Return the demo ``.h5`` path, downloading the dataset if it is absent.

Import LIBERO's benchmark module (raises if LIBERO is not installed).

Cache the active→sorted joint mapping the handler's ``set_dof_targets`` expects.

Run one closed-loop episode; return (success, steps_taken). If ``video_path`` is set, capture a world-up observer camera every action step and

Minimal stand-in with the fields get_ee_state_from_list reads.

Split a done env's cache into (completed_episode, next_episode_seed). ``reset_frame`` is the post-auto-reset state read on the done step; it is t

Get the shape of the first frame in the video.

Actor ids kept as foreground for the greenscreen overlay (robot + manipulated objects).

Iterate over frames in a video file.

Continuous random-noise rough terrain (mjlab-like) via scenario.ground. Same continuous-noise approach as go1 (zero-slope ``SlopeCfg`` + ``random

Generator: find episodes and parse them.

Generate default reset states

World-frame position of YAM end-effector via tcp_site.

Get positions of all five finger tips. Returns: finger_tips_pos: (B, 5, 3) tensor with positions of [thumb, index, middle, ring,

Return a list of per-env initial state dicts. Default builds a minimal robot-only template from ``scenario.robots[0]`` with its ``def

Return list of per-env initial states derived from config.

Get initial states for all environments.

Get masses for IsaacGym simulator. Note that the isaacgym handler only support 1 robot and multiple objects, currently.

Get masses for IsaacGym simulator. Note that the isaacgym handler only support 1 robot and multiple objects, currently.

Get masses for MuJoCo simulator.

Get masses for MuJoCo simulator.

:param batch_inds: :return:

Get the set shape indices.

Return the MuJoCo joint names sorted by id. Used for scene-MJCF assets that aren't registered as RobotCfg entities.

Continuous random-noise rough terrain (mjlab-like) via scenario.ground. A zero-slope ``SlopeCfg`` with ``random=True`` triggers ``random_unif

Discover and import every ``roboverse_pack.robots.*`` module so its subclasses register in ``RobotCfg.__subclasses__``.

Eagerly import task modules so @register_task runs, avoiding a full package crawl. We only import modules under `locomotion/` to register tasks w

Compute action dim. Subclasses may set ``self.num_actions`` themselves first.

Initialize LiDAR point cloud backend for Isaac Sim using Isaac Lab's LidarSensor. This attaches a LiDAR sensor (Livox/MID-360 pattern by defa

Initialize Warp-based LiDAR backend for IsaacGym / MuJoCo. Builds a static triangle mesh that includes: - Ground/terrain (from handle

Parse default joint positions and torque limits from cfg.

Set position limits, default joint positions, and default joint velocities.

Set the demo init state if dimensions match; else use the env reset state.

Method for subclasses to inherit to load in the policy

Initialize SmolVLA model.

Prepares a list of reward functions, which will be called to compute the total reward.

Whether ``self._initial_states`` carries any robot or object state. Scene-MJCF tasks (e.g. mjlab cartpole loaded as the scene rather

Force RoboTwin's object creation to reuse the bridge's recorded model_ids. RoboTwin tasks pick each object's ``model_id`` with ``np.random.choice

Install SimplerEnv lighting on the handler scene (default lights were skipped).

Record the exact mesh file RoboTwin loads for each named actor. ``create_actor`` resolves an object's mesh via the module-internal ``get_glb_

Smoothly interpolates between the current state and final predicted action of the chunk