aloha_incontext

A Mobile ALOHA robot manipulation dataset for in-context imitation learning. It contains human-teleoperated demonstrations of pick-and-place, pen uncapping, placing eggs in a box and closing it, and additional bimanual tasks.

1,328 episodes / 31 task configurations / 587,000 frames, recorded at 50 Hz. The task configurations are divided into 25 seen configurations (1,318 episodes) and 6 unseen configurations (10 episodes).

Observations and actions

• Robot: Mobile ALOHA with two arms and 14 joint/gripper dimensions.
• RGB cameras: cam_high, cam_left_wrist, and cam_right_wrist.
• State: observation.state, a 14-dimensional vector of joint and gripper positions.
• Additional observations: observation.velocity and observation.effort.
• Action: a 16-dimensional vector containing 14 joint/gripper targets and 2 base-velocity dimensions.
• Format: LeRobot v2.0, with one Parquet file per episode and camera images embedded in the Parquet data.

Tasks

Task family	Seen configurations	Seen episodes	Unseen configurations	Unseen episodes
Pick and place into a basket	14	527	4	6
Uncap a pen	6	338	1	1
Put an egg in a box and close it	1	199	1	3
Additional bimanual tasks	4	254	0	0
Total	25	1,318	6	10

The additional bimanual tasks are handover (104 episodes), cup_stack (50), stir (50), and water_wipe (50).

Pen uncapping

In every pen instruction, left/right identifies the hand that picks up the pen. The other hand grasps and removes the cap.

Task ID	Pen	Pickup hand	Episodes	Configuration
18	Gray pen	Right	143	Seen
19	Gray pen	Left	51	Seen
20	Red pen	Right	69	Seen
21	Blue pen	Left	25	Seen
22	Second blue pen	Right	25	Seen
23	Second blue pen	Left	25	Seen
24	Red pen	Left	1	Unseen

For example: "Pick up the gray pen with the right hand, grasp the cap with the other hand and uncap it."

For pick-and-place tasks, the hand named in the instruction picks up the object and places it in the basket.

Seen and unseen configurations

All episodes are packaged in a single Hugging Face train split. The seen/unseen division is defined by task configuration, not by separate Hugging Face splits.

The unseen configurations are:

Task ID	Task instruction	Episodes
14	Pick up the pear and place it in the basket with the left hand.	1
15	Pick up the orange juice and place it in the basket with the left hand.	1
16	Pick up the kiwi and place it in the basket with the right hand.	2
17	Pick up the banana and place it in the basket with the right hand.	2
24	Pick up the red pen with the left hand, grasp the cap with the other hand and uncap it.	1
26	Pick up the red egg with the right hand, place it in the box, and close the box.	3

For evaluation on unseen configurations, exclude these six configurations from training. This leaves 1,318 training episodes. Unseen demonstrations can be used as in-context examples at evaluation time.

File organization

• data/chunk-*/episode_*.parquet: observations, actions, timestamps, and frame, episode, and task indices.
• meta/info.json: dataset schema and summary statistics.
• meta/tasks.jsonl: task IDs and task instructions.
• meta/episodes.jsonl: episode IDs, task assignments, and episode lengths.
• meta/stats.json: normalization statistics. Numeric-feature statistics cover all 1,328 episodes, including unseen configurations. Image normalization uses fixed per-channel statistics.

License

Apache-2.0.