Per-body xy-speed (linear-vel norm in xy).

Per-site xy-speed (norm of linear xy velocity) read from physics.data. Returns ``(num_envs, len(site_names))``. Uses ``mj_objectVelocity`` for

Gaussian sigmoid tolerance: 1 at x=0, ``value_at_margin=0.1`` at ``|x|=margin``.

Return an activation function given a string

Compute net contact forces on each body. Returns: torch.Tensor: shape (nbody, 3), contact forces for each body

World-frame position of injected cube body.

Get initial states for all environments.

Get initial states for all environments. Must be implemented by subclasses.

Get current observations for all environments

Get sensor ID and address for a given sensor name.

Waypoint index = current episode step (optionally with foresight).

Wrap a mesh in a minimal single-link URDF (sapien loads GLB/OBJ via assimp). The collision geometry points at RoboTwin's dedicated ``collision/``

Extract an RGB frame (HxWx3 uint8) from a SimplerEnv obs dict.

True if ``name`` transitively subclasses a contract root (Tier-1).

Native/passthrough adapters are an explicitly-exempt compatibility tier. Exemption is PATH-based only (dedicated ``_native``/``_passthrough`` dir

Draw a caption banner at the top of a frame (PIL if available, else plain).

Linear interpolation between two tensors.

Load metadata JSON file.

Resolve a list of name patterns against a sorted available-name list. Match strategy (in order): 1. exact match (``pat in available``)

A handler whose module looks like the raw-MuJoCo backend.

Replay a demo episode in the native env from its initial state. Returns ``(frames, success_step, track_positions)`` — ``success_step`` is the ind

``(N,3)`` world position of the robot end-effector body on Newton. If ``body_suffix`` is given, match the body whose name ends with it; other

Flatten robosuite obs to the robomimic-style low-dim vector.

Compute observation for every configured group.

Concatenate ObsTerm outputs within one group, applying noise.

Issue low-level actions and simulate one physics step.

Build a minimal env_states for PickCube where the cube sits at the TCP. Fingers occupy sorted columns 0-1; ``arm_tail_val`` is written to columns

Name of the most-moved task object over the demo (the thing the task acts on).

Angle between two unit quaternions (wxyz). Shape preserves leading dims.

Perform data randomization based on the given operation and distribution. Args: data: The data tensor to be randomized. Shape is (dim_0,

Step ``env`` through ``acts`` and record (obs, reward, done) per step.

Map a local asset abs-path to its stable two-namespace vendored relpath.

Replay `rollout.qpos/qvel` onto a fresh model from `xml_path` and return frames + nq.

Resample the command.

Per-env reset (set initial state, run reset events, zero buffers).

Reset selected envs (defaults to all).

Reset environments based on specified indices. Args: env_ids: List of environment ids which must be reset