December 2004

AI Newsletter

• 22 min read

Introduction

Welcome to the December AI Expert. In this issue, I look at programming the Sony Aibo robot dog, the World Wide Mind, and (with some help from Hitch Hiker's Guide to the Galaxy author Douglas Adams), Ronald Reagan. But first, some truly remarkable neural network news from the University of Florida. It's probably not yet feasible to do this at home, but it would be great to see the technology taken up by private experimenters.

Somewhere in Florida, Disembodied Rat Neurons Dream of Flying a Fighter Jet

Thus begins a feature in Wired's on this month's lead news. As reported there, and in the University of Florida News for October 21, Thomas DeMarse, a professor of biomedical engineering at Florida, has cultured a collection of 25,000 rat neurons on an electrode array, and then trained them to fly a simulated F-22 fighter jet.

When DeMarse first put the neurons in the culture dish, they looked like little more than grains of sand sprinkled in water. However, individual neurons soon begin to extend microscopic lines toward each other and make synaptic connections:

You see one extend a process, pull it back, extend it out - and it may do that a couple of times, just sampling who’s next to it, until over time the connectivity starts to establish itself.

The electrode array on which the neurons are growing is connected to a desktop PC running the fighter simulator. This sends them information about flight conditions: whether the plane is flying straight and level or is tilted to the left or to the right. The electrodes pick up the neurons' responses and send them back to the simulator, thus controlling the fighter:

Initially when we hook up this brain to a flight simulator, it doesn’t know how to control the aircraft. So you hook it up and the aircraft simply drifts randomly. And as the data comes in, it slowly modifies the (neural) network so over time, the network gradually learns to fly the aircraft.

It's always risky to accept popular press accounts of research, and the peer-reviewed paper on this work has apparently not yet been published. However, some information can be found. A brief account of the hardware on the university site tells us that the electrode array, supplied by Multi Channel Systems, consists of 60 electrodes, each capable of both recording and stimulating neural activity. The electrodes, 30 microns wide, were spaced 200 microns apart in an 8-by-8 grid. This kind of array was not developed solely for probing neural activation: a Natural and Medical Sciences Institute page of the University of Tübingen indicates their use on heart and retinal cells. Pharmacology is one application, using the electrodes to examine how drugs affect the patterns of activity in the tissue.

How can the neurons know to make signals the fighter simulator can understand? If the setup is similar to that in a paper on previous research, The Neurally Controlled Animat: Biological Brains Acting with Simulated Bodies, then there is cooperation between the computer and the network. The paper describes how the neurons - derived from 18-day old rat embryos - were cultured, forming after a few days a complex network. After one month, this had become relatively stable in its activity. Via the electrodes, the neurons were connected to a simulated animal living within a room with four walls and a few internal barriers. The computer analysed neural activity recorded by the electrodes into sequences of action-potential spikes. These spike trains were then passed through a clustering algorithm to detect frequently-occurring spike patterns. Some of these patterns were allocated arbitrarily to motion commands: forward, back, and so on. When the simulated animal received sensory input, e.g. by hitting a wall, this was fed back to the electrodes and hence to the neurons, thus creating a continuously running cells-to-computer-to-cells feedback system.

The aim of this research is not to replace US Air Force pilots by dishfuls of rat cortical cells, but to help understand signals and coding in the nervous system. That said, the research is exciting purely for its technological potential. Most computing technology eventually moves out of the lab to the hackers, and if it can be done without hurting animals (one can envisage a neural-culture supply house, perhaps...) I would love to see this taken up and developed by home hobbyists.

www.wired.com/news/medtech/0,1286,65438,00.html - Wired's write-up. See www.napa.ufl.edu/2004news/braindish.htm for the original University of Florida news item.

www.bme.ufl.edu/people/detailperson.php?PEOPLE_id=2 - Home page for Thomas DeMarse.

www.bme.ufl.edu/research/projects/detailproject.php?RP_id=4 - Summary of how the multi-channel arrays were used, with a photograph.

www.multichannelsystems.com/ - Home page for Multi Channel Systems.

www.nmi.de/englisch/showprj.php3?id=22&typ=1 - Application of microelectrode array (MEA) technology in pharmaceutical biotechnology, University of Tübingen summary.

www.ecmjournal.org/journal/supplements/vol006supp01/pdf/38.pdf - Micropatterned Neural Nets on Different Solid Surfaces, poster summary from the Max Planck Institute for Polymer Research.

www.bme.ufl.edu/documents/the_neurally_10.pdf - The Neurally Controlled Animat: Biological Brains Acting with Simulated Bodies. This is a readable paper by DeMarse and colleagues on the earlier experiment in which neurons were used to control a simulated animal in its virtual world. It describes how signals from the neurons were interpreted by the software.

www.neuro.gatech.edu/groups/potter/potter.html - Home page for Steve Potter, one of DeMarse's collaborators. There are some interesting links on consciousness, artificial life, self-organising systems and other topics.

www.neurodudes.com/archives/000068.html - Posting (which introduced me to the word "hybrot") about this work from the Neurodude list, "at the intersection of neuroscience and AI". This list seems to carry interesting news from time to time.

http://science.slashdot.org/article.pl?sid=04/10/24/0024241&tid=191&tid=126&tid=14 - Copious Slashdot postings on the research.

World Wide Mind

The World Wide Mind is a project that aims to help people collaborate on AI by sharing components over the Web. As the originators say on their site,

This work proposes that the construction of advanced artificial minds may be too difficult for any single laboratory to complete. At the moment, no easy system exists whereby a working mind can be made from the components of two or more laboratories. This system aims to change that, and accelerate the growth of Artificial Intelligence, once the requirement that a single laboratory understand the entire system is removed.

This is something that looks valuable - and fun - for teaching AI, so let me say a bit about the implementation.

Mind servers and world servers

The idea is that collaborators provide either a mind (or part of a mind) or a world in which the mind can operate. Such a component is called a "service". Services use standard HTTP protocol: in other words, they run on Web servers, and can be implemented using CGI, servlets, and other standard authoring techniques.

To ensure smooth collaboration, services must follow a standard protocol. Each service can be asked to perform various actions and send back various types of information. As an example, to create a world, one sends its server a newrun message. The world-simulator on the server is duty-bound by the protocol to understand this message and to reply by creating a new instance of the world and sending back an ID for it.

A world must also understand the getstate message, which asks it to return its current state, and the takeaction message, which tells it to perform a given action - such as moving a block in a blocks world - and update its state accordingly. These ideas will be familiar to anyone who has built simulated environments for AIs. There are various other messages, all described in the documentation linked at the end of this article.

Not surprisingly, given their autonomy, minds respond to a different set of messages. It's still necessary to create an instance before use, but one then sends it a getaction message, together with the ID for a world. This instructs it to "think" about what it must do next. The action it returns can then be fed back to its world by a takeaction message.

In order that a user can access the same instance of a mind or world in each transaction, its author preserve state on the server across transactions. This involves session-handling, which all Web authoring systems now provide.

Society Of Mind Markup language

To accomodate arbitrarily complicated messages and replies, the World Wide Mind authors have defined an XML-based language called SOML, or "Society Of Mind Markup Language". This name is taken from Marvin Minsky's book The Society of Mind, which envisages human minds as a society of collaborating agents. For those who don't know XML, there's no mystery to it: it's just a way to encode feature-attribute tree structures in a standard text-based (and highly verbose) syntax. Any programming language these days will have tools for parsing XML. The World Wide mind site includes a definition of SOML. Note incidentally that although SOML is standardised, the way worlds and minds convey their states and actions isn't, as these are carried as data within the XML. One world might describe itself using Prolog predicates, while another sends back vectors of object colour-location-size triples.

Minds and subminds

World Wide Mind had the objective of making it easy to build AIs out of shared components, and so a mind can call upon any number of subminds, wherever they are in the world. To do so, it just uses SOML and messages as above, instantiating a submind and passing it a request for action.

Try it for yourself

Any system that uses the Web for its internal messaging is never going to push the envelope of speed - the phrase "vaster than continents and more slow" comes to mind. However, it does look extremely useful for teaching, enabling teachers and students to share components and build upon one another's work. It could also inspire some interesting competitions, so why not have a go for yourself? The site comes with two example worlds. One was built by ALifer Toby Tyrrell as part of his University of Edinburgh research into action-selection mechanisms, and models a complex environment inhabited by a creature with multiple conflicting goals. The other is a simple blocks world containing a robot arm and a few blocks it can move around, and uses Prolog as a state-description language. If you're giving an AI course and want to give your students something to get them started quickly, these would be good starting points.

http://w2mind.org/ - Home page for the World Wide Mind project.

http://w2m.comp.dit.ie/services/ - The documentation, SOML and message protocols included, and the example worlds, are all linked from this page, rather than w2mind.org.

Douglas Adams, interactive fiction, and Ronald Reagan v0.9

It occurred to me, watching the Mondale/Reagan debates on television, that, on one side at least, we were watching 2K man in action. Ol' Ronnie is not, let us be honest, about to surprise the world by coming up with, for instance, a Unified Field Theory. If he were to play chess against a large block of wood, then I for one would not know where to put my money. It also occurred to me that the people responsible for briefing Reagan for these confrontations have therefore to give him the bare minimum number of facts that he could get hold of, and the maximum number of ways in which he could get hold of them. Added to that he must have a lot of long stop responses with which to field questions which he did not know the answer to or simply did not understand. Which is exactly the way in which you set about writing a program to mimic conversation.

This quote comes from an article in the British humorous weekly Punch, published sometime in the early '80s. To explain it, I should point out that Adams, author of the Hitch Hiker's Guide to the Galaxy, was an unusually clear thinker, as the following excerpt from The Restaurant at the end of the Universe illustrates. Arthur Dent and Ford Prefect have leapt into a matter transporter to escape certain death, and have been teleported to a space ark which they will later discover to be packed with cryogenically frozen telephone-sanitisers and advertising executives. In this scene, just after they arrive, they find themselves captured by the ark's Number Two officer and dragged to see its Captain:

Number Two's eyes narrowed and became what are known in the Shooting and Killing People trade as cold slits, the idea presumably being to give your opponent the impression that you have lost your glasses or are having difficulty keeping awake. Why this is frightening is an, as yet, unresolved problem.

He advanced on the Captain, his (Number Two's) mouth a thin hard line. Again, tricky to know why this is understood as fighting behaviour. If, while wandering through the jungle of Traal, you were to come across the fabled Ravenous Bugblatter Beast, you would have reason to be grateful if its mouth was a thin hard line rather than, as it usually is, a gaping mass of slavering fangs.

It does seem to me that Adams has a point. And his clear thinking makes much sense when applied to Reagan, too.
Adams was a keen observer of computing and the software industry. In Dirk Gently's Holistic Detective Agency, he imagines an AI company, WayForward Technologies, who have invented a novel decision-support system. Instead of deciding for you which actions are best for your business, it asks which decision you would like to reach, and then outputs an impregnable argument with which to justify it to your critics. Now there's an interesting research project.

In the story, WayForward have sold their program to the Pentagon, who want to generate arguments justifying Star Wars (the missile defence, not the film). This becomes a bit worrying for one of the employees, who after reverse-engineering recent US policy statements, notices that while the US Navy is clearly using version 2.00, the arguments expressed by the Air Force are still generated by the algorithms from beta-test version 1.5. This notwithstanding, the Pentagon sale is just starting to make WayForward:

the only British software company that could be mentioned in the same sentence as such major U.S. companies as Microsoft or Lotus. The sentence would probably run along the lines of "WayForward Technologies, unlike such major U.S. companies as Microsoft or Lotus...", but it was a start.

In all, Dirk Gently's Holistic Detective Agency is a fun read - amongst its cast is a time-travelling Cambridge don who uses his time machine to go back and watch TV programs because he can't cope with his video-recorder's user-interface - and, 20 years later, it is interesting to relive the personal computer craze of the times.

Returning to Reagan, the Punch quote with which I started was inspired when Adams combined his observations of the software industry - and perhaps of US politics - with an interest in interactive fiction, itself inspired when he began writing a text-adventure version of Hitch Hiker's. Adams explains, in a quote from Neil Gaiman's biography of him Don't Panic, that as he watched Reagan and Mondale:

I thought, 'This is exactly the way you program a computer to appear to be taking part in a conversation.' So, with a friend in New York, I was going to do a program to emulate Reagan, so you could sit down and talk to a computer and it would respond as Reagan would. And then we could do a Thatcher one, and after a while you could do all the world leaders, and get all the various modules to talk to each other. [World Wide Mind, anyone?]

After that we were going to do a program called God, and program all God's attributes into it, and you'd have all the different denominations of God on it ... you know, a Methodist God, a Jewish God, and so on... I wanted to be the first person to have computer software burned in the Bible Belt, which I felt was a rite of passage that any young medium had to pass through.

However, with the recession in the American computer industry, all that came to nothing, largely because the people who wanted to do it with me discovered they didn't have cars or money or jobs.

Barry Goldwater and the Ideology Machine

Adams did not produce a Reagan game, and there is not, as far as I know, a Reagan simulator on the market. But believe it or not, there was once a Barry Goldwater simulator. In her Artificial Intelligence and Natural Man, still tbe best general book on pre-1980s AI, Margaret Boden describes Abelson's Ideology Machine, a program designed to model how mental belief structures determine one's reply to questions such as "If the Communists attack Thailand, what will happen?" (Abelson was working bang in the middle of the Cold War.)

Abelson chose Goldwater because of his exceptionally clearcut and unchanging (rigid...?) ideology. To the Thailand question, Ideology Machine would apparently reply:

If Communists attack Thailand, Communists take over unprepared nations unless Thailand ask-aid-from United States and United-States give-aid-to Thailand.

As Boden remarks, this answer appears reasonable enough as words to be put into the mouth of Goldwater, while appropriately inappropriate if suggested as coming from, for example, Tito. The point is that such a model must explain differences, i.e. what it is about Goldwater's conceptual structures that would cause him to answer differently from Tito. This it did by analysing the question to give a starting node in the belief-graph verbalised below, and then following the links round to generate its answer:

The Communists want to dominate the world and are continually using Communist schemes (Branch 5) to bring this about; these schemes when successful bring Communist victories (Branch 6) which will eventually fulfil their ultimate purpose; if on the other hand the Free World really uses its power (Branch 4), then Communist schemes will surely fail (Branch 7), and thus their ultimate purpose will be thwarted. However, the misguided policies of liberal dupes (Branch 2) result in inhibition of full use of Free World power (Branch 3); therefore it is necessary to enlighten all good Americans with the facts so that they may expose and overturn these misguided liberal policies (Branch 1).

Any relation to the policies of any current leader is entirely coincidental.
For further reading, I recommend Boden's section on Abelson. The political science link below surveys various projects, and it's interesting too to follow up later references to Abelson and Roger Schank, with whom Abelson began a long collaboration on modelling belief structures.

www.douglasadams.com/creations/infocomjava.html - Hitch Hiker's Guide to the Galaxy Infocom Adventure, including a Java applet that you can play online. It is claimed that this may be the only text-adventure that lies to its players!

www.ainewsletter.com/newsletters/aix_0406.htm - AI Expert Newsletter for June 2004 contains Dennis Merritt's nice explanation of interactive fiction, with numerous links.

http://web.syr.edu/~satucker/ai.html - A survey covering the Ideology Machine and other AI applications to political science, written by Gavan Duffy and Seth Tucker, Syracuse University Department of Political Science.

Programming the Sony Aibo robot dog

When Sony brought out Aibo in 1999, they sold 3,000 in a mere 20 minutes over the Internet. Since then, Aibo has already gone through five versions, the latest being the wireless-enabled ERS-7. In this feature, I recount what I disovered about how to program Aibo, and his uses in teaching.

Your little plastic pal who's fun to be with

If you're interested in programming an Aibo, the first step is to buy one - or to decide whether you can afford to, since Aibo is not cheap. Looking for online shops so I could check prices, I found a confusing tangle of URLs - sony.com, sony.net, sonystyle.com, aibo-europe.com, shop.sonystyle-europe.com. According to the sony.net shop, Aibo sells at £1,399 in the UK, $1,899 in the US, and 1,999€ in the Eurozone. (This may not be the best Sony site, since the Europe page lacks online shops for Greece and other countries, while the French will not be pleased to see that one of their links contains a dud country code, causing their page to appear in German.) Teachers and researchers should look out for educational discounts, and it's also worth checking for reconditioned second-hand Aibos.

Let me add that the price is actually excellent value when you consider the development and engineering that has gone into Aibo. He has become deservedly popular for teaching robotics, at school as well as in university.

What's in an Aibo?

Sony have numerous descriptions of Aibo scattered over their sites - there's a nice little tour via the Sonystyle "Product Tour" link on their Aibo learning centre page at www.sonystyle.com. I'd give the URL in my links list below, but it's far too big to fit on a line, being well over 100 characters long.

More information on Aibo's hardware can be gleaned from notes written by various universities for their robotics courses. According to the University of Pittsburgh, the ERS-7 contains:

  • 320 line CCD camera with color segmentation hardware.
  • 2 IR distance sensors.
  • Position encoders for all joints' actuator motors.
  • Joint actuation motors, as follows:
    • Tail: left/right;
    • Neck: up/down;
    • Head: left/right and up/down;
    • High shoulder, 2 per leg: in/out and forward/backward;
    • Knee: bend position;
    • Ankle: bend position.
  • Touch-sensor switches: 1 in the head (1), 3 in the back, and in the chin.
  • Accelerometer for detecting falls and pick-ups.
  • Switch sensors on each paw pad.
  • Stereo microphone.
  • Audio channel.
  • Wireless network channel.
  • Slot for removable memory stick.

The software is harder to find out about. Aibo's joints give it a good number of degrees of freedom, and its native software is geared towards using this for lifelike "pet emulation", as the list of primitive actions demonstrates:

  • Play Alone - things Aibo does by himself
  • Play with Human - things Aibo does with you
  • Play with Robot - things Aibo does with other Aibos
  • Play with Object - things Aibo does with things (usually pink balls)
  • Show emotions / moods (joy, fear, disgust, ...)
  • Show feelings (bored, embarrased, dislike, ...)
  • Physiological reactions (cold, hot, itchy, peeing, ...)
  • Requests (play with me)
  • Reflexes (react to noise or surprise)
  • Show Intentions (just say no)
  • Offensive (attack, provoke, escape)
  • Guard / Defensive (attack, cover, threaten)
  • Contact - Aibo raises paw to touch something (in different locations)
  • Tactile - Reactions to being touched (hate it, pressed, tickled)
  • Special case position changes (advanced)
  • Ball search and tracking (advanced)

As AI-ers, we want to know what Aibo can do with his senses. Sony say the ERS-7 recognises its owner's face, as well as the blocky black-and-white pattern on the marker pole of its recharging station. Aibo ships with two toys - a ball and a bone, the "Aibone" [oh dear] - and the recognition software apparently knows about these. That they're both coloured a lurid and unlikely pink suggests that Aibo's object-recognition abilities are not overly discriminating. This is backed up by a disappointed review from "etienne" of Brussels on the Home Robot News and Reviews site. Having said that, Sony have still packed a huge amount into Aibo.

YART, R-CODE, and OPEN-R

If you want to try some elementary programming, an easy way to begin is the YART graphical editor written by AiboPet. He, incidentally, is one of Aibo's most notorious hackers, having become famous when Sony invoked the US Digital Millennium Copyright Act against him for cracking Aibo's protection and reverse-engineering its software. Returning to YART, the tool lets you define responses that Aibo must execute when he detects certain conditions. For example, you can program Aibo to sit alone and groom himself when the touch sensor in his head is pressed for a long time, or even to dance. AiboPet's site contains a wealth of information on this.

YART works by expanding its user's condition-response pairs into R-CODE, Sony's BASIC-like scripting language. R-CODE is fine for extending Aibo's repertoire of pet-like behaviours, but it is limited and can't access all of his hardware, so I'll just mention it in passing. Sony implicitly admits this in that while they permit commercial use of their R-CODE development kit, they don't for OPEN-R, which I'll get onto next.

Lower-level than R-CODE, OPEN-R gives you more power. It implements an object-oriented interface based on C++, which allows you to control everything from the gain values of Aibo's actuators to getting data from his camera and communication via his LAN. One very important topic here is parallelism. The OPEN-R programming model views Aibo as a collection of concurrently executing OPEN-R objects (not the same as C++ objects, note) which communicate by message passing. You can't use OPEN-R unless you thoroughly understand this. The best help I've found is a OPEN-R tutorial by Francois Serra and Jean-Christophe Baillie from ENSTA. As they say in their introduction, Sony's official Web-based documentation is incomplete, and some examples, including an often-cited ball tracker, are not at all easy to understand. Because of this, they wrote this tutorial as a service to the research community. It looks like essential reading for anybody beginning OPEN-R.

Tekkotsu

OPEN-R is powerful, but for serious AI, I'd recommend Tekkotsu, and not merely because it is free and open-source. Tekkotsu is a rapid-application development framework developed at Carnegie Mellon University. Amongst the reasons for recommending it is that it removes some of the low-level tedium of robot programming, as well as the complications of OPEN-R parallelism. Amongst other things, it centralises sensory processing in one place, preventing code duplication, and making it possible for programmers to use an event-driven style - familiar to most C++ and Java users - for handling perceptions. This is nice: as an example, the Tekkotsu tutorial shows events being used, in less than a page of C++, to detect when Aibo's pink ball passes into or out of his field of vision.

Prolog, SWI-Prolog, and functional programming on Aibo

Computer scientists rightly attach importance to functional and logic programming languages, because of their clean semantics. There's been a lot of research on such languages applied to various kinds of parallel systems, so I was interested to see whether they'd been tried for Aibo. Disappointingly, a search turned up very little. Jelle Herold at the University of Utrecht has slides describing a reimplementation of Utrecht's 3APL cognitive agent language for Aibo. Another collection of pages on this work mentions that SWI-Prolog has been ported to Aibo, with the help of SWI's author Jan Wielemaker. A download is available; it would be a good start for anyone interested in logic programming and Aibo.

As far as functional languages go, Ian Horswill of Northwestern University describes a simple functional language called GRL, based on Lisp. This allows robot behaviours to be written and composed in a modular fashion, and has apparently been tried on prototypes of Aibo.

Aibo at school and university

Tekkotsu itself is an example of software developed for university work, and it has been used in numerous courses. Just as one example of a non-Tekkotsu course, I have linked at the end to a course at the University of New Orleans, which involved a mapping and path-planning project, using an interface between Java/R-CODE interface.

In schools, the Natural Object Categorisation Research Group at Plymouth University have been exploring how robotics can be incorporated into the UK National Curriculum. As they say,

Using AIBO robots in the class is easy and fun to do. You can extend pupil understanding of control and AI from traffic light controllers and state machines to state of art intelligent machines. AIBO robots can see, walk, navigate around obstacles and even talk to each other. These complex interactions with the real world provide an excellent introduction to today's state of art machines and to the future (their future) of intelligent machines. Using AIBO allows learners to be exposed to very high level concepts in machine control that is common in the latest intelligent machines, including cellnet phones and electronic games and the complex management control systems in every car.

They link to some interesting reports and videos about teaching. The youngest group taught was only between 7 and 10 years, and used flow charts to program simple behaviours into Aibo.

Warning! Aibo moves!

Although I once read about a DEC-10 program which made a disc drive "walk" across the room, most programmers can assume their computer will remain safely where they leave it. Not so Aibo. And since Aibo is expensive, and contains fine moving parts, some care is needed. As one user says, though it may be cute to make Aibo bob his head up and down in a greeting, letting the code loop until his battery runs down or joints seize up is not a good idea. Similarly, if you are working on a desk or table, remember that Aibo may walk off the edge and damage himself in a fall. He is fragile.

Conclusion

I've tried here to hack my way through the jungle of Aibo information. There's masses of stuff out there, from little R-CODE routines to make your doggie dance and sing, to experimental robot languages and Aibo application development frameworks. If you're interested in "pet emulation", either for yourself or perhaps in teaching children, AiboPet's graphical R-CODE interface YART has a lot of work behind it. Look also to see what Sony are currently offering in the way of graphical editors. The Plymouth Browsing the Web to get familiarised with Aibo is a nice little guide which links to many useful sites.

Serious AI-ers should look at Tekkotsu, a standardised and reliable, free, open-source, framework which avoids a lot of the low-level programming tedium. Read also the ENSTA tutorial on OPEN-R, which explains a lot that's hard to find out elsewhere. Researchers interested in language design could note that despite all the (justified) claims in favour of functional, logic and equational languages, these are yet to be proven on Aibo; though Jelle Herold's SWI-Prolog might offer one starting point. And how about seeing how those rat cortical neurons would fare when given control of a robot dog?

www.sony.net/Products/aibo/index.html - One of many Sony sales sites.

www.aibo-europe.com/1_1_3_aibo_story.asp - Brief history slide-show from Sony, with pictures of the prototypes.

http://www-2.cs.cmu.edu/~tekkotsu/AiboInfo.html - >From the people who brought you Tekkotsu, a summary of Aibo's hardware, with pictures. This URL, on a major hobby site, links to pictures of the innards from two early models: www.aibosite.com/sp/gen/index-2.html.

www.onrobo.com/cgi-bin/rms2/magpie/do/submit.cgi?product-sku=on00ers700rosny - Searching for "etienne" on this Home Robot News and Reviews page will find you a disappointed review of Aibo's sensory software. N.B. Always check model numbers when reading reviews, as there has been a lot of progress since the first model.

www.the-gadgeteer.com/sony-aibo-review.html - A rather amusing review of Aibo, from the point of view of a dog owner. To quote, "Supposedly the robot has 'free will'.... but, I think I'd like it better if it always did what I asked of it. I've already got a real dog with free will and it's not all it's cracked up to be let me tell ya. :-)". Incidentally, the link Dogs in Elk Carcass, www.farmount.org/nightshade/dogelk.html, is a hilarious account of what happens when free will in real dogs goes very badly awry.

www.aibohack.com/ - AiboPet's site. There are links to YART tutorials, and information on R-CODE and OPEN-R. At least with the latter two, it will help to have looked at the Sony software development site, http://openr.aibo.com/. Amongst other information, this contains FAQs on R-CODE and OPEN-R. The AiboLife site www.aibo-life.org contains, under the URL www.aibo-life.org/forums/cgi-bin/ultimatebb.cgi?ubb=get_topic;f=7;t=000053, some R-CODE snippets sent in by users. There is also an R-CODE reference here, www.etc.cmu.edu/bvw/aibo/docs/rcode-ers7-cmdref-20040501_E.txt, with short examples of each language construct.

www.ensta.fr/%7Ebaillie/tutorial_OPENR_ENSTA-1.0.pdf - The ENSTA OPEN-R tutorial. Recommended to all OPEN-R programmers.

http://www-2.cs.cmu.edu/~tekkotsu - Home page for Tekkotsu. Recommended. The page includes links to a beginner's tutorial, and to downloads.

www.cs.uu.nl/docs/vakken/aibop/6-AIBOP-aibo_3apl.pdf - As an example of work on language development, here are slides describing a reimplementation of the 3APL agent language for Aibo. For 3APL itself, see www.cs.uu.nl/3apl/. For an Aibo port of SWI-Prolog, with download, see http://defekt.nl/aibo3apl/moin.cgi/FrontPage?action=show&redirect=StartSeite.

www.cs.northwestern.edu/~ian/grl-paper.pdf - Ian Horswill's Lisp-based GRL language, which has been tested on Aibo prototypes.

http://cogrob.ensta.fr/publis/epirob2004-baillie.pdf - Grounding Symbols in Perception with two Interacting Autonomous Robots. This paper, by the ENSTA team who produced the OPEN-R tutorial, investigates the symbol-grounding problem using Aibos. It also describes URBI - Universal Robotic Body Interface - a layer that like Tekkotsu sits above OPEN-R in order to simplify programming.

www.cs.uno.edu/~starapor/research/project2/ - An example of university teaching using Aibo, a project that does mapping and path planning.

www.cis.plym.ac.uk/cis/projects/AIBO%20school/RM%20&%20TF_web%20sites%20report.pdf - Browsing the Web to get familiarised with Aibo. This is the summary of Aibo Web resources produced at Plymouth. I recommend it to school teachers and indeed, all Aibo users. This page, www.cis.plym.ac.uk/cis/projects/aiboschool.html, describes the various levels of school teaching done with Aibo.
www.sciam.com/article.cfm?articleID=0005510C-EABD-1CD6-B4A8809EC588EEDF - Scientific American article on AiboPet and the Digital Millennium Copyright Act. There is an article by Wired at www.wired.com/news/business/0,1367,48088,00.html?tw=wn_story_related, and a page on The DMCA vs. the First Amendment, by Tekkotsu researcher David Touretzky, at http://www-2.cs.cmu.edu/~dst/DMCA/Gallery/. It includes a link to the letter Sony sent AiboPet.

http://sitereview.org/?article=370 - Aibo security alert! Spoof on the buffer-overflow attack that invades Aibo with experimental PitBull code.

Acknowledgements

Thanks to Steven Green of Greenius, http://greenius.ltd.uk, for telling me about the University of Florida work.