem solving technique
rlm@441	187 to interpret the actions of a simple, worm-like creature.
rlm@437	188
rlm@441	189 #+caption: The worm performs many actions during free play such as
rlm@441	190 #+caption: curling, wiggling, and resting.
rlm@441	191 #+name: worm-intro
rlm@446	192 #+ATTR_LaTeX: :width 15cm
rlm@445	193 [[./images/worm-intro-white.png]]
rlm@437	194
rlm@462	195 #+caption: =EMPATH= recognized and classified each of these
rlm@462	196 #+caption: poses by inferring the complete sensory experience
rlm@462	197 #+caption: from proprioceptive data.
rlm@441	198 #+name: worm-recognition-intro
rlm@446	199 #+ATTR_LaTeX: :width 15cm
rlm@445	200 [[./images/worm-poses.png]]
rlm@441	201
rlm@441	202 One powerful advantage of empathic problem solving is that it
rlm@441	203 factors the action recognition problem into two easier problems. To
rlm@441	204 use empathy, you need an /aligner/, which takes the video and a
rlm@441	205 model of your body, and aligns the model with the video. Then, you
rlm@441	206 need a /recognizer/, which uses the aligned model to interpret the
rlm@441	207 action. The power in this method lies in the fact that you describe
rlm@448	208 all actions form a body-centered viewpoint. You are less tied to
rlm@447	209 the particulars of any visual representation of the actions. If you
rlm@441	210 teach the system what ``running'' is, and you have a good enough
rlm@441	211 aligner, the system will from then on be able to recognize running
rlm@441	212 from any point of view, even strange points of view like above or
rlm@441	213 underneath the runner. This is in contrast to action recognition
rlm@448	214 schemes that try to identify actions using a non-embodied approach.
rlm@448	215 If these systems learn about running as viewed from the side, they
rlm@448	216 will not automatically be able to recognize running from any other
rlm@448	217 viewpoint.
rlm@441	218
rlm@441	219 Another powerful advantage is that using the language of multiple
rlm@441	220 body-centered rich senses to describe body-centerd actions offers a
rlm@441	221 massive boost in descriptive capability. Consider how difficult it
rlm@441	222 would be to compose a set of HOG filters to describe the action of
rlm@447	223 a simple worm-creature ``curling'' so that its head touches its
rlm@447	224 tail, and then behold the simplicity of describing thus action in a
rlm@441	225 language designed for the task (listing \ref{grand-circle-intro}):
rlm@441	226
rlm@446	227 #+caption: Body-centerd actions are best expressed in a body-centered
rlm@446	228 #+caption: language. This code detects when the worm has curled into a
rlm@446	229 #+caption: full circle. Imagine how you would replicate this functionality
rlm@446	230 #+caption: using low-level pixel features such as HOG filters!
rlm@446	231 #+name: grand-circle-intro
rlm@452	232 #+attr_latex: [htpb]
rlm@452	233 #+begin_listing clojure
rlm@446	234 #+begin_src clojure
rlm@446	235 (defn grand-circle?
rlm@446	236 "Does the worm form a majestic circle (one end touching the other)?"
rlm@446	237 [experiences]
rlm@446	238 (and (curled? experiences)
rlm@446	239 (let [worm-touch (:touch (peek experiences))
rlm@446	240 tail-touch (worm-touch 0)
rlm@446	241 head-touch (worm-touch 4)]
rlm@462	242 (and (< 0.2 (contact worm-segment-bottom-tip tail-touch))
rlm@462	243 (< 0.2 (contact worm-segment-top-tip head-touch))))))
rlm@446	244 #+end_src
rlm@446	245 #+end_listing
rlm@446	246
rlm@435	247
rlm@449	248 ** =CORTEX= is a toolkit for building sensate creatures
rlm@435	249
rlm@448	250 I built =CORTEX= to be a general AI research platform for doing
rlm@448	251 experiments involving multiple rich senses and a wide variety and
rlm@448	252 number of creatures. I intend it to be useful as a library for many
rlm@462	253 more projects than just this thesis. =CORTEX= was necessary to meet
rlm@462	254 a need among AI researchers at CSAIL and beyond, which is that
rlm@462	255 people often will invent neat ideas that are best expressed in the
rlm@448	256 language of creatures and senses, but in order to explore those
rlm@448	257 ideas they must first build a platform in which they can create
rlm@448	258 simulated creatures with rich senses! There are many ideas that
rlm@448	259 would be simple to execute (such as =EMPATH=), but attached to them
rlm@448	260 is the multi-month effort to make a good creature simulator. Often,
rlm@448	261 that initial investment of time proves to be too much, and the
rlm@448	262 project must make do with a lesser environment.
rlm@435	263
rlm@448	264 =CORTEX= is well suited as an environment for embodied AI research
rlm@448	265 for three reasons:
rlm@448	266
rlm@448	267 - You can create new creatures using Blender, a popular 3D modeling
rlm@448	268 program. Each sense can be specified using special blender nodes
rlm@448	269 with biologically inspired paramaters. You need not write any
rlm@448	270 code to create a creature, and can use a wide library of
rlm@448	271 pre-existing blender models as a base for your own creatures.
rlm@448	272
rlm@448	273 - =CORTEX= implements a wide variety of senses, including touch,
rlm@448	274 proprioception, vision, hearing, and muscle tension. Complicated
rlm@448	275 senses like touch, and vision involve multiple sensory elements
rlm@448	276 embedded in a 2D surface. You have complete control over the
rlm@448	277 distribution of these sensor elements through the use of simple
rlm@448	278 png image files. In particular, =CORTEX= implements more
rlm@448	279 comprehensive hearing than any other creature simulation system
rlm@448	280 available.
rlm@448	281
rlm@448	282 - =CORTEX= supports any number of creatures and any number of
rlm@448	283 senses. Time in =CORTEX= dialates so that the simulated creatures
rlm@448	284 always precieve a perfectly smooth flow of time, regardless of
rlm@448	285 the actual computational load.
rlm@448	286
rlm@448	287 =CORTEX= is built on top of =jMonkeyEngine3=, which is a video game
rlm@448	288 engine designed to create cross-platform 3D desktop games. =CORTEX=
rlm@448	289 is mainly written in clojure, a dialect of =LISP= that runs on the
rlm@448	290 java virtual machine (JVM). The API for creating and simulating
rlm@449	291 creatures and senses is entirely expressed in clojure, though many
rlm@449	292 senses are implemented at the layer of jMonkeyEngine or below. For
rlm@449	293 example, for the sense of hearing I use a layer of clojure code on
rlm@449	294 top of a layer of java JNI bindings that drive a layer of =C++=
rlm@449	295 code which implements a modified version of =OpenAL= to support
rlm@449	296 multiple listeners. =CORTEX= is the only simulation environment
rlm@449	297 that I know of that can support multiple entities that can each
rlm@449	298 hear the world from their own perspective. Other senses also
rlm@449	299 require a small layer of Java code. =CORTEX= also uses =bullet=, a
rlm@449	300 physics simulator written in =C=.
rlm@448	301
rlm@448	302 #+caption: Here is the worm from above modeled in Blender, a free
rlm@448	303 #+caption: 3D-modeling program. Senses and joints are described
rlm@448	304 #+caption: using special nodes in Blender.
rlm@448	305 #+name: worm-recognition-intro
rlm@448	306 #+ATTR_LaTeX: :width 12cm
rlm@448	307 [[./images/blender-worm.png]]
rlm@448	308
rlm@449	309 Here are some thing I anticipate that =CORTEX= might be used for:
rlm@449	310
rlm@449	311 - exploring new ideas about sensory integration
rlm@449	312 - distributed communication among swarm creatures
rlm@449	313 - self-learning using free exploration,
rlm@449	314 - evolutionary algorithms involving creature construction
rlm@449	315 - exploration of exoitic senses and effectors that are not possible
rlm@449	316 in the real world (such as telekenisis or a semantic sense)
rlm@449	317 - imagination using subworlds
rlm@449	318
rlm@451	319 During one test with =CORTEX=, I created 3,000 creatures each with
rlm@448	320 their own independent senses and ran them all at only 1/80 real
rlm@448	321 time. In another test, I created a detailed model of my own hand,
rlm@448	322 equipped with a realistic distribution of touch (more sensitive at
rlm@448	323 the fingertips), as well as eyes and ears, and it ran at around 1/4
rlm@451	324 real time.
rlm@448	325
rlm@451	326 #+BEGIN_LaTeX
rlm@449	327 \begin{sidewaysfigure}
rlm@449	328 \includegraphics[width=9.5in]{images/full-hand.png}
rlm@451	329 \caption{
rlm@451	330 I modeled my own right hand in Blender and rigged it with all the
rlm@451	331 senses that {\tt CORTEX} supports. My simulated hand has a
rlm@451	332 biologically inspired distribution of touch sensors. The senses are
rlm@451	333 displayed on the right, and the simulation is displayed on the
rlm@451	334 left. Notice that my hand is curling its fingers, that it can see
rlm@451	335 its own finger from the eye in its palm, and that it can feel its
rlm@451	336 own thumb touching its palm.}
rlm@449	337 \end{sidewaysfigure}
rlm@451	338 #+END_LaTeX
rlm@448	339
rlm@437	340 ** Contributions
rlm@435	341
rlm@451	342 - I built =CORTEX=, a comprehensive platform for embodied AI
rlm@451	343 experiments. =CORTEX= supports many features lacking in other
rlm@451	344 systems, such proper simulation of hearing. It is easy to create
rlm@451	345 new =CORTEX= creatures using Blender, a free 3D modeling program.
rlm@449	346
rlm@451	347 - I built =EMPATH=, which uses =CORTEX= to identify the actions of
rlm@451	348 a worm-like creature using a computational model of empathy.
rlm@449	349
rlm@436	350 * Building =CORTEX=
rlm@435	351
rlm@462	352 I intend for =CORTEX= to be used as a general purpose library for
rlm@462	353 building creatures and outfitting them with senses, so that it will
rlm@462	354 be useful for other researchers who want to test out ideas of their
rlm@462	355 own. To this end, wherver I have had to make archetictural choices
rlm@462	356 about =CORTEX=, I have chosen to give as much freedom to the user as
rlm@462	357 possible, so that =CORTEX= may be used for things I have not
rlm@462	358 forseen.
rlm@462	359
rlm@465	360 ** COMMENT Simulation or Reality?
rlm@462	361
rlm@462	362 The most important archetictural decision of all is the choice to
rlm@462	363 use a computer-simulated environemnt in the first place! The world
rlm@462	364 is a vast and rich place, and for now simulations are a very poor
rlm@462	365 reflection of its complexity. It may be that there is a significant
rlm@462	366 qualatative difference between dealing with senses in the real
rlm@468	367 world and dealing with pale facilimilies of them in a simulation.
rlm@468	368 What are the advantages and disadvantages of a simulation vs.
rlm@468	369 reality?
rlm@462	370
rlm@462	371 *** Simulation
rlm@462	372
rlm@462	373 The advantages of virtual reality are that when everything is a
rlm@462	374 simulation, experiments in that simulation are absolutely
rlm@462	375 reproducible. It's also easier to change the character and world
rlm@462	376 to explore new situations and different sensory combinations.
rlm@462	377
rlm@462	378 If the world is to be simulated on a computer, then not only do
rlm@462	379 you have to worry about whether the character's senses are rich
rlm@462	380 enough to learn from the world, but whether the world itself is
rlm@462	381 rendered with enough detail and realism to give enough working
rlm@462	382 material to the character's senses. To name just a few
rlm@462	383 difficulties facing modern physics simulators: destructibility of
rlm@462	384 the environment, simulation of water/other fluids, large areas,
rlm@462	385 nonrigid bodies, lots of objects, smoke. I don't know of any
rlm@462	386 computer simulation that would allow a character to take a rock
rlm@462	387 and grind it into fine dust, then use that dust to make a clay
rlm@462	388 sculpture, at least not without spending years calculating the
rlm@462	389 interactions of every single small grain of dust. Maybe a
rlm@462	390 simulated world with today's limitations doesn't provide enough
rlm@462	391 richness for real intelligence to evolve.
rlm@462	392
rlm@462	393 *** Reality
rlm@462	394
rlm@462	395 The other approach for playing with senses is to hook your
rlm@462	396 software up to real cameras, microphones, robots, etc., and let it
rlm@462	397 loose in the real world. This has the advantage of eliminating
rlm@462	398 concerns about simulating the world at the expense of increasing
rlm@462	399 the complexity of implementing the senses. Instead of just
rlm@462	400 grabbing the current rendered frame for processing, you have to
rlm@462	401 use an actual camera with real lenses and interact with photons to
rlm@462	402 get an image. It is much harder to change the character, which is
rlm@462	403 now partly a physical robot of some sort, since doing so involves
rlm@462	404 changing things around in the real world instead of modifying
rlm@462	405 lines of code. While the real world is very rich and definitely
rlm@462	406 provides enough stimulation for intelligence to develop as
rlm@462	407 evidenced by our own existence, it is also uncontrollable in the
rlm@462	408 sense that a particular situation cannot be recreated perfectly or
rlm@462	409 saved for later use. It is harder to conduct science because it is
rlm@462	410 harder to repeat an experiment. The worst thing about using the
rlm@462	411 real world instead of a simulation is the matter of time. Instead
rlm@462	412 of simulated time you get the constant and unstoppable flow of
rlm@462	413 real time. This severely limits the sorts of software you can use
rlm@462	414 to program the AI because all sense inputs must be handled in real
rlm@462	415 time. Complicated ideas may have to be implemented in hardware or
rlm@462	416 may simply be impossible given the current speed of our
rlm@462	417 processors. Contrast this with a simulation, in which the flow of
rlm@462	418 time in the simulated world can be slowed down to accommodate the
rlm@462	419 limitations of the character's programming. In terms of cost,
rlm@462	420 doing everything in software is far cheaper than building custom
rlm@462	421 real-time hardware. All you need is a laptop and some patience.
rlm@435	422
rlm@465	423 ** COMMENT Because of Time, simulation is perferable to reality
rlm@435	424
rlm@462	425 I envision =CORTEX= being used to support rapid prototyping and
rlm@462	426 iteration of ideas. Even if I could put together a well constructed
rlm@462	427 kit for creating robots, it would still not be enough because of
rlm@462	428 the scourge of real-time processing. Anyone who wants to test their
rlm@462	429 ideas in the real world must always worry about getting their
rlm@465	430 algorithms to run fast enough to process information in real time.
rlm@465	431 The need for real time processing only increases if multiple senses
rlm@465	432 are involved. In the extreme case, even simple algorithms will have
rlm@465	433 to be accelerated by ASIC chips or FPGAs, turning what would
rlm@465	434 otherwise be a few lines of code and a 10x speed penality into a
rlm@465	435 multi-month ordeal. For this reason, =CORTEX= supports
rlm@462	436 /time-dialiation/, which scales back the framerate of the
rlm@465	437 simulation in proportion to the amount of processing each frame.
rlm@465	438 From the perspective of the creatures inside the simulation, time
rlm@465	439 always appears to flow at a constant rate, regardless of how
rlm@462	440 complicated the envorimnent becomes or how many creatures are in
rlm@462	441 the simulation. The cost is that =CORTEX= can sometimes run slower
rlm@462	442 than real time. This can also be an advantage, however ---
rlm@462	443 simulations of very simple creatures in =CORTEX= generally run at
rlm@462	444 40x on my machine!
rlm@462	445
rlm@469	446 ** COMMENT What is a sense?
rlm@468	447
rlm@468	448 If =CORTEX= is to support a wide variety of senses, it would help
rlm@468	449 to have a better understanding of what a ``sense'' actually is!
rlm@468	450 While vision, touch, and hearing all seem like they are quite
rlm@468	451 different things, I was supprised to learn during the course of
rlm@468	452 this thesis that they (and all physical senses) can be expressed as
rlm@468	453 exactly the same mathematical object due to a dimensional argument!
rlm@468	454
rlm@468	455 Human beings are three-dimensional objects, and the nerves that
rlm@468	456 transmit data from our various sense organs to our brain are
rlm@468	457 essentially one-dimensional. This leaves up to two dimensions in
rlm@468	458 which our sensory information may flow. For example, imagine your
rlm@468	459 skin: it is a two-dimensional surface around a three-dimensional
rlm@468	460 object (your body). It has discrete touch sensors embedded at
rlm@468	461 various points, and the density of these sensors corresponds to the
rlm@468	462 sensitivity of that region of skin. Each touch sensor connects to a
rlm@468	463 nerve, all of which eventually are bundled together as they travel
rlm@468	464 up the spinal cord to the brain. Intersect the spinal nerves with a
rlm@468	465 guillotining plane and you will see all of the sensory data of the
rlm@468	466 skin revealed in a roughly circular two-dimensional image which is
rlm@468	467 the cross section of the spinal cord. Points on this image that are
rlm@468	468 close together in this circle represent touch sensors that are
rlm@468	469 /probably/ close together on the skin, although there is of course
rlm@468	470 some cutting and rearrangement that has to be done to transfer the
rlm@468	471 complicated surface of the skin onto a two dimensional image.
rlm@468	472
rlm@468	473 Most human senses consist of many discrete sensors of various
rlm@468	474 properties distributed along a surface at various densities. For
rlm@468	475 skin, it is Pacinian corpuscles, Meissner's corpuscles, Merkel's
rlm@468	476 disks, and Ruffini's endings, which detect pressure and vibration
rlm@468	477 of various intensities. For ears, it is the stereocilia distributed
rlm@468	478 along the basilar membrane inside the cochlea; each one is
rlm@468	479 sensitive to a slightly different frequency of sound. For eyes, it
rlm@468	480 is rods and cones distributed along the surface of the retina. In
rlm@468	481 each case, we can describe the sense with a surface and a
rlm@468	482 distribution of sensors along that surface.
rlm@468	483
rlm@468	484 The neat idea is that every human sense can be effectively
rlm@468	485 described in terms of a surface containing embedded sensors. If the
rlm@468	486 sense had any more dimensions, then there wouldn't be enough room
rlm@468	487 in the spinal chord to transmit the information!
rlm@468	488
rlm@468	489 Therefore, =CORTEX= must support the ability to create objects and
rlm@468	490 then be able to ``paint'' points along their surfaces to describe
rlm@468	491 each sense.
rlm@468	492
rlm@468	493 Fortunately this idea is already a well known computer graphics
rlm@468	494 technique called called /UV-mapping/. The three-dimensional surface
rlm@468	495 of a model is cut and smooshed until it fits on a two-dimensional
rlm@468	496 image. You paint whatever you want on that image, and when the
rlm@468	497 three-dimensional shape is rendered in a game the smooshing and
rlm@468	498 cutting is reversed and the image appears on the three-dimensional
rlm@468	499 object.
rlm@468	500
rlm@468	501 To make a sense, interpret the UV-image as describing the
rlm@468	502 distribution of that senses sensors. To get different types of
rlm@468	503 sensors, you can either use a different color for each type of
rlm@468	504 sensor, or use multiple UV-maps, each labeled with that sensor
rlm@468	505 type. I generally use a white pixel to mean the presence of a
rlm@468	506 sensor and a black pixel to mean the absence of a sensor, and use
rlm@468	507 one UV-map for each sensor-type within a given sense.
rlm@468	508
rlm@468	509 #+CAPTION: The UV-map for an elongated icososphere. The white
rlm@468	510 #+caption: dots each represent a touch sensor. They are dense
rlm@468	511 #+caption: in the regions that describe the tip of the finger,
rlm@468	512 #+caption: and less dense along the dorsal side of the finger
rlm@468	513 #+caption: opposite the tip.
rlm@468	514 #+name: finger-UV
rlm@468	515 #+ATTR_latex: :width 10cm
rlm@468	516 [[./images/finger-UV.png]]
rlm@468	517
rlm@468	518 #+caption: Ventral side of the UV-mapped finger. Notice the
rlm@468	519 #+caption: density of touch sensors at the tip.
rlm@468	520 #+name: finger-side-view
rlm@468	521 #+ATTR_LaTeX: :width 10cm
rlm@468	522 [[./images/finger-1.png]]
rlm@468	523
rlm@465	524 ** COMMENT Video game engines are a great starting point
rlm@462	525
rlm@462	526 I did not need to write my own physics simulation code or shader to
rlm@462	527 build =CORTEX=. Doing so would lead to a system that is impossible
rlm@462	528 for anyone but myself to use anyway. Instead, I use a video game
rlm@462	529 engine as a base and modify it to accomodate the additional needs
rlm@462	530 of =CORTEX=. Video game engines are an ideal starting point to
rlm@462	531 build =CORTEX=, because they are not far from being creature
rlm@463	532 building systems themselves.
rlm@462	533
rlm@462	534 First off, general purpose video game engines come with a physics
rlm@462	535 engine and lighting / sound system. The physics system provides
rlm@462	536 tools that can be co-opted to serve as touch, proprioception, and
rlm@462	537 muscles. Since some games support split screen views, a good video
rlm@462	538 game engine will allow you to efficiently create multiple cameras
rlm@463	539 in the simulated world that can be used as eyes. Video game systems
rlm@463	540 offer integrated asset management for things like textures and
rlm@468	541 creatures models, providing an avenue for defining creatures. They
rlm@468	542 also understand UV-mapping, since this technique is used to apply a
rlm@468	543 texture to a model. Finally, because video game engines support a
rlm@468	544 large number of users, as long as =CORTEX= doesn't stray too far
rlm@468	545 from the base system, other researchers can turn to this community
rlm@468	546 for help when doing their research.
rlm@463	547
rlm@465	548 ** COMMENT =CORTEX= is based on jMonkeyEngine3
rlm@463	549
rlm@463	550 While preparing to build =CORTEX= I studied several video game
rlm@463	551 engines to see which would best serve as a base. The top contenders
rlm@463	552 were:
rlm@463	553
rlm@463	554 - [[http://www.idsoftware.com][Quake II]]/[[http://www.bytonic.de/html/jake2.html][Jake2]] :: The Quake II engine was designed by ID
rlm@463	555 software in 1997. All the source code was released by ID
rlm@463	556 software into the Public Domain several years ago, and as a
rlm@463	557 result it has been ported to many different languages. This
rlm@463	558 engine was famous for its advanced use of realistic shading
rlm@463	559 and had decent and fast physics simulation. The main advantage
rlm@463	560 of the Quake II engine is its simplicity, but I ultimately
rlm@463	561 rejected it because the engine is too tied to the concept of a
rlm@463	562 first-person shooter game. One of the problems I had was that
rlm@463	563 there does not seem to be any easy way to attach multiple
rlm@463	564 cameras to a single character. There are also several physics
rlm@463	565 clipping issues that are corrected in a way that only applies
rlm@463	566 to the main character and do not apply to arbitrary objects.
rlm@463	567
rlm@463	568 - [[http://source.valvesoftware.com/][Source Engine]] :: The Source Engine evolved from the Quake II
rlm@463	569 and Quake I engines and is used by Valve in the Half-Life
rlm@463	570 series of games. The physics simulation in the Source Engine
rlm@463	571 is quite accurate and probably the best out of all the engines
rlm@463	572 I investigated. There is also an extensive community actively
rlm@463	573 working with the engine. However, applications that use the
rlm@463	574 Source Engine must be written in C++, the code is not open, it
rlm@463	575 only runs on Windows, and the tools that come with the SDK to
rlm@463	576 handle models and textures are complicated and awkward to use.
rlm@463	577
rlm@463	578 - [[http://jmonkeyengine.com/][jMonkeyEngine3]] :: jMonkeyEngine3 is a new library for creating
rlm@463	579 games in Java. It uses OpenGL to render to the screen and uses
rlm@463	580 screengraphs to avoid drawing things that do not appear on the
rlm@463	581 screen. It has an active community and several games in the
rlm@463	582 pipeline. The engine was not built to serve any particular
rlm@463	583 game but is instead meant to be used for any 3D game.
rlm@463	584
rlm@463	585 I chose jMonkeyEngine3 because it because it had the most features
rlm@464	586 out of all the free projects I looked at, and because I could then
rlm@463	587 write my code in clojure, an implementation of =LISP= that runs on
rlm@463	588 the JVM.
rlm@435	589
rlm@469	590 ** COMMENT =CORTEX= uses Blender to create creature models
rlm@435	591
rlm@464	592 For the simple worm-like creatures I will use later on in this
rlm@464	593 thesis, I could define a simple API in =CORTEX= that would allow
rlm@464	594 one to create boxes, spheres, etc., and leave that API as the sole
rlm@464	595 way to create creatures. However, for =CORTEX= to truly be useful
rlm@468	596 for other projects, it needs a way to construct complicated
rlm@464	597 creatures. If possible, it would be nice to leverage work that has
rlm@464	598 already been done by the community of 3D modelers, or at least
rlm@464	599 enable people who are talented at moedling but not programming to
rlm@468	600 design =CORTEX= creatures.
rlm@464	601
rlm@464	602 Therefore, I use Blender, a free 3D modeling program, as the main
rlm@464	603 way to create creatures in =CORTEX=. However, the creatures modeled
rlm@464	604 in Blender must also be simple to simulate in jMonkeyEngine3's game
rlm@468	605 engine, and must also be easy to rig with =CORTEX='s senses. I
rlm@468	606 accomplish this with extensive use of Blender's ``empty nodes.''
rlm@464	607
rlm@468	608 Empty nodes have no mass, physical presence, or appearance, but
rlm@468	609 they can hold metadata and have names. I use a tree structure of
rlm@468	610 empty nodes to specify senses in the following manner:
rlm@468	611
rlm@468	612 - Create a single top-level empty node whose name is the name of
rlm@468	613 the sense.
rlm@468	614 - Add empty nodes which each contain meta-data relevant to the
rlm@468	615 sense, including a UV-map describing the number/distribution of
rlm@468	616 sensors if applicable.
rlm@468	617 - Make each empty-node the child of the top-level node.
rlm@468	618
rlm@468	619 #+caption: An example of annoting a creature model with empty
rlm@468	620 #+caption: nodes to describe the layout of senses. There are
rlm@468	621 #+caption: multiple empty nodes which each describe the position
rlm@468	622 #+caption: of muscles, ears, eyes, or joints.
rlm@468	623 #+name: sense-nodes
rlm@468	624 #+ATTR_LaTeX: :width 10cm
rlm@468	625 [[./images/empty-sense-nodes.png]]
rlm@468	626
rlm@469	627 ** COMMENT Bodies are composed of segments connected by joints
rlm@468	628
rlm@468	629 Blender is a general purpose animation tool, which has been used in
rlm@468	630 the past to create high quality movies such as Sintel
rlm@468	631 \cite{sintel}. Though Blender can model and render even complicated
rlm@468	632 things like water, it is crucual to keep models that are meant to
rlm@468	633 be simulated as creatures simple. =Bullet=, which =CORTEX= uses
rlm@468	634 though jMonkeyEngine3, is a rigid-body physics system. This offers
rlm@468	635 a compromise between the expressiveness of a game level and the
rlm@468	636 speed at which it can be simulated, and it means that creatures
rlm@468	637 should be naturally expressed as rigid components held together by
rlm@468	638 joint constraints.
rlm@468	639
rlm@468	640 But humans are more like a squishy bag with wrapped around some
rlm@468	641 hard bones which define the overall shape. When we move, our skin
rlm@468	642 bends and stretches to accomodate the new positions of our bones.
rlm@468	643
rlm@468	644 One way to make bodies composed of rigid pieces connected by joints
rlm@468	645 /seem/ more human-like is to use an /armature/, (or /rigging/)
rlm@468	646 system, which defines a overall ``body mesh'' and defines how the
rlm@468	647 mesh deforms as a function of the position of each ``bone'' which
rlm@468	648 is a standard rigid body. This technique is used extensively to
rlm@468	649 model humans and create realistic animations. It is not a good
rlm@468	650 technique for physical simulation, however because it creates a lie
rlm@468	651 -- the skin is not a physical part of the simulation and does not
rlm@468	652 interact with any objects in the world or itself. Objects will pass
rlm@468	653 right though the skin until they come in contact with the
rlm@468	654 underlying bone, which is a physical object. Whithout simulating
rlm@468	655 the skin, the sense of touch has little meaning, and the creature's
rlm@468	656 own vision will lie to it about the true extent of its body.
rlm@468	657 Simulating the skin as a physical object requires some way to
rlm@468	658 continuously update the physical model of the skin along with the
rlm@468	659 movement of the bones, which is unacceptably slow compared to rigid
rlm@468	660 body simulation.
rlm@468	661
rlm@468	662 Therefore, instead of using the human-like ``deformable bag of
rlm@468	663 bones'' approach, I decided to base my body plans on multiple solid
rlm@468	664 objects that are connected by joints, inspired by the robot =EVE=
rlm@468	665 from the movie WALL-E.
rlm@464	666
rlm@464	667 #+caption: =EVE= from the movie WALL-E. This body plan turns
rlm@464	668 #+caption: out to be much better suited to my purposes than a more
rlm@464	669 #+caption: human-like one.
rlm@465	670 #+ATTR_LaTeX: :width 10cm
rlm@464	671 [[./images/Eve.jpg]]
rlm@464	672
rlm@464	673 =EVE='s body is composed of several rigid components that are held
rlm@464	674 together by invisible joint constraints. This is what I mean by
rlm@464	675 ``eve-like''. The main reason that I use eve-style bodies is for
rlm@464	676 efficiency, and so that there will be correspondence between the
rlm@468	677 AI's semses and the physical presence of its body. Each individual
rlm@464	678 section is simulated by a separate rigid body that corresponds
rlm@464	679 exactly with its visual representation and does not change.
rlm@464	680 Sections are connected by invisible joints that are well supported
rlm@464	681 in jMonkeyEngine3. Bullet, the physics backend for jMonkeyEngine3,
rlm@464	682 can efficiently simulate hundreds of rigid bodies connected by
rlm@468	683 joints. Just because sections are rigid does not mean they have to
rlm@468	684 stay as one piece forever; they can be dynamically replaced with
rlm@468	685 multiple sections to simulate splitting in two. This could be used
rlm@468	686 to simulate retractable claws or =EVE='s hands, which are able to
rlm@468	687 coalesce into one object in the movie.
rlm@465	688
rlm@469	689 *** Solidifying/Connecting a body
rlm@465	690
rlm@469	691 =CORTEX= creates a creature in two steps: first, it traverses the
rlm@469	692 nodes in the blender file and creates physical representations for
rlm@469	693 any of them that have mass defined in their blender meta-data.
rlm@466	694
rlm@466	695 #+caption: Program for iterating through the nodes in a blender file
rlm@466	696 #+caption: and generating physical jMonkeyEngine3 objects with mass
rlm@466	697 #+caption: and a matching physics shape.
rlm@466	698 #+name: name
rlm@466	699 #+begin_listing clojure
rlm@466	700 #+begin_src clojure
rlm@466	701 (defn physical!
rlm@466	702 "Iterate through the nodes in creature and make them real physical
rlm@466	703 objects in the simulation."
rlm@466	704 [#^Node creature]
rlm@466	705 (dorun
rlm@466	706 (map
rlm@466	707 (fn [geom]
rlm@466	708 (let [physics-control
rlm@466	709 (RigidBodyControl.
rlm@466	710 (HullCollisionShape.
rlm@466	711 (.getMesh geom))
rlm@466	712 (if-let [mass (meta-data geom "mass")]
rlm@466	713 (float mass) (float 1)))]
rlm@466	714 (.addControl geom physics-control)))
rlm@466	715 (filter #(isa? (class %) Geometry )
rlm@466	716 (node-seq creature)))))
rlm@466	717 #+end_src
rlm@466	718 #+end_listing
rlm@465	719
rlm@469	720 The next step to making a proper body is to connect those pieces
rlm@469	721 together with joints. jMonkeyEngine has a large array of joints
rlm@469	722 available via =bullet=, such as Point2Point, Cone, Hinge, and a
rlm@469	723 generic Six Degree of Freedom joint, with or without spring
rlm@469	724 restitution.
rlm@465	725
rlm@469	726 Joints are treated a lot like proper senses, in that there is a
rlm@469	727 top-level empty node named ``joints'' whose children each
rlm@469	728 represent a joint.
rlm@466	729
rlm@469	730 #+caption: View of the hand model in Blender showing the main ``joints''
rlm@469	731 #+caption: node (highlighted in yellow) and its children which each
rlm@469	732 #+caption: represent a joint in the hand. Each joint node has metadata
rlm@469	733 #+caption: specifying what sort of joint it is.
rlm@469	734 #+name: blender-hand
rlm@469	735 #+ATTR_LaTeX: :width 10cm
rlm@469	736 [[./images/hand-screenshot1.png]]
rlm@469	737
rlm@469	738
rlm@469	739 =CORTEX='s procedure for binding the creature together with joints
rlm@469	740 is as follows:
rlm@469	741
rlm@469	742 - Find the children of the ``joints'' node.
rlm@469	743 - Determine the two spatials the joint is meant to connect.
rlm@469	744 - Create the joint based on the meta-data of the empty node.
rlm@469	745
rlm@469	746 The higher order function =sense-nodes= from =cortex.sense=
rlm@469	747 simplifies finding the joints based on their parent ``joints''
rlm@469	748 node.
rlm@466	749
rlm@466	750 #+caption: Retrieving the children empty nodes from a single
rlm@466	751 #+caption: named empty node is a common pattern in =CORTEX=
rlm@466	752 #+caption: further instances of this technique for the senses
rlm@466	753 #+caption: will be omitted
rlm@466	754 #+name: get-empty-nodes
rlm@466	755 #+begin_listing clojure
rlm@466	756 #+begin_src clojure
rlm@466	757 (defn sense-nodes
rlm@466	758 "For some senses there is a special empty blender node whose
rlm@466	759 children are considered markers for an instance of that sense. This
rlm@466	760 function generates functions to find those children, given the name
rlm@466	761 of the special parent node."
rlm@466	762 [parent-name]
rlm@466	763 (fn [#^Node creature]
rlm@466	764 (if-let [sense-node (.getChild creature parent-name)]
rlm@466	765 (seq (.getChildren sense-node)) [])))
rlm@466	766
rlm@466	767 (def
rlm@466	768 ^{:doc "Return the children of the creature's \"joints\" node."
rlm@466	769 :arglists '([creature])}
rlm@466	770 joints
rlm@466	771 (sense-nodes "joints"))
rlm@466	772 #+end_src
rlm@466	773 #+end_listing
rlm@466	774
rlm@469	775 To find a joint's targets, =CORTEX= creates a small cube, centered
rlm@469	776 around the empty-node, and grows the cube exponentially until it
rlm@469	777 intersects two physical objects. The objects are ordered according
rlm@469	778 to the joint's rotation, with the first one being the object that
rlm@469	779 has more negative coordinates in the joint's reference frame.
rlm@469	780 Since the objects must be physical, the empty-node itself escapes
rlm@469	781 detection. Because the objects must be physical, =joint-targets=
rlm@469	782 must be called /after/ =physical!= is called.
rlm@464	783
rlm@469	784 #+caption: Program to find the targets of a joint node by
rlm@469	785 #+caption: exponentiallly growth of a search cube.
rlm@469	786 #+name: joint-targets
rlm@469	787 #+begin_listing clojure
rlm@469	788 #+begin_src clojure
rlm@466	789 (defn joint-targets
rlm@466	790 "Return the two closest two objects to the joint object, ordered
rlm@466	791 from bottom to top according to the joint's rotation."
rlm@466	792 [#^Node parts #^Node joint]
rlm@466	793 (loop [radius (float 0.01)]
rlm@466	794 (let [results (CollisionResults.)]
rlm@466	795 (.collideWith
rlm@466	796 parts
rlm@466	797 (BoundingBox. (.getWorldTranslation joint)
rlm@466	798 radius radius radius) results)
rlm@466	799 (let [targets
rlm@466	800 (distinct
rlm@466	801 (map #(.getGeometry %) results))]
rlm@466	802 (if (>= (count targets) 2)
rlm@466	803 (sort-by
rlm@466	804 #(let [joint-ref-frame-position
rlm@466	805 (jme-to-blender
rlm@466	806 (.mult
rlm@466	807 (.inverse (.getWorldRotation joint))
rlm@466	808 (.subtract (.getWorldTranslation %)
rlm@466	809 (.getWorldTranslation joint))))]
rlm@466	810 (.dot (Vector3f. 1 1 1) joint-ref-frame-position))
rlm@466	811 (take 2 targets))
rlm@466	812 (recur (float (* radius 2))))))))
rlm@469	813 #+end_src
rlm@469	814 #+end_listing
rlm@464	815
rlm@469	816 Once =CORTEX= finds all joints and targets, it creates them using
rlm@469	817 a dispatch on the metadata of each joint node.
rlm@466	818
rlm@469	819 #+caption: Program to dispatch on blender metadata and create joints
rlm@469	820 #+caption: sutiable for physical simulation.
rlm@469	821 #+name: joint-dispatch
rlm@469	822 #+begin_listing clojure
rlm@469	823 #+begin_src clojure
rlm@466	824 (defmulti joint-dispatch
rlm@466	825 "Translate blender pseudo-joints into real JME joints."
rlm@466	826 (fn [constraints & _]
rlm@466	827 (:type constraints)))
rlm@466	828
rlm@466	829 (defmethod joint-dispatch :point
rlm@466	830 [constraints control-a control-b pivot-a pivot-b rotation]
rlm@466	831 (doto (SixDofJoint. control-a control-b pivot-a pivot-b false)
rlm@466	832 (.setLinearLowerLimit Vector3f/ZERO)
rlm@466	833 (.setLinearUpperLimit Vector3f/ZERO)))
rlm@466	834
rlm@466	835 (defmethod joint-dispatch :hinge
rlm@466	836 [constraints control-a control-b pivot-a pivot-b rotation]
rlm@466	837 (let [axis (if-let [axis (:axis constraints)] axis Vector3f/UNIT_X)
rlm@466	838 [limit-1 limit-2] (:limit constraints)
rlm@466	839 hinge-axis (.mult rotation (blender-to-jme axis))]
rlm@466	840 (doto (HingeJoint. control-a control-b pivot-a pivot-b
rlm@466	841 hinge-axis hinge-axis)
rlm@466	842 (.setLimit limit-1 limit-2))))
rlm@466	843
rlm@466	844 (defmethod joint-dispatch :cone
rlm@466	845 [constraints control-a control-b pivot-a pivot-b rotation]
rlm@466	846 (let [limit-xz (:limit-xz constraints)
rlm@466	847 limit-xy (:limit-xy constraints)
rlm@466	848 twist (:twist constraints)]
rlm@466	849 (doto (ConeJoint. control-a control-b pivot-a pivot-b
rlm@466	850 rotation rotation)
rlm@466	851 (.setLimit (float limit-xz) (float limit-xy)
rlm@466	852 (float twist)))))
rlm@469	853 #+end_src
rlm@469	854 #+end_listing
rlm@466	855
rlm@469	856 All that is left for joints it to combine the above pieces into a
rlm@469	857 something that can operate on the collection of nodes that a
rlm@469	858 blender file represents.
rlm@466	859
rlm@469	860 #+caption: Program to completely create a joint given information
rlm@469	861 #+caption: from a blender file.
rlm@469	862 #+name: connect
rlm@469	863 #+begin_listing clojure
rlm@466	864 #+begin_src clojure
rlm@466	865 (defn connect
rlm@466	866 "Create a joint between 'obj-a and 'obj-b at the location of
rlm@466	867 'joint. The type of joint is determined by the metadata on 'joint.
rlm@466	868
rlm@466	869 Here are some examples:
rlm@466	870 {:type :point}
rlm@466	871 {:type :hinge :limit [0 (/ Math/PI 2)] :axis (Vector3f. 0 1 0)}
rlm@466	872 (:axis defaults to (Vector3f. 1 0 0) if not provided for hinge joints)
rlm@466	873
rlm@466	874 {:type :cone :limit-xz 0]
rlm@466	875 :limit-xy 0]
rlm@466	876 :twist 0]} (use XZY rotation mode in blender!)"
rlm@466	877 [#^Node obj-a #^Node obj-b #^Node joint]
rlm@466	878 (let [control-a (.getControl obj-a RigidBodyControl)
rlm@466	879 control-b (.getControl obj-b RigidBodyControl)
rlm@466	880 joint-center (.getWorldTranslation joint)
rlm@466	881 joint-rotation (.toRotationMatrix (.getWorldRotation joint))
rlm@466	882 pivot-a (world-to-local obj-a joint-center)
rlm@466	883 pivot-b (world-to-local obj-b joint-center)]
rlm@466	884 (if-let
rlm@466	885 [constraints (map-vals eval (read-string (meta-data joint "joint")))]
rlm@466	886 ;; A side-effect of creating a joint registers
rlm@466	887 ;; it with both physics objects which in turn
rlm@466	888 ;; will register the joint with the physics system
rlm@466	889 ;; when the simulation is started.
rlm@466	890 (joint-dispatch constraints
rlm@466	891 control-a control-b
rlm@466	892 pivot-a pivot-b
rlm@466	893 joint-rotation))))
rlm@469	894 #+end_src
rlm@469	895 #+end_listing
rlm@466	896
rlm@469	897 In general, whenever =CORTEX= exposes a sense (or in this case
rlm@469	898 physicality), it provides a function of the type =sense!=, which
rlm@469	899 takes in a collection of nodes and augments it to support that
rlm@469	900 sense. The function returns any controlls necessary to use that
rlm@469	901 sense. In this case =body!= cerates a physical body and returns no
rlm@469	902 control functions.
rlm@466	903
rlm@469	904 #+caption: Program to give joints to a creature.
rlm@469	905 #+name: name
rlm@469	906 #+begin_listing clojure
rlm@469	907 #+begin_src clojure
rlm@466	908 (defn joints!
rlm@466	909 "Connect the solid parts of the creature with physical joints. The
rlm@466	910 joints are taken from the \"joints\" node in the creature."
rlm@466	911 [#^Node creature]
rlm@466	912 (dorun
rlm@466	913 (map
rlm@466	914 (fn [joint]
rlm@466	915 (let [[obj-a obj-b] (joint-targets creature joint)]
rlm@466	916 (connect obj-a obj-b joint)))
rlm@466	917 (joints creature))))
rlm@466	918 (defn body!
rlm@466	919 "Endow the creature with a physical body connected with joints. The
rlm@466	920 particulars of the joints and the masses of each body part are
rlm@466	921 determined in blender."
rlm@466	922 [#^Node creature]
rlm@466	923 (physical! creature)
rlm@466	924 (joints! creature))
rlm@469	925 #+end_src
rlm@469	926 #+end_listing
rlm@466	927
rlm@469	928 All of the code you have just seen amounts to only 130 lines, yet
rlm@469	929 because it builds on top of Blender and jMonkeyEngine3, those few
rlm@469	930 lines pack quite a punch!
rlm@466	931
rlm@469	932 The hand from figure \ref{blender-hand}, which was modeled after
rlm@469	933 my own right hand, can now be given joints and simulated as a
rlm@469	934 creature.
rlm@466	935
rlm@469	936 #+caption: With the ability to create physical creatures from blender,
rlm@469	937 #+caption: =CORTEX= gets one step closer to becomming a full creature
rlm@469	938 #+caption: simulation environment.
rlm@469	939 #+name: name
rlm@469	940 #+ATTR_LaTeX: :width 15cm
rlm@469	941 [[./images/physical-hand.png]]
rlm@468	942
rlm@436	943 ** Eyes reuse standard video game components
rlm@436	944
rlm@436	945 ** Hearing is hard; =CORTEX= does it right
rlm@436	946
rlm@436	947 ** Touch uses hundreds of hair-like elements
rlm@436	948
rlm@440	949 ** Proprioception is the sense that makes everything ``real''
rlm@436	950
rlm@436	951 ** Muscles are both effectors and sensors
rlm@436	952
rlm@436	953 ** =CORTEX= brings complex creatures to life!
rlm@436	954
rlm@436	955 ** =CORTEX= enables many possiblities for further research
rlm@435	956
rlm@465	957 * COMMENT Empathy in a simulated worm
rlm@435	958
rlm@449	959 Here I develop a computational model of empathy, using =CORTEX= as a
rlm@449	960 base. Empathy in this context is the ability to observe another
rlm@449	961 creature and infer what sorts of sensations that creature is
rlm@449	962 feeling. My empathy algorithm involves multiple phases. First is
rlm@449	963 free-play, where the creature moves around and gains sensory
rlm@449	964 experience. From this experience I construct a representation of the
rlm@449	965 creature's sensory state space, which I call \Phi-space. Using
rlm@449	966 \Phi-space, I construct an efficient function which takes the
rlm@449	967 limited data that comes from observing another creature and enriches
rlm@449	968 it full compliment of imagined sensory data. I can then use the
rlm@449	969 imagined sensory data to recognize what the observed creature is
rlm@449	970 doing and feeling, using straightforward embodied action predicates.
rlm@449	971 This is all demonstrated with using a simple worm-like creature, and
rlm@449	972 recognizing worm-actions based on limited data.
rlm@449	973
rlm@449	974 #+caption: Here is the worm with which we will be working.
rlm@449	975 #+caption: It is composed of 5 segments. Each segment has a
rlm@449	976 #+caption: pair of extensor and flexor muscles. Each of the
rlm@449	977 #+caption: worm's four joints is a hinge joint which allows
rlm@451	978 #+caption: about 30 degrees of rotation to either side. Each segment
rlm@449	979 #+caption: of the worm is touch-capable and has a uniform
rlm@449	980 #+caption: distribution of touch sensors on each of its faces.
rlm@449	981 #+caption: Each joint has a proprioceptive sense to detect
rlm@449	982 #+caption: relative positions. The worm segments are all the
rlm@449	983 #+caption: same except for the first one, which has a much
rlm@449	984 #+caption: higher weight than the others to allow for easy
rlm@449	985 #+caption: manual motor control.
rlm@449	986 #+name: basic-worm-view
rlm@449	987 #+ATTR_LaTeX: :width 10cm
rlm@449	988 [[./images/basic-worm-view.png]]
rlm@449	989
rlm@449	990 #+caption: Program for reading a worm from a blender file and
rlm@449	991 #+caption: outfitting it with the senses of proprioception,
rlm@449	992 #+caption: touch, and the ability to move, as specified in the
rlm@449	993 #+caption: blender file.
rlm@449	994 #+name: get-worm
rlm@449	995 #+begin_listing clojure
rlm@449	996 #+begin_src clojure
rlm@449	997 (defn worm []
rlm@449	998 (let [model (load-blender-model "Models/worm/worm.blend")]
rlm@449	999 {:body (doto model (body!))
rlm@449	1000 :touch (touch! model)
rlm@449	1001 :proprioception (proprioception! model)
rlm@449	1002 :muscles (movement! model)}))
rlm@449	1003 #+end_src
rlm@449	1004 #+end_listing
rlm@452	1005
rlm@436	1006 ** Embodiment factors action recognition into managable parts
rlm@435	1007
rlm@449	1008 Using empathy, I divide the problem of action recognition into a
rlm@449	1009 recognition process expressed in the language of a full compliment
rlm@449	1010 of senses, and an imaganitive process that generates full sensory
rlm@449	1011 data from partial sensory data. Splitting the action recognition
rlm@449	1012 problem in this manner greatly reduces the total amount of work to
rlm@449	1013 recognize actions: The imaganitive process is mostly just matching
rlm@449	1014 previous experience, and the recognition process gets to use all
rlm@449	1015 the senses to directly describe any action.
rlm@449	1016
rlm@436	1017 ** Action recognition is easy with a full gamut of senses
rlm@435	1018
rlm@449	1019 Embodied representations using multiple senses such as touch,
rlm@449	1020 proprioception, and muscle tension turns out be be exceedingly
rlm@449	1021 efficient at describing body-centered actions. It is the ``right
rlm@449	1022 language for the job''. For example, it takes only around 5 lines
rlm@449	1023 of LISP code to describe the action of ``curling'' using embodied
rlm@451	1024 primitives. It takes about 10 lines to describe the seemingly
rlm@449	1025 complicated action of wiggling.
rlm@449	1026
rlm@449	1027 The following action predicates each take a stream of sensory
rlm@449	1028 experience, observe however much of it they desire, and decide
rlm@449	1029 whether the worm is doing the action they describe. =curled?=
rlm@449	1030 relies on proprioception, =resting?= relies on touch, =wiggling?=
rlm@449	1031 relies on a fourier analysis of muscle contraction, and
rlm@449	1032 =grand-circle?= relies on touch and reuses =curled?= as a gaurd.
rlm@449	1033
rlm@449	1034 #+caption: Program for detecting whether the worm is curled. This is the
rlm@449	1035 #+caption: simplest action predicate, because it only uses the last frame
rlm@449	1036 #+caption: of sensory experience, and only uses proprioceptive data. Even
rlm@449	1037 #+caption: this simple predicate, however, is automatically frame
rlm@449	1038 #+caption: independent and ignores vermopomorphic differences such as
rlm@449	1039 #+caption: worm textures and colors.
rlm@449	1040 #+name: curled
rlm@452	1041 #+attr_latex: [htpb]
rlm@452	1042 #+begin_listing clojure
rlm@449	1043 #+begin_src clojure
rlm@449	1044 (defn curled?
rlm@449	1045 "Is the worm curled up?"
rlm@449	1046 [experiences]
rlm@449	1047 (every?
rlm@449	1048 (fn [[_ _ bend]]
rlm@449	1049 (> (Math/sin bend) 0.64))
rlm@449	1050 (:proprioception (peek experiences))))
rlm@449	1051 #+end_src
rlm@449	1052 #+end_listing
rlm@449	1053
rlm@449	1054 #+caption: Program for summarizing the touch information in a patch
rlm@449	1055 #+caption: of skin.
rlm@449	1056 #+name: touch-summary
rlm@452	1057 #+attr_latex: [htpb]
rlm@452	1058
rlm@452	1059 #+begin_listing clojure
rlm@449	1060 #+begin_src clojure
rlm@449	1061 (defn contact
rlm@449	1062 "Determine how much contact a particular worm segment has with
rlm@449	1063 other objects. Returns a value between 0 and 1, where 1 is full
rlm@449	1064 contact and 0 is no contact."
rlm@449	1065 [touch-region [coords contact :as touch]]
rlm@449	1066 (-> (zipmap coords contact)
rlm@449	1067 (select-keys touch-region)
rlm@449	1068 (vals)
rlm@449	1069 (#(map first %))
rlm@449	1070 (average)
rlm@449	1071 (* 10)
rlm@449	1072 (- 1)
rlm@449	1073 (Math/abs)))
rlm@449	1074 #+end_src
rlm@449	1075 #+end_listing
rlm@449	1076
rlm@449	1077
rlm@449	1078 #+caption: Program for detecting whether the worm is at rest. This program
rlm@449	1079 #+caption: uses a summary of the tactile information from the underbelly
rlm@449	1080 #+caption: of the worm, and is only true if every segment is touching the
rlm@449	1081 #+caption: floor. Note that this function contains no references to
rlm@449	1082 #+caption: proprioction at all.
rlm@449	1083 #+name: resting
rlm@452	1084 #+attr_latex: [htpb]
rlm@452	1085 #+begin_listing clojure
rlm@449	1086 #+begin_src clojure
rlm@449	1087 (def worm-segment-bottom (rect-region [8 15] [14 22]))
rlm@449	1088
rlm@449	1089 (defn resting?
rlm@449	1090 "Is the worm resting on the ground?"
rlm@449	1091 [experiences]
rlm@449	1092 (every?
rlm@449	1093 (fn [touch-data]
rlm@449	1094 (< 0.9 (contact worm-segment-bottom touch-data)))
rlm@449	1095 (:touch (peek experiences))))
rlm@449	1096 #+end_src
rlm@449	1097 #+end_listing
rlm@449	1098
rlm@449	1099 #+caption: Program for detecting whether the worm is curled up into a
rlm@449	1100 #+caption: full circle. Here the embodied approach begins to shine, as
rlm@449	1101 #+caption: I am able to both use a previous action predicate (=curled?=)
rlm@449	1102 #+caption: as well as the direct tactile experience of the head and tail.
rlm@449	1103 #+name: grand-circle
rlm@452	1104 #+attr_latex: [htpb]
rlm@452	1105 #+begin_listing clojure
rlm@449	1106 #+begin_src clojure
rlm@449	1107 (def worm-segment-bottom-tip (rect-region [15 15] [22 22]))
rlm@449	1108
rlm@449	1109 (def worm-segment-top-tip (rect-region [0 15] [7 22]))
rlm@449	1110
rlm@449	1111 (defn grand-circle?
rlm@449	1112 "Does the worm form a majestic circle (one end touching the other)?"
rlm@449	1113 [experiences]
rlm@449	1114 (and (curled? experiences)
rlm@449	1115 (let [worm-touch (:touch (peek experiences))
rlm@449	1116 tail-touch (worm-touch 0)
rlm@449	1117 head-touch (worm-touch 4)]
rlm@449	1118 (and (< 0.55 (contact worm-segment-bottom-tip tail-touch))
rlm@449	1119 (< 0.55 (contact worm-segment-top-tip head-touch))))))
rlm@449	1120 #+end_src
rlm@449	1121 #+end_listing
rlm@449	1122
rlm@449	1123
rlm@449	1124 #+caption: Program for detecting whether the worm has been wiggling for
rlm@449	1125 #+caption: the last few frames. It uses a fourier analysis of the muscle
rlm@449	1126 #+caption: contractions of the worm's tail to determine wiggling. This is
rlm@449	1127 #+caption: signigicant because there is no particular frame that clearly
rlm@449	1128 #+caption: indicates that the worm is wiggling --- only when multiple frames
rlm@449	1129 #+caption: are analyzed together is the wiggling revealed. Defining
rlm@449	1130 #+caption: wiggling this way also gives the worm an opportunity to learn
rlm@449	1131 #+caption: and recognize ``frustrated wiggling'', where the worm tries to
rlm@449	1132 #+caption: wiggle but can't. Frustrated wiggling is very visually different
rlm@449	1133 #+caption: from actual wiggling, but this definition gives it to us for free.
rlm@449	1134 #+name: wiggling
rlm@452	1135 #+attr_latex: [htpb]
rlm@452	1136 #+begin_listing clojure
rlm@449	1137 #+begin_src clojure
rlm@449	1138 (defn fft [nums]
rlm@449	1139 (map
rlm@449	1140 #(.getReal %)
rlm@449	1141 (.transform
rlm@449	1142 (FastFourierTransformer. DftNormalization/STANDARD)
rlm@449	1143 (double-array nums) TransformType/FORWARD)))
rlm@449	1144
rlm@449	1145 (def indexed (partial map-indexed vector))
rlm@449	1146
rlm@449	1147 (defn max-indexed [s]
rlm@449	1148 (first (sort-by (comp - second) (indexed s))))
rlm@449	1149
rlm@449	1150 (defn wiggling?
rlm@449	1151 "Is the worm wiggling?"
rlm@449	1152 [experiences]
rlm@449	1153 (let [analysis-interval 0x40]
rlm@449	1154 (when (> (count experiences) analysis-interval)
rlm@449	1155 (let [a-flex 3
rlm@449	1156 a-ex 2
rlm@449	1157 muscle-activity
rlm@449	1158 (map :muscle (vector:last-n experiences analysis-interval))
rlm@449	1159 base-activity
rlm@449	1160 (map #(- (% a-flex) (% a-ex)) muscle-activity)]
rlm@449	1161 (= 2
rlm@449	1162 (first
rlm@449	1163 (max-indexed
rlm@449	1164 (map #(Math/abs %)
rlm@449	1165 (take 20 (fft base-activity))))))))))
rlm@449	1166 #+end_src
rlm@449	1167 #+end_listing
rlm@449	1168
rlm@449	1169 With these action predicates, I can now recognize the actions of
rlm@449	1170 the worm while it is moving under my control and I have access to
rlm@449	1171 all the worm's senses.
rlm@449	1172
rlm@449	1173 #+caption: Use the action predicates defined earlier to report on
rlm@449	1174 #+caption: what the worm is doing while in simulation.
rlm@449	1175 #+name: report-worm-activity
rlm@452	1176 #+attr_latex: [htpb]
rlm@452	1177 #+begin_listing clojure
rlm@449	1178 #+begin_src clojure
rlm@449	1179 (defn debug-experience
rlm@449	1180 [experiences text]
rlm@449	1181 (cond
rlm@449	1182 (grand-circle? experiences) (.setText text "Grand Circle")
rlm@449	1183 (curled? experiences) (.setText text "Curled")
rlm@449	1184 (wiggling? experiences) (.setText text "Wiggling")
rlm@449	1185 (resting? experiences) (.setText text "Resting")))
rlm@449	1186 #+end_src
rlm@449	1187 #+end_listing
rlm@449	1188
rlm@449	1189 #+caption: Using =debug-experience=, the body-centered predicates
rlm@449	1190 #+caption: work together to classify the behaviour of the worm.
rlm@451	1191 #+caption: the predicates are operating with access to the worm's
rlm@451	1192 #+caption: full sensory data.
rlm@449	1193 #+name: basic-worm-view
rlm@449	1194 #+ATTR_LaTeX: :width 10cm
rlm@449	1195 [[./images/worm-identify-init.png]]
rlm@449	1196
rlm@449	1197 These action predicates satisfy the recognition requirement of an
rlm@451	1198 empathic recognition system. There is power in the simplicity of
rlm@451	1199 the action predicates. They describe their actions without getting
rlm@451	1200 confused in visual details of the worm. Each one is frame
rlm@451	1201 independent, but more than that, they are each indepent of
rlm@449	1202 irrelevant visual details of the worm and the environment. They
rlm@449	1203 will work regardless of whether the worm is a different color or
rlm@451	1204 hevaily textured, or if the environment has strange lighting.
rlm@449	1205
rlm@449	1206 The trick now is to make the action predicates work even when the
rlm@449	1207 sensory data on which they depend is absent. If I can do that, then
rlm@449	1208 I will have gained much,
rlm@435	1209
rlm@436	1210 ** \Phi-space describes the worm's experiences
rlm@449	1211
rlm@449	1212 As a first step towards building empathy, I need to gather all of
rlm@449	1213 the worm's experiences during free play. I use a simple vector to
rlm@449	1214 store all the experiences.
rlm@449	1215
rlm@449	1216 Each element of the experience vector exists in the vast space of
rlm@449	1217 all possible worm-experiences. Most of this vast space is actually
rlm@449	1218 unreachable due to physical constraints of the worm's body. For
rlm@449	1219 example, the worm's segments are connected by hinge joints that put
rlm@451	1220 a practical limit on the worm's range of motions without limiting
rlm@451	1221 its degrees of freedom. Some groupings of senses are impossible;
rlm@451	1222 the worm can not be bent into a circle so that its ends are
rlm@451	1223 touching and at the same time not also experience the sensation of
rlm@451	1224 touching itself.
rlm@449	1225
rlm@451	1226 As the worm moves around during free play and its experience vector
rlm@451	1227 grows larger, the vector begins to define a subspace which is all
rlm@451	1228 the sensations the worm can practicaly experience during normal
rlm@451	1229 operation. I call this subspace \Phi-space, short for
rlm@451	1230 physical-space. The experience vector defines a path through
rlm@451	1231 \Phi-space. This path has interesting properties that all derive
rlm@451	1232 from physical embodiment. The proprioceptive components are
rlm@451	1233 completely smooth, because in order for the worm to move from one
rlm@451	1234 position to another, it must pass through the intermediate
rlm@451	1235 positions. The path invariably forms loops as actions are repeated.
rlm@451	1236 Finally and most importantly, proprioception actually gives very
rlm@451	1237 strong inference about the other senses. For example, when the worm
rlm@451	1238 is flat, you can infer that it is touching the ground and that its
rlm@451	1239 muscles are not active, because if the muscles were active, the
rlm@451	1240 worm would be moving and would not be perfectly flat. In order to
rlm@451	1241 stay flat, the worm has to be touching the ground, or it would
rlm@451	1242 again be moving out of the flat position due to gravity. If the
rlm@451	1243 worm is positioned in such a way that it interacts with itself,
rlm@451	1244 then it is very likely to be feeling the same tactile feelings as
rlm@451	1245 the last time it was in that position, because it has the same body
rlm@451	1246 as then. If you observe multiple frames of proprioceptive data,
rlm@451	1247 then you can become increasingly confident about the exact
rlm@451	1248 activations of the worm's muscles, because it generally takes a
rlm@451	1249 unique combination of muscle contractions to transform the worm's
rlm@451	1250 body along a specific path through \Phi-space.
rlm@449	1251
rlm@449	1252 There is a simple way of taking \Phi-space and the total ordering
rlm@449	1253 provided by an experience vector and reliably infering the rest of
rlm@449	1254 the senses.
rlm@435	1255
rlm@436	1256 ** Empathy is the process of tracing though \Phi-space
rlm@449	1257
rlm@450	1258 Here is the core of a basic empathy algorithm, starting with an
rlm@451	1259 experience vector:
rlm@451	1260
rlm@451	1261 First, group the experiences into tiered proprioceptive bins. I use
rlm@451	1262 powers of 10 and 3 bins, and the smallest bin has an approximate
rlm@451	1263 size of 0.001 radians in all proprioceptive dimensions.
rlm@450	1264
rlm@450	1265 Then, given a sequence of proprioceptive input, generate a set of
rlm@451	1266 matching experience records for each input, using the tiered
rlm@451	1267 proprioceptive bins.
rlm@449	1268
rlm@450	1269 Finally, to infer sensory data, select the longest consective chain
rlm@451	1270 of experiences. Conecutive experience means that the experiences
rlm@451	1271 appear next to each other in the experience vector.
rlm@449	1272
rlm@450	1273 This algorithm has three advantages:
rlm@450	1274
rlm@450	1275 1. It's simple
rlm@450	1276
rlm@451	1277 3. It's very fast -- retrieving possible interpretations takes
rlm@451	1278 constant time. Tracing through chains of interpretations takes
rlm@451	1279 time proportional to the average number of experiences in a
rlm@451	1280 proprioceptive bin. Redundant experiences in \Phi-space can be
rlm@451	1281 merged to save computation.
rlm@450	1282
rlm@450	1283 2. It protects from wrong interpretations of transient ambiguous
rlm@451	1284 proprioceptive data. For example, if the worm is flat for just
rlm@450	1285 an instant, this flattness will not be interpreted as implying
rlm@450	1286 that the worm has its muscles relaxed, since the flattness is
rlm@450	1287 part of a longer chain which includes a distinct pattern of
rlm@451	1288 muscle activation. Markov chains or other memoryless statistical
rlm@451	1289 models that operate on individual frames may very well make this
rlm@451	1290 mistake.
rlm@450	1291
rlm@450	1292 #+caption: Program to convert an experience vector into a
rlm@450	1293 #+caption: proprioceptively binned lookup function.
rlm@450	1294 #+name: bin
rlm@452	1295 #+attr_latex: [htpb]
rlm@452	1296 #+begin_listing clojure
rlm@450	1297 #+begin_src clojure
rlm@449	1298 (defn bin [digits]
rlm@449	1299 (fn [angles]
rlm@449	1300 (->> angles
rlm@449	1301 (flatten)
rlm@449	1302 (map (juxt #(Math/sin %) #(Math/cos %)))
rlm@449	1303 (flatten)
rlm@449	1304 (mapv #(Math/round (* % (Math/pow 10 (dec digits))))))))
rlm@449	1305
rlm@449	1306 (defn gen-phi-scan
rlm@450	1307 "Nearest-neighbors with binning. Only returns a result if
rlm@450	1308 the propriceptive data is within 10% of a previously recorded
rlm@450	1309 result in all dimensions."
rlm@450	1310 [phi-space]
rlm@449	1311 (let [bin-keys (map bin [3 2 1])
rlm@449	1312 bin-maps
rlm@449	1313 (map (fn [bin-key]
rlm@449	1314 (group-by
rlm@449	1315 (comp bin-key :proprioception phi-space)
rlm@449	1316 (range (count phi-space)))) bin-keys)
rlm@449	1317 lookups (map (fn [bin-key bin-map]
rlm@450	1318 (fn [proprio] (bin-map (bin-key proprio))))
rlm@450	1319 bin-keys bin-maps)]
rlm@449	1320 (fn lookup [proprio-data]
rlm@449	1321 (set (some #(% proprio-data) lookups)))))
rlm@450	1322 #+end_src
rlm@450	1323 #+end_listing
rlm@449	1324
rlm@451	1325 #+caption: =longest-thread= finds the longest path of consecutive
rlm@451	1326 #+caption: experiences to explain proprioceptive worm data.
rlm@451	1327 #+name: phi-space-history-scan
rlm@451	1328 #+ATTR_LaTeX: :width 10cm
rlm@451	1329 [[./images/aurellem-gray.png]]
rlm@451	1330
rlm@451	1331 =longest-thread= infers sensory data by stitching together pieces
rlm@451	1332 from previous experience. It prefers longer chains of previous
rlm@451	1333 experience to shorter ones. For example, during training the worm
rlm@451	1334 might rest on the ground for one second before it performs its
rlm@451	1335 excercises. If during recognition the worm rests on the ground for
rlm@451	1336 five seconds, =longest-thread= will accomodate this five second
rlm@451	1337 rest period by looping the one second rest chain five times.
rlm@451	1338
rlm@451	1339 =longest-thread= takes time proportinal to the average number of
rlm@451	1340 entries in a proprioceptive bin, because for each element in the
rlm@451	1341 starting bin it performes a series of set lookups in the preceeding
rlm@451	1342 bins. If the total history is limited, then this is only a constant
rlm@451	1343 multiple times the number of entries in the starting bin. This
rlm@451	1344 analysis also applies even if the action requires multiple longest
rlm@451	1345 chains -- it's still the average number of entries in a
rlm@451	1346 proprioceptive bin times the desired chain length. Because
rlm@451	1347 =longest-thread= is so efficient and simple, I can interpret
rlm@451	1348 worm-actions in real time.
rlm@449	1349
rlm@450	1350 #+caption: Program to calculate empathy by tracing though \Phi-space
rlm@450	1351 #+caption: and finding the longest (ie. most coherent) interpretation
rlm@450	1352 #+caption: of the data.
rlm@450	1353 #+name: longest-thread
rlm@452	1354 #+attr_latex: [htpb]
rlm@452	1355 #+begin_listing clojure
rlm@450	1356 #+begin_src clojure
rlm@449	1357 (defn longest-thread
rlm@449	1358 "Find the longest thread from phi-index-sets. The index sets should
rlm@449	1359 be ordered from most recent to least recent."
rlm@449	1360 [phi-index-sets]
rlm@449	1361 (loop [result '()
rlm@449	1362 [thread-bases & remaining :as phi-index-sets] phi-index-sets]
rlm@449	1363 (if (empty? phi-index-sets)
rlm@449	1364 (vec result)
rlm@449	1365 (let [threads
rlm@449	1366 (for [thread-base thread-bases]
rlm@449	1367 (loop [thread (list thread-base)
rlm@449	1368 remaining remaining]
rlm@449	1369 (let [next-index (dec (first thread))]
rlm@449	1370 (cond (empty? remaining) thread
rlm@449	1371 (contains? (first remaining) next-index)
rlm@449	1372 (recur
rlm@449	1373 (cons next-index thread) (rest remaining))
rlm@449	1374 :else thread))))
rlm@449	1375 longest-thread
rlm@449	1376 (reduce (fn [thread-a thread-b]
rlm@449	1377 (if (> (count thread-a) (count thread-b))
rlm@449	1378 thread-a thread-b))
rlm@449	1379 '(nil)
rlm@449	1380 threads)]
rlm@449	1381 (recur (concat longest-thread result)
rlm@449	1382 (drop (count longest-thread) phi-index-sets))))))
rlm@450	1383 #+end_src
rlm@450	1384 #+end_listing
rlm@450	1385
rlm@451	1386 There is one final piece, which is to replace missing sensory data
rlm@451	1387 with a best-guess estimate. While I could fill in missing data by
rlm@451	1388 using a gradient over the closest known sensory data points,
rlm@451	1389 averages can be misleading. It is certainly possible to create an
rlm@451	1390 impossible sensory state by averaging two possible sensory states.
rlm@451	1391 Therefore, I simply replicate the most recent sensory experience to
rlm@451	1392 fill in the gaps.
rlm@449	1393
rlm@449	1394 #+caption: Fill in blanks in sensory experience by replicating the most
rlm@449	1395 #+caption: recent experience.
rlm@449	1396 #+name: infer-nils
rlm@452	1397 #+attr_latex: [htpb]
rlm@452	1398 #+begin_listing clojure
rlm@449	1399 #+begin_src clojure
rlm@449	1400 (defn infer-nils
rlm@449	1401 "Replace nils with the next available non-nil element in the
rlm@449	1402 sequence, or barring that, 0."
rlm@449	1403 [s]
rlm@449	1404 (loop [i (dec (count s))
rlm@449	1405 v (transient s)]
rlm@449	1406 (if (zero? i) (persistent! v)
rlm@449	1407 (if-let [cur (v i)]
rlm@449	1408 (if (get v (dec i) 0)
rlm@449	1409 (recur (dec i) v)
rlm@449	1410 (recur (dec i) (assoc! v (dec i) cur)))
rlm@449	1411 (recur i (assoc! v i 0))))))
rlm@449	1412 #+end_src
rlm@449	1413 #+end_listing
rlm@435	1414
rlm@441	1415 ** Efficient action recognition with =EMPATH=
rlm@451	1416
rlm@451	1417 To use =EMPATH= with the worm, I first need to gather a set of
rlm@451	1418 experiences from the worm that includes the actions I want to
rlm@452	1419 recognize. The =generate-phi-space= program (listing
rlm@451	1420 \ref{generate-phi-space} runs the worm through a series of
rlm@451	1421 exercices and gatheres those experiences into a vector. The
rlm@451	1422 =do-all-the-things= program is a routine expressed in a simple
rlm@452	1423 muscle contraction script language for automated worm control. It
rlm@452	1424 causes the worm to rest, curl, and wiggle over about 700 frames
rlm@452	1425 (approx. 11 seconds).
rlm@425	1426
rlm@451	1427 #+caption: Program to gather the worm's experiences into a vector for
rlm@451	1428 #+caption: further processing. The =motor-control-program= line uses
rlm@451	1429 #+caption: a motor control script that causes the worm to execute a series
rlm@451	1430 #+caption: of ``exercices'' that include all the action predicates.
rlm@451	1431 #+name: generate-phi-space
rlm@452	1432 #+attr_latex: [htpb]
rlm@452	1433 #+begin_listing clojure
rlm@451	1434 #+begin_src clojure
rlm@451	1435 (def do-all-the-things
rlm@451	1436 (concat
rlm@451	1437 curl-script
rlm@451	1438 [[300 :d-ex 40]
rlm@451	1439 [320 :d-ex 0]]
rlm@451	1440 (shift-script 280 (take 16 wiggle-script))))
rlm@451	1441
rlm@451	1442 (defn generate-phi-space []
rlm@451	1443 (let [experiences (atom [])]
rlm@451	1444 (run-world
rlm@451	1445 (apply-map
rlm@451	1446 worm-world
rlm@451	1447 (merge
rlm@451	1448 (worm-world-defaults)
rlm@451	1449 {:end-frame 700
rlm@451	1450 :motor-control
rlm@451	1451 (motor-control-program worm-muscle-labels do-all-the-things)
rlm@451	1452 :experiences experiences})))
rlm@451	1453 @experiences))
rlm@451	1454 #+end_src
rlm@451	1455 #+end_listing
rlm@451	1456
rlm@451	1457 #+caption: Use longest thread and a phi-space generated from a short
rlm@451	1458 #+caption: exercise routine to interpret actions during free play.
rlm@451	1459 #+name: empathy-debug
rlm@452	1460 #+attr_latex: [htpb]
rlm@452	1461 #+begin_listing clojure
rlm@451	1462 #+begin_src clojure
rlm@451	1463 (defn init []
rlm@451	1464 (def phi-space (generate-phi-space))
rlm@451	1465 (def phi-scan (gen-phi-scan phi-space)))
rlm@451	1466
rlm@451	1467 (defn empathy-demonstration []
rlm@451	1468 (let [proprio (atom ())]
rlm@451	1469 (fn
rlm@451	1470 [experiences text]
rlm@451	1471 (let [phi-indices (phi-scan (:proprioception (peek experiences)))]
rlm@451	1472 (swap! proprio (partial cons phi-indices))
rlm@451	1473 (let [exp-thread (longest-thread (take 300 @proprio))
rlm@451	1474 empathy (mapv phi-space (infer-nils exp-thread))]
rlm@451	1475 (println-repl (vector:last-n exp-thread 22))
rlm@451	1476 (cond
rlm@451	1477 (grand-circle? empathy) (.setText text "Grand Circle")
rlm@451	1478 (curled? empathy) (.setText text "Curled")
rlm@451	1479 (wiggling? empathy) (.setText text "Wiggling")
rlm@451	1480 (resting? empathy) (.setText text "Resting")
rlm@451	1481 :else (.setText text "Unknown")))))))
rlm@451	1482
rlm@451	1483 (defn empathy-experiment [record]
rlm@451	1484 (.start (worm-world :experience-watch (debug-experience-phi)
rlm@451	1485 :record record :worm worm*)))
rlm@451	1486 #+end_src
rlm@451	1487 #+end_listing
rlm@451	1488
rlm@451	1489 The result of running =empathy-experiment= is that the system is
rlm@451	1490 generally able to interpret worm actions using the action-predicates
rlm@451	1491 on simulated sensory data just as well as with actual data. Figure
rlm@451	1492 \ref{empathy-debug-image} was generated using =empathy-experiment=:
rlm@451	1493
rlm@451	1494 #+caption: From only proprioceptive data, =EMPATH= was able to infer
rlm@451	1495 #+caption: the complete sensory experience and classify four poses
rlm@451	1496 #+caption: (The last panel shows a composite image of \emph{wriggling},
rlm@451	1497 #+caption: a dynamic pose.)
rlm@451	1498 #+name: empathy-debug-image
rlm@451	1499 #+ATTR_LaTeX: :width 10cm :placement [H]
rlm@451	1500 [[./images/empathy-1.png]]
rlm@451	1501
rlm@451	1502 One way to measure the performance of =EMPATH= is to compare the
rlm@451	1503 sutiability of the imagined sense experience to trigger the same
rlm@451	1504 action predicates as the real sensory experience.
rlm@451	1505
rlm@451	1506 #+caption: Determine how closely empathy approximates actual
rlm@451	1507 #+caption: sensory data.
rlm@451	1508 #+name: test-empathy-accuracy
rlm@452	1509 #+attr_latex: [htpb]
rlm@452	1510 #+begin_listing clojure
rlm@451	1511 #+begin_src clojure
rlm@451	1512 (def worm-action-label
rlm@451	1513 (juxt grand-circle? curled? wiggling?))
rlm@451	1514
rlm@451	1515 (defn compare-empathy-with-baseline [matches]
rlm@451	1516 (let [proprio (atom ())]
rlm@451	1517 (fn
rlm@451	1518 [experiences text]
rlm@451	1519 (let [phi-indices (phi-scan (:proprioception (peek experiences)))]
rlm@451	1520 (swap! proprio (partial cons phi-indices))
rlm@451	1521 (let [exp-thread (longest-thread (take 300 @proprio))
rlm@451	1522 empathy (mapv phi-space (infer-nils exp-thread))
rlm@451	1523 experience-matches-empathy
rlm@451	1524 (= (worm-action-label experiences)
rlm@451	1525 (worm-action-label empathy))]
rlm@451	1526 (println-repl experience-matches-empathy)
rlm@451	1527 (swap! matches #(conj % experience-matches-empathy)))))))
rlm@451	1528
rlm@451	1529 (defn accuracy [v]
rlm@451	1530 (float (/ (count (filter true? v)) (count v))))
rlm@451	1531
rlm@451	1532 (defn test-empathy-accuracy []
rlm@451	1533 (let [res (atom [])]
rlm@451	1534 (run-world
rlm@451	1535 (worm-world :experience-watch
rlm@451	1536 (compare-empathy-with-baseline res)
rlm@451	1537 :worm worm*))
rlm@451	1538 (accuracy @res)))
rlm@451	1539 #+end_src
rlm@451	1540 #+end_listing
rlm@451	1541
rlm@451	1542 Running =test-empathy-accuracy= using the very short exercise
rlm@451	1543 program defined in listing \ref{generate-phi-space}, and then doing
rlm@451	1544 a similar pattern of activity manually yeilds an accuracy of around
rlm@451	1545 73%. This is based on very limited worm experience. By training the
rlm@451	1546 worm for longer, the accuracy dramatically improves.
rlm@451	1547
rlm@451	1548 #+caption: Program to generate \Phi-space using manual training.
rlm@451	1549 #+name: manual-phi-space
rlm@452	1550 #+attr_latex: [htpb]
rlm@451	1551 #+begin_listing clojure
rlm@451	1552 #+begin_src clojure
rlm@451	1553 (defn init-interactive []
rlm@451	1554 (def phi-space
rlm@451	1555 (let [experiences (atom [])]
rlm@451	1556 (run-world
rlm@451	1557 (apply-map
rlm@451	1558 worm-world
rlm@451	1559 (merge
rlm@451	1560 (worm-world-defaults)
rlm@451	1561 {:experiences experiences})))
rlm@451	1562 @experiences))
rlm@451	1563 (def phi-scan (gen-phi-scan phi-space)))
rlm@451	1564 #+end_src
rlm@451	1565 #+end_listing
rlm@451	1566
rlm@451	1567 After about 1 minute of manual training, I was able to achieve 95%
rlm@451	1568 accuracy on manual testing of the worm using =init-interactive= and
rlm@452	1569 =test-empathy-accuracy=. The majority of errors are near the
rlm@452	1570 boundaries of transitioning from one type of action to another.
rlm@452	1571 During these transitions the exact label for the action is more open
rlm@452	1572 to interpretation, and dissaggrement between empathy and experience
rlm@452	1573 is more excusable.
rlm@450	1574
rlm@449	1575 ** Digression: bootstrapping touch using free exploration
rlm@449	1576
rlm@452	1577 In the previous section I showed how to compute actions in terms of
rlm@452	1578 body-centered predicates which relied averate touch activation of
rlm@452	1579 pre-defined regions of the worm's skin. What if, instead of recieving
rlm@452	1580 touch pre-grouped into the six faces of each worm segment, the true
rlm@452	1581 topology of the worm's skin was unknown? This is more similiar to how
rlm@452	1582 a nerve fiber bundle might be arranged. While two fibers that are
rlm@452	1583 close in a nerve bundle /might/ correspond to two touch sensors that
rlm@452	1584 are close together on the skin, the process of taking a complicated
rlm@452	1585 surface and forcing it into essentially a circle requires some cuts
rlm@452	1586 and rerragenments.
rlm@452	1587
rlm@452	1588 In this section I show how to automatically learn the skin-topology of
rlm@452	1589 a worm segment by free exploration. As the worm rolls around on the
rlm@452	1590 floor, large sections of its surface get activated. If the worm has
rlm@452	1591 stopped moving, then whatever region of skin that is touching the
rlm@452	1592 floor is probably an important region, and should be recorded.
rlm@452	1593
rlm@452	1594 #+caption: Program to detect whether the worm is in a resting state
rlm@452	1595 #+caption: with one face touching the floor.
rlm@452	1596 #+name: pure-touch
rlm@452	1597 #+begin_listing clojure
rlm@452	1598 #+begin_src clojure
rlm@452	1599 (def full-contact [(float 0.0) (float 0.1)])
rlm@452	1600
rlm@452	1601 (defn pure-touch?
rlm@452	1602 "This is worm specific code to determine if a large region of touch
rlm@452	1603 sensors is either all on or all off."
rlm@452	1604 [[coords touch :as touch-data]]
rlm@452	1605 (= (set (map first touch)) (set full-contact)))
rlm@452	1606 #+end_src
rlm@452	1607 #+end_listing
rlm@452	1608
rlm@452	1609 After collecting these important regions, there will many nearly
rlm@452	1610 similiar touch regions. While for some purposes the subtle
rlm@452	1611 differences between these regions will be important, for my
rlm@452	1612 purposes I colapse them into mostly non-overlapping sets using
rlm@452	1613 =remove-similiar= in listing \ref{remove-similiar}
rlm@452	1614
rlm@452	1615 #+caption: Program to take a lits of set of points and ``collapse them''
rlm@452	1616 #+caption: so that the remaining sets in the list are siginificantly
rlm@452	1617 #+caption: different from each other. Prefer smaller sets to larger ones.
rlm@452	1618 #+name: remove-similiar
rlm@452	1619 #+begin_listing clojure
rlm@452	1620 #+begin_src clojure
rlm@452	1621 (defn remove-similar
rlm@452	1622 [coll]
rlm@452	1623 (loop [result () coll (sort-by (comp - count) coll)]
rlm@452	1624 (if (empty? coll) result
rlm@452	1625 (let [[x & xs] coll
rlm@452	1626 c (count x)]
rlm@452	1627 (if (some
rlm@452	1628 (fn [other-set]
rlm@452	1629 (let [oc (count other-set)]
rlm@452	1630 (< (- (count (union other-set x)) c) (* oc 0.1))))
rlm@452	1631 xs)
rlm@452	1632 (recur result xs)
rlm@452	1633 (recur (cons x result) xs))))))
rlm@452	1634 #+end_src
rlm@452	1635 #+end_listing
rlm@452	1636
rlm@452	1637 Actually running this simulation is easy given =CORTEX='s facilities.
rlm@452	1638
rlm@452	1639 #+caption: Collect experiences while the worm moves around. Filter the touch
rlm@452	1640 #+caption: sensations by stable ones, collapse similiar ones together,
rlm@452	1641 #+caption: and report the regions learned.
rlm@452	1642 #+name: learn-touch
rlm@452	1643 #+begin_listing clojure
rlm@452	1644 #+begin_src clojure
rlm@452	1645 (defn learn-touch-regions []
rlm@452	1646 (let [experiences (atom [])
rlm@452	1647 world (apply-map
rlm@452	1648 worm-world
rlm@452	1649 (assoc (worm-segment-defaults)
rlm@452	1650 :experiences experiences))]
rlm@452	1651 (run-world world)
rlm@452	1652 (->>
rlm@452	1653 @experiences
rlm@452	1654 (drop 175)
rlm@452	1655 ;; access the single segment's touch data
rlm@452	1656 (map (comp first :touch))
rlm@452	1657 ;; only deal with "pure" touch data to determine surfaces
rlm@452	1658 (filter pure-touch?)
rlm@452	1659 ;; associate coordinates with touch values
rlm@452	1660 (map (partial apply zipmap))
rlm@452	1661 ;; select those regions where contact is being made
rlm@452	1662 (map (partial group-by second))
rlm@452	1663 (map #(get % full-contact))
rlm@452	1664 (map (partial map first))
rlm@452	1665 ;; remove redundant/subset regions
rlm@452	1666 (map set)
rlm@452	1667 remove-similar)))
rlm@452	1668
rlm@452	1669 (defn learn-and-view-touch-regions []
rlm@452	1670 (map view-touch-region
rlm@452	1671 (learn-touch-regions)))
rlm@452	1672 #+end_src
rlm@452	1673 #+end_listing
rlm@452	1674
rlm@452	1675 The only thing remining to define is the particular motion the worm
rlm@452	1676 must take. I accomplish this with a simple motor control program.
rlm@452	1677
rlm@452	1678 #+caption: Motor control program for making the worm roll on the ground.
rlm@452	1679 #+caption: This could also be replaced with random motion.
rlm@452	1680 #+name: worm-roll
rlm@452	1681 #+begin_listing clojure
rlm@452	1682 #+begin_src clojure
rlm@452	1683 (defn touch-kinesthetics []
rlm@452	1684 [[170 :lift-1 40]
rlm@452	1685 [190 :lift-1 19]
rlm@452	1686 [206 :lift-1 0]
rlm@452	1687
rlm@452	1688 [400 :lift-2 40]
rlm@452	1689 [410 :lift-2 0]
rlm@452	1690
rlm@452	1691 [570 :lift-2 40]
rlm@452	1692 [590 :lift-2 21]
rlm@452	1693 [606 :lift-2 0]
rlm@452	1694
rlm@452	1695 [800 :lift-1 30]
rlm@452	1696 [809 :lift-1 0]
rlm@452	1697
rlm@452	1698 [900 :roll-2 40]
rlm@452	1699 [905 :roll-2 20]
rlm@452	1700 [910 :roll-2 0]
rlm@452	1701
rlm@452	1702 [1000 :roll-2 40]
rlm@452	1703 [1005 :roll-2 20]
rlm@452	1704 [1010 :roll-2 0]
rlm@452	1705
rlm@452	1706 [1100 :roll-2 40]
rlm@452	1707 [1105 :roll-2 20]
rlm@452	1708 [1110 :roll-2 0]
rlm@452	1709 ])
rlm@452	1710 #+end_src
rlm@452	1711 #+end_listing
rlm@452	1712
rlm@452	1713
rlm@452	1714 #+caption: The small worm rolls around on the floor, driven
rlm@452	1715 #+caption: by the motor control program in listing \ref{worm-roll}.
rlm@452	1716 #+name: worm-roll
rlm@452	1717 #+ATTR_LaTeX: :width 12cm
rlm@452	1718 [[./images/worm-roll.png]]
rlm@452	1719
rlm@452	1720
rlm@452	1721 #+caption: After completing its adventures, the worm now knows
rlm@452	1722 #+caption: how its touch sensors are arranged along its skin. These
rlm@452	1723 #+caption: are the regions that were deemed important by
rlm@452	1724 #+caption: =learn-touch-regions=. Note that the worm has discovered
rlm@452	1725 #+caption: that it has six sides.
rlm@452	1726 #+name: worm-touch-map
rlm@452	1727 #+ATTR_LaTeX: :width 12cm
rlm@452	1728 [[./images/touch-learn.png]]
rlm@452	1729
rlm@452	1730 While simple, =learn-touch-regions= exploits regularities in both
rlm@452	1731 the worm's physiology and the worm's environment to correctly
rlm@452	1732 deduce that the worm has six sides. Note that =learn-touch-regions=
rlm@452	1733 would work just as well even if the worm's touch sense data were
rlm@452	1734 completely scrambled. The cross shape is just for convienence. This
rlm@452	1735 example justifies the use of pre-defined touch regions in =EMPATH=.
rlm@452	1736
rlm@465	1737 * COMMENT Contributions
rlm@454	1738
rlm@461	1739 In this thesis you have seen the =CORTEX= system, a complete
rlm@461	1740 environment for creating simulated creatures. You have seen how to
rlm@461	1741 implement five senses including touch, proprioception, hearing,
rlm@461	1742 vision, and muscle tension. You have seen how to create new creatues
rlm@461	1743 using blender, a 3D modeling tool. I hope that =CORTEX= will be
rlm@461	1744 useful in further research projects. To this end I have included the
rlm@461	1745 full source to =CORTEX= along with a large suite of tests and
rlm@461	1746 examples. I have also created a user guide for =CORTEX= which is
rlm@461	1747 inculded in an appendix to this thesis.
rlm@447	1748
rlm@461	1749 You have also seen how I used =CORTEX= as a platform to attach the
rlm@461	1750 /action recognition/ problem, which is the problem of recognizing
rlm@461	1751 actions in video. You saw a simple system called =EMPATH= which
rlm@461	1752 ientifies actions by first describing actions in a body-centerd,
rlm@461	1753 rich sense language, then infering a full range of sensory
rlm@461	1754 experience from limited data using previous experience gained from
rlm@461	1755 free play.
rlm@447	1756
rlm@461	1757 As a minor digression, you also saw how I used =CORTEX= to enable a
rlm@461	1758 tiny worm to discover the topology of its skin simply by rolling on
rlm@461	1759 the ground.
rlm@461	1760
rlm@461	1761 In conclusion, the main contributions of this thesis are:
rlm@461	1762
rlm@461	1763 - =CORTEX=, a system for creating simulated creatures with rich
rlm@461	1764 senses.
rlm@461	1765 - =EMPATH=, a program for recognizing actions by imagining sensory
rlm@461	1766 experience.
rlm@447	1767
rlm@447	1768 # An anatomical joke:
rlm@447	1769 # - Training
rlm@447	1770 # - Skeletal imitation
rlm@447	1771 # - Sensory fleshing-out
rlm@447	1772 # - Classification

Mercurial > cortex

annotate thesis/cortex.org @ 469:ae10f35022ba