Mercurial > cortex
view thesis/cortex.org @ 513:4c4d45f6f30b
accept/reject changes
author | Robert McIntyre <rlm@mit.edu> |
---|---|
date | Sun, 30 Mar 2014 10:41:18 -0400 |
parents | 07c3feb32df3 |
children | 447c3c8405a2 |
line wrap: on
line source
1 #+title: =CORTEX=2 #+author: Robert McIntyre3 #+email: rlm@mit.edu4 #+description: Using embodied AI to facilitate Artificial Imagination.5 #+keywords: AI, clojure, embodiment6 #+LaTeX_CLASS_OPTIONS: [nofloat]8 * COMMENT templates9 #+caption:10 #+caption:11 #+caption:12 #+caption:13 #+name: name14 #+begin_listing clojure15 #+BEGIN_SRC clojure16 #+END_SRC17 #+end_listing19 #+caption:20 #+caption:21 #+caption:22 #+name: name23 #+ATTR_LaTeX: :width 10cm24 [[./images/aurellem-gray.png]]26 #+caption:27 #+caption:28 #+caption:29 #+caption:30 #+name: name31 #+begin_listing clojure32 #+BEGIN_SRC clojure33 #+END_SRC34 #+end_listing36 #+caption:37 #+caption:38 #+caption:39 #+name: name40 #+ATTR_LaTeX: :width 10cm41 [[./images/aurellem-gray.png]]44 * Empathy \& Embodiment: problem solving strategies46 ** The problem: recognizing actions in video is extremely difficult47 # developing / requires useful representations49 Examine the following collection of images. As you, and indeed very50 young children, can easily determine, each one is a picture of51 someone drinking.53 # dxh: cat, cup, drinking fountain, rain, straw, coconut54 #+caption: A cat drinking some water. Identifying this action is55 #+caption: beyond the capabilities of existing computer vision systems.56 #+ATTR_LaTeX: :width 7cm57 [[./images/cat-drinking.jpg]]59 Nevertheless, it is beyond the state of the art for a computer60 vision program to describe what's happening in each of these61 images, or what's common to them. Part of the problem is that many62 computer vision systems focus on pixel-level details or probability63 distributions of pixels, with little focus on [...]66 In fact, the contents of scene may have much less to do with pixel67 probabilities than with recognizing various affordances: things you68 can move, objects you can grasp, spaces that can be filled69 (Gibson). For example, what processes might enable you to see the70 chair in figure \ref{hidden-chair}?71 # Or suppose that you are building a program that recognizes chairs.72 # How could you ``see'' the chair ?74 # dxh: blur chair75 #+caption: The chair in this image is quite obvious to humans, but I76 #+caption: doubt that any modern computer vision program can find it.77 #+name: hidden-chair78 #+ATTR_LaTeX: :width 10cm79 [[./images/fat-person-sitting-at-desk.jpg]]85 Finally, how is it that you can easily tell the difference between86 how the girls /muscles/ are working in figure \ref{girl}?88 #+caption: The mysterious ``common sense'' appears here as you are able89 #+caption: to discern the difference in how the girl's arm muscles90 #+caption: are activated between the two images.91 #+name: girl92 #+ATTR_LaTeX: :width 7cm93 [[./images/wall-push.png]]98 Each of these examples tells us something about what might be going99 on in our minds as we easily solve these recognition problems.101 The hidden chair shows us that we are strongly triggered by cues102 relating to the position of human bodies, and that we can determine103 the overall physical configuration of a human body even if much of104 that body is occluded.106 The picture of the girl pushing against the wall tells us that we107 have common sense knowledge about the kinetics of our own bodies.108 We know well how our muscles would have to work to maintain us in109 most positions, and we can easily project this self-knowledge to110 imagined positions triggered by images of the human body.112 ** A step forward: the sensorimotor-centered approach113 # ** =EMPATH= recognizes what creatures are doing114 # neatly solves recognition problems115 In this thesis, I explore the idea that our knowledge of our own116 bodies enables us to recognize the actions of others.118 First, I built a system for constructing virtual creatures with119 physiologically plausible sensorimotor systems and detailed120 environments. The result is =CORTEX=, which is described in section121 \ref{sec-2}. (=CORTEX= was built to be flexible and useful to other122 AI researchers; it is provided in full with detailed instructions123 on the web [here].)125 Next, I wrote routines which enabled a simple worm-like creature to126 infer the actions of a second worm-like creature, using only its127 own prior sensorimotor experiences and knowledge of the second128 worm's joint positions. This program, =EMPATH=, is described in129 section \ref{sec-3}, and the key results of this experiment are130 summarized below.132 #+caption: From only \emph{proprioceptive} data, =EMPATH= was able to infer133 #+caption: the complete sensory experience and classify these four poses.134 #+caption: The last image is a composite, depicting the intermediate stages of \emph{wriggling}.135 #+name: worm-recognition-intro-2136 #+ATTR_LaTeX: :width 15cm137 [[./images/empathy-1.png]]139 # =CORTEX= provides a language for describing the sensorimotor140 # experiences of various creatures.142 # Next, I developed an experiment to test the power of =CORTEX='s143 # sensorimotor-centered language for solving recognition problems. As144 # a proof of concept, I wrote routines which enabled a simple145 # worm-like creature to infer the actions of a second worm-like146 # creature, using only its own previous sensorimotor experiences and147 # knowledge of the second worm's joints (figure148 # \ref{worm-recognition-intro-2}). The result of this proof of149 # concept was the program =EMPATH=, described in section150 # \ref{sec-3}. The key results of this152 # Using only first-person sensorimotor experiences and third-person153 # proprioceptive data,155 *** Key results156 - After one-shot supervised training, =EMPATH= was able recognize a157 wide variety of static poses and dynamic actions---ranging from158 curling in a circle to wriggling with a particular frequency ---159 with 95\% accuracy.160 - These results were completely independent of viewing angle161 because the underlying body-centered language fundamentally is162 independent; once an action is learned, it can be recognized163 equally well from any viewing angle.164 - =EMPATH= is surprisingly short; the sensorimotor-centered165 language provided by =CORTEX= resulted in extremely economical166 recognition routines --- about 0000 lines in all --- suggesting167 that such representations are very powerful, and often168 indispensible for the types of recognition tasks considered here.169 - Although for expediency's sake, I relied on direct knowledge of170 joint positions in this proof of concept, it would be171 straightforward to extend =EMPATH= so that it (more172 realistically) infers joint positions from its visual data.174 # because the underlying language is fundamentally orientation-independent176 # recognize the actions of a worm with 95\% accuracy. The177 # recognition tasks182 [Talk about these results and what you find promising about them]184 ** Roadmap185 [I'm going to explain how =CORTEX= works, then break down how186 =EMPATH= does its thing. Because the details reveal such-and-such187 about the approach.]189 # The success of this simple proof-of-concept offers a tantalizing192 # explore the idea193 # The key contribution of this thesis is the idea that body-centered194 # representations (which express197 # the198 # body-centered approach --- in which I try to determine what's199 # happening in a scene by bringing it into registration with my own200 # bodily experiences --- are indispensible for recognizing what201 # creatures are doing in a scene.203 * COMMENT204 # body-centered language206 In this thesis, I'll describe =EMPATH=, which solves a certain207 class of recognition problems209 The key idea is to use self-centered (or first-person) language.211 I have built a system that can express the types of recognition212 problems in a form amenable to computation. It is split into213 four parts:215 - Free/Guided Play :: The creature moves around and experiences the216 world through its unique perspective. Many otherwise217 complicated actions are easily described in the language of a218 full suite of body-centered, rich senses. For example,219 drinking is the feeling of water sliding down your throat, and220 cooling your insides. It's often accompanied by bringing your221 hand close to your face, or bringing your face close to water.222 Sitting down is the feeling of bending your knees, activating223 your quadriceps, then feeling a surface with your bottom and224 relaxing your legs. These body-centered action descriptions225 can be either learned or hard coded.226 - Posture Imitation :: When trying to interpret a video or image,227 the creature takes a model of itself and aligns it with228 whatever it sees. This alignment can even cross species, as229 when humans try to align themselves with things like ponies,230 dogs, or other humans with a different body type.231 - Empathy :: The alignment triggers associations with232 sensory data from prior experiences. For example, the233 alignment itself easily maps to proprioceptive data. Any234 sounds or obvious skin contact in the video can to a lesser235 extent trigger previous experience. Segments of previous236 experiences are stitched together to form a coherent and237 complete sensory portrait of the scene.238 - Recognition :: With the scene described in terms of first239 person sensory events, the creature can now run its240 action-identification programs on this synthesized sensory241 data, just as it would if it were actually experiencing the242 scene first-hand. If previous experience has been accurately243 retrieved, and if it is analogous enough to the scene, then244 the creature will correctly identify the action in the scene.246 For example, I think humans are able to label the cat video as247 ``drinking'' because they imagine /themselves/ as the cat, and248 imagine putting their face up against a stream of water and249 sticking out their tongue. In that imagined world, they can feel250 the cool water hitting their tongue, and feel the water entering251 their body, and are able to recognize that /feeling/ as drinking.252 So, the label of the action is not really in the pixels of the253 image, but is found clearly in a simulation inspired by those254 pixels. An imaginative system, having been trained on drinking and255 non-drinking examples and learning that the most important256 component of drinking is the feeling of water sliding down one's257 throat, would analyze a video of a cat drinking in the following258 manner:260 1. Create a physical model of the video by putting a ``fuzzy''261 model of its own body in place of the cat. Possibly also create262 a simulation of the stream of water.264 2. Play out this simulated scene and generate imagined sensory265 experience. This will include relevant muscle contractions, a266 close up view of the stream from the cat's perspective, and most267 importantly, the imagined feeling of water entering the268 mouth. The imagined sensory experience can come from a269 simulation of the event, but can also be pattern-matched from270 previous, similar embodied experience.272 3. The action is now easily identified as drinking by the sense of273 taste alone. The other senses (such as the tongue moving in and274 out) help to give plausibility to the simulated action. Note that275 the sense of vision, while critical in creating the simulation,276 is not critical for identifying the action from the simulation.278 For the chair examples, the process is even easier:280 1. Align a model of your body to the person in the image.282 2. Generate proprioceptive sensory data from this alignment.284 3. Use the imagined proprioceptive data as a key to lookup related285 sensory experience associated with that particular proproceptive286 feeling.288 4. Retrieve the feeling of your bottom resting on a surface, your289 knees bent, and your leg muscles relaxed.291 5. This sensory information is consistent with the =sitting?=292 sensory predicate, so you (and the entity in the image) must be293 sitting.295 6. There must be a chair-like object since you are sitting.297 Empathy offers yet another alternative to the age-old AI298 representation question: ``What is a chair?'' --- A chair is the299 feeling of sitting.301 My program, =EMPATH= uses this empathic problem solving technique302 to interpret the actions of a simple, worm-like creature.304 #+caption: The worm performs many actions during free play such as305 #+caption: curling, wiggling, and resting.306 #+name: worm-intro307 #+ATTR_LaTeX: :width 15cm308 [[./images/worm-intro-white.png]]310 #+caption: =EMPATH= recognized and classified each of these311 #+caption: poses by inferring the complete sensory experience312 #+caption: from proprioceptive data.313 #+name: worm-recognition-intro314 #+ATTR_LaTeX: :width 15cm315 [[./images/worm-poses.png]]317 One powerful advantage of empathic problem solving is that it318 factors the action recognition problem into two easier problems. To319 use empathy, you need an /aligner/, which takes the video and a320 model of your body, and aligns the model with the video. Then, you321 need a /recognizer/, which uses the aligned model to interpret the322 action. The power in this method lies in the fact that you describe323 all actions form a body-centered viewpoint. You are less tied to324 the particulars of any visual representation of the actions. If you325 teach the system what ``running'' is, and you have a good enough326 aligner, the system will from then on be able to recognize running327 from any point of view, even strange points of view like above or328 underneath the runner. This is in contrast to action recognition329 schemes that try to identify actions using a non-embodied approach.330 If these systems learn about running as viewed from the side, they331 will not automatically be able to recognize running from any other332 viewpoint.334 Another powerful advantage is that using the language of multiple335 body-centered rich senses to describe body-centerd actions offers a336 massive boost in descriptive capability. Consider how difficult it337 would be to compose a set of HOG filters to describe the action of338 a simple worm-creature ``curling'' so that its head touches its339 tail, and then behold the simplicity of describing thus action in a340 language designed for the task (listing \ref{grand-circle-intro}):342 #+caption: Body-centerd actions are best expressed in a body-centered343 #+caption: language. This code detects when the worm has curled into a344 #+caption: full circle. Imagine how you would replicate this functionality345 #+caption: using low-level pixel features such as HOG filters!346 #+name: grand-circle-intro347 #+begin_listing clojure348 #+begin_src clojure349 (defn grand-circle?350 "Does the worm form a majestic circle (one end touching the other)?"351 [experiences]352 (and (curled? experiences)353 (let [worm-touch (:touch (peek experiences))354 tail-touch (worm-touch 0)355 head-touch (worm-touch 4)]356 (and (< 0.2 (contact worm-segment-bottom-tip tail-touch))357 (< 0.2 (contact worm-segment-top-tip head-touch))))))358 #+end_src359 #+end_listing361 ** =CORTEX= is a toolkit for building sensate creatures363 I built =CORTEX= to be a general AI research platform for doing364 experiments involving multiple rich senses and a wide variety and365 number of creatures. I intend it to be useful as a library for many366 more projects than just this thesis. =CORTEX= was necessary to meet367 a need among AI researchers at CSAIL and beyond, which is that368 people often will invent neat ideas that are best expressed in the369 language of creatures and senses, but in order to explore those370 ideas they must first build a platform in which they can create371 simulated creatures with rich senses! There are many ideas that372 would be simple to execute (such as =EMPATH=), but attached to them373 is the multi-month effort to make a good creature simulator. Often,374 that initial investment of time proves to be too much, and the375 project must make do with a lesser environment.377 =CORTEX= is well suited as an environment for embodied AI research378 for three reasons:380 - You can create new creatures using Blender, a popular 3D modeling381 program. Each sense can be specified using special blender nodes382 with biologically inspired paramaters. You need not write any383 code to create a creature, and can use a wide library of384 pre-existing blender models as a base for your own creatures.386 - =CORTEX= implements a wide variety of senses: touch,387 proprioception, vision, hearing, and muscle tension. Complicated388 senses like touch, and vision involve multiple sensory elements389 embedded in a 2D surface. You have complete control over the390 distribution of these sensor elements through the use of simple391 png image files. In particular, =CORTEX= implements more392 comprehensive hearing than any other creature simulation system393 available.395 - =CORTEX= supports any number of creatures and any number of396 senses. Time in =CORTEX= dialates so that the simulated creatures397 always precieve a perfectly smooth flow of time, regardless of398 the actual computational load.400 =CORTEX= is built on top of =jMonkeyEngine3=, which is a video game401 engine designed to create cross-platform 3D desktop games. =CORTEX=402 is mainly written in clojure, a dialect of =LISP= that runs on the403 java virtual machine (JVM). The API for creating and simulating404 creatures and senses is entirely expressed in clojure, though many405 senses are implemented at the layer of jMonkeyEngine or below. For406 example, for the sense of hearing I use a layer of clojure code on407 top of a layer of java JNI bindings that drive a layer of =C++=408 code which implements a modified version of =OpenAL= to support409 multiple listeners. =CORTEX= is the only simulation environment410 that I know of that can support multiple entities that can each411 hear the world from their own perspective. Other senses also412 require a small layer of Java code. =CORTEX= also uses =bullet=, a413 physics simulator written in =C=.415 #+caption: Here is the worm from above modeled in Blender, a free416 #+caption: 3D-modeling program. Senses and joints are described417 #+caption: using special nodes in Blender.418 #+name: worm-recognition-intro419 #+ATTR_LaTeX: :width 12cm420 [[./images/blender-worm.png]]422 Here are some thing I anticipate that =CORTEX= might be used for:424 - exploring new ideas about sensory integration425 - distributed communication among swarm creatures426 - self-learning using free exploration,427 - evolutionary algorithms involving creature construction428 - exploration of exoitic senses and effectors that are not possible429 in the real world (such as telekenisis or a semantic sense)430 - imagination using subworlds432 During one test with =CORTEX=, I created 3,000 creatures each with433 their own independent senses and ran them all at only 1/80 real434 time. In another test, I created a detailed model of my own hand,435 equipped with a realistic distribution of touch (more sensitive at436 the fingertips), as well as eyes and ears, and it ran at around 1/4437 real time.439 #+BEGIN_LaTeX440 \begin{sidewaysfigure}441 \includegraphics[width=9.5in]{images/full-hand.png}442 \caption{443 I modeled my own right hand in Blender and rigged it with all the444 senses that {\tt CORTEX} supports. My simulated hand has a445 biologically inspired distribution of touch sensors. The senses are446 displayed on the right, and the simulation is displayed on the447 left. Notice that my hand is curling its fingers, that it can see448 its own finger from the eye in its palm, and that it can feel its449 own thumb touching its palm.}450 \end{sidewaysfigure}451 #+END_LaTeX453 ** Road map455 By the end of this thesis, you will have seen a novel approach to456 interpreting video using embodiment and empathy. You will have also457 seen one way to efficiently implement empathy for embodied458 creatures. Finally, you will become familiar with =CORTEX=, a system459 for designing and simulating creatures with rich senses, which you460 may choose to use in your own research.462 This is the core vision of my thesis: That one of the important ways463 in which we understand others is by imagining ourselves in their464 position and emphatically feeling experiences relative to our own465 bodies. By understanding events in terms of our own previous466 corporeal experience, we greatly constrain the possibilities of what467 would otherwise be an unwieldy exponential search. This extra468 constraint can be the difference between easily understanding what469 is happening in a video and being completely lost in a sea of470 incomprehensible color and movement.472 - I built =CORTEX=, a comprehensive platform for embodied AI473 experiments. =CORTEX= supports many features lacking in other474 systems, such proper simulation of hearing. It is easy to create475 new =CORTEX= creatures using Blender, a free 3D modeling program.477 - I built =EMPATH=, which uses =CORTEX= to identify the actions of478 a worm-like creature using a computational model of empathy.481 * Designing =CORTEX=482 In this section, I outline the design decisions that went into483 making =CORTEX=, along with some details about its484 implementation. (A practical guide to getting started with =CORTEX=,485 which skips over the history and implementation details presented486 here, is provided in an appendix \ref{} at the end of this paper.)488 Throughout this project, I intended for =CORTEX= to be flexible and489 extensible enough to be useful for other researchers who want to490 test out ideas of their own. To this end, wherver I have had to make491 archetictural choices about =CORTEX=, I have chosen to give as much492 freedom to the user as possible, so that =CORTEX= may be used for493 things I have not forseen.495 ** Building in simulation versus reality496 The most important archetictural decision of all is the choice to497 use a computer-simulated environemnt in the first place! The world498 is a vast and rich place, and for now simulations are a very poor499 reflection of its complexity. It may be that there is a significant500 qualatative difference between dealing with senses in the real501 world and dealing with pale facilimilies of them in a simulation.502 What are the advantages and disadvantages of a simulation vs.503 reality?505 *** Simulation507 The advantages of virtual reality are that when everything is a508 simulation, experiments in that simulation are absolutely509 reproducible. It's also easier to change the character and world510 to explore new situations and different sensory combinations.512 If the world is to be simulated on a computer, then not only do513 you have to worry about whether the character's senses are rich514 enough to learn from the world, but whether the world itself is515 rendered with enough detail and realism to give enough working516 material to the character's senses. To name just a few517 difficulties facing modern physics simulators: destructibility of518 the environment, simulation of water/other fluids, large areas,519 nonrigid bodies, lots of objects, smoke. I don't know of any520 computer simulation that would allow a character to take a rock521 and grind it into fine dust, then use that dust to make a clay522 sculpture, at least not without spending years calculating the523 interactions of every single small grain of dust. Maybe a524 simulated world with today's limitations doesn't provide enough525 richness for real intelligence to evolve.527 *** Reality529 The other approach for playing with senses is to hook your530 software up to real cameras, microphones, robots, etc., and let it531 loose in the real world. This has the advantage of eliminating532 concerns about simulating the world at the expense of increasing533 the complexity of implementing the senses. Instead of just534 grabbing the current rendered frame for processing, you have to535 use an actual camera with real lenses and interact with photons to536 get an image. It is much harder to change the character, which is537 now partly a physical robot of some sort, since doing so involves538 changing things around in the real world instead of modifying539 lines of code. While the real world is very rich and definitely540 provides enough stimulation for intelligence to develop as541 evidenced by our own existence, it is also uncontrollable in the542 sense that a particular situation cannot be recreated perfectly or543 saved for later use. It is harder to conduct science because it is544 harder to repeat an experiment. The worst thing about using the545 real world instead of a simulation is the matter of time. Instead546 of simulated time you get the constant and unstoppable flow of547 real time. This severely limits the sorts of software you can use548 to program the AI because all sense inputs must be handled in real549 time. Complicated ideas may have to be implemented in hardware or550 may simply be impossible given the current speed of our551 processors. Contrast this with a simulation, in which the flow of552 time in the simulated world can be slowed down to accommodate the553 limitations of the character's programming. In terms of cost,554 doing everything in software is far cheaper than building custom555 real-time hardware. All you need is a laptop and some patience.557 ** Simulated time enables rapid prototyping and complex scenes559 I envision =CORTEX= being used to support rapid prototyping and560 iteration of ideas. Even if I could put together a well constructed561 kit for creating robots, it would still not be enough because of562 the scourge of real-time processing. Anyone who wants to test their563 ideas in the real world must always worry about getting their564 algorithms to run fast enough to process information in real time.565 The need for real time processing only increases if multiple senses566 are involved. In the extreme case, even simple algorithms will have567 to be accelerated by ASIC chips or FPGAs, turning what would568 otherwise be a few lines of code and a 10x speed penality into a569 multi-month ordeal. For this reason, =CORTEX= supports570 /time-dialiation/, which scales back the framerate of the571 simulation in proportion to the amount of processing each frame.572 From the perspective of the creatures inside the simulation, time573 always appears to flow at a constant rate, regardless of how574 complicated the envorimnent becomes or how many creatures are in575 the simulation. The cost is that =CORTEX= can sometimes run slower576 than real time. This can also be an advantage, however ---577 simulations of very simple creatures in =CORTEX= generally run at578 40x on my machine!580 ** All sense organs are two-dimensional surfaces581 # What is a sense?582 If =CORTEX= is to support a wide variety of senses, it would help583 to have a better understanding of what a ``sense'' actually is!584 While vision, touch, and hearing all seem like they are quite585 different things, I was supprised to learn during the course of586 this thesis that they (and all physical senses) can be expressed as587 exactly the same mathematical object due to a dimensional argument!589 Human beings are three-dimensional objects, and the nerves that590 transmit data from our various sense organs to our brain are591 essentially one-dimensional. This leaves up to two dimensions in592 which our sensory information may flow. For example, imagine your593 skin: it is a two-dimensional surface around a three-dimensional594 object (your body). It has discrete touch sensors embedded at595 various points, and the density of these sensors corresponds to the596 sensitivity of that region of skin. Each touch sensor connects to a597 nerve, all of which eventually are bundled together as they travel598 up the spinal cord to the brain. Intersect the spinal nerves with a599 guillotining plane and you will see all of the sensory data of the600 skin revealed in a roughly circular two-dimensional image which is601 the cross section of the spinal cord. Points on this image that are602 close together in this circle represent touch sensors that are603 /probably/ close together on the skin, although there is of course604 some cutting and rearrangement that has to be done to transfer the605 complicated surface of the skin onto a two dimensional image.607 Most human senses consist of many discrete sensors of various608 properties distributed along a surface at various densities. For609 skin, it is Pacinian corpuscles, Meissner's corpuscles, Merkel's610 disks, and Ruffini's endings, which detect pressure and vibration611 of various intensities. For ears, it is the stereocilia distributed612 along the basilar membrane inside the cochlea; each one is613 sensitive to a slightly different frequency of sound. For eyes, it614 is rods and cones distributed along the surface of the retina. In615 each case, we can describe the sense with a surface and a616 distribution of sensors along that surface.618 The neat idea is that every human sense can be effectively619 described in terms of a surface containing embedded sensors. If the620 sense had any more dimensions, then there wouldn't be enough room621 in the spinal chord to transmit the information!623 Therefore, =CORTEX= must support the ability to create objects and624 then be able to ``paint'' points along their surfaces to describe625 each sense.627 Fortunately this idea is already a well known computer graphics628 technique called called /UV-mapping/. The three-dimensional surface629 of a model is cut and smooshed until it fits on a two-dimensional630 image. You paint whatever you want on that image, and when the631 three-dimensional shape is rendered in a game the smooshing and632 cutting is reversed and the image appears on the three-dimensional633 object.635 To make a sense, interpret the UV-image as describing the636 distribution of that senses sensors. To get different types of637 sensors, you can either use a different color for each type of638 sensor, or use multiple UV-maps, each labeled with that sensor639 type. I generally use a white pixel to mean the presence of a640 sensor and a black pixel to mean the absence of a sensor, and use641 one UV-map for each sensor-type within a given sense.643 #+CAPTION: The UV-map for an elongated icososphere. The white644 #+caption: dots each represent a touch sensor. They are dense645 #+caption: in the regions that describe the tip of the finger,646 #+caption: and less dense along the dorsal side of the finger647 #+caption: opposite the tip.648 #+name: finger-UV649 #+ATTR_latex: :width 10cm650 [[./images/finger-UV.png]]652 #+caption: Ventral side of the UV-mapped finger. Notice the653 #+caption: density of touch sensors at the tip.654 #+name: finger-side-view655 #+ATTR_LaTeX: :width 10cm656 [[./images/finger-1.png]]658 ** Video game engines provide ready-made physics and shading660 I did not need to write my own physics simulation code or shader to661 build =CORTEX=. Doing so would lead to a system that is impossible662 for anyone but myself to use anyway. Instead, I use a video game663 engine as a base and modify it to accomodate the additional needs664 of =CORTEX=. Video game engines are an ideal starting point to665 build =CORTEX=, because they are not far from being creature666 building systems themselves.668 First off, general purpose video game engines come with a physics669 engine and lighting / sound system. The physics system provides670 tools that can be co-opted to serve as touch, proprioception, and671 muscles. Since some games support split screen views, a good video672 game engine will allow you to efficiently create multiple cameras673 in the simulated world that can be used as eyes. Video game systems674 offer integrated asset management for things like textures and675 creatures models, providing an avenue for defining creatures. They676 also understand UV-mapping, since this technique is used to apply a677 texture to a model. Finally, because video game engines support a678 large number of users, as long as =CORTEX= doesn't stray too far679 from the base system, other researchers can turn to this community680 for help when doing their research.682 ** =CORTEX= is based on jMonkeyEngine3684 While preparing to build =CORTEX= I studied several video game685 engines to see which would best serve as a base. The top contenders686 were:688 - [[http://www.idsoftware.com][Quake II]]/[[http://www.bytonic.de/html/jake2.html][Jake2]] :: The Quake II engine was designed by ID689 software in 1997. All the source code was released by ID690 software into the Public Domain several years ago, and as a691 result it has been ported to many different languages. This692 engine was famous for its advanced use of realistic shading693 and had decent and fast physics simulation. The main advantage694 of the Quake II engine is its simplicity, but I ultimately695 rejected it because the engine is too tied to the concept of a696 first-person shooter game. One of the problems I had was that697 there does not seem to be any easy way to attach multiple698 cameras to a single character. There are also several physics699 clipping issues that are corrected in a way that only applies700 to the main character and do not apply to arbitrary objects.702 - [[http://source.valvesoftware.com/][Source Engine]] :: The Source Engine evolved from the Quake II703 and Quake I engines and is used by Valve in the Half-Life704 series of games. The physics simulation in the Source Engine705 is quite accurate and probably the best out of all the engines706 I investigated. There is also an extensive community actively707 working with the engine. However, applications that use the708 Source Engine must be written in C++, the code is not open, it709 only runs on Windows, and the tools that come with the SDK to710 handle models and textures are complicated and awkward to use.712 - [[http://jmonkeyengine.com/][jMonkeyEngine3]] :: jMonkeyEngine3 is a new library for creating713 games in Java. It uses OpenGL to render to the screen and uses714 screengraphs to avoid drawing things that do not appear on the715 screen. It has an active community and several games in the716 pipeline. The engine was not built to serve any particular717 game but is instead meant to be used for any 3D game.719 I chose jMonkeyEngine3 because it because it had the most features720 out of all the free projects I looked at, and because I could then721 write my code in clojure, an implementation of =LISP= that runs on722 the JVM.724 ** =CORTEX= uses Blender to create creature models726 For the simple worm-like creatures I will use later on in this727 thesis, I could define a simple API in =CORTEX= that would allow728 one to create boxes, spheres, etc., and leave that API as the sole729 way to create creatures. However, for =CORTEX= to truly be useful730 for other projects, it needs a way to construct complicated731 creatures. If possible, it would be nice to leverage work that has732 already been done by the community of 3D modelers, or at least733 enable people who are talented at moedling but not programming to734 design =CORTEX= creatures.736 Therefore, I use Blender, a free 3D modeling program, as the main737 way to create creatures in =CORTEX=. However, the creatures modeled738 in Blender must also be simple to simulate in jMonkeyEngine3's game739 engine, and must also be easy to rig with =CORTEX='s senses. I740 accomplish this with extensive use of Blender's ``empty nodes.''742 Empty nodes have no mass, physical presence, or appearance, but743 they can hold metadata and have names. I use a tree structure of744 empty nodes to specify senses in the following manner:746 - Create a single top-level empty node whose name is the name of747 the sense.748 - Add empty nodes which each contain meta-data relevant to the749 sense, including a UV-map describing the number/distribution of750 sensors if applicable.751 - Make each empty-node the child of the top-level node.753 #+caption: An example of annoting a creature model with empty754 #+caption: nodes to describe the layout of senses. There are755 #+caption: multiple empty nodes which each describe the position756 #+caption: of muscles, ears, eyes, or joints.757 #+name: sense-nodes758 #+ATTR_LaTeX: :width 10cm759 [[./images/empty-sense-nodes.png]]761 ** Bodies are composed of segments connected by joints763 Blender is a general purpose animation tool, which has been used in764 the past to create high quality movies such as Sintel765 \cite{blender}. Though Blender can model and render even complicated766 things like water, it is crucual to keep models that are meant to767 be simulated as creatures simple. =Bullet=, which =CORTEX= uses768 though jMonkeyEngine3, is a rigid-body physics system. This offers769 a compromise between the expressiveness of a game level and the770 speed at which it can be simulated, and it means that creatures771 should be naturally expressed as rigid components held together by772 joint constraints.774 But humans are more like a squishy bag with wrapped around some775 hard bones which define the overall shape. When we move, our skin776 bends and stretches to accomodate the new positions of our bones.778 One way to make bodies composed of rigid pieces connected by joints779 /seem/ more human-like is to use an /armature/, (or /rigging/)780 system, which defines a overall ``body mesh'' and defines how the781 mesh deforms as a function of the position of each ``bone'' which782 is a standard rigid body. This technique is used extensively to783 model humans and create realistic animations. It is not a good784 technique for physical simulation, however because it creates a lie785 -- the skin is not a physical part of the simulation and does not786 interact with any objects in the world or itself. Objects will pass787 right though the skin until they come in contact with the788 underlying bone, which is a physical object. Whithout simulating789 the skin, the sense of touch has little meaning, and the creature's790 own vision will lie to it about the true extent of its body.791 Simulating the skin as a physical object requires some way to792 continuously update the physical model of the skin along with the793 movement of the bones, which is unacceptably slow compared to rigid794 body simulation.796 Therefore, instead of using the human-like ``deformable bag of797 bones'' approach, I decided to base my body plans on multiple solid798 objects that are connected by joints, inspired by the robot =EVE=799 from the movie WALL-E.801 #+caption: =EVE= from the movie WALL-E. This body plan turns802 #+caption: out to be much better suited to my purposes than a more803 #+caption: human-like one.804 #+ATTR_LaTeX: :width 10cm805 [[./images/Eve.jpg]]807 =EVE='s body is composed of several rigid components that are held808 together by invisible joint constraints. This is what I mean by809 ``eve-like''. The main reason that I use eve-style bodies is for810 efficiency, and so that there will be correspondence between the811 AI's semses and the physical presence of its body. Each individual812 section is simulated by a separate rigid body that corresponds813 exactly with its visual representation and does not change.814 Sections are connected by invisible joints that are well supported815 in jMonkeyEngine3. Bullet, the physics backend for jMonkeyEngine3,816 can efficiently simulate hundreds of rigid bodies connected by817 joints. Just because sections are rigid does not mean they have to818 stay as one piece forever; they can be dynamically replaced with819 multiple sections to simulate splitting in two. This could be used820 to simulate retractable claws or =EVE='s hands, which are able to821 coalesce into one object in the movie.823 *** Solidifying/Connecting a body825 =CORTEX= creates a creature in two steps: first, it traverses the826 nodes in the blender file and creates physical representations for827 any of them that have mass defined in their blender meta-data.829 #+caption: Program for iterating through the nodes in a blender file830 #+caption: and generating physical jMonkeyEngine3 objects with mass831 #+caption: and a matching physics shape.832 #+name: name833 #+begin_listing clojure834 #+begin_src clojure835 (defn physical!836 "Iterate through the nodes in creature and make them real physical837 objects in the simulation."838 [#^Node creature]839 (dorun840 (map841 (fn [geom]842 (let [physics-control843 (RigidBodyControl.844 (HullCollisionShape.845 (.getMesh geom))846 (if-let [mass (meta-data geom "mass")]847 (float mass) (float 1)))]848 (.addControl geom physics-control)))849 (filter #(isa? (class %) Geometry )850 (node-seq creature)))))851 #+end_src852 #+end_listing854 The next step to making a proper body is to connect those pieces855 together with joints. jMonkeyEngine has a large array of joints856 available via =bullet=, such as Point2Point, Cone, Hinge, and a857 generic Six Degree of Freedom joint, with or without spring858 restitution.860 Joints are treated a lot like proper senses, in that there is a861 top-level empty node named ``joints'' whose children each862 represent a joint.864 #+caption: View of the hand model in Blender showing the main ``joints''865 #+caption: node (highlighted in yellow) and its children which each866 #+caption: represent a joint in the hand. Each joint node has metadata867 #+caption: specifying what sort of joint it is.868 #+name: blender-hand869 #+ATTR_LaTeX: :width 10cm870 [[./images/hand-screenshot1.png]]873 =CORTEX='s procedure for binding the creature together with joints874 is as follows:876 - Find the children of the ``joints'' node.877 - Determine the two spatials the joint is meant to connect.878 - Create the joint based on the meta-data of the empty node.880 The higher order function =sense-nodes= from =cortex.sense=881 simplifies finding the joints based on their parent ``joints''882 node.884 #+caption: Retrieving the children empty nodes from a single885 #+caption: named empty node is a common pattern in =CORTEX=886 #+caption: further instances of this technique for the senses887 #+caption: will be omitted888 #+name: get-empty-nodes889 #+begin_listing clojure890 #+begin_src clojure891 (defn sense-nodes892 "For some senses there is a special empty blender node whose893 children are considered markers for an instance of that sense. This894 function generates functions to find those children, given the name895 of the special parent node."896 [parent-name]897 (fn [#^Node creature]898 (if-let [sense-node (.getChild creature parent-name)]899 (seq (.getChildren sense-node)) [])))901 (def902 ^{:doc "Return the children of the creature's \"joints\" node."903 :arglists '([creature])}904 joints905 (sense-nodes "joints"))906 #+end_src907 #+end_listing909 To find a joint's targets, =CORTEX= creates a small cube, centered910 around the empty-node, and grows the cube exponentially until it911 intersects two physical objects. The objects are ordered according912 to the joint's rotation, with the first one being the object that913 has more negative coordinates in the joint's reference frame.914 Since the objects must be physical, the empty-node itself escapes915 detection. Because the objects must be physical, =joint-targets=916 must be called /after/ =physical!= is called.918 #+caption: Program to find the targets of a joint node by919 #+caption: exponentiallly growth of a search cube.920 #+name: joint-targets921 #+begin_listing clojure922 #+begin_src clojure923 (defn joint-targets924 "Return the two closest two objects to the joint object, ordered925 from bottom to top according to the joint's rotation."926 [#^Node parts #^Node joint]927 (loop [radius (float 0.01)]928 (let [results (CollisionResults.)]929 (.collideWith930 parts931 (BoundingBox. (.getWorldTranslation joint)932 radius radius radius) results)933 (let [targets934 (distinct935 (map #(.getGeometry %) results))]936 (if (>= (count targets) 2)937 (sort-by938 #(let [joint-ref-frame-position939 (jme-to-blender940 (.mult941 (.inverse (.getWorldRotation joint))942 (.subtract (.getWorldTranslation %)943 (.getWorldTranslation joint))))]944 (.dot (Vector3f. 1 1 1) joint-ref-frame-position))945 (take 2 targets))946 (recur (float (* radius 2))))))))947 #+end_src948 #+end_listing950 Once =CORTEX= finds all joints and targets, it creates them using951 a dispatch on the metadata of each joint node.953 #+caption: Program to dispatch on blender metadata and create joints954 #+caption: sutiable for physical simulation.955 #+name: joint-dispatch956 #+begin_listing clojure957 #+begin_src clojure958 (defmulti joint-dispatch959 "Translate blender pseudo-joints into real JME joints."960 (fn [constraints & _]961 (:type constraints)))963 (defmethod joint-dispatch :point964 [constraints control-a control-b pivot-a pivot-b rotation]965 (doto (SixDofJoint. control-a control-b pivot-a pivot-b false)966 (.setLinearLowerLimit Vector3f/ZERO)967 (.setLinearUpperLimit Vector3f/ZERO)))969 (defmethod joint-dispatch :hinge970 [constraints control-a control-b pivot-a pivot-b rotation]971 (let [axis (if-let [axis (:axis constraints)] axis Vector3f/UNIT_X)972 [limit-1 limit-2] (:limit constraints)973 hinge-axis (.mult rotation (blender-to-jme axis))]974 (doto (HingeJoint. control-a control-b pivot-a pivot-b975 hinge-axis hinge-axis)976 (.setLimit limit-1 limit-2))))978 (defmethod joint-dispatch :cone979 [constraints control-a control-b pivot-a pivot-b rotation]980 (let [limit-xz (:limit-xz constraints)981 limit-xy (:limit-xy constraints)982 twist (:twist constraints)]983 (doto (ConeJoint. control-a control-b pivot-a pivot-b984 rotation rotation)985 (.setLimit (float limit-xz) (float limit-xy)986 (float twist)))))987 #+end_src988 #+end_listing990 All that is left for joints it to combine the above pieces into a991 something that can operate on the collection of nodes that a992 blender file represents.994 #+caption: Program to completely create a joint given information995 #+caption: from a blender file.996 #+name: connect997 #+begin_listing clojure998 #+begin_src clojure999 (defn connect1000 "Create a joint between 'obj-a and 'obj-b at the location of1001 'joint. The type of joint is determined by the metadata on 'joint.1003 Here are some examples:1004 {:type :point}1005 {:type :hinge :limit [0 (/ Math/PI 2)] :axis (Vector3f. 0 1 0)}1006 (:axis defaults to (Vector3f. 1 0 0) if not provided for hinge joints)1008 {:type :cone :limit-xz 0]1009 :limit-xy 0]1010 :twist 0]} (use XZY rotation mode in blender!)"1011 [#^Node obj-a #^Node obj-b #^Node joint]1012 (let [control-a (.getControl obj-a RigidBodyControl)1013 control-b (.getControl obj-b RigidBodyControl)1014 joint-center (.getWorldTranslation joint)1015 joint-rotation (.toRotationMatrix (.getWorldRotation joint))1016 pivot-a (world-to-local obj-a joint-center)1017 pivot-b (world-to-local obj-b joint-center)]1018 (if-let1019 [constraints (map-vals eval (read-string (meta-data joint "joint")))]1020 ;; A side-effect of creating a joint registers1021 ;; it with both physics objects which in turn1022 ;; will register the joint with the physics system1023 ;; when the simulation is started.1024 (joint-dispatch constraints1025 control-a control-b1026 pivot-a pivot-b1027 joint-rotation))))1028 #+end_src1029 #+end_listing1031 In general, whenever =CORTEX= exposes a sense (or in this case1032 physicality), it provides a function of the type =sense!=, which1033 takes in a collection of nodes and augments it to support that1034 sense. The function returns any controlls necessary to use that1035 sense. In this case =body!= cerates a physical body and returns no1036 control functions.1038 #+caption: Program to give joints to a creature.1039 #+name: name1040 #+begin_listing clojure1041 #+begin_src clojure1042 (defn joints!1043 "Connect the solid parts of the creature with physical joints. The1044 joints are taken from the \"joints\" node in the creature."1045 [#^Node creature]1046 (dorun1047 (map1048 (fn [joint]1049 (let [[obj-a obj-b] (joint-targets creature joint)]1050 (connect obj-a obj-b joint)))1051 (joints creature))))1052 (defn body!1053 "Endow the creature with a physical body connected with joints. The1054 particulars of the joints and the masses of each body part are1055 determined in blender."1056 [#^Node creature]1057 (physical! creature)1058 (joints! creature))1059 #+end_src1060 #+end_listing1062 All of the code you have just seen amounts to only 130 lines, yet1063 because it builds on top of Blender and jMonkeyEngine3, those few1064 lines pack quite a punch!1066 The hand from figure \ref{blender-hand}, which was modeled after1067 my own right hand, can now be given joints and simulated as a1068 creature.1070 #+caption: With the ability to create physical creatures from blender,1071 #+caption: =CORTEX= gets one step closer to becomming a full creature1072 #+caption: simulation environment.1073 #+name: name1074 #+ATTR_LaTeX: :width 15cm1075 [[./images/physical-hand.png]]1077 ** Sight reuses standard video game components...1079 Vision is one of the most important senses for humans, so I need to1080 build a simulated sense of vision for my AI. I will do this with1081 simulated eyes. Each eye can be independently moved and should see1082 its own version of the world depending on where it is.1084 Making these simulated eyes a reality is simple because1085 jMonkeyEngine already contains extensive support for multiple views1086 of the same 3D simulated world. The reason jMonkeyEngine has this1087 support is because the support is necessary to create games with1088 split-screen views. Multiple views are also used to create1089 efficient pseudo-reflections by rendering the scene from a certain1090 perspective and then projecting it back onto a surface in the 3D1091 world.1093 #+caption: jMonkeyEngine supports multiple views to enable1094 #+caption: split-screen games, like GoldenEye, which was one of1095 #+caption: the first games to use split-screen views.1096 #+name: name1097 #+ATTR_LaTeX: :width 10cm1098 [[./images/goldeneye-4-player.png]]1100 *** A Brief Description of jMonkeyEngine's Rendering Pipeline1102 jMonkeyEngine allows you to create a =ViewPort=, which represents a1103 view of the simulated world. You can create as many of these as you1104 want. Every frame, the =RenderManager= iterates through each1105 =ViewPort=, rendering the scene in the GPU. For each =ViewPort= there1106 is a =FrameBuffer= which represents the rendered image in the GPU.1108 #+caption: =ViewPorts= are cameras in the world. During each frame,1109 #+caption: the =RenderManager= records a snapshot of what each view1110 #+caption: is currently seeing; these snapshots are =FrameBuffer= objects.1111 #+name: rendermanagers1112 #+ATTR_LaTeX: :width 10cm1113 [[./images/diagram_rendermanager2.png]]1115 Each =ViewPort= can have any number of attached =SceneProcessor=1116 objects, which are called every time a new frame is rendered. A1117 =SceneProcessor= receives its =ViewPort's= =FrameBuffer= and can do1118 whatever it wants to the data. Often this consists of invoking GPU1119 specific operations on the rendered image. The =SceneProcessor= can1120 also copy the GPU image data to RAM and process it with the CPU.1122 *** Appropriating Views for Vision1124 Each eye in the simulated creature needs its own =ViewPort= so1125 that it can see the world from its own perspective. To this1126 =ViewPort=, I add a =SceneProcessor= that feeds the visual data to1127 any arbitrary continuation function for further processing. That1128 continuation function may perform both CPU and GPU operations on1129 the data. To make this easy for the continuation function, the1130 =SceneProcessor= maintains appropriately sized buffers in RAM to1131 hold the data. It does not do any copying from the GPU to the CPU1132 itself because it is a slow operation.1134 #+caption: Function to make the rendered secne in jMonkeyEngine1135 #+caption: available for further processing.1136 #+name: pipeline-11137 #+begin_listing clojure1138 #+begin_src clojure1139 (defn vision-pipeline1140 "Create a SceneProcessor object which wraps a vision processing1141 continuation function. The continuation is a function that takes1142 [#^Renderer r #^FrameBuffer fb #^ByteBuffer b #^BufferedImage bi],1143 each of which has already been appropriately sized."1144 [continuation]1145 (let [byte-buffer (atom nil)1146 renderer (atom nil)1147 image (atom nil)]1148 (proxy [SceneProcessor] []1149 (initialize1150 [renderManager viewPort]1151 (let [cam (.getCamera viewPort)1152 width (.getWidth cam)1153 height (.getHeight cam)]1154 (reset! renderer (.getRenderer renderManager))1155 (reset! byte-buffer1156 (BufferUtils/createByteBuffer1157 (* width height 4)))1158 (reset! image (BufferedImage.1159 width height1160 BufferedImage/TYPE_4BYTE_ABGR))))1161 (isInitialized [] (not (nil? @byte-buffer)))1162 (reshape [_ _ _])1163 (preFrame [_])1164 (postQueue [_])1165 (postFrame1166 [#^FrameBuffer fb]1167 (.clear @byte-buffer)1168 (continuation @renderer fb @byte-buffer @image))1169 (cleanup []))))1170 #+end_src1171 #+end_listing1173 The continuation function given to =vision-pipeline= above will be1174 given a =Renderer= and three containers for image data. The1175 =FrameBuffer= references the GPU image data, but the pixel data1176 can not be used directly on the CPU. The =ByteBuffer= and1177 =BufferedImage= are initially "empty" but are sized to hold the1178 data in the =FrameBuffer=. I call transferring the GPU image data1179 to the CPU structures "mixing" the image data.1181 *** Optical sensor arrays are described with images and referenced with metadata1183 The vision pipeline described above handles the flow of rendered1184 images. Now, =CORTEX= needs simulated eyes to serve as the source1185 of these images.1187 An eye is described in blender in the same way as a joint. They1188 are zero dimensional empty objects with no geometry whose local1189 coordinate system determines the orientation of the resulting eye.1190 All eyes are children of a parent node named "eyes" just as all1191 joints have a parent named "joints". An eye binds to the nearest1192 physical object with =bind-sense=.1194 #+caption: Here, the camera is created based on metadata on the1195 #+caption: eye-node and attached to the nearest physical object1196 #+caption: with =bind-sense=1197 #+name: add-eye1198 #+begin_listing clojure1199 (defn add-eye!1200 "Create a Camera centered on the current position of 'eye which1201 follows the closest physical node in 'creature. The camera will1202 point in the X direction and use the Z vector as up as determined1203 by the rotation of these vectors in blender coordinate space. Use1204 XZY rotation for the node in blender."1205 [#^Node creature #^Spatial eye]1206 (let [target (closest-node creature eye)1207 [cam-width cam-height]1208 ;;[640 480] ;; graphics card on laptop doesn't support1209 ;; arbitray dimensions.1210 (eye-dimensions eye)1211 cam (Camera. cam-width cam-height)1212 rot (.getWorldRotation eye)]1213 (.setLocation cam (.getWorldTranslation eye))1214 (.lookAtDirection1215 cam ; this part is not a mistake and1216 (.mult rot Vector3f/UNIT_X) ; is consistent with using Z in1217 (.mult rot Vector3f/UNIT_Y)) ; blender as the UP vector.1218 (.setFrustumPerspective1219 cam (float 45)1220 (float (/ (.getWidth cam) (.getHeight cam)))1221 (float 1)1222 (float 1000))1223 (bind-sense target cam) cam))1224 #+end_listing1226 *** Simulated Retina1228 An eye is a surface (the retina) which contains many discrete1229 sensors to detect light. These sensors can have different1230 light-sensing properties. In humans, each discrete sensor is1231 sensitive to red, blue, green, or gray. These different types of1232 sensors can have different spatial distributions along the retina.1233 In humans, there is a fovea in the center of the retina which has1234 a very high density of color sensors, and a blind spot which has1235 no sensors at all. Sensor density decreases in proportion to1236 distance from the fovea.1238 I want to be able to model any retinal configuration, so my1239 eye-nodes in blender contain metadata pointing to images that1240 describe the precise position of the individual sensors using1241 white pixels. The meta-data also describes the precise sensitivity1242 to light that the sensors described in the image have. An eye can1243 contain any number of these images. For example, the metadata for1244 an eye might look like this:1246 #+begin_src clojure1247 {0xFF0000 "Models/test-creature/retina-small.png"}1248 #+end_src1250 #+caption: An example retinal profile image. White pixels are1251 #+caption: photo-sensitive elements. The distribution of white1252 #+caption: pixels is denser in the middle and falls off at the1253 #+caption: edges and is inspired by the human retina.1254 #+name: retina1255 #+ATTR_LaTeX: :width 7cm1256 [[./images/retina-small.png]]1258 Together, the number 0xFF0000 and the image image above describe1259 the placement of red-sensitive sensory elements.1261 Meta-data to very crudely approximate a human eye might be1262 something like this:1264 #+begin_src clojure1265 (let [retinal-profile "Models/test-creature/retina-small.png"]1266 {0xFF0000 retinal-profile1267 0x00FF00 retinal-profile1268 0x0000FF retinal-profile1269 0xFFFFFF retinal-profile})1270 #+end_src1272 The numbers that serve as keys in the map determine a sensor's1273 relative sensitivity to the channels red, green, and blue. These1274 sensitivity values are packed into an integer in the order1275 =|_|R|G|B|= in 8-bit fields. The RGB values of a pixel in the1276 image are added together with these sensitivities as linear1277 weights. Therefore, 0xFF0000 means sensitive to red only while1278 0xFFFFFF means sensitive to all colors equally (gray).1280 #+caption: This is the core of vision in =CORTEX=. A given eye node1281 #+caption: is converted into a function that returns visual1282 #+caption: information from the simulation.1283 #+name: vision-kernel1284 #+begin_listing clojure1285 #+BEGIN_SRC clojure1286 (defn vision-kernel1287 "Returns a list of functions, each of which will return a color1288 channel's worth of visual information when called inside a running1289 simulation."1290 [#^Node creature #^Spatial eye & {skip :skip :or {skip 0}}]1291 (let [retinal-map (retina-sensor-profile eye)1292 camera (add-eye! creature eye)1293 vision-image1294 (atom1295 (BufferedImage. (.getWidth camera)1296 (.getHeight camera)1297 BufferedImage/TYPE_BYTE_BINARY))1298 register-eye!1299 (runonce1300 (fn [world]1301 (add-camera!1302 world camera1303 (let [counter (atom 0)]1304 (fn [r fb bb bi]1305 (if (zero? (rem (swap! counter inc) (inc skip)))1306 (reset! vision-image1307 (BufferedImage! r fb bb bi))))))))]1308 (vec1309 (map1310 (fn [[key image]]1311 (let [whites (white-coordinates image)1312 topology (vec (collapse whites))1313 sensitivity (sensitivity-presets key key)]1314 (attached-viewport.1315 (fn [world]1316 (register-eye! world)1317 (vector1318 topology1319 (vec1320 (for [[x y] whites]1321 (pixel-sense1322 sensitivity1323 (.getRGB @vision-image x y))))))1324 register-eye!)))1325 retinal-map))))1326 #+END_SRC1327 #+end_listing1329 Note that since each of the functions generated by =vision-kernel=1330 shares the same =register-eye!= function, the eye will be1331 registered only once the first time any of the functions from the1332 list returned by =vision-kernel= is called. Each of the functions1333 returned by =vision-kernel= also allows access to the =Viewport=1334 through which it receives images.1336 All the hard work has been done; all that remains is to apply1337 =vision-kernel= to each eye in the creature and gather the results1338 into one list of functions.1341 #+caption: With =vision!=, =CORTEX= is already a fine simulation1342 #+caption: environment for experimenting with different types of1343 #+caption: eyes.1344 #+name: vision!1345 #+begin_listing clojure1346 #+BEGIN_SRC clojure1347 (defn vision!1348 "Returns a list of functions, each of which returns visual sensory1349 data when called inside a running simulation."1350 [#^Node creature & {skip :skip :or {skip 0}}]1351 (reduce1352 concat1353 (for [eye (eyes creature)]1354 (vision-kernel creature eye))))1355 #+END_SRC1356 #+end_listing1358 #+caption: Simulated vision with a test creature and the1359 #+caption: human-like eye approximation. Notice how each channel1360 #+caption: of the eye responds differently to the differently1361 #+caption: colored balls.1362 #+name: worm-vision-test.1363 #+ATTR_LaTeX: :width 13cm1364 [[./images/worm-vision.png]]1366 The vision code is not much more complicated than the body code,1367 and enables multiple further paths for simulated vision. For1368 example, it is quite easy to create bifocal vision -- you just1369 make two eyes next to each other in blender! It is also possible1370 to encode vision transforms in the retinal files. For example, the1371 human like retina file in figure \ref{retina} approximates a1372 log-polar transform.1374 This vision code has already been absorbed by the jMonkeyEngine1375 community and is now (in modified form) part of a system for1376 capturing in-game video to a file.1378 ** ...but hearing must be built from scratch1379 # is hard; =CORTEX= does it right1380 At the end of this section I will have simulated ears that work the1381 same way as the simulated eyes in the last section. I will be able to1382 place any number of ear-nodes in a blender file, and they will bind to1383 the closest physical object and follow it as it moves around. Each ear1384 will provide access to the sound data it picks up between every frame.1386 Hearing is one of the more difficult senses to simulate, because there1387 is less support for obtaining the actual sound data that is processed1388 by jMonkeyEngine3. There is no "split-screen" support for rendering1389 sound from different points of view, and there is no way to directly1390 access the rendered sound data.1392 =CORTEX='s hearing is unique because it does not have any1393 limitations compared to other simulation environments. As far as I1394 know, there is no other system that supports multiple listerers,1395 and the sound demo at the end of this section is the first time1396 it's been done in a video game environment.1398 *** Brief Description of jMonkeyEngine's Sound System1400 jMonkeyEngine's sound system works as follows:1402 - jMonkeyEngine uses the =AppSettings= for the particular1403 application to determine what sort of =AudioRenderer= should be1404 used.1405 - Although some support is provided for multiple AudioRendering1406 backends, jMonkeyEngine at the time of this writing will either1407 pick no =AudioRenderer= at all, or the =LwjglAudioRenderer=.1408 - jMonkeyEngine tries to figure out what sort of system you're1409 running and extracts the appropriate native libraries.1410 - The =LwjglAudioRenderer= uses the [[http://lwjgl.org/][=LWJGL=]] (LightWeight Java Game1411 Library) bindings to interface with a C library called [[http://kcat.strangesoft.net/openal.html][=OpenAL=]]1412 - =OpenAL= renders the 3D sound and feeds the rendered sound1413 directly to any of various sound output devices with which it1414 knows how to communicate.1416 A consequence of this is that there's no way to access the actual1417 sound data produced by =OpenAL=. Even worse, =OpenAL= only supports1418 one /listener/ (it renders sound data from only one perspective),1419 which normally isn't a problem for games, but becomes a problem1420 when trying to make multiple AI creatures that can each hear the1421 world from a different perspective.1423 To make many AI creatures in jMonkeyEngine that can each hear the1424 world from their own perspective, or to make a single creature with1425 many ears, it is necessary to go all the way back to =OpenAL= and1426 implement support for simulated hearing there.1428 *** Extending =OpenAl=1430 Extending =OpenAL= to support multiple listeners requires 5001431 lines of =C= code and is too hairy to mention here. Instead, I1432 will show a small amount of extension code and go over the high1433 level stragety. Full source is of course available with the1434 =CORTEX= distribution if you're interested.1436 =OpenAL= goes to great lengths to support many different systems,1437 all with different sound capabilities and interfaces. It1438 accomplishes this difficult task by providing code for many1439 different sound backends in pseudo-objects called /Devices/.1440 There's a device for the Linux Open Sound System and the Advanced1441 Linux Sound Architecture, there's one for Direct Sound on Windows,1442 and there's even one for Solaris. =OpenAL= solves the problem of1443 platform independence by providing all these Devices.1445 Wrapper libraries such as LWJGL are free to examine the system on1446 which they are running and then select an appropriate device for1447 that system.1449 There are also a few "special" devices that don't interface with1450 any particular system. These include the Null Device, which1451 doesn't do anything, and the Wave Device, which writes whatever1452 sound it receives to a file, if everything has been set up1453 correctly when configuring =OpenAL=.1455 Actual mixing (doppler shift and distance.environment-based1456 attenuation) of the sound data happens in the Devices, and they1457 are the only point in the sound rendering process where this data1458 is available.1460 Therefore, in order to support multiple listeners, and get the1461 sound data in a form that the AIs can use, it is necessary to1462 create a new Device which supports this feature.1464 Adding a device to OpenAL is rather tricky -- there are five1465 separate files in the =OpenAL= source tree that must be modified1466 to do so. I named my device the "Multiple Audio Send" Device, or1467 =Send= Device for short, since it sends audio data back to the1468 calling application like an Aux-Send cable on a mixing board.1470 The main idea behind the Send device is to take advantage of the1471 fact that LWJGL only manages one /context/ when using OpenAL. A1472 /context/ is like a container that holds samples and keeps track1473 of where the listener is. In order to support multiple listeners,1474 the Send device identifies the LWJGL context as the master1475 context, and creates any number of slave contexts to represent1476 additional listeners. Every time the device renders sound, it1477 synchronizes every source from the master LWJGL context to the1478 slave contexts. Then, it renders each context separately, using a1479 different listener for each one. The rendered sound is made1480 available via JNI to jMonkeyEngine.1482 Switching between contexts is not the normal operation of a1483 Device, and one of the problems with doing so is that a Device1484 normally keeps around a few pieces of state such as the1485 =ClickRemoval= array above which will become corrupted if the1486 contexts are not rendered in parallel. The solution is to create a1487 copy of this normally global device state for each context, and1488 copy it back and forth into and out of the actual device state1489 whenever a context is rendered.1491 The core of the =Send= device is the =syncSources= function, which1492 does the job of copying all relevant data from one context to1493 another.1495 #+caption: Program for extending =OpenAL= to support multiple1496 #+caption: listeners via context copying/switching.1497 #+name: sync-openal-sources1498 #+begin_listing c1499 #+BEGIN_SRC c1500 void syncSources(ALsource *masterSource, ALsource *slaveSource,1501 ALCcontext *masterCtx, ALCcontext *slaveCtx){1502 ALuint master = masterSource->source;1503 ALuint slave = slaveSource->source;1504 ALCcontext *current = alcGetCurrentContext();1506 syncSourcef(master,slave,masterCtx,slaveCtx,AL_PITCH);1507 syncSourcef(master,slave,masterCtx,slaveCtx,AL_GAIN);1508 syncSourcef(master,slave,masterCtx,slaveCtx,AL_MAX_DISTANCE);1509 syncSourcef(master,slave,masterCtx,slaveCtx,AL_ROLLOFF_FACTOR);1510 syncSourcef(master,slave,masterCtx,slaveCtx,AL_REFERENCE_DISTANCE);1511 syncSourcef(master,slave,masterCtx,slaveCtx,AL_MIN_GAIN);1512 syncSourcef(master,slave,masterCtx,slaveCtx,AL_MAX_GAIN);1513 syncSourcef(master,slave,masterCtx,slaveCtx,AL_CONE_OUTER_GAIN);1514 syncSourcef(master,slave,masterCtx,slaveCtx,AL_CONE_INNER_ANGLE);1515 syncSourcef(master,slave,masterCtx,slaveCtx,AL_CONE_OUTER_ANGLE);1516 syncSourcef(master,slave,masterCtx,slaveCtx,AL_SEC_OFFSET);1517 syncSourcef(master,slave,masterCtx,slaveCtx,AL_SAMPLE_OFFSET);1518 syncSourcef(master,slave,masterCtx,slaveCtx,AL_BYTE_OFFSET);1520 syncSource3f(master,slave,masterCtx,slaveCtx,AL_POSITION);1521 syncSource3f(master,slave,masterCtx,slaveCtx,AL_VELOCITY);1522 syncSource3f(master,slave,masterCtx,slaveCtx,AL_DIRECTION);1524 syncSourcei(master,slave,masterCtx,slaveCtx,AL_SOURCE_RELATIVE);1525 syncSourcei(master,slave,masterCtx,slaveCtx,AL_LOOPING);1527 alcMakeContextCurrent(masterCtx);1528 ALint source_type;1529 alGetSourcei(master, AL_SOURCE_TYPE, &source_type);1531 // Only static sources are currently synchronized!1532 if (AL_STATIC == source_type){1533 ALint master_buffer;1534 ALint slave_buffer;1535 alGetSourcei(master, AL_BUFFER, &master_buffer);1536 alcMakeContextCurrent(slaveCtx);1537 alGetSourcei(slave, AL_BUFFER, &slave_buffer);1538 if (master_buffer != slave_buffer){1539 alSourcei(slave, AL_BUFFER, master_buffer);1540 }1541 }1543 // Synchronize the state of the two sources.1544 alcMakeContextCurrent(masterCtx);1545 ALint masterState;1546 ALint slaveState;1548 alGetSourcei(master, AL_SOURCE_STATE, &masterState);1549 alcMakeContextCurrent(slaveCtx);1550 alGetSourcei(slave, AL_SOURCE_STATE, &slaveState);1552 if (masterState != slaveState){1553 switch (masterState){1554 case AL_INITIAL : alSourceRewind(slave); break;1555 case AL_PLAYING : alSourcePlay(slave); break;1556 case AL_PAUSED : alSourcePause(slave); break;1557 case AL_STOPPED : alSourceStop(slave); break;1558 }1559 }1560 // Restore whatever context was previously active.1561 alcMakeContextCurrent(current);1562 }1563 #+END_SRC1564 #+end_listing1566 With this special context-switching device, and some ugly JNI1567 bindings that are not worth mentioning, =CORTEX= gains the ability1568 to access multiple sound streams from =OpenAL=.1570 #+caption: Program to create an ear from a blender empty node. The ear1571 #+caption: follows around the nearest physical object and passes1572 #+caption: all sensory data to a continuation function.1573 #+name: add-ear1574 #+begin_listing clojure1575 #+BEGIN_SRC clojure1576 (defn add-ear!1577 "Create a Listener centered on the current position of 'ear1578 which follows the closest physical node in 'creature and1579 sends sound data to 'continuation."1580 [#^Application world #^Node creature #^Spatial ear continuation]1581 (let [target (closest-node creature ear)1582 lis (Listener.)1583 audio-renderer (.getAudioRenderer world)1584 sp (hearing-pipeline continuation)]1585 (.setLocation lis (.getWorldTranslation ear))1586 (.setRotation lis (.getWorldRotation ear))1587 (bind-sense target lis)1588 (update-listener-velocity! target lis)1589 (.addListener audio-renderer lis)1590 (.registerSoundProcessor audio-renderer lis sp)))1591 #+END_SRC1592 #+end_listing1594 The =Send= device, unlike most of the other devices in =OpenAL=,1595 does not render sound unless asked. This enables the system to1596 slow down or speed up depending on the needs of the AIs who are1597 using it to listen. If the device tried to render samples in1598 real-time, a complicated AI whose mind takes 100 seconds of1599 computer time to simulate 1 second of AI-time would miss almost1600 all of the sound in its environment!1602 #+caption: Program to enable arbitrary hearing in =CORTEX=1603 #+name: hearing1604 #+begin_listing clojure1605 #+BEGIN_SRC clojure1606 (defn hearing-kernel1607 "Returns a function which returns auditory sensory data when called1608 inside a running simulation."1609 [#^Node creature #^Spatial ear]1610 (let [hearing-data (atom [])1611 register-listener!1612 (runonce1613 (fn [#^Application world]1614 (add-ear!1615 world creature ear1616 (comp #(reset! hearing-data %)1617 byteBuffer->pulse-vector))))]1618 (fn [#^Application world]1619 (register-listener! world)1620 (let [data @hearing-data1621 topology1622 (vec (map #(vector % 0) (range 0 (count data))))]1623 [topology data]))))1625 (defn hearing!1626 "Endow the creature in a particular world with the sense of1627 hearing. Will return a sequence of functions, one for each ear,1628 which when called will return the auditory data from that ear."1629 [#^Node creature]1630 (for [ear (ears creature)]1631 (hearing-kernel creature ear)))1632 #+END_SRC1633 #+end_listing1635 Armed with these functions, =CORTEX= is able to test possibly the1636 first ever instance of multiple listeners in a video game engine1637 based simulation!1639 #+caption: Here a simple creature responds to sound by changing1640 #+caption: its color from gray to green when the total volume1641 #+caption: goes over a threshold.1642 #+name: sound-test1643 #+begin_listing java1644 #+BEGIN_SRC java1645 /**1646 * Respond to sound! This is the brain of an AI entity that1647 * hears its surroundings and reacts to them.1648 */1649 public void process(ByteBuffer audioSamples,1650 int numSamples, AudioFormat format) {1651 audioSamples.clear();1652 byte[] data = new byte[numSamples];1653 float[] out = new float[numSamples];1654 audioSamples.get(data);1655 FloatSampleTools.1656 byte2floatInterleaved1657 (data, 0, out, 0, numSamples/format.getFrameSize(), format);1659 float max = Float.NEGATIVE_INFINITY;1660 for (float f : out){if (f > max) max = f;}1661 audioSamples.clear();1663 if (max > 0.1){1664 entity.getMaterial().setColor("Color", ColorRGBA.Green);1665 }1666 else {1667 entity.getMaterial().setColor("Color", ColorRGBA.Gray);1668 }1669 #+END_SRC1670 #+end_listing1672 #+caption: First ever simulation of multiple listerners in =CORTEX=.1673 #+caption: Each cube is a creature which processes sound data with1674 #+caption: the =process= function from listing \ref{sound-test}.1675 #+caption: the ball is constantally emiting a pure tone of1676 #+caption: constant volume. As it approaches the cubes, they each1677 #+caption: change color in response to the sound.1678 #+name: sound-cubes.1679 #+ATTR_LaTeX: :width 10cm1680 [[./images/java-hearing-test.png]]1682 This system of hearing has also been co-opted by the1683 jMonkeyEngine3 community and is used to record audio for demo1684 videos.1686 ** Hundreds of hair-like elements provide a sense of touch1688 Touch is critical to navigation and spatial reasoning and as such I1689 need a simulated version of it to give to my AI creatures.1691 Human skin has a wide array of touch sensors, each of which1692 specialize in detecting different vibrational modes and pressures.1693 These sensors can integrate a vast expanse of skin (i.e. your1694 entire palm), or a tiny patch of skin at the tip of your finger.1695 The hairs of the skin help detect objects before they even come1696 into contact with the skin proper.1698 However, touch in my simulated world can not exactly correspond to1699 human touch because my creatures are made out of completely rigid1700 segments that don't deform like human skin.1702 Instead of measuring deformation or vibration, I surround each1703 rigid part with a plenitude of hair-like objects (/feelers/) which1704 do not interact with the physical world. Physical objects can pass1705 through them with no effect. The feelers are able to tell when1706 other objects pass through them, and they constantly report how1707 much of their extent is covered. So even though the creature's body1708 parts do not deform, the feelers create a margin around those body1709 parts which achieves a sense of touch which is a hybrid between a1710 human's sense of deformation and sense from hairs.1712 Implementing touch in jMonkeyEngine follows a different technical1713 route than vision and hearing. Those two senses piggybacked off1714 jMonkeyEngine's 3D audio and video rendering subsystems. To1715 simulate touch, I use jMonkeyEngine's physics system to execute1716 many small collision detections, one for each feeler. The placement1717 of the feelers is determined by a UV-mapped image which shows where1718 each feeler should be on the 3D surface of the body.1720 *** Defining Touch Meta-Data in Blender1722 Each geometry can have a single UV map which describes the1723 position of the feelers which will constitute its sense of touch.1724 This image path is stored under the ``touch'' key. The image itself1725 is black and white, with black meaning a feeler length of 0 (no1726 feeler is present) and white meaning a feeler length of =scale=,1727 which is a float stored under the key "scale".1729 #+caption: Touch does not use empty nodes, to store metadata,1730 #+caption: because the metadata of each solid part of a1731 #+caption: creature's body is sufficient.1732 #+name: touch-meta-data1733 #+begin_listing clojure1734 #+BEGIN_SRC clojure1735 (defn tactile-sensor-profile1736 "Return the touch-sensor distribution image in BufferedImage format,1737 or nil if it does not exist."1738 [#^Geometry obj]1739 (if-let [image-path (meta-data obj "touch")]1740 (load-image image-path)))1742 (defn tactile-scale1743 "Return the length of each feeler. Default scale is 0.011744 jMonkeyEngine units."1745 [#^Geometry obj]1746 (if-let [scale (meta-data obj "scale")]1747 scale 0.1))1748 #+END_SRC1749 #+end_listing1751 Here is an example of a UV-map which specifies the position of1752 touch sensors along the surface of the upper segment of a fingertip.1754 #+caption: This is the tactile-sensor-profile for the upper segment1755 #+caption: of a fingertip. It defines regions of high touch sensitivity1756 #+caption: (where there are many white pixels) and regions of low1757 #+caption: sensitivity (where white pixels are sparse).1758 #+name: fingertip-UV1759 #+ATTR_LaTeX: :width 13cm1760 [[./images/finger-UV.png]]1762 *** Implementation Summary1764 To simulate touch there are three conceptual steps. For each solid1765 object in the creature, you first have to get UV image and scale1766 parameter which define the position and length of the feelers.1767 Then, you use the triangles which comprise the mesh and the UV1768 data stored in the mesh to determine the world-space position and1769 orientation of each feeler. Then once every frame, update these1770 positions and orientations to match the current position and1771 orientation of the object, and use physics collision detection to1772 gather tactile data.1774 Extracting the meta-data has already been described. The third1775 step, physics collision detection, is handled in =touch-kernel=.1776 Translating the positions and orientations of the feelers from the1777 UV-map to world-space is itself a three-step process.1779 - Find the triangles which make up the mesh in pixel-space and in1780 world-space. \\(=triangles=, =pixel-triangles=).1782 - Find the coordinates of each feeler in world-space. These are1783 the origins of the feelers. (=feeler-origins=).1785 - Calculate the normals of the triangles in world space, and add1786 them to each of the origins of the feelers. These are the1787 normalized coordinates of the tips of the feelers.1788 (=feeler-tips=).1790 *** Triangle Math1792 The rigid objects which make up a creature have an underlying1793 =Geometry=, which is a =Mesh= plus a =Material= and other1794 important data involved with displaying the object.1796 A =Mesh= is composed of =Triangles=, and each =Triangle= has three1797 vertices which have coordinates in world space and UV space.1799 Here, =triangles= gets all the world-space triangles which1800 comprise a mesh, while =pixel-triangles= gets those same triangles1801 expressed in pixel coordinates (which are UV coordinates scaled to1802 fit the height and width of the UV image).1804 #+caption: Programs to extract triangles from a geometry and get1805 #+caption: their verticies in both world and UV-coordinates.1806 #+name: get-triangles1807 #+begin_listing clojure1808 #+BEGIN_SRC clojure1809 (defn triangle1810 "Get the triangle specified by triangle-index from the mesh."1811 [#^Geometry geo triangle-index]1812 (triangle-seq1813 (let [scratch (Triangle.)]1814 (.getTriangle (.getMesh geo) triangle-index scratch) scratch)))1816 (defn triangles1817 "Return a sequence of all the Triangles which comprise a given1818 Geometry."1819 [#^Geometry geo]1820 (map (partial triangle geo) (range (.getTriangleCount (.getMesh geo)))))1822 (defn triangle-vertex-indices1823 "Get the triangle vertex indices of a given triangle from a given1824 mesh."1825 [#^Mesh mesh triangle-index]1826 (let [indices (int-array 3)]1827 (.getTriangle mesh triangle-index indices)1828 (vec indices)))1830 (defn vertex-UV-coord1831 "Get the UV-coordinates of the vertex named by vertex-index"1832 [#^Mesh mesh vertex-index]1833 (let [UV-buffer1834 (.getData1835 (.getBuffer1836 mesh1837 VertexBuffer$Type/TexCoord))]1838 [(.get UV-buffer (* vertex-index 2))1839 (.get UV-buffer (+ 1 (* vertex-index 2)))]))1841 (defn pixel-triangle [#^Geometry geo image index]1842 (let [mesh (.getMesh geo)1843 width (.getWidth image)1844 height (.getHeight image)]1845 (vec (map (fn [[u v]] (vector (* width u) (* height v)))1846 (map (partial vertex-UV-coord mesh)1847 (triangle-vertex-indices mesh index))))))1849 (defn pixel-triangles1850 "The pixel-space triangles of the Geometry, in the same order as1851 (triangles geo)"1852 [#^Geometry geo image]1853 (let [height (.getHeight image)1854 width (.getWidth image)]1855 (map (partial pixel-triangle geo image)1856 (range (.getTriangleCount (.getMesh geo))))))1857 #+END_SRC1858 #+end_listing1860 *** The Affine Transform from one Triangle to Another1862 =pixel-triangles= gives us the mesh triangles expressed in pixel1863 coordinates and =triangles= gives us the mesh triangles expressed1864 in world coordinates. The tactile-sensor-profile gives the1865 position of each feeler in pixel-space. In order to convert1866 pixel-space coordinates into world-space coordinates we need1867 something that takes coordinates on the surface of one triangle1868 and gives the corresponding coordinates on the surface of another1869 triangle.1871 Triangles are [[http://mathworld.wolfram.com/AffineTransformation.html ][affine]], which means any triangle can be transformed1872 into any other by a combination of translation, scaling, and1873 rotation. The affine transformation from one triangle to another1874 is readily computable if the triangle is expressed in terms of a1875 $4x4$ matrix.1877 #+BEGIN_LaTeX1878 $$1879 \begin{bmatrix}1880 x_1 & x_2 & x_3 & n_x \\1881 y_1 & y_2 & y_3 & n_y \\1882 z_1 & z_2 & z_3 & n_z \\1883 1 & 1 & 1 & 11884 \end{bmatrix}1885 $$1886 #+END_LaTeX1888 Here, the first three columns of the matrix are the vertices of1889 the triangle. The last column is the right-handed unit normal of1890 the triangle.1892 With two triangles $T_{1}$ and $T_{2}$ each expressed as a1893 matrix like above, the affine transform from $T_{1}$ to $T_{2}$1894 is $T_{2}T_{1}^{-1}$.1896 The clojure code below recapitulates the formulas above, using1897 jMonkeyEngine's =Matrix4f= objects, which can describe any affine1898 transformation.1900 #+caption: Program to interpert triangles as affine transforms.1901 #+name: triangle-affine1902 #+begin_listing clojure1903 #+BEGIN_SRC clojure1904 (defn triangle->matrix4f1905 "Converts the triangle into a 4x4 matrix: The first three columns1906 contain the vertices of the triangle; the last contains the unit1907 normal of the triangle. The bottom row is filled with 1s."1908 [#^Triangle t]1909 (let [mat (Matrix4f.)1910 [vert-1 vert-2 vert-3]1911 (mapv #(.get t %) (range 3))1912 unit-normal (do (.calculateNormal t)(.getNormal t))1913 vertices [vert-1 vert-2 vert-3 unit-normal]]1914 (dorun1915 (for [row (range 4) col (range 3)]1916 (do1917 (.set mat col row (.get (vertices row) col))1918 (.set mat 3 row 1)))) mat))1920 (defn triangles->affine-transform1921 "Returns the affine transformation that converts each vertex in the1922 first triangle into the corresponding vertex in the second1923 triangle."1924 [#^Triangle tri-1 #^Triangle tri-2]1925 (.mult1926 (triangle->matrix4f tri-2)1927 (.invert (triangle->matrix4f tri-1))))1928 #+END_SRC1929 #+end_listing1931 *** Triangle Boundaries1933 For efficiency's sake I will divide the tactile-profile image into1934 small squares which inscribe each pixel-triangle, then extract the1935 points which lie inside the triangle and map them to 3D-space using1936 =triangle-transform= above. To do this I need a function,1937 =convex-bounds= which finds the smallest box which inscribes a 2D1938 triangle.1940 =inside-triangle?= determines whether a point is inside a triangle1941 in 2D pixel-space.1943 #+caption: Program to efficiently determine point includion1944 #+caption: in a triangle.1945 #+name: in-triangle1946 #+begin_listing clojure1947 #+BEGIN_SRC clojure1948 (defn convex-bounds1949 "Returns the smallest square containing the given vertices, as a1950 vector of integers [left top width height]."1951 [verts]1952 (let [xs (map first verts)1953 ys (map second verts)1954 x0 (Math/floor (apply min xs))1955 y0 (Math/floor (apply min ys))1956 x1 (Math/ceil (apply max xs))1957 y1 (Math/ceil (apply max ys))]1958 [x0 y0 (- x1 x0) (- y1 y0)]))1960 (defn same-side?1961 "Given the points p1 and p2 and the reference point ref, is point p1962 on the same side of the line that goes through p1 and p2 as ref is?"1963 [p1 p2 ref p]1964 (<=1965 01966 (.dot1967 (.cross (.subtract p2 p1) (.subtract p p1))1968 (.cross (.subtract p2 p1) (.subtract ref p1)))))1970 (defn inside-triangle?1971 "Is the point inside the triangle?"1972 {:author "Dylan Holmes"}1973 [#^Triangle tri #^Vector3f p]1974 (let [[vert-1 vert-2 vert-3] [(.get1 tri) (.get2 tri) (.get3 tri)]]1975 (and1976 (same-side? vert-1 vert-2 vert-3 p)1977 (same-side? vert-2 vert-3 vert-1 p)1978 (same-side? vert-3 vert-1 vert-2 p))))1979 #+END_SRC1980 #+end_listing1982 *** Feeler Coordinates1984 The triangle-related functions above make short work of1985 calculating the positions and orientations of each feeler in1986 world-space.1988 #+caption: Program to get the coordinates of ``feelers '' in1989 #+caption: both world and UV-coordinates.1990 #+name: feeler-coordinates1991 #+begin_listing clojure1992 #+BEGIN_SRC clojure1993 (defn feeler-pixel-coords1994 "Returns the coordinates of the feelers in pixel space in lists, one1995 list for each triangle, ordered in the same way as (triangles) and1996 (pixel-triangles)."1997 [#^Geometry geo image]1998 (map1999 (fn [pixel-triangle]2000 (filter2001 (fn [coord]2002 (inside-triangle? (->triangle pixel-triangle)2003 (->vector3f coord)))2004 (white-coordinates image (convex-bounds pixel-triangle))))2005 (pixel-triangles geo image)))2007 (defn feeler-world-coords2008 "Returns the coordinates of the feelers in world space in lists, one2009 list for each triangle, ordered in the same way as (triangles) and2010 (pixel-triangles)."2011 [#^Geometry geo image]2012 (let [transforms2013 (map #(triangles->affine-transform2014 (->triangle %1) (->triangle %2))2015 (pixel-triangles geo image)2016 (triangles geo))]2017 (map (fn [transform coords]2018 (map #(.mult transform (->vector3f %)) coords))2019 transforms (feeler-pixel-coords geo image))))2020 #+END_SRC2021 #+end_listing2023 #+caption: Program to get the position of the base and tip of2024 #+caption: each ``feeler''2025 #+name: feeler-tips2026 #+begin_listing clojure2027 #+BEGIN_SRC clojure2028 (defn feeler-origins2029 "The world space coordinates of the root of each feeler."2030 [#^Geometry geo image]2031 (reduce concat (feeler-world-coords geo image)))2033 (defn feeler-tips2034 "The world space coordinates of the tip of each feeler."2035 [#^Geometry geo image]2036 (let [world-coords (feeler-world-coords geo image)2037 normals2038 (map2039 (fn [triangle]2040 (.calculateNormal triangle)2041 (.clone (.getNormal triangle)))2042 (map ->triangle (triangles geo)))]2044 (mapcat (fn [origins normal]2045 (map #(.add % normal) origins))2046 world-coords normals)))2048 (defn touch-topology2049 [#^Geometry geo image]2050 (collapse (reduce concat (feeler-pixel-coords geo image))))2051 #+END_SRC2052 #+end_listing2054 *** Simulated Touch2056 Now that the functions to construct feelers are complete,2057 =touch-kernel= generates functions to be called from within a2058 simulation that perform the necessary physics collisions to2059 collect tactile data, and =touch!= recursively applies it to every2060 node in the creature.2062 #+caption: Efficient program to transform a ray from2063 #+caption: one position to another.2064 #+name: set-ray2065 #+begin_listing clojure2066 #+BEGIN_SRC clojure2067 (defn set-ray [#^Ray ray #^Matrix4f transform2068 #^Vector3f origin #^Vector3f tip]2069 ;; Doing everything locally reduces garbage collection by enough to2070 ;; be worth it.2071 (.mult transform origin (.getOrigin ray))2072 (.mult transform tip (.getDirection ray))2073 (.subtractLocal (.getDirection ray) (.getOrigin ray))2074 (.normalizeLocal (.getDirection ray)))2075 #+END_SRC2076 #+end_listing2078 #+caption: This is the core of touch in =CORTEX= each feeler2079 #+caption: follows the object it is bound to, reporting any2080 #+caption: collisions that may happen.2081 #+name: touch-kernel2082 #+begin_listing clojure2083 #+BEGIN_SRC clojure2084 (defn touch-kernel2085 "Constructs a function which will return tactile sensory data from2086 'geo when called from inside a running simulation"2087 [#^Geometry geo]2088 (if-let2089 [profile (tactile-sensor-profile geo)]2090 (let [ray-reference-origins (feeler-origins geo profile)2091 ray-reference-tips (feeler-tips geo profile)2092 ray-length (tactile-scale geo)2093 current-rays (map (fn [_] (Ray.)) ray-reference-origins)2094 topology (touch-topology geo profile)2095 correction (float (* ray-length -0.2))]2096 ;; slight tolerance for very close collisions.2097 (dorun2098 (map (fn [origin tip]2099 (.addLocal origin (.mult (.subtract tip origin)2100 correction)))2101 ray-reference-origins ray-reference-tips))2102 (dorun (map #(.setLimit % ray-length) current-rays))2103 (fn [node]2104 (let [transform (.getWorldMatrix geo)]2105 (dorun2106 (map (fn [ray ref-origin ref-tip]2107 (set-ray ray transform ref-origin ref-tip))2108 current-rays ray-reference-origins2109 ray-reference-tips))2110 (vector2111 topology2112 (vec2113 (for [ray current-rays]2114 (do2115 (let [results (CollisionResults.)]2116 (.collideWith node ray results)2117 (let [touch-objects2118 (filter #(not (= geo (.getGeometry %)))2119 results)2120 limit (.getLimit ray)]2121 [(if (empty? touch-objects)2122 limit2123 (let [response2124 (apply min (map #(.getDistance %)2125 touch-objects))]2126 (FastMath/clamp2127 (float2128 (if (> response limit) (float 0.0)2129 (+ response correction)))2130 (float 0.0)2131 limit)))2132 limit])))))))))))2133 #+END_SRC2134 #+end_listing2136 Armed with the =touch!= function, =CORTEX= becomes capable of2137 giving creatures a sense of touch. A simple test is to create a2138 cube that is outfitted with a uniform distrubition of touch2139 sensors. It can feel the ground and any balls that it touches.2141 #+caption: =CORTEX= interface for creating touch in a simulated2142 #+caption: creature.2143 #+name: touch2144 #+begin_listing clojure2145 #+BEGIN_SRC clojure2146 (defn touch!2147 "Endow the creature with the sense of touch. Returns a sequence of2148 functions, one for each body part with a tactile-sensor-profile,2149 each of which when called returns sensory data for that body part."2150 [#^Node creature]2151 (filter2152 (comp not nil?)2153 (map touch-kernel2154 (filter #(isa? (class %) Geometry)2155 (node-seq creature)))))2156 #+END_SRC2157 #+end_listing2159 The tactile-sensor-profile image for the touch cube is a simple2160 cross with a unifom distribution of touch sensors:2162 #+caption: The touch profile for the touch-cube. Each pure white2163 #+caption: pixel defines a touch sensitive feeler.2164 #+name: touch-cube-uv-map2165 #+ATTR_LaTeX: :width 7cm2166 [[./images/touch-profile.png]]2168 #+caption: The touch cube reacts to canonballs. The black, red,2169 #+caption: and white cross on the right is a visual display of2170 #+caption: the creature's touch. White means that it is feeling2171 #+caption: something strongly, black is not feeling anything,2172 #+caption: and gray is in-between. The cube can feel both the2173 #+caption: floor and the ball. Notice that when the ball causes2174 #+caption: the cube to tip, that the bottom face can still feel2175 #+caption: part of the ground.2176 #+name: touch-cube-uv-map2177 #+ATTR_LaTeX: :width 15cm2178 [[./images/touch-cube.png]]2180 ** Proprioception provides knowledge of your own body's position2182 Close your eyes, and touch your nose with your right index finger.2183 How did you do it? You could not see your hand, and neither your2184 hand nor your nose could use the sense of touch to guide the path2185 of your hand. There are no sound cues, and Taste and Smell2186 certainly don't provide any help. You know where your hand is2187 without your other senses because of Proprioception.2189 Humans can sometimes loose this sense through viral infections or2190 damage to the spinal cord or brain, and when they do, they loose2191 the ability to control their own bodies without looking directly at2192 the parts they want to move. In [[http://en.wikipedia.org/wiki/The_Man_Who_Mistook_His_Wife_for_a_Hat][The Man Who Mistook His Wife for a2193 Hat]], a woman named Christina looses this sense and has to learn how2194 to move by carefully watching her arms and legs. She describes2195 proprioception as the "eyes of the body, the way the body sees2196 itself".2198 Proprioception in humans is mediated by [[http://en.wikipedia.org/wiki/Articular_capsule][joint capsules]], [[http://en.wikipedia.org/wiki/Muscle_spindle][muscle2199 spindles]], and the [[http://en.wikipedia.org/wiki/Golgi_tendon_organ][Golgi tendon organs]]. These measure the relative2200 positions of each body part by monitoring muscle strain and length.2202 It's clear that this is a vital sense for fluid, graceful movement.2203 It's also particularly easy to implement in jMonkeyEngine.2205 My simulated proprioception calculates the relative angles of each2206 joint from the rest position defined in the blender file. This2207 simulates the muscle-spindles and joint capsules. I will deal with2208 Golgi tendon organs, which calculate muscle strain, in the next2209 section.2211 *** Helper functions2213 =absolute-angle= calculates the angle between two vectors,2214 relative to a third axis vector. This angle is the number of2215 radians you have to move counterclockwise around the axis vector2216 to get from the first to the second vector. It is not commutative2217 like a normal dot-product angle is.2219 The purpose of these functions is to build a system of angle2220 measurement that is biologically plausable.2222 #+caption: Program to measure angles along a vector2223 #+name: helpers2224 #+begin_listing clojure2225 #+BEGIN_SRC clojure2226 (defn right-handed?2227 "true iff the three vectors form a right handed coordinate2228 system. The three vectors do not have to be normalized or2229 orthogonal."2230 [vec1 vec2 vec3]2231 (pos? (.dot (.cross vec1 vec2) vec3)))2233 (defn absolute-angle2234 "The angle between 'vec1 and 'vec2 around 'axis. In the range2235 [0 (* 2 Math/PI)]."2236 [vec1 vec2 axis]2237 (let [angle (.angleBetween vec1 vec2)]2238 (if (right-handed? vec1 vec2 axis)2239 angle (- (* 2 Math/PI) angle))))2240 #+END_SRC2241 #+end_listing2243 *** Proprioception Kernel2245 Given a joint, =proprioception-kernel= produces a function that2246 calculates the Euler angles between the the objects the joint2247 connects. The only tricky part here is making the angles relative2248 to the joint's initial ``straightness''.2250 #+caption: Program to return biologially reasonable proprioceptive2251 #+caption: data for each joint.2252 #+name: proprioception2253 #+begin_listing clojure2254 #+BEGIN_SRC clojure2255 (defn proprioception-kernel2256 "Returns a function which returns proprioceptive sensory data when2257 called inside a running simulation."2258 [#^Node parts #^Node joint]2259 (let [[obj-a obj-b] (joint-targets parts joint)2260 joint-rot (.getWorldRotation joint)2261 x0 (.mult joint-rot Vector3f/UNIT_X)2262 y0 (.mult joint-rot Vector3f/UNIT_Y)2263 z0 (.mult joint-rot Vector3f/UNIT_Z)]2264 (fn []2265 (let [rot-a (.clone (.getWorldRotation obj-a))2266 rot-b (.clone (.getWorldRotation obj-b))2267 x (.mult rot-a x0)2268 y (.mult rot-a y0)2269 z (.mult rot-a z0)2271 X (.mult rot-b x0)2272 Y (.mult rot-b y0)2273 Z (.mult rot-b z0)2274 heading (Math/atan2 (.dot X z) (.dot X x))2275 pitch (Math/atan2 (.dot X y) (.dot X x))2277 ;; rotate x-vector back to origin2278 reverse2279 (doto (Quaternion.)2280 (.fromAngleAxis2281 (.angleBetween X x)2282 (let [cross (.normalize (.cross X x))]2283 (if (= 0 (.length cross)) y cross))))2284 roll (absolute-angle (.mult reverse Y) y x)]2285 [heading pitch roll]))))2287 (defn proprioception!2288 "Endow the creature with the sense of proprioception. Returns a2289 sequence of functions, one for each child of the \"joints\" node in2290 the creature, which each report proprioceptive information about2291 that joint."2292 [#^Node creature]2293 ;; extract the body's joints2294 (let [senses (map (partial proprioception-kernel creature)2295 (joints creature))]2296 (fn []2297 (map #(%) senses))))2298 #+END_SRC2299 #+end_listing2301 =proprioception!= maps =proprioception-kernel= across all the2302 joints of the creature. It uses the same list of joints that2303 =joints= uses. Proprioception is the easiest sense to implement in2304 =CORTEX=, and it will play a crucial role when efficiently2305 implementing empathy.2307 #+caption: In the upper right corner, the three proprioceptive2308 #+caption: angle measurements are displayed. Red is yaw, Green is2309 #+caption: pitch, and White is roll.2310 #+name: proprio2311 #+ATTR_LaTeX: :width 11cm2312 [[./images/proprio.png]]2314 ** Muscles contain both sensors and effectors2316 Surprisingly enough, terrestrial creatures only move by using2317 torque applied about their joints. There's not a single straight2318 line of force in the human body at all! (A straight line of force2319 would correspond to some sort of jet or rocket propulsion.)2321 In humans, muscles are composed of muscle fibers which can contract2322 to exert force. The muscle fibers which compose a muscle are2323 partitioned into discrete groups which are each controlled by a2324 single alpha motor neuron. A single alpha motor neuron might2325 control as little as three or as many as one thousand muscle2326 fibers. When the alpha motor neuron is engaged by the spinal cord,2327 it activates all of the muscle fibers to which it is attached. The2328 spinal cord generally engages the alpha motor neurons which control2329 few muscle fibers before the motor neurons which control many2330 muscle fibers. This recruitment strategy allows for precise2331 movements at low strength. The collection of all motor neurons that2332 control a muscle is called the motor pool. The brain essentially2333 says "activate 30% of the motor pool" and the spinal cord recruits2334 motor neurons until 30% are activated. Since the distribution of2335 power among motor neurons is unequal and recruitment goes from2336 weakest to strongest, the first 30% of the motor pool might be 5%2337 of the strength of the muscle.2339 My simulated muscles follow a similar design: Each muscle is2340 defined by a 1-D array of numbers (the "motor pool"). Each entry in2341 the array represents a motor neuron which controls a number of2342 muscle fibers equal to the value of the entry. Each muscle has a2343 scalar strength factor which determines the total force the muscle2344 can exert when all motor neurons are activated. The effector2345 function for a muscle takes a number to index into the motor pool,2346 and then "activates" all the motor neurons whose index is lower or2347 equal to the number. Each motor-neuron will apply force in2348 proportion to its value in the array. Lower values cause less2349 force. The lower values can be put at the "beginning" of the 1-D2350 array to simulate the layout of actual human muscles, which are2351 capable of more precise movements when exerting less force. Or, the2352 motor pool can simulate more exotic recruitment strategies which do2353 not correspond to human muscles.2355 This 1D array is defined in an image file for ease of2356 creation/visualization. Here is an example muscle profile image.2358 #+caption: A muscle profile image that describes the strengths2359 #+caption: of each motor neuron in a muscle. White is weakest2360 #+caption: and dark red is strongest. This particular pattern2361 #+caption: has weaker motor neurons at the beginning, just2362 #+caption: like human muscle.2363 #+name: muscle-recruit2364 #+ATTR_LaTeX: :width 7cm2365 [[./images/basic-muscle.png]]2367 *** Muscle meta-data2369 #+caption: Program to deal with loading muscle data from a blender2370 #+caption: file's metadata.2371 #+name: motor-pool2372 #+begin_listing clojure2373 #+BEGIN_SRC clojure2374 (defn muscle-profile-image2375 "Get the muscle-profile image from the node's blender meta-data."2376 [#^Node muscle]2377 (if-let [image (meta-data muscle "muscle")]2378 (load-image image)))2380 (defn muscle-strength2381 "Return the strength of this muscle, or 1 if it is not defined."2382 [#^Node muscle]2383 (if-let [strength (meta-data muscle "strength")]2384 strength 1))2386 (defn motor-pool2387 "Return a vector where each entry is the strength of the \"motor2388 neuron\" at that part in the muscle."2389 [#^Node muscle]2390 (let [profile (muscle-profile-image muscle)]2391 (vec2392 (let [width (.getWidth profile)]2393 (for [x (range width)]2394 (- 2552395 (bit-and2396 0x0000FF2397 (.getRGB profile x 0))))))))2398 #+END_SRC2399 #+end_listing2401 Of note here is =motor-pool= which interprets the muscle-profile2402 image in a way that allows me to use gradients between white and2403 red, instead of shades of gray as I've been using for all the2404 other senses. This is purely an aesthetic touch.2406 *** Creating muscles2408 #+caption: This is the core movement functoion in =CORTEX=, which2409 #+caption: implements muscles that report on their activation.2410 #+name: muscle-kernel2411 #+begin_listing clojure2412 #+BEGIN_SRC clojure2413 (defn movement-kernel2414 "Returns a function which when called with a integer value inside a2415 running simulation will cause movement in the creature according2416 to the muscle's position and strength profile. Each function2417 returns the amount of force applied / max force."2418 [#^Node creature #^Node muscle]2419 (let [target (closest-node creature muscle)2420 axis2421 (.mult (.getWorldRotation muscle) Vector3f/UNIT_Y)2422 strength (muscle-strength muscle)2424 pool (motor-pool muscle)2425 pool-integral (reductions + pool)2426 forces2427 (vec (map #(float (* strength (/ % (last pool-integral))))2428 pool-integral))2429 control (.getControl target RigidBodyControl)]2430 ;;(println-repl (.getName target) axis)2431 (fn [n]2432 (let [pool-index (max 0 (min n (dec (count pool))))2433 force (forces pool-index)]2434 (.applyTorque control (.mult axis force))2435 (float (/ force strength))))))2437 (defn movement!2438 "Endow the creature with the power of movement. Returns a sequence2439 of functions, each of which accept an integer value and will2440 activate their corresponding muscle."2441 [#^Node creature]2442 (for [muscle (muscles creature)]2443 (movement-kernel creature muscle)))2444 #+END_SRC2445 #+end_listing2448 =movement-kernel= creates a function that will move the nearest2449 physical object to the muscle node. The muscle exerts a rotational2450 force dependent on it's orientation to the object in the blender2451 file. The function returned by =movement-kernel= is also a sense2452 function: it returns the percent of the total muscle strength that2453 is currently being employed. This is analogous to muscle tension2454 in humans and completes the sense of proprioception begun in the2455 last section.2457 ** =CORTEX= brings complex creatures to life!2459 The ultimate test of =CORTEX= is to create a creature with the full2460 gamut of senses and put it though its paces.2462 With all senses enabled, my right hand model looks like an2463 intricate marionette hand with several strings for each finger:2465 #+caption: View of the hand model with all sense nodes. You can see2466 #+caption: the joint, muscle, ear, and eye nodess here.2467 #+name: hand-nodes-12468 #+ATTR_LaTeX: :width 11cm2469 [[./images/hand-with-all-senses2.png]]2471 #+caption: An alternate view of the hand.2472 #+name: hand-nodes-22473 #+ATTR_LaTeX: :width 15cm2474 [[./images/hand-with-all-senses3.png]]2476 With the hand fully rigged with senses, I can run it though a test2477 that will test everything.2479 #+caption: A full test of the hand with all senses. Note expecially2480 #+caption: the interactions the hand has with itself: it feels2481 #+caption: its own palm and fingers, and when it curls its fingers,2482 #+caption: it sees them with its eye (which is located in the center2483 #+caption: of the palm. The red block appears with a pure tone sound.2484 #+caption: The hand then uses its muscles to launch the cube!2485 #+name: integration2486 #+ATTR_LaTeX: :width 16cm2487 [[./images/integration.png]]2489 ** =CORTEX= enables many possiblities for further research2491 Often times, the hardest part of building a system involving2492 creatures is dealing with physics and graphics. =CORTEX= removes2493 much of this initial difficulty and leaves researchers free to2494 directly pursue their ideas. I hope that even undergrads with a2495 passing curiosity about simulated touch or creature evolution will2496 be able to use cortex for experimentation. =CORTEX= is a completely2497 simulated world, and far from being a disadvantage, its simulated2498 nature enables you to create senses and creatures that would be2499 impossible to make in the real world.2501 While not by any means a complete list, here are some paths2502 =CORTEX= is well suited to help you explore:2504 - Empathy :: my empathy program leaves many areas for2505 improvement, among which are using vision to infer2506 proprioception and looking up sensory experience with imagined2507 vision, touch, and sound.2508 - Evolution :: Karl Sims created a rich environment for2509 simulating the evolution of creatures on a connection2510 machine. Today, this can be redone and expanded with =CORTEX=2511 on an ordinary computer.2512 - Exotic senses :: Cortex enables many fascinating senses that are2513 not possible to build in the real world. For example,2514 telekinesis is an interesting avenue to explore. You can also2515 make a ``semantic'' sense which looks up metadata tags on2516 objects in the environment the metadata tags might contain2517 other sensory information.2518 - Imagination via subworlds :: this would involve a creature with2519 an effector which creates an entire new sub-simulation where2520 the creature has direct control over placement/creation of2521 objects via simulated telekinesis. The creature observes this2522 sub-world through it's normal senses and uses its observations2523 to make predictions about its top level world.2524 - Simulated prescience :: step the simulation forward a few ticks,2525 gather sensory data, then supply this data for the creature as2526 one of its actual senses. The cost of prescience is slowing2527 the simulation down by a factor proportional to however far2528 you want the entities to see into the future. What happens2529 when two evolved creatures that can each see into the future2530 fight each other?2531 - Swarm creatures :: Program a group of creatures that cooperate2532 with each other. Because the creatures would be simulated, you2533 could investigate computationally complex rules of behavior2534 which still, from the group's point of view, would happen in2535 ``real time''. Interactions could be as simple as cellular2536 organisms communicating via flashing lights, or as complex as2537 humanoids completing social tasks, etc.2538 - =HACKER= for writing muscle-control programs :: Presented with2539 low-level muscle control/ sense API, generate higher level2540 programs for accomplishing various stated goals. Example goals2541 might be "extend all your fingers" or "move your hand into the2542 area with blue light" or "decrease the angle of this joint".2543 It would be like Sussman's HACKER, except it would operate2544 with much more data in a more realistic world. Start off with2545 "calisthenics" to develop subroutines over the motor control2546 API. This would be the "spinal chord" of a more intelligent2547 creature. The low level programming code might be a turning2548 machine that could develop programs to iterate over a "tape"2549 where each entry in the tape could control recruitment of the2550 fibers in a muscle.2551 - Sense fusion :: There is much work to be done on sense2552 integration -- building up a coherent picture of the world and2553 the things in it with =CORTEX= as a base, you can explore2554 concepts like self-organizing maps or cross modal clustering2555 in ways that have never before been tried.2556 - Inverse kinematics :: experiments in sense guided motor control2557 are easy given =CORTEX='s support -- you can get right to the2558 hard control problems without worrying about physics or2559 senses.2561 * =EMPATH=: the simulated worm experiment2562 # Empathy in a simulated worm2564 Here I develop a computational model of empathy, using =CORTEX= as a2565 base. Empathy in this context is the ability to observe another2566 creature and infer what sorts of sensations that creature is2567 feeling. My empathy algorithm involves multiple phases. First is2568 free-play, where the creature moves around and gains sensory2569 experience. From this experience I construct a representation of the2570 creature's sensory state space, which I call \Phi-space. Using2571 \Phi-space, I construct an efficient function which takes the2572 limited data that comes from observing another creature and enriches2573 it full compliment of imagined sensory data. I can then use the2574 imagined sensory data to recognize what the observed creature is2575 doing and feeling, using straightforward embodied action predicates.2576 This is all demonstrated with using a simple worm-like creature, and2577 recognizing worm-actions based on limited data.2579 #+caption: Here is the worm with which we will be working.2580 #+caption: It is composed of 5 segments. Each segment has a2581 #+caption: pair of extensor and flexor muscles. Each of the2582 #+caption: worm's four joints is a hinge joint which allows2583 #+caption: about 30 degrees of rotation to either side. Each segment2584 #+caption: of the worm is touch-capable and has a uniform2585 #+caption: distribution of touch sensors on each of its faces.2586 #+caption: Each joint has a proprioceptive sense to detect2587 #+caption: relative positions. The worm segments are all the2588 #+caption: same except for the first one, which has a much2589 #+caption: higher weight than the others to allow for easy2590 #+caption: manual motor control.2591 #+name: basic-worm-view2592 #+ATTR_LaTeX: :width 10cm2593 [[./images/basic-worm-view.png]]2595 #+caption: Program for reading a worm from a blender file and2596 #+caption: outfitting it with the senses of proprioception,2597 #+caption: touch, and the ability to move, as specified in the2598 #+caption: blender file.2599 #+name: get-worm2600 #+begin_listing clojure2601 #+begin_src clojure2602 (defn worm []2603 (let [model (load-blender-model "Models/worm/worm.blend")]2604 {:body (doto model (body!))2605 :touch (touch! model)2606 :proprioception (proprioception! model)2607 :muscles (movement! model)}))2608 #+end_src2609 #+end_listing2611 ** Embodiment factors action recognition into managable parts2613 Using empathy, I divide the problem of action recognition into a2614 recognition process expressed in the language of a full compliment2615 of senses, and an imaganitive process that generates full sensory2616 data from partial sensory data. Splitting the action recognition2617 problem in this manner greatly reduces the total amount of work to2618 recognize actions: The imaganitive process is mostly just matching2619 previous experience, and the recognition process gets to use all2620 the senses to directly describe any action.2622 ** Action recognition is easy with a full gamut of senses2624 Embodied representations using multiple senses such as touch,2625 proprioception, and muscle tension turns out be be exceedingly2626 efficient at describing body-centered actions. It is the ``right2627 language for the job''. For example, it takes only around 5 lines2628 of LISP code to describe the action of ``curling'' using embodied2629 primitives. It takes about 10 lines to describe the seemingly2630 complicated action of wiggling.2632 The following action predicates each take a stream of sensory2633 experience, observe however much of it they desire, and decide2634 whether the worm is doing the action they describe. =curled?=2635 relies on proprioception, =resting?= relies on touch, =wiggling?=2636 relies on a fourier analysis of muscle contraction, and2637 =grand-circle?= relies on touch and reuses =curled?= as a gaurd.2639 #+caption: Program for detecting whether the worm is curled. This is the2640 #+caption: simplest action predicate, because it only uses the last frame2641 #+caption: of sensory experience, and only uses proprioceptive data. Even2642 #+caption: this simple predicate, however, is automatically frame2643 #+caption: independent and ignores vermopomorphic differences such as2644 #+caption: worm textures and colors.2645 #+name: curled2646 #+begin_listing clojure2647 #+begin_src clojure2648 (defn curled?2649 "Is the worm curled up?"2650 [experiences]2651 (every?2652 (fn [[_ _ bend]]2653 (> (Math/sin bend) 0.64))2654 (:proprioception (peek experiences))))2655 #+end_src2656 #+end_listing2658 #+caption: Program for summarizing the touch information in a patch2659 #+caption: of skin.2660 #+name: touch-summary2661 #+begin_listing clojure2662 #+begin_src clojure2663 (defn contact2664 "Determine how much contact a particular worm segment has with2665 other objects. Returns a value between 0 and 1, where 1 is full2666 contact and 0 is no contact."2667 [touch-region [coords contact :as touch]]2668 (-> (zipmap coords contact)2669 (select-keys touch-region)2670 (vals)2671 (#(map first %))2672 (average)2673 (* 10)2674 (- 1)2675 (Math/abs)))2676 #+end_src2677 #+end_listing2680 #+caption: Program for detecting whether the worm is at rest. This program2681 #+caption: uses a summary of the tactile information from the underbelly2682 #+caption: of the worm, and is only true if every segment is touching the2683 #+caption: floor. Note that this function contains no references to2684 #+caption: proprioction at all.2685 #+name: resting2686 #+begin_listing clojure2687 #+begin_src clojure2688 (def worm-segment-bottom (rect-region [8 15] [14 22]))2690 (defn resting?2691 "Is the worm resting on the ground?"2692 [experiences]2693 (every?2694 (fn [touch-data]2695 (< 0.9 (contact worm-segment-bottom touch-data)))2696 (:touch (peek experiences))))2697 #+end_src2698 #+end_listing2700 #+caption: Program for detecting whether the worm is curled up into a2701 #+caption: full circle. Here the embodied approach begins to shine, as2702 #+caption: I am able to both use a previous action predicate (=curled?=)2703 #+caption: as well as the direct tactile experience of the head and tail.2704 #+name: grand-circle2705 #+begin_listing clojure2706 #+begin_src clojure2707 (def worm-segment-bottom-tip (rect-region [15 15] [22 22]))2709 (def worm-segment-top-tip (rect-region [0 15] [7 22]))2711 (defn grand-circle?2712 "Does the worm form a majestic circle (one end touching the other)?"2713 [experiences]2714 (and (curled? experiences)2715 (let [worm-touch (:touch (peek experiences))2716 tail-touch (worm-touch 0)2717 head-touch (worm-touch 4)]2718 (and (< 0.55 (contact worm-segment-bottom-tip tail-touch))2719 (< 0.55 (contact worm-segment-top-tip head-touch))))))2720 #+end_src2721 #+end_listing2724 #+caption: Program for detecting whether the worm has been wiggling for2725 #+caption: the last few frames. It uses a fourier analysis of the muscle2726 #+caption: contractions of the worm's tail to determine wiggling. This is2727 #+caption: signigicant because there is no particular frame that clearly2728 #+caption: indicates that the worm is wiggling --- only when multiple frames2729 #+caption: are analyzed together is the wiggling revealed. Defining2730 #+caption: wiggling this way also gives the worm an opportunity to learn2731 #+caption: and recognize ``frustrated wiggling'', where the worm tries to2732 #+caption: wiggle but can't. Frustrated wiggling is very visually different2733 #+caption: from actual wiggling, but this definition gives it to us for free.2734 #+name: wiggling2735 #+begin_listing clojure2736 #+begin_src clojure2737 (defn fft [nums]2738 (map2739 #(.getReal %)2740 (.transform2741 (FastFourierTransformer. DftNormalization/STANDARD)2742 (double-array nums) TransformType/FORWARD)))2744 (def indexed (partial map-indexed vector))2746 (defn max-indexed [s]2747 (first (sort-by (comp - second) (indexed s))))2749 (defn wiggling?2750 "Is the worm wiggling?"2751 [experiences]2752 (let [analysis-interval 0x40]2753 (when (> (count experiences) analysis-interval)2754 (let [a-flex 32755 a-ex 22756 muscle-activity2757 (map :muscle (vector:last-n experiences analysis-interval))2758 base-activity2759 (map #(- (% a-flex) (% a-ex)) muscle-activity)]2760 (= 22761 (first2762 (max-indexed2763 (map #(Math/abs %)2764 (take 20 (fft base-activity))))))))))2765 #+end_src2766 #+end_listing2768 With these action predicates, I can now recognize the actions of2769 the worm while it is moving under my control and I have access to2770 all the worm's senses.2772 #+caption: Use the action predicates defined earlier to report on2773 #+caption: what the worm is doing while in simulation.2774 #+name: report-worm-activity2775 #+begin_listing clojure2776 #+begin_src clojure2777 (defn debug-experience2778 [experiences text]2779 (cond2780 (grand-circle? experiences) (.setText text "Grand Circle")2781 (curled? experiences) (.setText text "Curled")2782 (wiggling? experiences) (.setText text "Wiggling")2783 (resting? experiences) (.setText text "Resting")))2784 #+end_src2785 #+end_listing2787 #+caption: Using =debug-experience=, the body-centered predicates2788 #+caption: work together to classify the behaviour of the worm.2789 #+caption: the predicates are operating with access to the worm's2790 #+caption: full sensory data.2791 #+name: basic-worm-view2792 #+ATTR_LaTeX: :width 10cm2793 [[./images/worm-identify-init.png]]2795 These action predicates satisfy the recognition requirement of an2796 empathic recognition system. There is power in the simplicity of2797 the action predicates. They describe their actions without getting2798 confused in visual details of the worm. Each one is frame2799 independent, but more than that, they are each indepent of2800 irrelevant visual details of the worm and the environment. They2801 will work regardless of whether the worm is a different color or2802 hevaily textured, or if the environment has strange lighting.2804 The trick now is to make the action predicates work even when the2805 sensory data on which they depend is absent. If I can do that, then2806 I will have gained much,2808 ** \Phi-space describes the worm's experiences2810 As a first step towards building empathy, I need to gather all of2811 the worm's experiences during free play. I use a simple vector to2812 store all the experiences.2814 Each element of the experience vector exists in the vast space of2815 all possible worm-experiences. Most of this vast space is actually2816 unreachable due to physical constraints of the worm's body. For2817 example, the worm's segments are connected by hinge joints that put2818 a practical limit on the worm's range of motions without limiting2819 its degrees of freedom. Some groupings of senses are impossible;2820 the worm can not be bent into a circle so that its ends are2821 touching and at the same time not also experience the sensation of2822 touching itself.2824 As the worm moves around during free play and its experience vector2825 grows larger, the vector begins to define a subspace which is all2826 the sensations the worm can practicaly experience during normal2827 operation. I call this subspace \Phi-space, short for2828 physical-space. The experience vector defines a path through2829 \Phi-space. This path has interesting properties that all derive2830 from physical embodiment. The proprioceptive components are2831 completely smooth, because in order for the worm to move from one2832 position to another, it must pass through the intermediate2833 positions. The path invariably forms loops as actions are repeated.2834 Finally and most importantly, proprioception actually gives very2835 strong inference about the other senses. For example, when the worm2836 is flat, you can infer that it is touching the ground and that its2837 muscles are not active, because if the muscles were active, the2838 worm would be moving and would not be perfectly flat. In order to2839 stay flat, the worm has to be touching the ground, or it would2840 again be moving out of the flat position due to gravity. If the2841 worm is positioned in such a way that it interacts with itself,2842 then it is very likely to be feeling the same tactile feelings as2843 the last time it was in that position, because it has the same body2844 as then. If you observe multiple frames of proprioceptive data,2845 then you can become increasingly confident about the exact2846 activations of the worm's muscles, because it generally takes a2847 unique combination of muscle contractions to transform the worm's2848 body along a specific path through \Phi-space.2850 There is a simple way of taking \Phi-space and the total ordering2851 provided by an experience vector and reliably infering the rest of2852 the senses.2854 ** ``Empathy'' requires retracing steps though \Phi-space2856 Here is the core of a basic empathy algorithm, starting with an2857 experience vector:2859 First, group the experiences into tiered proprioceptive bins. I use2860 powers of 10 and 3 bins, and the smallest bin has an approximate2861 size of 0.001 radians in all proprioceptive dimensions.2863 Then, given a sequence of proprioceptive input, generate a set of2864 matching experience records for each input, using the tiered2865 proprioceptive bins.2867 Finally, to infer sensory data, select the longest consective chain2868 of experiences. Conecutive experience means that the experiences2869 appear next to each other in the experience vector.2871 This algorithm has three advantages:2873 1. It's simple2875 3. It's very fast -- retrieving possible interpretations takes2876 constant time. Tracing through chains of interpretations takes2877 time proportional to the average number of experiences in a2878 proprioceptive bin. Redundant experiences in \Phi-space can be2879 merged to save computation.2881 2. It protects from wrong interpretations of transient ambiguous2882 proprioceptive data. For example, if the worm is flat for just2883 an instant, this flattness will not be interpreted as implying2884 that the worm has its muscles relaxed, since the flattness is2885 part of a longer chain which includes a distinct pattern of2886 muscle activation. Markov chains or other memoryless statistical2887 models that operate on individual frames may very well make this2888 mistake.2890 #+caption: Program to convert an experience vector into a2891 #+caption: proprioceptively binned lookup function.2892 #+name: bin2893 #+begin_listing clojure2894 #+begin_src clojure2895 (defn bin [digits]2896 (fn [angles]2897 (->> angles2898 (flatten)2899 (map (juxt #(Math/sin %) #(Math/cos %)))2900 (flatten)2901 (mapv #(Math/round (* % (Math/pow 10 (dec digits))))))))2903 (defn gen-phi-scan2904 "Nearest-neighbors with binning. Only returns a result if2905 the propriceptive data is within 10% of a previously recorded2906 result in all dimensions."2907 [phi-space]2908 (let [bin-keys (map bin [3 2 1])2909 bin-maps2910 (map (fn [bin-key]2911 (group-by2912 (comp bin-key :proprioception phi-space)2913 (range (count phi-space)))) bin-keys)2914 lookups (map (fn [bin-key bin-map]2915 (fn [proprio] (bin-map (bin-key proprio))))2916 bin-keys bin-maps)]2917 (fn lookup [proprio-data]2918 (set (some #(% proprio-data) lookups)))))2919 #+end_src2920 #+end_listing2922 #+caption: =longest-thread= finds the longest path of consecutive2923 #+caption: experiences to explain proprioceptive worm data.2924 #+name: phi-space-history-scan2925 #+ATTR_LaTeX: :width 10cm2926 [[./images/aurellem-gray.png]]2928 =longest-thread= infers sensory data by stitching together pieces2929 from previous experience. It prefers longer chains of previous2930 experience to shorter ones. For example, during training the worm2931 might rest on the ground for one second before it performs its2932 excercises. If during recognition the worm rests on the ground for2933 five seconds, =longest-thread= will accomodate this five second2934 rest period by looping the one second rest chain five times.2936 =longest-thread= takes time proportinal to the average number of2937 entries in a proprioceptive bin, because for each element in the2938 starting bin it performes a series of set lookups in the preceeding2939 bins. If the total history is limited, then this is only a constant2940 multiple times the number of entries in the starting bin. This2941 analysis also applies even if the action requires multiple longest2942 chains -- it's still the average number of entries in a2943 proprioceptive bin times the desired chain length. Because2944 =longest-thread= is so efficient and simple, I can interpret2945 worm-actions in real time.2947 #+caption: Program to calculate empathy by tracing though \Phi-space2948 #+caption: and finding the longest (ie. most coherent) interpretation2949 #+caption: of the data.2950 #+name: longest-thread2951 #+begin_listing clojure2952 #+begin_src clojure2953 (defn longest-thread2954 "Find the longest thread from phi-index-sets. The index sets should2955 be ordered from most recent to least recent."2956 [phi-index-sets]2957 (loop [result '()2958 [thread-bases & remaining :as phi-index-sets] phi-index-sets]2959 (if (empty? phi-index-sets)2960 (vec result)2961 (let [threads2962 (for [thread-base thread-bases]2963 (loop [thread (list thread-base)2964 remaining remaining]2965 (let [next-index (dec (first thread))]2966 (cond (empty? remaining) thread2967 (contains? (first remaining) next-index)2968 (recur2969 (cons next-index thread) (rest remaining))2970 :else thread))))2971 longest-thread2972 (reduce (fn [thread-a thread-b]2973 (if (> (count thread-a) (count thread-b))2974 thread-a thread-b))2975 '(nil)2976 threads)]2977 (recur (concat longest-thread result)2978 (drop (count longest-thread) phi-index-sets))))))2979 #+end_src2980 #+end_listing2982 There is one final piece, which is to replace missing sensory data2983 with a best-guess estimate. While I could fill in missing data by2984 using a gradient over the closest known sensory data points,2985 averages can be misleading. It is certainly possible to create an2986 impossible sensory state by averaging two possible sensory states.2987 Therefore, I simply replicate the most recent sensory experience to2988 fill in the gaps.2990 #+caption: Fill in blanks in sensory experience by replicating the most2991 #+caption: recent experience.2992 #+name: infer-nils2993 #+begin_listing clojure2994 #+begin_src clojure2995 (defn infer-nils2996 "Replace nils with the next available non-nil element in the2997 sequence, or barring that, 0."2998 [s]2999 (loop [i (dec (count s))3000 v (transient s)]3001 (if (zero? i) (persistent! v)3002 (if-let [cur (v i)]3003 (if (get v (dec i) 0)3004 (recur (dec i) v)3005 (recur (dec i) (assoc! v (dec i) cur)))3006 (recur i (assoc! v i 0))))))3007 #+end_src3008 #+end_listing3010 ** =EMPATH= recognizes actions efficiently3012 To use =EMPATH= with the worm, I first need to gather a set of3013 experiences from the worm that includes the actions I want to3014 recognize. The =generate-phi-space= program (listing3015 \ref{generate-phi-space} runs the worm through a series of3016 exercices and gatheres those experiences into a vector. The3017 =do-all-the-things= program is a routine expressed in a simple3018 muscle contraction script language for automated worm control. It3019 causes the worm to rest, curl, and wiggle over about 700 frames3020 (approx. 11 seconds).3022 #+caption: Program to gather the worm's experiences into a vector for3023 #+caption: further processing. The =motor-control-program= line uses3024 #+caption: a motor control script that causes the worm to execute a series3025 #+caption: of ``exercices'' that include all the action predicates.3026 #+name: generate-phi-space3027 #+begin_listing clojure3028 #+begin_src clojure3029 (def do-all-the-things3030 (concat3031 curl-script3032 [[300 :d-ex 40]3033 [320 :d-ex 0]]3034 (shift-script 280 (take 16 wiggle-script))))3036 (defn generate-phi-space []3037 (let [experiences (atom [])]3038 (run-world3039 (apply-map3040 worm-world3041 (merge3042 (worm-world-defaults)3043 {:end-frame 7003044 :motor-control3045 (motor-control-program worm-muscle-labels do-all-the-things)3046 :experiences experiences})))3047 @experiences))3048 #+end_src3049 #+end_listing3051 #+caption: Use longest thread and a phi-space generated from a short3052 #+caption: exercise routine to interpret actions during free play.3053 #+name: empathy-debug3054 #+begin_listing clojure3055 #+begin_src clojure3056 (defn init []3057 (def phi-space (generate-phi-space))3058 (def phi-scan (gen-phi-scan phi-space)))3060 (defn empathy-demonstration []3061 (let [proprio (atom ())]3062 (fn3063 [experiences text]3064 (let [phi-indices (phi-scan (:proprioception (peek experiences)))]3065 (swap! proprio (partial cons phi-indices))3066 (let [exp-thread (longest-thread (take 300 @proprio))3067 empathy (mapv phi-space (infer-nils exp-thread))]3068 (println-repl (vector:last-n exp-thread 22))3069 (cond3070 (grand-circle? empathy) (.setText text "Grand Circle")3071 (curled? empathy) (.setText text "Curled")3072 (wiggling? empathy) (.setText text "Wiggling")3073 (resting? empathy) (.setText text "Resting")3074 :else (.setText text "Unknown")))))))3076 (defn empathy-experiment [record]3077 (.start (worm-world :experience-watch (debug-experience-phi)3078 :record record :worm worm*)))3079 #+end_src3080 #+end_listing3082 The result of running =empathy-experiment= is that the system is3083 generally able to interpret worm actions using the action-predicates3084 on simulated sensory data just as well as with actual data. Figure3085 \ref{empathy-debug-image} was generated using =empathy-experiment=:3087 #+caption: From only proprioceptive data, =EMPATH= was able to infer3088 #+caption: the complete sensory experience and classify four poses3089 #+caption: (The last panel shows a composite image of \emph{wriggling},3090 #+caption: a dynamic pose.)3091 #+name: empathy-debug-image3092 #+ATTR_LaTeX: :width 10cm :placement [H]3093 [[./images/empathy-1.png]]3095 One way to measure the performance of =EMPATH= is to compare the3096 sutiability of the imagined sense experience to trigger the same3097 action predicates as the real sensory experience.3099 #+caption: Determine how closely empathy approximates actual3100 #+caption: sensory data.3101 #+name: test-empathy-accuracy3102 #+begin_listing clojure3103 #+begin_src clojure3104 (def worm-action-label3105 (juxt grand-circle? curled? wiggling?))3107 (defn compare-empathy-with-baseline [matches]3108 (let [proprio (atom ())]3109 (fn3110 [experiences text]3111 (let [phi-indices (phi-scan (:proprioception (peek experiences)))]3112 (swap! proprio (partial cons phi-indices))3113 (let [exp-thread (longest-thread (take 300 @proprio))3114 empathy (mapv phi-space (infer-nils exp-thread))3115 experience-matches-empathy3116 (= (worm-action-label experiences)3117 (worm-action-label empathy))]3118 (println-repl experience-matches-empathy)3119 (swap! matches #(conj % experience-matches-empathy)))))))3121 (defn accuracy [v]3122 (float (/ (count (filter true? v)) (count v))))3124 (defn test-empathy-accuracy []3125 (let [res (atom [])]3126 (run-world3127 (worm-world :experience-watch3128 (compare-empathy-with-baseline res)3129 :worm worm*))3130 (accuracy @res)))3131 #+end_src3132 #+end_listing3134 Running =test-empathy-accuracy= using the very short exercise3135 program defined in listing \ref{generate-phi-space}, and then doing3136 a similar pattern of activity manually yeilds an accuracy of around3137 73%. This is based on very limited worm experience. By training the3138 worm for longer, the accuracy dramatically improves.3140 #+caption: Program to generate \Phi-space using manual training.3141 #+name: manual-phi-space3142 #+begin_listing clojure3143 #+begin_src clojure3144 (defn init-interactive []3145 (def phi-space3146 (let [experiences (atom [])]3147 (run-world3148 (apply-map3149 worm-world3150 (merge3151 (worm-world-defaults)3152 {:experiences experiences})))3153 @experiences))3154 (def phi-scan (gen-phi-scan phi-space)))3155 #+end_src3156 #+end_listing3158 After about 1 minute of manual training, I was able to achieve 95%3159 accuracy on manual testing of the worm using =init-interactive= and3160 =test-empathy-accuracy=. The majority of errors are near the3161 boundaries of transitioning from one type of action to another.3162 During these transitions the exact label for the action is more open3163 to interpretation, and dissaggrement between empathy and experience3164 is more excusable.3166 ** Digression: Learn touch sensor layout through haptic experimentation, instead3167 # Boostraping touch using free exploration3168 In the previous section I showed how to compute actions in terms of3169 body-centered predicates which relied averate touch activation of3170 pre-defined regions of the worm's skin. What if, instead of recieving3171 touch pre-grouped into the six faces of each worm segment, the true3172 topology of the worm's skin was unknown? This is more similiar to how3173 a nerve fiber bundle might be arranged. While two fibers that are3174 close in a nerve bundle /might/ correspond to two touch sensors that3175 are close together on the skin, the process of taking a complicated3176 surface and forcing it into essentially a circle requires some cuts3177 and rerragenments.3179 In this section I show how to automatically learn the skin-topology of3180 a worm segment by free exploration. As the worm rolls around on the3181 floor, large sections of its surface get activated. If the worm has3182 stopped moving, then whatever region of skin that is touching the3183 floor is probably an important region, and should be recorded.3185 #+caption: Program to detect whether the worm is in a resting state3186 #+caption: with one face touching the floor.3187 #+name: pure-touch3188 #+begin_listing clojure3189 #+begin_src clojure3190 (def full-contact [(float 0.0) (float 0.1)])3192 (defn pure-touch?3193 "This is worm specific code to determine if a large region of touch3194 sensors is either all on or all off."3195 [[coords touch :as touch-data]]3196 (= (set (map first touch)) (set full-contact)))3197 #+end_src3198 #+end_listing3200 After collecting these important regions, there will many nearly3201 similiar touch regions. While for some purposes the subtle3202 differences between these regions will be important, for my3203 purposes I colapse them into mostly non-overlapping sets using3204 =remove-similiar= in listing \ref{remove-similiar}3206 #+caption: Program to take a lits of set of points and ``collapse them''3207 #+caption: so that the remaining sets in the list are siginificantly3208 #+caption: different from each other. Prefer smaller sets to larger ones.3209 #+name: remove-similiar3210 #+begin_listing clojure3211 #+begin_src clojure3212 (defn remove-similar3213 [coll]3214 (loop [result () coll (sort-by (comp - count) coll)]3215 (if (empty? coll) result3216 (let [[x & xs] coll3217 c (count x)]3218 (if (some3219 (fn [other-set]3220 (let [oc (count other-set)]3221 (< (- (count (union other-set x)) c) (* oc 0.1))))3222 xs)3223 (recur result xs)3224 (recur (cons x result) xs))))))3225 #+end_src3226 #+end_listing3228 Actually running this simulation is easy given =CORTEX='s facilities.3230 #+caption: Collect experiences while the worm moves around. Filter the touch3231 #+caption: sensations by stable ones, collapse similiar ones together,3232 #+caption: and report the regions learned.3233 #+name: learn-touch3234 #+begin_listing clojure3235 #+begin_src clojure3236 (defn learn-touch-regions []3237 (let [experiences (atom [])3238 world (apply-map3239 worm-world3240 (assoc (worm-segment-defaults)3241 :experiences experiences))]3242 (run-world world)3243 (->>3244 @experiences3245 (drop 175)3246 ;; access the single segment's touch data3247 (map (comp first :touch))3248 ;; only deal with "pure" touch data to determine surfaces3249 (filter pure-touch?)3250 ;; associate coordinates with touch values3251 (map (partial apply zipmap))3252 ;; select those regions where contact is being made3253 (map (partial group-by second))3254 (map #(get % full-contact))3255 (map (partial map first))3256 ;; remove redundant/subset regions3257 (map set)3258 remove-similar)))3260 (defn learn-and-view-touch-regions []3261 (map view-touch-region3262 (learn-touch-regions)))3263 #+end_src3264 #+end_listing3266 The only thing remining to define is the particular motion the worm3267 must take. I accomplish this with a simple motor control program.3269 #+caption: Motor control program for making the worm roll on the ground.3270 #+caption: This could also be replaced with random motion.3271 #+name: worm-roll3272 #+begin_listing clojure3273 #+begin_src clojure3274 (defn touch-kinesthetics []3275 [[170 :lift-1 40]3276 [190 :lift-1 19]3277 [206 :lift-1 0]3279 [400 :lift-2 40]3280 [410 :lift-2 0]3282 [570 :lift-2 40]3283 [590 :lift-2 21]3284 [606 :lift-2 0]3286 [800 :lift-1 30]3287 [809 :lift-1 0]3289 [900 :roll-2 40]3290 [905 :roll-2 20]3291 [910 :roll-2 0]3293 [1000 :roll-2 40]3294 [1005 :roll-2 20]3295 [1010 :roll-2 0]3297 [1100 :roll-2 40]3298 [1105 :roll-2 20]3299 [1110 :roll-2 0]3300 ])3301 #+end_src3302 #+end_listing3305 #+caption: The small worm rolls around on the floor, driven3306 #+caption: by the motor control program in listing \ref{worm-roll}.3307 #+name: worm-roll3308 #+ATTR_LaTeX: :width 12cm3309 [[./images/worm-roll.png]]3312 #+caption: After completing its adventures, the worm now knows3313 #+caption: how its touch sensors are arranged along its skin. These3314 #+caption: are the regions that were deemed important by3315 #+caption: =learn-touch-regions=. Note that the worm has discovered3316 #+caption: that it has six sides.3317 #+name: worm-touch-map3318 #+ATTR_LaTeX: :width 12cm3319 [[./images/touch-learn.png]]3321 While simple, =learn-touch-regions= exploits regularities in both3322 the worm's physiology and the worm's environment to correctly3323 deduce that the worm has six sides. Note that =learn-touch-regions=3324 would work just as well even if the worm's touch sense data were3325 completely scrambled. The cross shape is just for convienence. This3326 example justifies the use of pre-defined touch regions in =EMPATH=.3328 * Contributions3330 In this thesis you have seen the =CORTEX= system, a complete3331 environment for creating simulated creatures. You have seen how to3332 implement five senses: touch, proprioception, hearing, vision, and3333 muscle tension. You have seen how to create new creatues using3334 blender, a 3D modeling tool. I hope that =CORTEX= will be useful in3335 further research projects. To this end I have included the full3336 source to =CORTEX= along with a large suite of tests and examples. I3337 have also created a user guide for =CORTEX= which is inculded in an3338 appendix to this thesis \ref{}.3339 # dxh: todo reference appendix3341 You have also seen how I used =CORTEX= as a platform to attach the3342 /action recognition/ problem, which is the problem of recognizing3343 actions in video. You saw a simple system called =EMPATH= which3344 ientifies actions by first describing actions in a body-centerd,3345 rich sense language, then infering a full range of sensory3346 experience from limited data using previous experience gained from3347 free play.3349 As a minor digression, you also saw how I used =CORTEX= to enable a3350 tiny worm to discover the topology of its skin simply by rolling on3351 the ground.3353 In conclusion, the main contributions of this thesis are:3355 - =CORTEX=, a system for creating simulated creatures with rich3356 senses.3357 - =EMPATH=, a program for recognizing actions by imagining sensory3358 experience.3360 # An anatomical joke:3361 # - Training3362 # - Skeletal imitation3363 # - Sensory fleshing-out3364 # - Classification3365 #+BEGIN_LaTeX3366 \appendix3367 #+END_LaTeX3368 * Appendix: =CORTEX= User Guide3370 Those who write a thesis should endeavor to make their code not only3371 accessable, but actually useable, as a way to pay back the community3372 that made the thesis possible in the first place. This thesis would3373 not be possible without Free Software such as jMonkeyEngine3,3374 Blender, clojure, emacs, ffmpeg, and many other tools. That is why I3375 have included this user guide, in the hope that someone else might3376 find =CORTEX= useful.3378 ** Obtaining =CORTEX=3380 You can get cortex from its mercurial repository at3381 http://hg.bortreb.com/cortex. You may also download =CORTEX=3382 releases at http://aurellem.org/cortex/releases/. As a condition of3383 making this thesis, I have also provided Professor Winston the3384 =CORTEX= source, and he knows how to run the demos and get started.3385 You may also email me at =cortex@aurellem.org= and I may help where3386 I can.3388 ** Running =CORTEX=3390 =CORTEX= comes with README and INSTALL files that will guide you3391 through installation and running the test suite. In particular you3392 should look at test =cortex.test= which contains test suites that3393 run through all senses and multiple creatures.3395 ** Creating creatures3397 Creatures are created using /Blender/, a free 3D modeling program.3398 You will need Blender version 2.6 when using the =CORTEX= included3399 in this thesis. You create a =CORTEX= creature in a similiar manner3400 to modeling anything in Blender, except that you also create3401 several trees of empty nodes which define the creature's senses.3403 *** Mass3405 To give an object mass in =CORTEX=, add a ``mass'' metadata label3406 to the object with the mass in jMonkeyEngine units. Note that3407 setting the mass to 0 causes the object to be immovable.3409 *** Joints3411 Joints are created by creating an empty node named =joints= and3412 then creating any number of empty child nodes to represent your3413 creature's joints. The joint will automatically connect the3414 closest two physical objects. It will help to set the empty node's3415 display mode to ``Arrows'' so that you can clearly see the3416 direction of the axes.3418 Joint nodes should have the following metadata under the ``joint''3419 label:3421 #+BEGIN_SRC clojure3422 ;; ONE OF the following, under the label "joint":3423 {:type :point}3425 ;; OR3427 {:type :hinge3428 :limit [<limit-low> <limit-high>]3429 :axis (Vector3f. <x> <y> <z>)}3430 ;;(:axis defaults to (Vector3f. 1 0 0) if not provided for hinge joints)3432 ;; OR3434 {:type :cone3435 :limit-xz <lim-xz>3436 :limit-xy <lim-xy>3437 :twist <lim-twist>} ;(use XZY rotation mode in blender!)3438 #+END_SRC3440 *** Eyes3442 Eyes are created by creating an empty node named =eyes= and then3443 creating any number of empty child nodes to represent your3444 creature's eyes.3446 Eye nodes should have the following metadata under the ``eye''3447 label:3449 #+BEGIN_SRC clojure3450 {:red <red-retina-definition>3451 :blue <blue-retina-definition>3452 :green <green-retina-definition>3453 :all <all-retina-definition>3454 (<0xrrggbb> <custom-retina-image>)...3455 }3456 #+END_SRC3458 Any of the color channels may be omitted. You may also include3459 your own color selectors, and in fact :red is equivalent to3460 0xFF0000 and so forth. The eye will be placed at the same position3461 as the empty node and will bind to the neatest physical object.3462 The eye will point outward from the X-axis of the node, and ``up''3463 will be in the direction of the X-axis of the node. It will help3464 to set the empty node's display mode to ``Arrows'' so that you can3465 clearly see the direction of the axes.3467 Each retina file should contain white pixels whever you want to be3468 sensitive to your chosen color. If you want the entire field of3469 view, specify :all of 0xFFFFFF and a retinal map that is entirely3470 white.3472 Here is a sample retinal map:3474 #+caption: An example retinal profile image. White pixels are3475 #+caption: photo-sensitive elements. The distribution of white3476 #+caption: pixels is denser in the middle and falls off at the3477 #+caption: edges and is inspired by the human retina.3478 #+name: retina3479 #+ATTR_LaTeX: :width 7cm :placement [H]3480 [[./images/retina-small.png]]3482 *** Hearing3484 Ears are created by creating an empty node named =ears= and then3485 creating any number of empty child nodes to represent your3486 creature's ears.3488 Ear nodes do not require any metadata.3490 The ear will bind to and follow the closest physical node.3492 *** Touch3494 Touch is handled similarly to mass. To make a particular object3495 touch sensitive, add metadata of the following form under the3496 object's ``touch'' metadata field:3498 #+BEGIN_EXAMPLE3499 <touch-UV-map-file-name>3500 #+END_EXAMPLE3502 You may also include an optional ``scale'' metadata number to3503 specifiy the length of the touch feelers. The default is $0.1$,3504 and this is generally sufficient.3506 The touch UV should contain white pixels for each touch sensor.3508 Here is an example touch-uv map that approximates a human finger,3509 and its corresponding model.3511 #+caption: This is the tactile-sensor-profile for the upper segment3512 #+caption: of a fingertip. It defines regions of high touch sensitivity3513 #+caption: (where there are many white pixels) and regions of low3514 #+caption: sensitivity (where white pixels are sparse).3515 #+name: guide-fingertip-UV3516 #+ATTR_LaTeX: :width 9cm :placement [H]3517 [[./images/finger-UV.png]]3519 #+caption: The fingertip UV-image form above applied to a simple3520 #+caption: model of a fingertip.3521 #+name: guide-fingertip3522 #+ATTR_LaTeX: :width 9cm :placement [H]3523 [[./images/finger-2.png]]3525 *** Propriocepotion3527 Proprioception is tied to each joint node -- nothing special must3528 be done in a blender model to enable proprioception other than3529 creating joint nodes.3531 *** Muscles3533 Muscles are created by creating an empty node named =muscles= and3534 then creating any number of empty child nodes to represent your3535 creature's muscles.3538 Muscle nodes should have the following metadata under the3539 ``muscle'' label:3541 #+BEGIN_EXAMPLE3542 <muscle-profile-file-name>3543 #+END_EXAMPLE3545 Muscles should also have a ``strength'' metadata entry describing3546 the muscle's total strength at full activation.3548 Muscle profiles are simple images that contain the relative amount3549 of muscle power in each simulated alpha motor neuron. The width of3550 the image is the total size of the motor pool, and the redness of3551 each neuron is the relative power of that motor pool.3553 While the profile image can have any dimensions, only the first3554 line of pixels is used to define the muscle. Here is a sample3555 muscle profile image that defines a human-like muscle.3557 #+caption: A muscle profile image that describes the strengths3558 #+caption: of each motor neuron in a muscle. White is weakest3559 #+caption: and dark red is strongest. This particular pattern3560 #+caption: has weaker motor neurons at the beginning, just3561 #+caption: like human muscle.3562 #+name: muscle-recruit3563 #+ATTR_LaTeX: :width 7cm :placement [H]3564 [[./images/basic-muscle.png]]3566 Muscles twist the nearest physical object about the muscle node's3567 Z-axis. I recommend using the ``Single Arrow'' display mode for3568 muscles and using the right hand rule to determine which way the3569 muscle will twist. To make a segment that can twist in multiple3570 directions, create multiple, differently aligned muscles.3572 ** =CORTEX= API3574 These are the some functions exposed by =CORTEX= for creating3575 worlds and simulating creatures. These are in addition to3576 jMonkeyEngine3's extensive library, which is documented elsewhere.3578 *** Simulation3579 - =(world root-node key-map setup-fn update-fn)= :: create3580 a simulation.3581 - /root-node/ :: a =com.jme3.scene.Node= object which3582 contains all of the objects that should be in the3583 simulation.3585 - /key-map/ :: a map from strings describing keys to3586 functions that should be executed whenever that key is3587 pressed. the functions should take a SimpleApplication3588 object and a boolean value. The SimpleApplication is the3589 current simulation that is running, and the boolean is true3590 if the key is being pressed, and false if it is being3591 released. As an example,3592 #+BEGIN_SRC clojure3593 {"key-j" (fn [game value] (if value (println "key j pressed")))}3594 #+END_SRC3595 is a valid key-map which will cause the simulation to print3596 a message whenever the 'j' key on the keyboard is pressed.3598 - /setup-fn/ :: a function that takes a =SimpleApplication=3599 object. It is called once when initializing the simulation.3600 Use it to create things like lights, change the gravity,3601 initialize debug nodes, etc.3603 - /update-fn/ :: this function takes a =SimpleApplication=3604 object and a float and is called every frame of the3605 simulation. The float tells how many seconds is has been3606 since the last frame was rendered, according to whatever3607 clock jme is currently using. The default is to use IsoTimer3608 which will result in this value always being the same.3610 - =(position-camera world position rotation)= :: set the position3611 of the simulation's main camera.3613 - =(enable-debug world)= :: turn on debug wireframes for each3614 simulated object.3616 - =(set-gravity world gravity)= :: set the gravity of a running3617 simulation.3619 - =(box length width height & {options})= :: create a box in the3620 simulation. Options is a hash map specifying texture, mass,3621 etc. Possible options are =:name=, =:color=, =:mass=,3622 =:friction=, =:texture=, =:material=, =:position=,3623 =:rotation=, =:shape=, and =:physical?=.3625 - =(sphere radius & {options})= :: create a sphere in the simulation.3626 Options are the same as in =box=.3628 - =(load-blender-model file-name)= :: create a node structure3629 representing that described in a blender file.3631 - =(light-up-everything world)= :: distribute a standard compliment3632 of lights throught the simulation. Should be adequate for most3633 purposes.3635 - =(node-seq node)= :: return a recursuve list of the node's3636 children.3638 - =(nodify name children)= :: construct a node given a node-name and3639 desired children.3641 - =(add-element world element)= :: add an object to a running world3642 simulation.3644 - =(set-accuracy world accuracy)= :: change the accuracy of the3645 world's physics simulator.3647 - =(asset-manager)= :: get an /AssetManager/, a jMonkeyEngine3648 construct that is useful for loading textures and is required3649 for smooth interaction with jMonkeyEngine library functions.3651 - =(load-bullet)= :: unpack native libraries and initialize3652 blender. This function is required before other world building3653 functions are called.3655 *** Creature Manipulation / Import3657 - =(body! creature)= :: give the creature a physical body.3659 - =(vision! creature)= :: give the creature a sense of vision.3660 Returns a list of functions which will each, when called3661 during a simulation, return the vision data for the channel of3662 one of the eyes. The functions are ordered depending on the3663 alphabetical order of the names of the eye nodes in the3664 blender file. The data returned by the functions is a vector3665 containing the eye's /topology/, a vector of coordinates, and3666 the eye's /data/, a vector of RGB values filtered by the eye's3667 sensitivity.3669 - =(hearing! creature)= :: give the creature a sense of hearing.3670 Returns a list of functions, one for each ear, that when3671 called will return a frame's worth of hearing data for that3672 ear. The functions are ordered depending on the alphabetical3673 order of the names of the ear nodes in the blender file. The3674 data returned by the functions is an array PCM encoded wav3675 data.3677 - =(touch! creature)= :: give the creature a sense of touch. Returns3678 a single function that must be called with the /root node/ of3679 the world, and which will return a vector of /touch-data/3680 one entry for each touch sensitive component, each entry of3681 which contains a /topology/ that specifies the distribution of3682 touch sensors, and the /data/, which is a vector of3683 =[activation, length]= pairs for each touch hair.3685 - =(proprioception! creature)= :: give the creature the sense of3686 proprioception. Returns a list of functions, one for each3687 joint, that when called during a running simulation will3688 report the =[headnig, pitch, roll]= of the joint.3690 - =(movement! creature)= :: give the creature the power of movement.3691 Creates a list of functions, one for each muscle, that when3692 called with an integer, will set the recruitment of that3693 muscle to that integer, and will report the current power3694 being exerted by the muscle. Order of muscles is determined by3695 the alphabetical sort order of the names of the muscle nodes.3697 *** Visualization/Debug3699 - =(view-vision)= :: create a function that when called with a list3700 of visual data returned from the functions made by =vision!=,3701 will display that visual data on the screen.3703 - =(view-hearing)= :: same as =view-vision= but for hearing.3705 - =(view-touch)= :: same as =view-vision= but for touch.3707 - =(view-proprioception)= :: same as =view-vision= but for3708 proprioception.3710 - =(view-movement)= :: same as =view-vision= but for3711 proprioception.3713 - =(view anything)= :: =view= is a polymorphic function that allows3714 you to inspect almost anything you could reasonably expect to3715 be able to ``see'' in =CORTEX=.3717 - =(text anything)= :: =text= is a polymorphic function that allows3718 you to convert practically anything into a text string.3720 - =(println-repl anything)= :: print messages to clojure's repl3721 instead of the simulation's terminal window.3723 - =(mega-import-jme3)= :: for experimenting at the REPL. This3724 function will import all jMonkeyEngine3 classes for immediate3725 use.3727 - =(display-dialated-time world timer)= :: Shows the time as it is3728 flowing in the simulation on a HUD display.