Perceptive Locomotion

A legged robot using cameras or lidar to see the terrain ahead and plan its steps accordingly.

Perceptive locomotion means a legged robot uses both proprioception (internal body signals such as joint angles and IMU readings) and exteroception (external sensing, such as elevation maps built from depth cameras or lidar) to spot steps, ditches, and obstacles ahead of time and adjust its gait — in contrast to blind locomotion, which “feels” its way using proprioception alone. The difficulty is that exteroception is not always reliable: snow, tall grass, and reflective surfaces can all corrupt the map. Work by Miki and colleagues at ETH Zurich, published in Science Robotics in 2022, fused both kinds of input with an attention-based recurrent encoder, trained in simulation and deployed on the ANYmal quadruped; when perception is distorted, the controller automatically leans more on proprioception. Most stair-climbing and parkour behavior seen in today's humanoids and quadrupeds falls under perceptive locomotion.

ExampleMiki and colleagues' controller drove an ANYmal quadruped through roughly an hour-long alpine hiking trail, finishing in about the time recommended for human hikers.

Also called

Vision-based Locomotion, Perceptive Legged Locomotion

Related

Blind Locomotion · Legged Locomotion · Rough-terrain Locomotion · Elevation Map · Proprioception · Exteroception

Sources

Learning robust perceptive locomotion for quadrupedal robots in the wild (Science Robotics, 2022)