Michigan researchers test a robot controller that learns physical interactions

University of Michigan researchers have developed a neural-network component called PRISM that learns interactions between a robot's movements, sensor readings and instructions. They report improvements in simulated walking and manipulation, as well as a test with a physical robotic arm.

In the hardware test, an SO-101 arm picked up a roll of tape successfully in 94% of 50 attempts, compared with 68% for a conventional controller trained on the same demonstrations. Those results concern one task, not a general robot success rate.

In simulated humanoid walking, the team reports completing 92.5% of 24-second trials without a fall, compared with 51% for a conventional network. Enlarging the conventional network to match PRISM's size brought its result to 52.3%.

PRISM replaces the first processing stage of a controller rather than requiring new sensors or different training goals. The work is a preprint and was presented at the Conference on Robot Learning.