Move all tensor values to device (HF BatchEncoding compat).

Try to resolve *raw_key* to a special-token string. Looks up ``raw_key`` in the registered ``tokens`` dict first, then tries ``<raw_k

Create a LeRobot Dataset from scratch in order to record data.

Additionally require non-empty bbox JSON; warning_lite batches may contain '{}'.

Map a global sample index back to its dataset directory and local index (O(log N) bisect).

Emit a structured box with logger.info, prefixed by a blank visual spacer. Args: rows: terminal display rows, truncated by inner_width.

Clear without returning, for exception paths or manual reset.

Move nn.Module attributes to device and call `.to(...)` on action_tokenizer if possible.

Register all resolvers commonly used across entrypoints. Safe to call multiple times.

Idempotently register a resolver, replacing any existing one.

Returns False if user pressed 'q' (server may ignore).

Print a single tree node: dim connector + bold key + description.

Eval path: preserve non-empty bbox JSON check. Empty JSON '{}' is meaningless for inference. BBoxSubtaskCoTBuilder (eval_required_fie

proprio tensor/ndarray/dict → "idx idx idx ..." string.

Return the current time in seconds.

Bytes → human-readable string (KB / MB / GB).

Build processor(s) from Hydra config. Mixture case (cfg.data.processors exists): For each embodiment_type, merge per-type config with cfg

Process instructions and update observation data. Args: obs: observation data dictionary, mutated in place.

Drop-in replacement for accelerate.logging.get_logger: - No implicit barriers. - Only the main process emits log records by default.

Main loop: gather obs → server → single action → publish → repeat. Uses a fixed-period timer so that the control frequency stays stable

Minimal obs dict that InstructionManager expects.

Concatenate a list of identically-keyed dicts along axis=0. Example: [{k: (B1,T,D)}, {k: (B2,T,D)}] -> {k: (B1+B2, T, D)}

Keep only datasets/processors matching the specified embodiment_type.

Find the run directory containing .hydra/config.yaml by searching upward from ckpt_path.

Load config from a run directory's .hydra/config.yaml and patch Hydra-specific interpolations.

Convert rotations given as Euler angles in radians to rotation matrices. Args: euler_angles: Euler angles in radians as tensor of sh

Return micro-batch means for each field, then clear. Fields with no data are omitted from the returned dict. Callers should use check

Return the matching sub-processor from MixtureProcessor, or single directly. Supports both raw_obs (key="embodiment_type") and obs_dict (key="emb

Unwrap model from the DDP wrapper to access the underlying module.

Subclass override: return non-None to replace the command slot text. The command slot is the injected value for <command_text_!_200>.

Random absolute sample_idx for retry fallback. Subclasses with index filtering (e.g. DroidLerobotDataset's _valid_local_indices) over

Encode service-side errors in a stable payload.

Reverse decode, needed by processor_inout.py backward.

Convert rotations given as rotation matrices to Euler angles in radians. Args: matrix: Rotation matrices as tensor of shape (..., 3,

Convert rotations given as quaternions to rotation matrices. Args: quaternions: quaternions with real part first, as ten

Rotates half the hidden dims of the input.

Configure the logging system for the entire codebase with Rich formatting. **Single source of truth**: all Rich logging configuration is del

Close and cleanup the parquet writer if it exists.