The process is two-step, and dubbed ‘intelligent trial and error’.
Step one is performed once for each type of robot: Computer modelling techniques are used to try a huge number of possible behaviours and rate them – in the research case: leg movements rated against the forward speed they produce.
The most successful strategies – hundreds of them – are stored for future reference in a multi-dimensional map.
This map is now a resource, given to every physical robot of the modelled type.
Damage is not detected by sensing, for example, locked leg joints. Instead, it is deduced when the current walking behaviour no longer yields the forward speed it once did.
When this happens, step two is initiated.
The damaged robot tries the most successful of the stored simulated strategies – even though they were simulations of a fully functioning robot.
“A key assumption is that information about many different behaviours of the undamaged robot will still be useful after damage, because some of these behaviours will still be functional despite the damage,” said the team in a Nature paper: ‘Robots that can adapt like animals‘.
Unsuccessful strategies – those which result in slow movement or move in the wrong direction – are discarded. When it finds a strategy that is more successful, it searches the map – updating it from recent experience as it goes – for similar strategies that were even more successful in simulation.
Eventually, it determines further tries are unlikely to provide improvements and stops searching.
Recovery gaits were three to seven times more efficient than the reference gait after damage.
Using the map avoids trying millions of behaviours that were likely to fail, and searching it strategically rapidly finds something that works.
“Each behaviour it tries is like an experiment and, if one behaviour doesn’t work, the robot is smart enough to rule out that entire type of behaviour and try a new type,” said Antoine Cully of the Pierre and Marie Curie University in France. “For example, if walking mostly on hind legs does not work well, it will next try walking mostly on its front legs. What’s surprising, is how quickly it can learn a new way to walk. It’s amazing to watch a robot go from crippled and flailing around to efficiently limping away in about two minutes.”
The university, part of the Sorbonne, worked with the University of Wyoming.
According to the team, it can also allow a robot to adapt to unforeseen situations, a new environment perhaps, or improve its undamaged behaviour. Initially given a hand-crafted walking gait, the research hexapod improved its own undamaged forward speed by 30% using the technique.
Intelligent trial and error has proved versatile, also working on a multi-jointed robot arm that had to drop a ball into a cup.
The all-important map in the first step is created with a new type of survival-of-the-fittest evolutionary algorithm called ‘MAP-Elites’ – read the paper for an understanding of this process. Adaptation in the second step involves a ‘Bayesian optimisation’ algorithm that takes advantage of the prior knowledge provided by the map to efficiently search for a behaviour that works.
from News http://ift.tt/1eDu7ZT
via Yuichun
沒有留言:
張貼留言