data_juicer.ops.mapper.video_hand_action_compute_mapper module#

class data_juicer.ops.mapper.video_hand_action_compute_mapper.VideoHandActionComputeMapper(*args, **kwargs)[source]#

Bases: Mapper

Compute 7-DoF actions and 8-dim states from hand reconstruction and camera pose results.

Reads hand MANO parameters (from VideoHandReconstructionHaworMapper) and camera-to-world transforms (from VideoCameraPoseMegaSaMMapper), then produces per-frame state [x,y,z,roll,pitch,yaw,pad,gripper] and per-frame action [dx,dy,dz,droll,dpitch,dyaw,gripper] compatible with LIBERO / StarVLA LeRobot format.

__init__(hand_reconstruction_field: str = 'hand_reconstruction_hawor_tags', camera_pose_field: str = 'video_camera_pose_tags', tag_field_name: str = 'hand_action_tags', hand_type: str = 'both', *args, **kwargs)[source]#

Initialization method.

Parameters:

  • hand_reconstruction_field – Meta field storing HaWoR hand reconstruction results.

  • camera_pose_field – Meta field storing camera pose (cam_c2w) results.

  • tag_field_name – Output field name in Fields.meta.

  • hand_type – Which hand to compute actions for. ‘right’, ‘left’, or ‘both’. Default is ‘both’.

process_single(sample=None, rank=None)[source]#

For sample level, sample –> sample

Parameters:

sample – sample to process

Returns:

processed sample