Researchers led by Owen Sullivan have built CADENCE, a learned step planner for four-legged robots that does not follow a fixed gait, according to a paper on arXiv. Model-based controllers usually lift the legs in a set order at a set rhythm, wherever the safe footholds are.

Every 40 milliseconds CADENCE picks one of up to 65 options, either no step or one leg's step to a safe foothold, and predicts how long the swing takes. It was trained with reinforcement learning on top of a model predictive controller, with no gait in the reward, only a cost per step and a limit of two legs in the air. The policy developed a gait with no fixed rhythm or order that still paired diagonal legs.

Moved unchanged onto the OCS2 controller, with only a small footstep model refitted, it succeeded in 85.1% of simulated trials on terrain with 26 to 49% gaps, against 67.8% for OCS2's trot. On the terrain with 49% gaps it succeeded in 72.9%, against 43.0%.

Source: arXiv Robotics (cs.RO)

Background

Owen Sullivan

Robotics researcher

Robotics researcher.