The env var above is only useful if submit() actually sets it.

End to end without the Hub or a GPU: an ONNX in, the three repo files out.

Without this, [tool.uv.sources] for torch is a silent no-op.

The fix changes only the wheel's SOURCE, not its version: the CUDA index carries newer builds than the PyPI pin, so a `>=` drags torch 2.9.1 -

Catches the vanished-script failure directly: `train` must be ours or mjlab's, never whatever else is installed on the machine.

Removing it uninstalls bin/train and recreates nothing (see docstring).

mjlab résout et MUTE les SceneEntityCfg en place : un objet partagé entre plusieurs termes provoque des indices périmés. Chaque terme doit avoir l

The interception rides on this entry point; without it, no flag.

La pièce nouvelle : roues FREINÉES → LIBRES. Les roues roulent, donc il n'y a aucune adhérence longitudinale pour pousser sur le sol. On boot

HF Jobs run on x86_64: their resolution must not move.

Penalty on downward trunk velocity beyond ``max_down_vel``. Caps descent SPEED, which ``trunk_vertical_accel_penalty`` alone cannot: a fast c

Penalty on upward trunk velocity beyond ``max_up_vel``. Mirror of ``trunk_downward_velocity_penalty`` for the rise: charges every step of a t

Penalty proportional to ``|a_z|`` of the trunk (finite-diff of v_z). Captures hard impacts (large deceleration spike on landing) AND incentivises

``body_upright_gaussian`` weighted by smoothstep on trunk z. Full Gaussian-upright reward when ``z >= height_high``, zero when ``z <= height_

Potential-based upright shaping: Δcos(tilt) per step. Pays for PROGRESS toward upright, charges for progress toward fallen, and pays exactly

Linear upright reward weighted by a smoothstep on trunk z. Returns ``body_upright_linear * smoothstep((z - low)/(high - low))`` so the uprigh

Refuse what the daemon would refuse, plus the mistakes it would load and run wrongly. Accepts both shapes and any schema version, because absence

Update velocity command ranges based on training progress. Gradually increases the commanded velocity ranges to allow the robot to learn high

Update velocity tracking std parameter based on training progress. Starts with loose std (easy rewards) to learn basic walking, then gradually

Update wheel friction based on training step stages.

Reward wheel spin proportional to commanded push. All 4 wheels spin positive for forward motion (verified visually). tanh saturation at vel_s

Constant-zero obs term of width `dim`. Used by envs that don't actively track head/body commands (e.g. sitstand, ground_pick) but still need