Robots can be surprisingly fragile. A breakdown in a key component can leave even the most advanced and expensive machines disabled or functioning below peak performance.
While the smartest of self-learning machines can adapt to the breakdowns and resume their normal function, this has traditionally been a slow process, as the robots’ programmes work through thousands upon thousands of options. Now though researchers have developed algorithms which speed the learning process, cutting time for adaptation to minutes instead of hours.
Detailed in the journal Nature, researchers have shown how giving some additional guidance to a trial-and-error algorithm can slash the time it takes a robot to figure out how to get back to work. In theory, robots could be taught ‘previous experience’ helping them to eliminate the myriad of bad options from the choices they consider as they adapt to any damage.
Trials of the algorithm with a six-legged hexapod robot showed how a damaged machine could rapidly figure out how it was affected by the damage and find an alternative way to work. If the work continues to be successful it could represent a major step forward in creating adaptable robots at a much lower price point than current levels.