meta/modality_config.py

# Auto-generated by ur5-gr00t — drop this file into
# Isaac-GR00T/gr00t/data/embodiment_configs/ and import it
# in gr00t/data/embodiment_tags.py (add UR5_LAB to the enum).

from gr00t.data.types import (
    ActionConfig, ActionFormat, ActionRepresentation,
    ActionType, ModalityConfig,
)

# Camera view keys must match keys in meta/modality.json
_CAM_KEYS = ['cam0', 'cam1']
_HORIZON  = 16

UR5_LAB_MODALITY_CONFIG = {
    "video": ModalityConfig(
        delta_indices=[0],
        modality_keys=_CAM_KEYS,
    ),
    "state": ModalityConfig(
        delta_indices=[0],
        modality_keys=["eef", "gripper"],
    ),
    "action": ModalityConfig(
        delta_indices=list(range(_HORIZON)),
        modality_keys=["eef", "gripper"],
        action_configs=[
            # EEF pose (xyz + axis-angle): stored absolute, applied absolute
            ActionConfig(
                rep=ActionRepresentation.ABSOLUTE,
                type=ActionType.EEF,
                format=ActionFormat.XYZ_ROTVEC,
            ),
            # Gripper: stored as [0,1] float, applied absolute
            ActionConfig(
                rep=ActionRepresentation.ABSOLUTE,
                type=ActionType.NON_EEF,
                format=ActionFormat.DEFAULT,
            ),
        ],
    ),
    "language": ModalityConfig(
        delta_indices=[0],
        modality_keys=["annotation.human.task_description"],
    ),
}