A Tale of Two Robots

The real test for any good idea in science is to use it as a guide when you build something. If you build something according to your theory, and it works just like the real thing, then you can take the hint that you're on the right track. For cognitive science, the rubber meets the road like this in the fields of robotics and artificial intelligence. AI has been a dream of the field since the dawn of the cognitive revolution; if we can uncover the algorithms of intelligence, people thought, then we can run those on anything, even a computer. This kind of 'good old fashioned artificial intelligence' (or GOFAI, as it's known) has had some notable successes. The massive supercomputer Deep Blue beat chess Grand Master Garry Kasparov in 1997, and Watson recently defeated several former champions at the difficult quiz show, Jeopardy!.

But more recently, these success have been overshadowed by robots built using more embodied principles, who are suddenly capable of a wide range of astonishingly biological looking behaviour. This is the tale of two robots: Asimo and Big Dog.

Robot #1: Asimo

The best known GOFAI robot is Honda's Asimo. His basic design is a complex computational "mind' which contains algorithms to solve problems of vision, navigation and balance, and these algorithms provide detailed instructions for his body to follow. Honda likes to showcase him running, jumping and interacting with objects (video).

Asimo has two kinds of problems that hint he is not quite up to the same things that we are, however. First, he is tremendously energy inefficient. In fact, his walking is 16 times less efficient than humans, who are able to take advantage of the passive dynamics of our limbs (their springiness, and how they're put together; see this work at Cornell for examples) to walk with nearly ideal efficiency. All the work of generating and controlling Asimo's walking is happening in his head, and his legs don't provide any help at all.

Meet Asimo, Honda's GOFAI robot

The second problem is that his behaviour is unstable. If you allow Asimo to execute his prepared programme in a safe environment, he typically does just fine. If, however, there is anything complicated in his world, or if he miscalculates by just a tiny bit, then he fails. Completely. Organisms out in the real world need some tolerance for error; we have to be able to fail as gracefully as possible if we don't want every little problem to become a disaster.

Asimo doesn't embody his cognition. All the hard work of making him go happens in his central processers and his body just happens to come along for the ride. The end result is an impressive feat of software engineering that can't get around the world better than a 4 year old human. Is this just a matter of computing power, or can we perhaps do a little better?

Robot #2: Big Dog

The MIT Leg Lab has long taken an embodied approach to robotics, focusing their attention on how the robot was built. The basic design principle was to build robots with very particular dynamical properties - springy limbs, joints in particular places, tails for balance that moved correctly because of how they were built and not in response to a computed command. Over the years, they have produced many promising test systems that tested various design ideas (see videos of them in action here) and often produced surprisingly biological looking motion (my favourite is Troody the dinosaur, which sometimes just looks like it's alive!).

Several years ago, Marc Raibert spun out the company Boston Dynamics, and this company builds robots for military and other applications. Their brief has been to create robots that can operate on battlefields and in unpredictable terrain; clearly something more than Asimo is required.

Meet Big Dog. This robot was built using the design principles pioneered by Raibert at MIT, and it has many astonishing capabilites:

Pay close attention from about 30s in; Big Dog is walking on an icy surface and is pushed, hard. Watch how it catches itself - it's fast, effective, and it looks very natural. Most impressively, that reaction is not the result of any fast computation, but is what happens when something built like Big Dog is knocked off balance. It's body solves the problem for it! If you watch the video all the way through, you'll see Big Dog walking up muddy hills, over ice and rocks as well as a messy pile of cinder blocks.

The Lesson for Cognitive Science

The difference between the two robots is this: Asimo mentally represents his abilities, while Big Dog embodies its. Big Dog doesn't need any complex computational machinery to produce stable, robust performance, and, if Asimo is anything to go by, such machinery might just get in its way.

Artificial Intelligence Essential Reads

Integrating AI Into Work: Start-Ups May Have an Advantage

The New Rules for AI

Most importantly, from a cognitive science perspective, is the fact that Asimo is bad at things we are very good at, while Big Dog does those things with ease. Remember, cognitive scientists are interested in robots because you can build one according to your theory of how people work and look to see how it goes. If it does well, then your theory is a good one. Asimo's good-old-fashioned AI design just doesn't seem to be a good description of how we work, while Big Dog's dynamical, embodied approach moves and behaves much more like us. It looks like, for cognitive science anyway, that embodiment is the right direction to be heading.

Further reading

Pfeifer, R., and Bongard, J. (2007). How the Body Shapes the Way We Think. MIT Press, Cambridge, MA.

Raibert, M. (2000). Legged robots that balance. MIT Press, Cambridge MA.

Robots, Representations & Dynamical Systems (Notes From Two Scientific Psychologists)