Env « se relever sur rollers » : départ au sol, cible = debout sur roues.

Commande du slot telle que la voit la policy : [cos(2*pi*phi), sin(...), 0].

Env spin sur rollers, piloté par la phase du slot ground-pick.

Update self.command (fed into obs) based on current policy and commands. Legacy mode (new_cmd_obs=False): self.command is 3D. New mod

Integrate forward pitch rate into the per-env rotation accumulator. Step-guarded so that multiple reward terms reading the accumulator in the

Smoothstep on the progress frontier: 0 below gate_lo rad, 1 above gate_hi. State-based replacement for the old phase-clock landing window — it ca

Entity-local indices of the servo (non-``passive_``) joints, cached. All joint-index-based reward/event params in this module (``joint_indices``,

Create Microduck ground pick environment configuration.

Create Microduck roller skate velocity tracking environment configuration.

Angle de rampe (radians) interpolé linéairement par la difficulté [0,1].

Per-env float mask: 1.0 where the robot counts as FALLEN — trunk height below `gate_z_below` OR tilt beyond `gate_tilt_above_deg`. Used to gate th

Env crouch-glide sur rollers, piloté par la phase du slot ground-pick.

Create Microduck velocity tracking environment configuration.

Blend 0..1 le long de la phase [0,1) — 0 = pose STAND, 1 = pose DOWN. [0, descent_end) : 0 -> 1 (se baisser) [descent_end, hold_end):

Masques de promotion/rétrogradation du curriculum de pente. move_up : a parcouru plus de 40% de la tuile → il a dévalé la rampe,

(slewed target trunk z, actual trunk z) per env.

A one-layer 'policy' with the daemon's shape, so the ONNX checks run without torch.

Récompense la vitesse d'avance vers le BAS de la pente (monde +x). La rampe descend en +x, donc la vitesse linéaire monde en x mesure la prog

Terminate environments where MuJoCo produced NaN joint positions. MuJoCo's contact solver can overflow to NaN under extreme penetration or im

Récompense le ROULEMENT des roues vers l'avant (glisse), plafonné. Contrairement à descent_speed (vitesse de la BASE, qu'on peut atteindre en

Torque off means limp, not frozen. Refusing to command a fallen robot only freezes it in the pose it fell in — which is why `RobotIo:

Per-env world-frame kick direction (XY unit vector), lazily allocated. Set by ``reset_ball_in_front_of_foot`` to the robot's forward direction at

0→1 ramp in body forward speed: 0 when standing still, 1 at/above v_ref. Used to gate stride-shaping rewards so that stepping which does NOT prop

Target blend ∈ [0, 1] (0 = STAND, 1 = SIT) for the posture rewards. Uses the SitStandCommand's slewed ``alpha`` (the moving setpoint) when the

(target blend, per-env joint target) for the commanded posture. STAND target = default_joint_pos (HOME); SIT target = HOME with the keyframe

Create Microduck stand environment configuration (sit-keyframe start).

Return one target angle per control step.

Gaussienne sur la pose articulaire vs cible interpolée STAND<->DOWN. Reward directif : indique la config articulaire exacte à chaque phase. Se

Rotate a vector by the inverse of a quaternion [w, x, y, z].

Set velocity command (used for walking / policy switching).

Récupère la phase [0,1) depuis la commande [cos(2πφ), sin(2πφ), 0] du slot.

(main_ids, backlash_ids, mask) — cached per (entity, joint selection). mask is 1.0 where a matching passive_<name>_backlash joint exists, so the

(cur, target) joint tensors for the phase-interpolated crouch pose. Target interpolates per joint STAND <-> crouch_pose by the 4-segment blend

True where the head-top axis points at the floor (dot with -z > min).

Per-env constant IMU mounting-misalignment rotation (sampled once). Models a fixed small mounting/calibration error of the IMU on each robot.

(cur, target) pour le geste de shoot, joints résolus PAR NOM. Les 3 poses partagent les mêmes clés (14 joints). L'ordre des noms est donné pa

World-z component of the body's lateral (y) axis. 0 = flat/sagittal.

Same logic as _multistage_target_pose but for trunk z height. waypoints: [{"frac": float, "height": float}, ...].

Compute the time-interpolated joint target across N waypoints. waypoints: ordered list of dicts {"frac": float in [0,1],

(cur, target) pour la pose interpolée par la phase, résolue PAR NOM. Cible = source + blend(phase)·(target_pose - source), source = STAND (`s

Apply action to MuJoCo controls with optional delay.

Blend 0..1 le long de la phase [0,1) — 0 = pose debout, 1 = pose accroupie. [0, descent_end) : 0 -> 1 (se baisser) [descent_end, hold_e

`""` for the first duck — it is the scene's own — and `d1_`, `d2_` … for attached ones.

Collect observations matching policy input. Order for velocity/standing task: 1. base_ang_vel (3D) 2. raw_accelerometer OR pr

The graph's single input and output widths, as the daemon checks them at load.

Roller swizzle env: the stride env minus its anti-swizzle terms, plus symmetry and grounded rewards. Everything else (robot, obs, command, DR) is

Bootstrap L1 vers la cible interpolée (pénalité négative). Gradient constant partout — donne une direction vers la cible même quand la gaussi

Gate montante pour le RETOUR : 0 avant hold_end, 0->1 sur [hold_end, rise_end), 1 après (repos debout).

Send joint position targets to the actuators. BAM: the firmware position loop lives in the controller (ctrl is the motor TORQUE it wr

Vitesse de lacet cible (rad/s, positive = anti-horaire) le long de la phase.

Parse submission args from ``argv`` and launch the HF job.

The entry whose resolution-markers SELECT linux-aarch64.

Poll until the job leaves SCHEDULING (image pull / queue / mounts). This phase legitimately takes minutes (the pytorch image pull alone is ~5

Map an axis name to the body_cmd index, depending on the active mode.

Create a tarball of HEAD + uncommitted tracked changes. Returns short SHA.

Switch back after a ground pick cycle completes.

True when argv[0] is mjlab's trainer (console script or `-m`). `play --hf-jobs` must NOT submit a training job; let that command's own parser