Researchers led by Raphaël Ségas have trained a walking policy for a Unitree Go2 quadruped that keeps its joint motors from overheating, according to a paper on arXiv. On legged robots, long walks often end in thermal shutdown well before the battery is empty, because the heat of walking concentrates in a few motors.

The team built a thermal model that predicts all twelve motor temperatures with a mean error below 1 °C across robots, gaits, loads and speeds. The reinforcement learning reward pushes down the temperatures the hottest motors would settle at, so the policy learns to lower and spread the heat without ever reading a temperature.

Carrying 5 kg at 1 m/s, a standard learned policy hit thermal shutdown after 19 minutes and 1.1 km. The new policy kept going until the battery was empty after 174 minutes and 10.0 km, and its hottest motor stayed 14 °C below the limit. On a factory circuit with stairs and ramps, the baseline overheated after 25 laps (6.2 km), and the new policy completed 47 laps (11.2 km) on one battery.

Source: arXiv Robotics (cs.RO)

Background

Raphaël Ségas

Robotics researcher

Robotics researcher.