Ahaskar04/cola-2arm-handover

CoLA two-arm handover

Two ALOHA arms. Arm A hands a red box to arm B. The same demonstrations train both the wrist-camera-only and the wrist-plus-overhead variants in the paper.

• 213 episodes, all successful scripted demonstrations (failures discarded).
• Episode length 93-132 steps (mean 100.1).
• The paper uses 150 of them (120 / 15 / 15), listed in splits.json; the remaining 63 are extra successful demonstrations not used in the paper.

Format

One HDF5 file per episode (episode_NNNN.h5), recorded at 50 Hz control (MuJoCo, timestep 2 ms x 10 substeps), plus a rendered review video per episode (episode_NNNN_review.mp4). Datasets inside each .h5, T = episode length:

key	shape	dtype	meaning
image_wrist_a	(T, 256, 256, 3)	uint8	arm A wrist camera
image_wrist_b	(T, 256, 256, 3)	uint8	arm B wrist camera
image_overhead	(T, 256, 256, 3)	uint8	fixed overhead camera
state_a	(T, 7)	float64	arm A joint state
action_a	(T, 7)	float64	arm A action
state_b	(T, 7)	float64	arm B joint state
action_b	(T, 7)	float64	arm B action
box_pos	(T, 3)	float64	box position (m)
phase	(T,)	bytes	scripted-expert phase label

state_* is the arm's 6 joint positions plus finger position; action_* is the 6 absolute joint-position targets plus the gripper command. Images are 256x256 RGB. File attributes: success, handover_only, policy_module, scan_found.

Every episode starts with arm A already holding the box and was produced by a scripted expert; only successful episodes are included.

Splits

splits.json gives the train/val/test split used in the paper (80/10/10, seed 42, stratified by success). Each entry is a filename in this repo.