Start an episodic behavior (kick_left / kick_right / roulade). The behavior policies were trained to run from a standing start with an

Start one ground pick cycle. Automatically returns to walking when done.

Advance the behavior timer; hand back to walking/standing when done.

Advance the ground pick phase; auto-exit when one full cycle completes.

Soften terrain box geom contacts to reduce edge-contact NaN instability. Box terrains place adjacent geoms at different heights. The hard edges w

Penalise commanding a joint target beyond its hard limit (+ overshoot). Policy-side deterrent against over-driving a joint onto its mechanical st

Reduce pose tracking weight when robot has high velocity (recovering from push). This gives the robot freedom to deviate from the standing pose w

Air-time reward with separate swing-time windows for walking vs running. - command < command_threshold → 0 (standing, no reward) - command_t

Slewed target blend: 0 = STAND target, 1 = SIT target.

Hook par-step (utilisé comme reward de poids 0) : applique le POIDS de l'objet tenu dans la bouche comme force externe verticale au mouth_tip, gat

Ball XY velocity along the per-env kick direction, clamped to [0, max]. Dense and linear-in-speed up to ``max_speed``: every extra bit of forward

Ball position relative to the robot root, in the robot's base frame. CRITIC-ONLY observation (asymmetric actor-critic): the deployed policy has

Ball forward speed in excess of ``target_speed`` (linear, ≥ 0). Companion to ``ball_forward_velocity`` for a target-speed kick: below the tar

Ball linear velocity in the robot's base frame. CRITIC-ONLY (see above).

base angular velocity with the SAME per-env IMU misalignment as gravity.

L1 penalty on left/right leg asymmetry. For a bilaterally-symmetric robot the leg HOME and any symmetric target (FOLD, SIT) satisfy ``q_left

Trunk ``sum(ω_xy²)`` penalty gated by trunk z (and optionally tilt). Height-gated arrival damper: zero below ``height_low`` (ground recovery —

Penalize terrain contact forces above a threshold on protected body parts. Used to discourage slamming the trunk shell or head into the ground du

Mean of 6 per-axis Gaussian rewards for tracking commanded body pose. cmd has shape (N, 6) = [x, y, z, roll, pitch, yaw] all as deltas from the

Locomotion-aware 6D body pose tracking. Same shape as body_pose_tracking_6d (6D cmd, mean of 6 Gaussians), but x/y/yaw are measured *relative

Gaussian reward on tilt magnitude — sharp pull toward fully vertical. Complements ``body_upright_linear`` (which is ``cos(tilt)`` and whose g

Linear reward for body uprightness — provides gradient at every tilt angle. Returns +1 when fully upright, 0 when horizontal (prone/supine), -1 w

Reward coming to a stop when cmd_x < 0 (brake commanded). Returns clamp(-cmd_x, 0) * exp(-fwd_vel² / vel_std²). - Silent when cmd_x ≥ 0 (coas