A single-policy manifest the daemon loads without surprises. Only the constant-command family is publishable from here — a skill's network is fed

Refuse a file the daemon would refuse at load: wrong widths, or one that is not 61 -> 14.

Reward coasting: low leg-joint velocity while at target speed. Returns exp(-vel_error / vel_std²) × exp(-sum(joint_vel²) / stillness_std²). B

Reward for keeping the center of mass within a target height range. Returns positive reward when in range, negative penalty when outside.

Reward gaussien : projection horizontale du CoM proche du pied d'appui, gaté sur la phase de frappe (kick_engagement). Apprend le transfert l

Update CoM randomization range based on training progress. Gradually increases the CoM offset range so the robot first learns to walk with a

Reward upward CoM velocity to incentivize dynamic standup motion. Gated by height: only active while the CoM is below `max_height` (the stand

Penalize high frequency of contact changes to encourage slower stepping. Tracks the number of contact state changes per second and penalizes

Léger penché AVANT du tronc pendant l'accroupi (gaté par le blend crouch). Contre la bascule arrière induite par la flexion rapide des hanches. P

Reward principale : suit la cible de hauteur du tronc le long de la phase. La hauteur du CoM est calculée comme dans `com_height_target` (world z

Gaussian match to a phase-interpolated joint pose (stand <-> crouch). Directive reward: tells the robot the exact joint configuration to be in at

L1 bootstrap toward the phase-interpolated crouch pose (negative penalty). Constant gradient everywhere — gives the policy a direction to the tar

Mutate an event term's params at scheduled steps. Mirror of termination_param_curriculum but for events. Uses the live EventManager term cfg

No-op startup event whose only purpose is the decorator above. bam's BamActuator (mjlab_frictionloss branch) writes a per-env friction budget

Ramp face_down_prob on a reset event over training. Args: event_name: name of the event term using set_random_prone_orientation p

1.0 while FALLEN (weight it negative): a flat per-step tax on staying down. Without it, lying still is ~0/step while attempting recovery costs

Terminate envs that have been continuously FALLEN for `max_duration_s`. For envs that mix walking with fall recovery (velstand): the fell_over

velocity template feet_air_time, zeroed while FALLEN (tilt > gate). velstand: a robot lying on its trunk can still tap its feet rhythmically

Penalize the feet getting too close to each other in the horizontal plane. Returns ``clamp(min_dist - d, min=0)`` per env (use with a negative we

Penalize foot sites not being parallel to the ground. The foot site frame has Z+ pointing up when flat. We project a unit gravity vector (poi

Positive reward for feet contacting the ground (0, +0.5, or +1.0). Uses the contact sensor's `found` field. For the feet_ground_contact sensor

Reward each foot site's local x-axis pointing downward — tiptoe stance. When flat, foot site x points roughly forward (horizontal). Pitching the

NaN-safe `foot_air_time` (see note above).

NaN-safe `foot_contact_forces` (see note above).

NaN-safe `foot_height` (see note above).

Penalise stepping when at zero command and the body is not being pushed. Symmetric counterpart to the air_time reward: - air_time gives +rew

Reward leaning slightly forward when pushing, to counteract the backward torque from skating strokes. Uses projected_gravity_b x-component as

Récompense la vitesse avant du tronc (conserver l'élan / ne pas freiner). Indépendante de la commande (la commande porte la phase, pas la vitesse

Penalize lopsided left/right foot usage (one blade doing most of the work). With symmetry augmentation OFF, nothing stops the policy learning an

`commit`, `branch`, `dirty` of the checkout the export ran from, or `{}` outside git.

Reward the GLIDE phase of a stride: coast on ONE blade with quiet legs. Nothing else rewards gliding — skating_air_time pays each swing, so the p

Reward for returning to the standing pose after ground pick, weighted by the return phase. The return phase is the second half-cycle (sin < 0, ph

ground_pick_return_pose gaté par la up-gate segmentée (remontée+repos).

Reward trunk verticality, weighted by the RETURN phase (stand-up aid). Same return weighting as ``ground_pick_return_pose`` (``max(0, -sin(2π·pha

ground_pick_return_upright gaté par la up-gate segmentée.

Reward BOTH blades in contact — a classic swizzle stays grounded (no lifting). Mirror of single_support_reward but rewarding double support (n_co

Penalize the time-averaged (DC) neck/head tracking error: -mean(|EMA(err)|). Companion to ``head_pose_tracking``, which scores the INSTANTANEOUS

Reward holding the SPAWN heading (go straight) — corrective, angle-based. Rewards the yaw ANGLE staying near the heading captured at reset:

Reward for reducing heading error when cmd[2] encodes heading error. Returns exp(-cmd[2]² / std²). - At error = 0 (on heading): reward = 1.0.

L1 companion to ``height_target_gaussian``.

Potential-based height shaping: Δ min(trunk z, ceiling) per step. The z-axis companion to ``upright_progress`` (velstand crouch-endpoint less

Gaussian on trunk z against a single fixed target.

Penalize hip_pitch and knee joint velocities (L2 squared). Walking requires rapid oscillation of these sagittal-plane joints. Skating uses hi

L1 distance from a time-interpolated target height (negative — penalty). Same role as ``interpolated_pose_l1_penalty`` but on trunk z. Provides a

Gaussian on trunk z vs a time-interpolated target height. Companion to ``interpolated_pose_target_match`` — same time-interpolation logic app

L1 distance from a time-interpolated target pose (negative — used as penalty). Same interpolation schedule as ``interpolated_pose_target_match``

Gaussian on joint positions vs a time-interpolated target pose. Tracks a target that linearly interpolates from a source pose to a target pos

Reward for staying alive (not terminated) Args: env: The environment Returns: Reward tensor of shape (num_envs,) - ones

Penalize joint accelerations using L2 squared norm. Joint accelerations are computed using finite differences of joint velocities. Args:

L1 penalty for joint positions deviating from their default (HOME). Returns sum of |joint_pos - default| over the selected joints. Unlike the

L1 penalty for joint positions entering a ``margin`` (rad) band next to their *hard* range limits. The base ``joint_pos_limits`` reward only

joint_pos_rel where the encoder reads through the backlash hinge. Returns (qpos[servo] + qpos[backlash]) - default[servo]. With biased=True t

Penalize rate of change in actuator torques (proxy for gearbox shock). Sudden torque spikes occur when the robot impacts the ground and actuators

Penalize actuator forces (torques) to encourage energy-efficient motion. Args: env: The environment asset_cfg: Asset configu

Penalise leg joint velocities only when command is near zero. Targets the standing-shake problem: the policy makes rapid oscillating correcti

joint_vel_rel where the encoder reads through the backlash hinge. The firmware derives present_velocity from encoder positions, so it also se

Gaussienne sur la pose articulaire vs cible interpolée du shoot. Reward directif et symétrique : chaque phase impose la config articulaire ex

Bootstrap L1 vers la cible interpolée (gradient constant, pénalité<=0).

Penalize leg action accelerations (action_t - 2*action_{t-1} + action_{t-2}). Leg joints are indices 0-4 and 9-13 (10 joints total). Arg

Penalize the rate of change of leg actions (action_t - action_{t-1}). Leg joints are indices 0-4 and 9-13 (10 joints total). Args:

Amorce le CISEAU des jambes (une avant / une arrière) pendant le spin. Le robot a des conventions de signe MIROIR gauche/droite : une pose sy

Penalize leg joint velocities to encourage smoother, less dynamic motion. Leg joints are indices 0-4 and 9-13 (10 joints total). Args:

Reward left/right legs mirroring — the swizzle's defining symmetry. The robot uses mirrored L/R sign conventions, so a bilaterally-symmetric conf

Reset event that overrides orientation to prone with probability `prone_prob`. With prob `prone_prob`, replaces the upright orientation (already

Bilateral symmetry augmentation / mirror function for the microduck vel env. Returns [original, mirrored] concatenated along the batch dimension.

Reward for mouth tip approaching the ground, weighted by the approach phase. The command for the ground pick task is [cos(2π*phase), sin(2π*phase

mouth_ground_proximity gaté par la down-gate segmentée (descente+palier).