Rough-terrain Locomotion

A legged robot's ability to move stably over uneven ground such as stairs, gravel, grass, or snow.

Rough-terrain locomotion means a legged robot walks without falling over uneven, or even deformable, ground such as stairs, slopes, gravel, mud, and snow. Traditional approaches relied on hand-built models and manual tuning, and tended to fail as soon as the terrain changed. In 2020, a team at ETH Zurich used reinforcement learning to train an ANYmal quadruped in simulation: a “teacher” policy that could see the true terrain learned to walk first, then was distilled into a “student” policy that relies only on proprioception, internal body sensors such as joint angles and the IMU, which was deployed directly outdoors. A 2022 follow-up added external perception such as depth sensing, and reported completing about an hour-long Alpine hiking route in a time comparable to the human-recommended time. Rough-terrain locomotion is now a basic test of legged control, and it commonly appears alongside terrain curricula and domain randomization.

ExampleIn a 2022 Science Robotics paper, an ANYmal quadruped, using a policy that fused depth perception with proprioception, completed a roughly hour-long Alpine hiking route.

Related

Legged Locomotion · Perceptive Locomotion · Blind Locomotion · Terrain Curriculum · Teacher-Student Distillation · Sim-to-Real Transfer

Sources

Learning Quadrupedal Locomotion over Challenging Terrain (Lee et al., Science Robotics 2020)
Learning robust perceptive locomotion for quadrupedal robots in the wild (Miki et al., Science Robotics 2022)
Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning (Rudin et al., CoRL 2021)

As of

2022-01