LittleDog video – 2010

The size of the bot compared to a human hand is shown near the end of the video. Sign me up for a robotic-pet version. The summarized specs of LittleDog as described on Boston Dynamics’ site:

LittleDog has four legs, each powered by three electric motors. The legs have a large range of motion. The robot is strong enough for climbing and dynamic locomotion gaits. The onboard PC-level computer does sensing, actuator control and communications. LittleDog’s sensors measure joint angles, motor currents, body orientation and foot/ground contact. Control programs access the robot through the Boston Dynamics Robot API. Onboard lithium polymer batteries allow for 30 minutes of continuous operation without recharging. Wireless communications and data logging support remote operation and data analysis. LittleDog development is funded by the DARPA Information Processing Technology Office.

For detailed specifications check out CLMC’s paper on LittleDog research (PDF), published at the 2010 IEEE International Conference on Robotics and Automation (one of four Best Paper Award finalists).

It’s worth mentioning that USC was the only team to outperform the government program’s guidelines on 5 out of 6 tests, achieving an average speed of 8.7 cm per second. (Congratulations to the team—Mrinal Kalakrishnan, Peter Pastor, Michael Mistry and Jonas Buchli).

Cry ‘havoc’ and let slip the dogs of war

The research is funded by DARPA and carried out in collaboration with MIT, Stanford, Carnegie Mellon, USC, Univ. of Pennsylvania and IHMC.

While the scientists working at Boston Dynamics and collaborating institutes envision the ‘bots being used for various purposes, DARPA’s point of view is to use them to haul equipment and luggage wherever needed on a battlefield. The first versions for fieldwork are called Legged Squad Support Systems (LS3); a type of completely autonomous robots, very similar to BigDog, that will be able to follow a human leader through rough terrain using computer vision and GPS systems. The LS3 program is jointly funded by the US Marine Corps.

The first “walk out” for LS3s is scheduled for 2012.

Automatic recommendations:

1. Boston Dynamics Unveil BigDog Progress
2. Modular Snake Robots From CMU’s Biorobotics Lab [Video]

The image recognition tech sounds exciting—image search and recognition in real time! I wouldn’t be surprised to see Google and Apple go heads on in a bloodsport match as they race towards the AR advertising market (incidentally bringing with them a wave of exciting apps and even AR goggle interfaces. Real ones.).

But, it’s best to let the video do the talking (read: I’m lazy). Here’s Google Goggles.

Automatic recommendations:

1. Wikitude AR – Augmented reality on Google Android
2. Layar’s Augmented Reality Web for mobiles: finally layered information

The Advanced Human Interface Group (AHIG), MIT Media Lab. The ICONC System, demonstrated by Hrafn Th. Thorisson, Summer 1993. The system enabled the use of co-occurring, natural speech and gesture to interactively describe the arrangement and movements of objects in a room. The computer would interpret the user’s actions and figure out which objects the user was talking about and how to arrange them based on spatial information in the user’s speech and gesture. [Excerpt from the YouTube description, continued below]

[If you can't see the video, click here to visit the post]

The main authors of this work were David Koons (spatial knowledge, multimodal integration) and Carlton Sparrell (gesture recognition), directed by Richard A. Bolt. This technology is described in part in the paper “Integrating simultaneous input from speech, gaze, and hand gestures” by D B Koons, C J Sparrell, K R Thorisson (1993).

UPDATE (Oct. 30th, 2009): The article stated, wrongly, that this took place in 1994. This has been corrected.

Links & references

• Kristinn R. Thórisson’s website at Reykjavik University & at the MIT Media Lab
• YouTube page for this video

Automatic recommendations:

1. Emergence of Creativity in Intelligent Complex Adaptive Systems
2. Wii Real-Time Head Tracking [Video]

To answer this bluntly: Game AI is very different from it’s non-game counterpart, and it’s not my field of study. I’ve only compared modern games through a window. However, Alex of AIGameDev has superb coverage of AI in games and the top AI games of 2007, by community vote. The top of the line are Half-Life-2.ep.2 and BioShock.

But regarding Game AI in general: modern games are horribly void of intelligence. It depends on where you set the bar, certainly. There’s tons of AI in modern games compared to 5 years ago. But the first thing to note is that Game AI is not the same as AI. It’s a subset of it. Just like discrete mathematics are a subset of mathematics. And moreover, Game AI is a very specialized subset—it has well defined goals, models for construction and limitations.

Games are governed by laws of commerce first, then innovation.

A game is governed by different laws than academic, general AI R&D. It’s a commercial product, and commercial products depend on older methods wherever possible—methods that have proven successful. Most (sane) business men do not put all their money on a new and untried idea because they don’t know if it will succeed. So, most of commercial products are bulked up with a lot of things that’ve been successful in the past and then leave a breadcrumb for innovation.

This is very different from academic AI research where the point is to do things that’ve not been done before.

Now aside from these drastically different goals of commercial ventures and academic ones, a game’s purpose is to entertain. As long as the player is entertained it doesn’t really matter what goes on under the hood. Because of this, there’s a certain witch hunt that takes place in the game industry:

A modern game developer is on a mission to slaughter innocent intelligent processes wherever possible.

A game AI developer tries as hard as he can (usually at the bidding of a project manager) to minimize intelligence. One reason is that intelligent processes are massive processing-power hogs. Thus, like an obese overeater the systems must forcibly give away every other meal to accommodate an average person’s desktop PC — and those machines don’t have much elbow-room to replicate the massive crimson jelly residing in the heads of animals. Human or other.

And then there are graphics, another obese overeater, who also need a place at the table. And because games are governed by the laws of commerce, Game AI must leave at least five chicken wings more than it ate itself for its obese, graphics rendering sibling. Beautiful games get a lot of coverage and attention, and developing graphics is a question of engineering. In a business plan it’s therefore rational to emphasize graphics. Both in terms of predicting the amount of effort required to implement it and the potential payoff.

Because of these severe limitations on how much processing power the intelligence is allowed, developers are forced to dumb-down the processing and make their AI appear intelligent instead.

To some it may not be clear what the difference is between making something appear intelligent and actually making it intelligent. After all, there has to be some amount of intelligence if something’s intended to keep its appearances. Right?

An intelligent system is expected to produce solutions to problems, uncertainty and often in complex situations. Appearances, however, are concerned with making an observer believe they are intelligent. To accomplish this in games the environment (the input to the AI) is kept controlled and limited. For a vivid example of this:

• NPCs in games don’t use computer vision to perceive where the player is, instead they get fed (X, Y, Z) coordinates, giving the appearances of eyesight and visual processing capabilities.

The game designers tailor the environment and its limits to make sure that the intelligent processes can handle them, and vice versa. It’s the lifelike gatekeeper who doesn’t need to know how to find his way home because he has no home. It’s the terrorist that can pull a trigger but couldn’t count his fingers if you took his shotgun and held it to his head (and you can’t).

For a different and real-life example of appearances versus actual intelligence, to show how diversified the game AI “trickery” can get:

• The developers of Halo 3 found a correlation between how smart the AI was and how tough it was. If they made the NPCs smarter the game became tougher. If they increased the health of the NPCs, making them tougher, the human players perceived them as more intelligent.

The first rule of AI in commerce is that you don’t talk about AI in commerce

Like I’ve stated in another post, the gaming industry is pretty much the only industry that dares market products using the term artificial intelligence. The reason for that is complicated and I won’t talk about it from all perspectives. But what I will tell you is that its partly due to people knowing what to expect from games.

The gaming industry builds from the sets of platform resources and devteam innovation. These are sets that the consumers know. It’s the set that the elite gamer knows because he understands the nature of programs and what the required Hz’s actually stand for. And it’s a set that the average gamer knows because he’s fought space aliens so many times that he’s learned what they’re capable of. With consumers that know what to expect, the term can be used without people boiling a can of hype.

In contrast, Academic AI builds with a set that appears to the consumer as one of infinite possibilities: Because people don’t know (exactly) how the mind works, they/we can’t evaluate how far we are from recreating it in machines. It’s unknown. And because the average consumer can’t accurately evaluate the unknown, it doesn’t matter if yesterday’s AI was primitive; most will still anticipate it advancing leap-years overnight. Just like kids in the backseat of a car asking if it’s: “human now? it must be human now? how about now?“.

Thus, when the term AI is used in areas where the limitations and previous products don’t foretell the nature of the next, people start imagining Terminators all over again. And then they get incredibly disappointed when they realize all the AI can do is answer questions about farm animals. (And only when you begin a sentence with “What is…”).

That’s obviously not a scenario a businessman would like to encounter, so it’s best just to focus on something else when marketing, like Apple does when presenting its operating systems to the public.

In summary

Game AI is just one of many different subfields of AI and is governed heavily by the laws of commerce, entertainment value and modern desktop computing resources. While it may seem it must intersect with other subfields, the truth is that its a somewhat isolated field with its own sets of tricks and tools. (Mostly tricks.)

At times, games may also feel like they are the only commercial products successfully employing AI. But this is largely because of too high consumer expectations to AI in other products, and the consequent fact that companies don’t like mentioning that their product uses, what is by definition, artificial intelligence.

And finally, a difference between apparent intelligence and actual intelligence is that the latter figures out solutions to problems, while the former doesn’t care about what happens under the hood as long as an observer thinks it’s intelligent. A lot of the times—that doesn’t involve intelligence at all.

Links & references

• Halo 3 AI “Trick” example from Teaming up with Halo’s AI

Automatic recommendations:

1. Leopard Sports Superb AI, Apple Doesn’t Mention It
2. My Robot is Your Congressman

But it certainly looks like they’ve made significant advances judging from some test queries.

The engine is similar to Twine in that it utilizes semantic metadata to process queries and reason. But that’s about all they have in common — they use the metadata to provide very different services.

Powerset’s first product is a search and discovery experience for Wikipedia, launched in May 2008. Powerset’s technology improves the entire search process. In the search box, you can express yourself in keywords, phrases, or simple questions. On the search results page, Powerset gives more accurate results, often answering questions directly, and aggregates information from across multiple articles.
[About Powerset]

It will be exciting to see how the titans of this first wave semantic web startups will band together in the near future. Powerset and Twine both utilize Wikipedia, and Powerset uses Freebase information. If I’m not mistaken I seem to remember the CEOs of Twine and Freebase mentioning potential collaboration at the Web 2.0 Summit in San Francisco last year; and its not too hard to envision how that could be mutually beneficial.

Here are some interesting test queries and commentary. But I recommend you head over there yourself to try it out; share with us the voes and virtues. Note that these are only the ones that worked!

When did Albert Einstein die?

This one came as a nice surprise because I remembered having seen a beta bug report where “die” returned results in German pages. But this time it resulted in the nice date icon and a list of Wikipedia pages.

Who played Doctor Who?

A little play with words resulted in a nice, AJAX rich list of all the Doctors! The same list popped up when I added “… in 2004?”.

Neal Stephenson

Here’s a nice and data rich result from Freebase showing birthdate, place of birth, books published, etc.

Fractal

We can see the different types of fields and related concepts. “Complex systems” returned a similar summary but included another tab titled “Journal” that displayed information on a scientific journal with this name.

A handful of results weren’t quite as elegant, but it’s safe to say that this is a great start for Powerset.

Links & References

• Powerset’s Launchpost

Automatic recommendations:

1. Best Technology Innovation, 5 Crunchies Finalists
2. Why I Migrated Over to Twine (And Other Social Services Bit the Dust)

Chapter V, titled Emergence of Creativity: A Simulation Approach, presents my latest research on the emergence of creativity in natural and artificial organisms, a theory of its origins and potential grounds for future artificial implementations. The book is distributed internationally.

Intelligent Complex Adaptive Systems.
ISBN: 978-1-59904-717-1; IGI Publishing.

An official description of the research can be found below, but in short my approach does not exclude creative behavior in animals other than ourselves and explores a possible foundation of creative systems. Using computer simulations for support, the research combines several scientific concepts, techniques and methods of simulation; including evolutionary computation, cellular automata, emergence and self-organization.

Related Wikipedia articles: Emergence, Self organization, Computational creativity, Complex systems.

Emergence of Creativity – Chapter Introduction

The following is an excerpt from the preface to ICAS provided by the editors; Dr. Ang Yang and Dr. Yin Shan.

The third section of this book features four chapters on the use of modern computing techniques to study the emergence of creativity, emergent specialisation, information bottleneck to central processing in adaptive systems and the role of barriers to information flows in the robustness of complex systems.

Creativity has been a difficult concept to define and its exact relationship with intelligence remains to be explained. In the first of the four chapters on computing techniques, Thórisson presents a theory of natural creativity and its relation to certain features of intelligence and cognitive faculties. To test the theory, the author employs simulated worlds of varying complexity that are inhabited by creatures with a genetically evolving mental model. Planmaking strategies are compared between creatures in each world. This shows that creative behaviours are governed by the world’s structural coherence and complexity. The theoretical framework presented in this chapter may serve as a foundation and tool to improve our understanding of natural creativity and to help develop creative artificially intelligent systems.

Interested can look to Amazon in the UK and US, and Barnes & Noble. You can also try finding a local store via the used & new links on Amazon, or Google’s Product Search.

About the Book

The following is the official & compressed introduction to the book’s content and purpose.

As the world currently subsists as a platform for exchange among complex, intelligent systems that are constantly adapting and evolving to suit the surrounding physical, sociological, emotional, and sensory environment, understanding the theory and emergence of complex adaptive systems is of paramount importance.

Intelligent Complex Adaptive Systems explores the foundation, history, and theory of intelligent adaptive systems, providing scholars, researchers, and practitioners with a fundamental resource on topics such as the emergence of intelligent adaptive systems in social sciences, biologically inspired artificial social systems, sensory information processing, as well as the conceptual and methodological issues and approaches to intelligent adaptive systems.

Note these links to a the book detail page at IGI, comprehensive introduction to the table of contents, and an introductory brochure (PDF).

Automatic recommendations:

1. The 5th International Workshop on Computational Creativity

Taking its cues from Google News Alerts on “robotics”, the system analyzes news articles — identifying and extracting relevant lines of text and generating a shorter version of the article. The text is then paraphrased using preset tunings to mimic my writing style. The ultimate result is an article that only needs my one-click administrator approval to be published.

The TOWTAW System (“Toto”)

The system, named The One Who Thinks Artificially & Writes (TOWTAW, pronounced ‘Toto’, as in OZ), is an example of AI systems integration, comprising several software modules that communicate through a central blackboard system.

Relevance ranking of paragraphs and sentences (understanding the text) is accomplished by a custom made module that makes use of the OpenCyc reasoning engine and the Open Mind Common Sense Net (a.k.a. ConceptNet). The excerpted text is paraphrased by another module that again makes use of these two systems, whose combined strength is hundreds of thousands of terms and logical expressions. This entire process is complimented by a third, custom software module that accesses the Wordpress database and tries to find sentences written by me that can be re-used. This is done to further improve writing style similarity.

Finally, the system makes use of Wordpress’ Post-by-email option, simply using my machine to mail in the post to await my admin approval. It’s worth noting that I rarely have to retouch the text.

So Long, and Thanks for All the Fish

With the continued success of the system, and the many great comments on the articles (it has even started debates!) — I’ve decided to leave Think Artificial entirely in the hands of artificial authors. The site, starting next week, will be renamed Think Artificial and Writes. Additionally, I’m opening several new blogs — including one called TOWTAW Row-baw-toh about Western musical adaptations of Asian robot culture.

I thank you very much for your support and compliments on my writings, and will announce at a later date where I plan on start writing online again.

Best regards,
-Hrafn Th.

[[This was April Fools' Day joke for 2008. Read this for the hidden joke]]

No related posts.

The New BigDog Video

Loving the Machine

It’s so interesting how most people I’ve talked to (and me myself) feel an emotional response when BigDog is slipping or being kicked. I felt like giving it a hand to balance on the ice — although on second thought I’m sure he’d crush me with his frantic maneuvers. The version of BigDog in the video weighs 106Kg (235 lbs).

The speed and agility of the feet give the fellow such natural (“messy instincts”) appearance that it’s hard to keep in mind it’s a machine. Unbelievable.

The control system depends heavily on internal state, awareness of joint positions, acceleration et cetera — which enables intelligent, real-time control. I hope the technologies spread to consumer markets, rather than being exclusively used for military purposes (this is what the AIBO should’ve looked like).

BigDog has an on-board computer that controls locomotion, servos the legs and handles a wide variety of sensors. BigDog’s control system manages the dynamics of its behavior to keep it balanced, steer, navigate, and regulate energetics as conditions vary. Sensors for locomotion include joint position, joint force, ground contact, ground load, a laser gyroscope, and a stereo vision system. Other sensors focus on the internal state of BigDog, monitoring the hydraulic pressure, oil temperature, engine temperature, rpm, battery charge and others. [BostonDynamics]

Boston Dynamics are also the creators of RiSE — an insect-like hexapod robot that climbs; and whose video demo ranked 10th in the most popular robot videos. A hat tip to the Boston Dynamics team for their achievements.

Links & References

• Boston Dynamics’ official site
• Hat tip for infosharing to Think Artificial reader Alebotta.

Automatic recommendations:

1. Modular Snake Robots From CMU’s Biorobotics Lab [Video]

Creators and Project Origins

The snakes are a project of CMUs Biorobotics lab, named Modular Snake Robots (or Modsnake for short). See also the list of project members.

Snake robots can use their many internal degrees of freedom to thread through tightly packed volumes and access locations that people and machinery otherwise cannot. These highly articulated devices can coordinate their internal degrees of freedom to perform a variety of locomotive gaits that go beyond the capabilities of conventional wheeled and legged robots. The true power of these devices is their versatility; they can crawl, climb, swim, and scale flights of stairs. [Official description from CMU]

Project Goals and Design

The issues tackled in the project are twofold, according to the information provided on the project’s site. The first is to achieve snake-like locomotion and the second is to have the artificial being’s design modular. The robots are a chain of modules, each of which have one degree of freedom. They’re powered by low-cost hobby servos (named SuperServos by CMU, or more recently SuperServo2), and an encapsulating mechanism designed to optimize efficiency and robustness.

Each robot is currently controlled by a tethered cable and are embedded with a camera on the snakes’ edge module.

[[Subscribers, please visit the site if you can't see the video above]]

Speed of Locomotion in Complex Environments

I’m very impressed by the speed of the robots, and you can see that CMUs team is proud as well, sporting the “real time” tag in the video. Lack of speed is one of the things I’ve touched on regarding modern robotics. It’s hard to deal with complexities of normal, or even catastrophic environments (for rescue-type, or military robots). If a robot moves fast its intelligence must be capable of making good judgement calls in real time, fast and efficiently.

Our (humanity’s) physical machinery is advancing rapidly and hopefully these superb platforms will help us excel development of systems that endow them with higher intelligence.

Related AI Projects & Commentary

From what I gathered on CMUs information I believe an emphasis is currently being put on engineering and physical versatility. But there are of course several projects focused on AI, specifically regarding path planning and the snakes’ ability to coordinate each of its joints by itself. As noted on the CMUs Snake Motion Planning Page (see also more of the snakes’ Control Theory projects):

Snake robots have many applications, but are hard to control. A person cannot simply operate each joint of a snake individually because there are too many. These robots require a motion planning algorithm. Motion planning for snake robots is difficult because the robots have many internal degrees of freedom that have to be coordinated to achieve purposeful motion. In motion planning jargon, this means the snake robots exist in large dimensional configuration spaces. Our work will make it possible for the robots to operate in several different modes from fully autonomous to human-guided. The robot will be able to optimize its own path based on a range of cost functions from power consumption to safety or even stealth.

[...]

We are working on two approaches for snake motion planning right now. The first introduces a new topological decomposition of space called WAFT (WAve Front Topology) This decomposition is well suited to tethered robots, like a snake that is fixed to the ground at one end. This decomposition of space allows for path planning that is path-dependent, meaning that veering left around the first obstacle encountered will change the options that are available in the future as compared to the options had the path to the right of the first obstacle been chosen. This WAFT planner can be implemented in a sensor based way and can be used to guarantee complete coverage of the work space. This is important for exploring unknown spaces when searching for survivors in collapsed building or when doing surveillance operations.

Very interesting and impressive work and I personally look forward to seeing how the projects progress.

Links & References

• The video and images are Copyright 2008 Biorobotics Lab, Carnegie Mellon University. All Rights Reserved.
• See the list and specifications of the Modsnake robots that have been created
• See links to the BioRobotics lab, Modsnakes and relevant AI projects in the article

Automatic recommendations:

1. Anthill Metropolis Filled With Cement and Excavated (Video)
2. Machine Interpretes Your Dreams, Robot Enacts Them [Art]

Bark, Bark

The new system uses artificial neural networks which are trained to discern different types of sound patterns, and can identify the 6 different situations mentioned above with 43% accuracy. Admittedly, that doesn’t sound like an ideal number but as stated by the researchers it’s a lot higher than we humans can achieve.

Komondor, a type of Hungarian sheepdog
(I don’t know the specific type used in the research)

Additionally, the system can correctly identify individual dogs by their bark 52% of the time during play, but interestingly only 30% when aggressively barking at strangers. The R&D was carried out by a team led by Csaba Molnár of Eötvös Loránd University in Hungary.

Arf Arf Arf

A few years back I heard about a simpler variation of a dog translator. A certain pitch in a dog’s bark would trigger the playing of recorded statements such as ‘I’m hungry’ or ‘I want to go outside’. And here’s the brilliant part: That’s all the device did! The dog would soon enough learn that a certain bark would result in his masters giving him food. However, this technique obviously didn’t pan out very well as evident by the lack of talking dogs.

Links & References

• Scientists develop computer that can translate a dog’s bark (DailyMail)
• Komondor image CC whartonds (via Wikipedia).

No related posts.

I’ve played games with floor-pads as controllers, like Dance Dance Revolution. I’ll be frank: it sucks. I’ve also had a go with Playstation’s EyeToy. It sucked too. The reason they suck is the limited freedom they provide. In essence, they’re still stuck in the prison-like dimension of finite button-pushing: The dance-pads are basically just a joystick you jump on, with the same monotonous button combinations as a typical handheld controller. Similarly, the EyeToy has extremely limited understanding of what it’s looking at and thus it basically has to split video input into sectors, where motion in a certain sector basically equals a push of a button. This gives the EyeToy’s entertainment value an eerie resemblance to the satisfaction I get from wiping my laptop screen clean.

We Need More Flexible Controls

What I’m looking forward to are interfaces that give you more dynamic control. When I first heard of the Nintendo Wii’s motion-controller (Wiimote) I got excited because I envisioned, for example, sword fighting where the slashes of the sword were not pre-determined “action sequences”, but real time actions performed by the user swinging the stick. Unfortunately, while the controller hardware certainly opens up this option — the software implementation is harder and thus the Wii games are still somewhat stuck in the same button-pushing dimension.

But that’s not to say things aren’t moving along. Nintendo made a major splash in the non-conventional controller department, and we’re seeing the effects of this reverberate throughout the commercial world. Portable media devices like the iPod are also giving wearable machine R&D a kick in the ass (something like a head-mounted display suddenly has obvious commercial potential for watching movies on your iPod). And then of course, there’s the constant ongoing research at MIT.

The MIT Whole-Body Sensor Suit

In collaboration with the Swiss Federal Institute of Technology (ETH) and Mitsubishi Electric Research Laboratories, they’ve produced a new kind of suit that can capture the motions of your entire body — and you don’t need a studio or lab environment to use it. It even works for people driving or playing ping-pong. A New Scientist article discusses the suit, explaining how it uses ultra-sonic beeps, gyroscopes and accelerometers:

Several sensors measuring about 2.5 centimetres on each side are attached to a person’s legs and arms. The sensors detect movement in two different ways: accelerometers and gyroscopes measure motion, but ultrasonic beeps are also emitted.

Tiny microphones mounted on the torso pick up these beeps, allowing a laptop computer, carried in a backpack, to calculate the distance to the sensor. The system is similar to, albeit much simpler than, bats’ ultrasonic echolocation, and together with the motion sensors provides a more accurate overall picture of body movement. The small backpack also holds the batteries that power the system.

Originally designed to allow MIT students to give Clippy a virtual beating, the system could be utilized for various purposes. A New Scientist article mentions the potential for making animated movies more realistic and help doctors analyze movements of patients undergoing physical therapy. The following video was uploaded by the wonderful people at CSAIL, explaining the system and showing it in use.

This is awesome. Is there anything more to this? Yes. The price. The NS article quotes Rolf Adelsberger from ETH on that: The suit was made from off-the-shelf components and is much cheaper than similar systems used in the past. Currently, it costs about $3,000 — but Adelsberger imagines that the price could go down to a few hundred dollars if mass-produced.

Combining Almost-On-the-Market Technologies for the Ultimate Control Suit

What I find very interesting is the prospect of integrating some of these new groundbreaking machine interfaces. Especially for immersive, 3D environments. For example, the MIT suit is excellent work — but complemented with a head-mounted display? Oh man, that’s a winning combination right there. But there’s still something missing … oh yeah, I can’t really use the MIT suit to walk around inside a virtual world: If I can see and use my entire body, an attempt to make my avatar duck the fiery breath of a dragon would probably leave me with one foot in my cat’s food-plate and the other on the cat.

So, its use is limited to an immobile, vitruvian sphere, if you will. No worries, let’s combine the suit with Project Epoc from Emotiv — a non-invasive brain-computer interface where the system detects user intentions. Ah-hah, now we’re getting somewhere: Using the suit to detect limb motions, a HMD for visuals and the Epoc headgear to track intentions to walk-forward, backwards or turn, we’ve got ourselves a damn nifty setup. Additionally, it wouldn’t be bad if the Epoc headgear detects facial expressions like the Emotiv team claims (they haven’t released new information for a while, I’m waiting for some answers in the mail). The only thing that’d be missing would be force-feedback, but I’m not aware of any plans to make a commercial product of out of those.

Can you imagine the potential such a suit would give, say, teleconferencing? gaming? virtual world communities? learning Tai Chi? cybersex? Hot stuff. And here we are, born about 50 years too early. Waiting for these devices to debut one at a time. Then waiting for the prices to come down. Then waiting for someone to combine all the systems into one. Finally, waiting for the price to come down on the combined system.

Well, I’m being pessimistic. Technology is really improving and at an ever increasing rate. Hell, we were using black and white laptops 15 years ago. But just in case immersive virtual reality technologies take a while to show up on our doorstep — we’re making headway in life-extension technologies as well, so no worries.

Links and References

• Practical Motion Capture in Everyday Surroundings, paper published in SIGGRAPH 2007 on the MIT suit (PDF)
• Cheap sensors could capture your every move, New Scientist on the motion suit

No related posts.

Ridiculously Brief Introduction to Lisp

If you’re not familiar with Lisp, it’s a functional programming language whose name stands for LISt Processing. It was created by John McCarthy in 1958, mainly to serve as a programming language for artificial intelligence. It was and is still used in various AI research. What’s perhaps the most prototypical example Lisp’s power is that the programs themselves can write their own, new code — which has obvious benefits for AI. The language is often symbolized with the greek lowercase Lambda letter, seen above, due to the fact that it’s based on Lambda Calculus.

MIT Lectures on Programming and Lisp

First up, courtesy of MITs excellent CSAIL and a part of MITs free OpenCourseWare collection, we get free access to course materials for Structure and Interpretation of Computer Programs:

Structure and Interpretation of Computer Programs has been MIT’s introductory pre-professional computer science subject since 1981. It emphasizes the role of computer languages as vehicles for expressing knowledge and it presents basic principles of abstraction and modularity, together with essential techniques for designing and implementing computer languages.

It’s an entire MIT course for free online, complete with an online version of the textbook as well as funky downloadable lectures. They give an introduction to Lisp in the first lecture and use Scheme throughout the course. Both the book and videos are Creative Commons licensed, and the videos are so retro that the rims of my glasses grew 2cm thicker (hot stuff regardless, mind you, this is MIT!). Live free, study hard, rock on ’80s lectures!

DrScheme Programming Environment

If you want to try your hand at Scheme, check out DrScheme — a graphical programming environment available for OS X, Unix and Windows. The screenshot below shows the main terminal.

It might not look like much if you’re used to something like Eclipse. I’d say it ranks 0.4 on the retro-glasses rim-thickness scale, but it’s really easy to install and get started, and it has some surprisingly fun features such as drawing arrows on the code for debugging. They also have good introductory articles to DrScheme.

Quick Introductory Articles and References

In case you don’t exactly have the time to take an entire course, like you probably don’t, then no worries. Here are a few other introductions and examples that come in handy as reference material and quick reads.

• An Introduction to Scheme by Shriram Krishnamurthi. Contained in one webpage with Scheme code examples and minimal text. Good for quick code reference.
• Teach Yourself Scheme in Fixnum Days by Dorai Sitaram. A bit tedious to browse for quick reference, but still a nice quick-start guide with Scheme code examples.
• Scheme Tutorial by James D. Arthur at Virginia Tech. A concise overview of the fundamentals, contained on one webpage. Good for quick conceptual+code reference.
• Game programming in Scheme Downloadable eBook (pdf), code examples and tutorial.
• Scheme Tutorial by someone unknown. Categorized text on Scheme. (Didn’t use this one much but may come in handy).



• Alligator Eggs! is a pretty funny game created by Brett Victor, based on Lambda Calculus (the alligator pictures are from the game).
• Lisp.org, the Association of Lisp Users. Supporting developer community with a lot of resources.
• Practical Common Lisp is a book available online (and in print) by Peter Leibel. While it’s about Common Lisp and not Scheme, it’s still worth listing as an excellent resource on Lisp.
• MITs OpenCourseWare. Worth mentioning twice. Loads of free interesting lectures. Truly a goldmine of knowledge.

No related posts.

The solution we’re creating is simple: an open-source filter software that can detect rampant stupidity in written English. This will be accomplished with weighted Bayesian or similar analysis and some rules-based processing, similar to spam detection engines. The primary challenge inherent in our task is that stupidity is not a binary distinction, but rather a matter of degree. To this end, we’re collecting a ranked corpus of stupid text, gleaned from user comments on public websites and ranked on a five-point scale. [StupidFilter About page]

So basically, you’ll be able to set the bar for how stupid or naive comments can sound. (I’m guessing bar-one will be YouTubian-like “ur fkn asshat! LOLZ!!!PO!!OP1!”). In a CNNmoney interview with the project’s leader, Gabriel Ortiz, a basic use-case scenario is described: “If cnnmoney had the filter installed on its servers, it would intercept the comment just before it was published and flash a little alert at the author that reads: “This comment is more or less unintelligible. Please try to restate it.“.

It sounds pretty elegant and I look forward to seeing how well it works. The project is in a design and analysis phase and an alpha release is planned for December. The core code is intended to be integrateable into standard content management platforms, but additionally they intend to release Wordpress and Firefox plugins which will probably arrive at a later date. If all goes as planned and described, you can be fairly certain the plugin will be put to the test here on Think Artificial (it’s only appropriate, considering the blog’s focus!).

Be sure to check out their FAQ, it’s pretty funny — including golden answers such as this one:

Q: Do you really expect to be able to detect and filter anything that’s conceivably stupid?
A: No, of course not. You’d need real AI for that, and beyond a certain point it’s simply subjective; after all, a sufficiently advanced AI would probably filter out the whole of human discourse, which isn’t the idea.

As mentioned in the description I quoted from their about page, a method of training the engine to recognize stupid comments is ranking them on a scale from 1-5. They pull comments from YouTube and then have human moderators rank stupidity. If you want to help speed the project along, you can apply to become a moderator of the stupidity corpus.

Links and References

• StupidFilter Project Home
• CNNMoney interview with Gabriel Ortiz
• Bayesian spam filtering on Wikipedia
• The StupidFilter reminds me of the Religious Speech Sensor

No related posts.

Leopard’s Features

What struck me the most during the (somewhat silly) guided-tour was how prevalent AI based features had become. What was that? AI? There’s no AI! Yes there is, they just don’t call it that. Most of Apple’s software is now more enabled by the use of techniques, algorithms and general software categorized under artificial intelligence. Here’s an example list.

• Alex — A new English male voice that uses advanced technologies to deliver natural breathing and intonation, even at fast speaking rates.
• Misspelled Word Detection — Hear when a word is misspelled while reading text. Choose a tone or a spoken description.
• iChat’s Computer Vision — iChat can learn the difference between the user and background, and visually integrate a user into a given picture or video
• Data Detectors — Mail automatically detects text fragments like appointments and addresses, and lets you choose smart actions (the guided tour shows Mail understanding which date the sentence “This Tuesday” refers to).



• Improved Search — Smarter relevance ranking in Spotlight.
• Spotlight Language Support — More language support. A new Chinese tokenizer intelligently parses the search characters to factor in their relationship and meaning with one another. Also improved support for German and Thai and faster indexing in Japanese. (They actually mention intelligence in this description)
• Dynamic Web Filter — Automatically tries to detect inappropriate content and prevents those web pages from appearing.
• Advanced Searches — The Spotlight search field now supports Boolean logic, search category labels such as “author” or “width.”, The use of ranges in your search including “greater than” and “less than.” plus it understands quoted phrases and dates.
• Grammar Check — English language grammar checker helps ensure that you don’t make errors in grammar.
• Research Assistant — View context-sensitive documentation as you code, automatically suggested to you in a heads-up display. The assistant provides relevant information, including API overviews, source code references, APIs, and more.

If you’re thinking that calling some of these features “intelligent” is pushing it — and that it should rather be referred to as automatic — then I have a question for you: where does automation end and intelligence begin? Do many stupid, automatic things make something intelligent? No? Then how do you explain the neurons in your brain giving rise to your thoughts?

I’m not going to convince you that automation is intelligence, but I am going to entice you to consider the blurred line between the two.

Why Apple Doesn’t Use the Term “AI”

Ironically, if you do a search for “Leopard A.I.”, one of the articles that’ll come up is an interview with Steve Wozniak where he claims the following:

[Modern robots] do one thing well, but we never will see a robot that makes a cup of coffee, never. I don’t believe we will ever see it. [...]

Think of the steps that a human being has to do to make a cup of coffee and you have covered basically 10, 20 years of your lifetime just to learn it. [...] You can’t program these things, you have to learn it, and you have to watch how other people make coffee. … This is a kind of logic that the human brain does just to make a cup of coffee. We will never ever have artificial intelligence. Your pet, for example, your pet is smarter than any computer.

“Never ever“? Well I never. Regular readers of Think Artificial can guess what my opinion is. I think Wozniak is off; there’s no reason at all to think we can’t continue to improve learning software (evidence is to the contrary). But I’m not going to debate that here, there’s another reason I quoted this article.

Did anything strike you as odd regarding how Wozniak refers to artificial intelligence? He says, and I requote: “We will never ever have artificial intelligence“. That’s right. He’s saying we don’t have it today.

(Artificial Intelligence is Controversial)

This a problem for modern AI products: the term’s meaning is convoluted and controversial amongst the public. People generally don’t know the details of formal AI science. When abrubtly asked to explain what AI means, most people will immediately think of conscious, self aware machines and disregard everything else.

As such, any company that dares use the term AI to describe their products must answer for it in the fiery pits of public hype and consequent disappointment. Mention your product having AI and you’ve increased the risk of bad reviews due to people’s high expectations. (Unless you do a superior job explaining what you mean, which costs more money). The only industry that regularly pulls off using the term AI for marketing is the gaming industry, which is largely due to people expecting such software to play intelligently — not actually be intelligently.

A View Askew

So that’s why Apple doesn’t mention AI: The term is surrounded by hype which makes using it not worth the marketing effort. Instead they use more neutral words such as, for example, “data detectors” or “dynamic web filtering”. Words that won’t make people think the operating system was cast in a giant metal machine and sent back in time to kill their unborn son.

Does this matter for AI development? Not necessarily. Marketing and AI R&D are separate things, and Hollywood inspired definitions of AI are bound to receive a pounding as intelligent machines become more prevalent in society and everyday lives.

Automatic recommendations:

1. My Robot is Your Congressman
2. Mass Producing Intelligence

Twine is a sort of knowledge management tool for the masses. Each user’s Twine home page is a sort of personal dashboard—its central feature is a list of updates not unlike the Facebook News Feed—that allows a user to import any memo, website, video, or photo from anywhere on the desktop or internet. Twine then uses semantic web technology to organize automatically all of your information by theme and then infer what other information might also interest you.

So says Wired News after a preview of Twine and an interview with Radar Networks’ founder Nova Spivack, who’s giving the public unveiling of Twine today at the Web 2.0 Summit in sunny San Francisco. Twine is a social network, slash collaborative editing tools, slash personal organizer, slash central information hub powered by AI tech to help cope with information overload. This is consistent with what I’d already gathered from hints in previous media coverage, but the full extent of the system becomes clearer in Tim O’Reilly’s article, published a few hours ago.

Underlying twine is Radar’s semantic engine, trained to do what is called entity extraction from documents. Put in plain language, the semantic engine auto-tags each document, turning each entity into what looks like a web link as well as a tag in the sidebar. Type a note in twine, and it picks out all of the people, places, companies, books, and other types of information contained in the note, separating them out by type.

Assortment of AI tech

Several screenshots from varying phases of Twine.com’s development have been released. O’Reilly has one demonstrating what he describes in the quote above, where keywords and information has been extracted from a document. How do they do that? Well, with an amazing assortment of technologies, it seems.

For one Twine mines whatever metadata may already be present in the documents it’s working with. Secondly, it runs natural language processing on the content to analyze entities and convert them to semantic tags. User tagging is also there, and according to Spivack around 80% of tags are automatically found and suggested for the user to minimize work. Additionally, Twine uses machine learning to categorize data using the 300,000 taxonomic categories of the Wikipedia for reference. Of more importance, Spivack mentions that imported data will not be locked in, it’ll be exportable with the added semantic information so that it can be reused somewhere else.

Will the Color of our Brick Road Change?

Twine.com will be in an invite-only beta phase for some time after today, but I’m sure we’ll start seeing hands-on reviews soon enough. I for one hope the system works as well as it sounds; from all the wonderful web applications that’ve debuted in the past two years, I’m continually frustrated over not having one that combines them in one place.

If Twine delivers it means we’ll be getting that cake with intelligent frosting plus a nice warm cup of IQ Cappuccino. Not to forget that its success would make a splash in semantic web development, adding a yellow brick to the road leading away from the ever-less-productive methods of modern search and data organization.

Links and References

• Keep Your Eye on Radar Networks
• Radar Networks to Unveil Its Semantic Web App, Twine, Wired News
• Web2Summit: Radar Networks Unveils twine.com, Tim O’Reilly
• Radar Networks‘ official company web
• Radar Networks on the Wikipedia

No related posts.