Organised chaos gets robots going
- 09:45 01 November 2004
- Exclusive from New Scientist Print Edition. Subscribe and get 4 free issues
- Will Knight
A control system based on chaos has made a simulated, multi-legged robot walk successfully. The researchers behind the feat say it may have brought us closer to understanding how people and animals learn to move.
Standard robots control their leg motion either through complex computer programs or by using so-called genetic algorithms to “evolve” a successful walking strategy. Both these options are time-consuming and require a lot of computer power.
Roboticists Yasuo Kuniyoshi and Shinsuke Suzuki wondered whether chaotic systems might also generate efficient walking behaviour. Chaotic systems behave in a way that means that small effects are amplified so rapidly that the systems’ behaviour becomes impossible to predict more than a short time ahead. Such chaotic systems are behind a number of phenomena, including the weather and the performance of financial markets.
The Tokyo University pair reasoned that just as the chaotic maths that determines the weather can produce clear patterns such as hurricanes and weather fronts, similar systems might underlie the movement patterns involved in locomotion. “We, and animals, seem to be able to work out how to move in different situations without going through thousands of trial-and-error situations like today’s robot-control software does,” says Kuniyoshi.
To test their idea, Kuniyoshi and Suzuki devised a computer simulation of a 12-legged machine in which each leg was controlled by a chaotic mathematical function. The functions were initially fed 12 parameters chosen at random. From then on, sensory information from each limb was fed back into the chaotic function that controlled it.
The team found that certain combinations of starting parameters made the robot’s limbs rapidly adopt “walking-on-the-spot” behaviour, but the machine did not get anywhere. However, when they placed a weight at one end of the simulated robot (see graphic) they found that four of the legs seized up, allowing the front and back legs to dominate movement and let the robot scamper along.
The robot could also negotiate obstacles in its path. After scuttling about for a few seconds, its mode of locomotion would change to allow it to scramble over whatever was in the way. Although it was just a simulation, the software mimicked the robot’s performance in fine detail. Kuniyoshi is confident that the trick will work in a real robot.
Remarkably, the robot performed these tricks without any conventional programming. And its behaviour emerged far more quickly than it would if it had used genetic algorithms. Kuniyoshi suggests that his chaotic approach may have similarities to the way that biological systems learn to move.
“Many findings point to the presence of chaotic patterns in general in the human brain,” says Max Lungarella, who researches artificial intelligence at the University of Tokyo. But Kuniyoshi and Suzuki’s approach is still unconventional, he says. “It diverges radically from the traditional way of thinking about intelligence.”
Roberto Fernández Galán, a biophysicist at Carnegie Mellon University in Pittsburgh, Pennsylvania, also finds the approach intriguing, but he is sceptical about the Japanese team’s idea that chaos plays a role in animal locomotion. “It is surprising to achieve what they call goal-directedness with a chaotic robot,” he says.