Researchers led by Dohyeon Kim have trained a policy that makes the arm of a quadruped robot follow a path more precisely, using an inertial sensor at the end of the arm, according to a paper on arXiv. Errors from the motors, the contact of the feet and the robot's dimensions add up along the chain of legs and arm, and joint encoders cannot see most of them.

The reinforcement learning policy uses the inertial readings to correct these errors while the body keeps its pose and the feet stay where they are.

In real tests, the error measured by motion capture fell by 17.3% and 28.0% under two command conditions compared with the standard controller, and by a further 12.6% and 15.9% compared with a policy that used only the robot's own joint readings. Positions worked out from the joint angles differed from the measured ones by 12.61 mm on average, the team found.

Source: arXiv Robotics (cs.RO)

Background

Dohyeon Kim

Robotics researcher

Robotics researcher.