Configuration for a modality.

Transforms states and actions to tensors.

Class for state or action transform. Args: apply_to (list[str]): The keys in the modality to load and transform. normalizati

Vision Transformer

Wrapper for OmegaConf to track accessed parameters. Only saves configuration items that were actually accessed during execution.

Compose multiple modality transforms.

Configuration for Agilex dual-arm robot with rich state/action information

Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).

LAM model with a shared vision encoder, latent action bottleneck, and state decoder.

Build LatentWorldPolicyConfig from global training config with strict field validation.

Mixture dataset that reuses starVLA's LeRobot loaders but emits LAM-ready raw samples. Output sample: { "frames": torch.

A mixture of multiple datasets. This class samples a single dataset based on the dataset weights and then calls the `__getitem__` method of the s

Type of rotation representation

Class for state or action sin-cos transform. Args: apply_to (list[str]): The keys in the modality to load and transform.

Implements the Policy interface by communicating with a server over websocket. See WebsocketPolicyServer for a corresponding server implementatio

Serves a policy using the websocket protocol. See websocket_client_policy.py for a client implementation. Currently only implements the `load` an

Simple linear bottleneck used when vq_type='ae'. - Keeps interface compatible with VQ/VAE modules. - Now operates directly in code_dim sp

Pre-norm Transformer block with Adaptive Layer Normalization (AdaLN) conditioning. Each block receives a conditioning vector c ~ [B, D] (act

Configuration for AgiBOT humanoid robot with dual arms

Concatenate modality keys according to specified order. Note: State/action padding and mask generation are intentionally not handled

Maps continuous time in seconds to sinusoidal embeddings.

Metadata of the trainable dataset Changes: - Update to use the new RawCommitHashMetadataMetadata_V1_2

Configuration for the real dual-arm fold-towel teleoperation dataset.

Independent LatentWorldPolicyBackend config.

Policy-side adapter for Qwen-VL prompt/processor encoding.

Base dataset class for LeRobot that supports sharding.

Configuration for LIBERO dataset with Franka robot