Velocity Command Tracking

The standard training task where a legged robot walks according to a given forward, lateral, and turning speed.

Velocity command tracking is the most basic task setup in reinforcement-learning locomotion for legged robots: at intervals, the robot is randomly given a target velocity — typically forward body speed vx, lateral speed vy, and yaw rate ωz about the vertical axis — and the policy outputs joint actions so the body's actual velocity matches the command as closely as possible. The reward is often written as exp(−‖v_command − v_actual‖²/σ), approaching 1 as the error shrinks, with σ controlling the tolerance. Once trained, the deployed policy takes velocity commands from a joystick or a higher-level navigation module, which only decides ‘which way, how fast,’ while the policy itself handles gait and balance. It's different from motor-level ‘velocity control,’ which governs a single joint's rotational speed.

Examplelegged_gym's default configuration uses a 4-dimensional command (vx, vy, ωz, and heading), resampled every 10 seconds with linear velocity ranging ±1 m/s; the linear-velocity tracking reward has weight 1.0 and the angular-velocity reward 0.5, with σ = 0.25, and commanded lateral velocities under 0.2 m/s are zeroed out.

Also called

Velocity Tracking, Command Tracking

Related

RL-based Locomotion Control · Reward Function · cmd_vel Topic (geometry_msgs/Twist velocity command) · legged_gym · Walk These Ways · Terrain Curriculum

Sources

legged_gym: legged_robot.py（_reward_tracking_lin_vel / _resample_commands）
legged_gym: legged_robot_config.py（commands / rewards）
Walk These Ways (arXiv:2212.03238)