Locomotion Control

The control technology that keeps a robot moving stably as intended; in embodied AI it usually means legged walking and balance.

Motion control is originally a general term in automation, meaning making a machine's parts move as required, with position, velocity, or force as the controlled quantity, implemented as a closed loop of controller, driver, motor, and encoder. In the context of embodied AI and humanoid robots, 'locomotion control' (a shortened term often just called 'motion control' in Chinese) more specifically refers to a legged robot's ability to walk, run, balance, and get up after a fall — the core capability of the 'cerebellum.' There are two main approaches. Model-based methods use a simplified dynamics model together with model predictive control and whole-body control, solved in real time. Learning-based methods train a policy with reinforcement learning inside massively parallel GPU simulation, then transfer it to the real robot. In recent years the learning-based approach has become increasingly common for both quadrupeds and humanoids, and is often combined with the model-based one. Locomotion control handles 'how to move stably'; motion planning handles 'which path to take' — the two divide the work differently.

ExampleRudin and colleagues trained an ANYmal quadruped's walking policy in 2021 using thousands of parallel simulated robots on a single GPU, finishing training on flat ground in under 4 minutes and on rough terrain in about 20 minutes, and then deployed it to the real robot.

Also called

Motion Control, Locomotion Policy

Related

RL-based Locomotion Control · Whole-Body Control · Model Predictive Control · Balance Control · Robot Cerebellum · Motion Planning

Sources

Wikipedia: Motion control
Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning (arXiv:2109.11978)
qiayuanl/legged_control（NMPC + WBC 足式运控框架） (Chinese)