Motion command

Motion command#

MotionCommand (env/mdp/commands.py) is the central command term. It loads a multi-clip motion bundle (NPZ), plays reference kinematics each step, and drives RSI resampling at episode boundaries.

Configuration#

MotionCommandCfg fields (set per robot entity in <robot>_base_cfg):

Field

Meaning

motion_file

Dataset folder of source clips (played online, How the library is built), a source clip, or converted NPZ (a .npz bundle or a folder of npz/*.npz)

motion_source

auto (online when there are source clips) · online · npz

mirror · debias_z

Online: add mirrored clips (virtual, needs the robot’s symmetry config) · shift each clip so its lowest foot is on the ground

source_format · source_fps

Online: source format name and rate, when the extension / file doesn’t tell

motion_storage

Online: source frames on the env device (default) or in pinned host memory

robot_id

Robot of the motion (symmetry, feet); set by the robot’s env builder

anchor_body_name

Body used for anchor-frame errors and assistive wrench (set in robot entity)

body_names

Keybodies tracked in rewards / RSI / critic obs

actuated_joint_names

Optional subset for joint tracking metrics and RSI

pose_range / velocity_range / joint_position_range

DR on reference root pose and joint offsets at resample

rsi

RSI configuration settings (RsiCfg)

assistive_wrench_enabled / assistive_beta_max / assistive_eta

Bin-conditioned assist on the anchor (Zest-style curriculum)

Runtime behavior#

Each environment step:

  1. Advance time_steps along the active clip segment.

  2. Expose reference body/joint states to reward and observation terms via command_name="motion" params.

  3. Apply assistive wrench (when enabled) scaled by the current bin’s failure level.

At episode reset (_resample_command):

  1. Pick (trajectory, bin, frame) from RSI policy (adaptive / uniform / start).

  2. Write reference root + joint state into sim (with pose/joint DR).

  3. Initialize episode RSI bookkeeping (start bin, similarity denominator, assist gain).

Multi-clip playback#

MotionLoader indexes a bundle with multiple segments. Each parallel env can be on a different clip. resampling_time_range in the base template is effectively infinite — episodes end via terminations, not command timeout.

Reference vs robot state#

Tracking rewards compare robot bodies/joints to reference values from the command. The action space adds residual joint deltas on top of reference joint positions (ReferenceJointPositionActionCfg), so the policy tracks the command rather than absolute pose targets.

Assistive wrench#

When assistive_wrench_enabled=True, compute_assist_gain_matrix maps bin failure levels to a gain β ∈ [0, assistive_beta_max]. Hard bins get more assist during early training. Play mode disables assist (see Architecture).

Visualization#

MotionCommandCfg.viz selects ghost vs frame debug (see Evaluate and visualize). WBC extends mjlab’s Viser play viewer with reference alignment (anchor-relative xy/yaw overlay), tracking-error coloring, and RSI bin panels during wbc-mjlab-play --viewer viser.

API: MDP terms (Motion command).

See also