Theoretical neuroscientists and mathematicians gathered at SFI recently to explore new ways to let “embodied intelligent systems” – that’s robots – learn coordination.
At the small working group October 8-11, 2013 at SFI, nine researchers began hammering out a new theoretical approach for enabling robots to learn to walk, for example, in much the same way infants do.
“It saves computational power,” explains SFI External Professor and working group co-organizer Nihat Ay, “because you let the intelligent system express its own walking behavior specific to its own body” rather than try to copy another’s walk. Just as people learn to walk with various body shapes, levels of agility, or injuries, the idea is to give intelligent systems coupled with their mechanical selves the algorithms they need to learn how their own bodies walk.
Walking, of course, is only representative of the group’s goal to enable such systems to explore and later exploit their individual bodies to accomplish a variety of movements and tasks.
“But there’s no general theory that can be used to design intelligent systems that exploit their bodies,” says Ay, who proposed the working group along with co-organizer Fritz Sommer of UC Berkeley in an effort to combine their two groups’ complementary algorithmic approaches and so start generating such a theory.
Sommer’s approach helps the robotic system choose the best actions to improve its internal world model, which is needed for it to respond to perceptions of the world with an appropriate behavior. Ay’s approach helps the robot produce coordinated behavior among physically coupled parts such as the lower leg and foot, but requires an accurate internal world model to make sense of sensory inputs in order to, say, take the next step without falling.
At the conclusion of the meeting, participants had combined the two complementary theoretic approaches and implemented and experimented with them in a virtual environment.
The next step, says Ay, is to test the efficiency of the combined approach and publish the findings. “We have made a commitment,” he says, “so we’ll be meeting again in the near future.”
The German science foundation DFG supported this event within its Autonomous Learning program.