Choose the namespace (personal account or org) the job runs under. Interactive prompt unless --namespace was given or there is nothing to cho

Teleport the ball in front of the kicking foot, matching training's reset_ball_in_front_of_foot (offset in the robot's yaw frame).

Read one byte from stdin, or None on timeout. os.read (unbuffered): buffered sys.stdin.read would swallow escape-sequence bytes past what

Repo root of the CURRENT directory — worktree-aware. (The old scripts/hf/train_hf.py used the script file's location; resolving from cwd inst

Difficulté de play : env SLOPE_PLAY_DIFFICULTY sinon la constante.

Proportion de départs sur le dos au play : env STANDUP_PLAY_FACE_UP sinon la constante.

The ONNX to publish and the provenance it carries. Exports when given a checkpoint.

Switch between walking and standing sessions based on vel_cmd magnitude.

Best-effort lookup of the user's wandb API key. Order: WANDB_API_KEY env -> ~/.netrc (machine api.wandb.ai).

Helper to add real and sim traces to a subplot.

Reference: BAM's own Python simulator.

Construit le modèle MuJoCo du terrain seul (rows rampes de raideur croissante).

One model holding `count` ducks, so they share a floor and can bump into each other. The scene already contains one duck; the rest are attached t

M6 friction computation matching our kernel (and BAM's model.py).

Récompense gaussienne du suivi de la cible de hauteur (fonction pure). Décode la phase depuis [cos, sin] puis compare la hauteur mesurée à la

Cible de hauteur du tronc « en trapèze » le long de la phase [0,1). phase ∈ [0, hold_lo) : descente height_high -> height_low phase ∈

One 8x8 depth frame, in the units `tofd` publishes. Sixty-four ray casts, so this is the most expensive thing here — asked for at the sensor'

Return the most recent run whose metadata email contains *user_filter*.

Return up to *n* most recent runs whose metadata email contains *user_filter* and whose task_id (metadata args[0]) satisfies *task_match*. ta

One capture: distances in millimetres and a status per zone. Self-hits are reported, not filtered. A real sensor sees the duck's own beak whe

Get base angular velocity from IMU gyro sensor.

Get joint positions relative to default pose.

Get joint velocities.

Drain and return all pending keys (symbolic names / characters).

Get projected gravity in body frame.

Get raw accelerometer reading from MuJoCo sensor.

World gravity in the trunk frame. Upright is `[0, 0, -1]`. What the policy actually observes, and it is here rather than in the daemon because th

Per-joint Gaussian reward for matching commanded neck/head deltas. Mean over the 4 neck/head joints of exp(-(err/std)^2). Result is (N,) in [

Run policy inference and return action.

The `robotctl` lines that put this policy on a robot — one story per shape. Episodic: a skill, length from the manifest. Perpetual with `unwind_s

Gate d'engagement du geste ∈ [0,1] (pur) — pour pondérer les rewards d'équilibre unipède qui ne doivent s'appliquer que hors du repos STAND.

Cible articulaire interpolée d'un geste de shoot à 4 keyframes. phase (B,) ∈ [0,1). stand/back/forward (k,) ou (1,k). Retour (B,k). [0, wind

Build the BAM M6 model + XL330 voltage-controlled actuator.

Load the scene and hand every non-passive actuator to bam.mujoco.MujocoController. Mirrors bam.mjlab.BamActuator.edit_spec (what warp does at tra

Load observations from pickle file.

Convert a microduck env cfg (velocity/velstand/standup/...) to backlash.

Create the Microduck BallKick environment configuration. ``kick_foot`` overrides the module-level KICK_FOOT flag (used by tests); normal trai

Create Microduck forward-roll environment configuration.

Create Microduck sitstand environment configuration.

Consume `--hf-jobs` and exit the process; a no-op without the flag. Called at import time of `mjlab_microduck.tasks`, so it runs inside mjlab's

Run the testbench in MuJoCo with our BAM M6 actuator logic.

Convert a rollout .npz (written by rollout_sim/rollout_real) to a BAM log json. BAM log format (see ~/Rhoban/bam/bam/logs.py): top-level: m

Plot comparison between real and simulated observations using Plotly. If sim_obs is None, only plots real data.

A named pose from the scene's keyframes, as joint name → angle. Read from a single-duck model and applied by name, because attaching a robot does

Print run details and return dict with env_name, run_path, checkpoints.

Set the trunk's world-frame xy velocity to a random vector of magnitude PUSH_MAX, simulating the push_by_setting_velocity training eve

Find latest run for user (optionally filtered to a task-id substring), print info, return (run, info). Exits on failure.

Put this duck back where it was, for the one that has not been enabled yet.

Step in real time. **Real time, not as fast as possible.** The daemon's loop is wall-clock and its health gate fails below 45 of 50 Hz, so a

Print and optionally execute a command from the project root.

Reset selected envs into a random mid-recovery crouch. Reverse curriculum for the recovery last mile (velstand run-5 lesson): prone-init epis

Randomly initialize each env as face-down (belly) or face-up (back), with random yaw. Face-down: +90° pitch → quat = [s*cy, -s*sy, s*cy, s*sy]

Porte de shaping dans [0,1] = enveloppe normalisée. Vaut 0 sur tout le segment de repos : les amorces (ciseau des jambes, différentiel des ro

Gaussienne sur l'erreur de vitesse de lacet (fonction pure, testable).

Fonction pure : tanh du différentiel de roues, portée par gate, clampée ≥ 0.

Smooth multiplicative goal-state score (product of three Gaussians). Returns ``height_score * upright_score * pose_score``, each ∈ [0, 1]. Be

Advance the world, taking the lock once for the whole batch. **Batched because the lock is the bottleneck, not the solver.** Four daemons at

Toggle body pose control mode on/off.

Toggle head control mode on/off.

Toggle sitting on/off (Y key). Old one-way sit policy (--sit): Y off switches back to the standing/ walking session, which does the s

Toggle slope policy mode on/off (passive descent, zero twist command).