Perform a per-image random crop over a batch of images in a vectorized way. (Same as shown previously.)

Builds a 'Read' status packet (also works for 'Sync Read') https://emanual.robotis.com/docs/en/dxl/protocol2/#read-0x02 https

This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch, num_key_value_heads, seqlen, he

Reset the tracker, clearing all recorded steps. Does nothing if tracker is None.

Resolve frame-level `collector_policy_id` from control mode and source.

PEFT training needs pretrained models, this finds the pretrained model of a policy type for PEFT training.

Rotates half the hidden dims of the input.

Helper function to save a tensor as an image file.

Set random seed for all RNG sources to ensure reproducibility.

Set random seed for all RNG sources to ensure reproducibility.

Optionally let robots skip observation-side work during pure teleoperation send loops.

Processes data step-by-step, yielding the transition at each stage. This is a generator method useful for debugging and inspecting the interm

Stop the controller and release resources.

Stop the robot client

Returns the current metric values (or averages if `use_avg=True`) as a dict.

Unflatten a dictionary with delimited keys into a nested dictionary. Example: >>> flat_dct = {"a/b": 1, "a/c/d": 2, "e": 3} >>> p

Unregisters a 'before_step' hook. Args: fn: The exact function object that was previously registered. Raises:

Write a value to a single motor's register. Contrary to :pymeth:`sync_write`, this expects a response status packet emitted by the motor, whi

Predict state value (V). Args: z: (*, latent_dim) tensor for the current state's latent representation. Returns:

Serialize and send a command message to the robot.

Initialize the controller. Args: x_step_size: Base movement step size in meters y_step_size: Base movement s

Finds the same timestep actions in the queue and aggregates them using the aggregate_fn

Helper method to apply gradient checkpointing if enabled.

Explicitly destroy a policy and free all associated memory. This method performs aggressive cleanup to ensure maximum memory is freed,

Helper to encode control parameters into 8 bytes for MIT mode.

Write all buffered episode metadata to parquet file.

Get the target dtype based on config.

Initialize a single policy instance with specified RTC configuration. Args: name: Name identifier for logging purposes

Helper to extract just the keys logged to rr.log

Create a TimedObservation with raw robot observation format.

Pad 7 → 20 dims (zeros for the dummy channels).

Pad real_dim → max_dim (zeros for the dummy channels).

Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape `(batch_size, key_value_l

Helper method to prepare 4D attention masks for transformer.

Helper method to prepare 4D attention masks for transformer.

Return the most recent (single) values from self.last_d, optionally applying calibration.

Return the most recent (single) values from self.last_d, optionally applying calibration.

Flags when the client is ready to send an observation

Efficiently receive responses from multiple motors at once. Uses the OpenArms pattern: collect all available messages within timeout.

Receive a response from a motor. Args: expected_recv_id: If provided, only return messages from this CAN ID

Raise migration error when we detect JSON files but no processor configs. This method is called when migration detection determines that a mo

Convert unsigned integer from CAN to float.

Write DataFrame to parquet This ensures images are properly embedded and the file can be loaded correctly by HF datasets.

Aggregates stats for a single feature.

Applies Rotary Position Embedding with Multimodal Sections to the query and key tensors (https://qwenlm.github.io/blog/qwen2-vl/). Explanation:

Convert rotation to rotation vector. Returns: Rotation vector [x, y, z] where magnitude is angle in radians

Test that the content of both datasets is preserved correctly in the aggregated dataset.

Test that total number of episodes and frames are correctly aggregated.

Test that episode indices are correctly updated after aggregation.

Test that metadata is correctly aggregated.

Test that all video timestamps are within valid bounds for their respective video files. This catches bugs where timestamps point to frames beyon

Reads the latest available frame asynchronously. This method retrieves the most recent frame captured by the background read

Calculate rewards for given text, video, and state representations. This is the canonical method for SARM reward computation, used f

Return a group name for logging. Optionally returns group name as list.

Check if CAN interface is UP and configured.

Check for NaN values in transition data. Args: observations: Dictionary of observation tensors actions: Action tensor

Clears all processor steps from the registry.

Convert image-to-video dataset. Creates a new LeRobotDataset with images encoded as videos, following the proper LeRobot dataset structure wi