Robot spy can survive battlefield damage
- 19:00 20 August 2003
- Exclusive from New Scientist Print Edition. Subscribe and get 4 free issues.
- Duncan Graham-Rowe
The serpentine spy is a research project funded by aerospace company BAE Systems to make a low-cost military robot that can be dropped out of helicopters to carry out reconnaissance missions. Because it is not wheeled, the low-profile, ground-hugging snakebot should make a versatile battlefield spy. The team behind it has also developed a shape-changing antenna that broadcasts high-quality video and audio.
A self-healing robot has long been a dream of robotics engineers, not least because the machines are notoriously unreliable and absolutely terrible at dealing with unforeseen circumstances.
"When a dog loses a leg it's got a clever enough brain to allow it to adapt," says computer scientist Peter Bentley at University College London. But robots still lack this adaptive ability and so tend to give up the ghost when circumstances change.
Bentley and his colleague Siavash Haroun Mahdavi borrowed a trick from evolution to allow their robot to adapt to damage. The snakebot is made up of modular vertebral units that "snap" together to form a snake-like body (see graphic).
Each unit contains three separate "muscles" running down its length. The muscles are made out of wires of a shape-memory alloy called nitinol, an alloy of nickel and titanium whose crystal structure shrinks when an electric current is applied to it. Usefully, it regains its original shape and length once the current is removed.
To make the snakebot move in a particular direction, a current is applied to certain wires. When the current is removed, the wires spring back and the robot will jump forward.
The software for making a robot wriggle like a snake is fairly straightforward. But ensuring that the snake will keep moving even if a segment is damaged is trickier, and relies on different segments taking over from the damaged ones.
So Bentley and Mahdavi have created a genetic algorithm (GA) - a software routine that takes a "survival of the fittest" approach - to produce a system that continually evolves to improve itself.
The program starts off with a population of 20 digital chromosomes, with each consisting of an initially random binary digit that corresponds to a muscle wire - where a 1 represents its activation and a 0 its deactivation. Each of these chromosomes forms the basis of a series of movements in the robot.
"You end up with a cyclic pattern of muscle activation," says Bentley. Some may result in the robot moving and some will not. The GA tries them all out and awards them a fitness rating, depending on how far it makes the snake move.
The two best chromosomes are then saved, the remainder are mixed up or randomly mutated and the process is repeated. After a number of generations, the amount of improvement finally tends to taper off, says Mahdavi, indicating that the GA has reached a performance plateau.
Once the robot was mobile, the team disabled some of its segments to see if it could adapt to injury. Initially it was immobilised, says Bentley, but as the GA continued to try to improve the locomotion, it gradually worked out how to move again, albeit more awkwardly and with an ungainly, dragging gait - but it was still good enough to get the robot to its destination.