Locomotion

Introduction

Locomotion is how a humanoid moves its whole body through the world: walking, turning, running, climbing and recovering its footing. Most platforms in this database walk on two legs; some, such as the Reachy 2, put a humanoid upper body on a wheeled base. Each robot's specification table below records how it moves and, where the maker publishes it, how fast.

On developer platforms the usual starting point is now a learned controller. Policies are trained with reinforcement learning on thousands of simulated robots in parallel on a single GPU, then transferred to the hardware ^[2]. Isaac Lab ships ready-made humanoid tasks that train a policy to follow a commanded velocity ^[1], and Unitree documents a Train, Play, Sim2Sim, Sim2Real workflow for its humanoids, deploying a policy to a second simulator before the robot ^[3].

Robot models matter as much as the algorithm: a policy is only as good as the model it was trained on. Manufacturers and projects publish URDF and MJCF models for that reason; MuJoCo loads both ^[4].