Get frames from a video at specified timestamps. Args: video_path (str): Path to the video file. timestamps (list[int] | np.ndarr

Get the metadata for a key in the LeRobot modality metadata. Args: key (str): The key to get the metadata for. Returns:

Returns: first model parameter's device

Get the parquet path for a trajectory.

Get the RAW data for a single step in a trajectory. No transforms are applied. Args: trajectory_id (int): The name of the traject

Print trainable parameters of the diffusion model.

Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with the multistep UniPC.

Implement blockwise causal attention using flash_attention. Matches the pattern from _prepare_blockwise_causal_attn_mask:

Calculate relative action statistics for each delta index (horizon step) separately. Unlike `_calculate_relative_stats_for_key` which

Lazily move the VAE to the correct device/dtype on first use.

repeat vars like kernel and stride to match dim

Extract specific frames using ffmpeg.

Get the interpolation mode for the given backend. Args: interpolation (str): The interpolation mode. backend

Load trajectory data from parquet file. Args: traj_id: The trajectory ID to load. Returns:

Build an observation dict from real video frames. For 1 frame: each image key is (H, W, 3). For 4 frames: each image key is (4, H, W, 3).

Internal method to reset policy state. Args: save_video: Whether to save accumulated video before reset.

"Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the prediction of x_0 at t

Worker loop for non-rank-0 processes to participate in distributed inference.

This function only adds the frame to the episode_buffer. Apart from images — which are written in a temporary directory — nothing is

Cache the next shard in a background thread.

Computes overall statistics from per-task statistics using dataset sample weights. Args: per_task_stats: List of per-tas

Create tasks.jsonl file with unique task names and action texts.

Detect whether the dataset follows old or new format structure

Faster map_structure implementation which skips some error checking.

imgs: [B, 3, H, W] of torch.float32. - mean: [0.48145466, 0.4578275, 0.40821073] - std: [0.26862954, 0.2613025

Get DREAM action data for a trajectory by step indices. Args: trajectory_id (int): The ID of the trajectory. key (str

Get the chunk index for an episode index.

Load initial actions from the dataset if available. Returns: list: List containing initial actions if the file exists, empty list

Get the language annotation data for a trajectory by step indices. Args: trajectory_id (int): The ID of the trajectory.

Get LAPA action data for a trajectory by step indices. Args: trajectory_id (int): The ID of the trajectory. key (str)

Get the reward data for a trajectory by step indices. If the step indices are out of range, pad with first/last step data. Args:

Get the dimension of a state or action key from the dataset metadata.

This function is used to get the dims of the state/action keys after transform is applied. It is different from the `get_state_action

Get task language instruction

Try this to automate the name: https://gist.github.com/techtonik/2151727#gistcomment-2333747

Check if this HDF5 group is a subset of another HDF5 group.

Inject LoRA adapters after pretrained weights have been loaded. This should be called when defer_lora_injection=True.

log.txt -> log.ep50.txt