Legged Locomotion

The control problem of getting quadruped, biped, and other legged robots to walk, run, jump, and climb slopes stably.

Legged locomotion studies how legged robots — quadrupeds, bipeds, and others — move by alternately placing their feet on the ground. Compared with wheels, legs can cross discontinuous terrain such as steps, loose gravel, and grass, but the robot must continuously manage balance, switching foot-ground contacts, and the risk of falling. Traditional methods relied on dynamics models and model predictive control (MPC), which optimizes a short segment of future action at every step. In the last few years, reinforcement learning has become the mainstream approach: Joonho Lee and colleagues at ETH Zurich (Science Robotics, 2020) trained using only proprioceptive signals such as joint and inertial data in simulation, then transferred zero-shot to a real ANYmal quadruped, which could walk on mud, snow, and gravel; Nikita Rudin and colleagues (2021) ran thousands of simulated robots in parallel on a single GPU, training a flat-terrain policy in under 4 minutes and a rough-terrain policy in about 20. The field breaks down into sub-areas such as bipedal walking, blind locomotion, perceptive locomotion, and parkour, and is the most fundamental capability for both quadruped and humanoid robots.

ExampleAn ANYmal quadruped, using only its own joint and IMU signals and no camera, keeps walking through mud, snow, and running water (Lee and colleagues, 2020).

Related

Locomotion · Bipedal Locomotion · Quadruped Robot · RL-based Locomotion Control · Perceptive Locomotion · Sim-to-Real Transfer

Sources

Learning Quadrupedal Locomotion over Challenging Terrain (Lee et al., Science Robotics 2020)
Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning (Rudin et al.)