README.md

metadata

license: apache-2.0
task_categories:
  - robotics
tags:
  - LeRobot
  - mobile_aloha
  - imitation-learning
configs:
  - config_name: default
    data_files: data/*/*.parquet

aloha_incontext

A paper-consistent reorganization of vo2yager/aloha_data_unique, whose task list matches the ContextFlow paper in both training and testing.

1,349 episodes / 31 task configurations / 595,900 frames — 25 seen configurations (1,318 episodes) and 6 unseen configurations (31 episodes).

Mobile ALOHA, bimanual 14-DoF, 3 RGB cameras (cam_high, cam_left_wrist, cam_right_wrist) at 480×640, 50 fps. action is 16-D (14 joint targets + 2 base velocity dimensions); observation.state is 14-D.

Relationship to aloha_data_unique

This is a regrouping, not new data. Every episode is copied from the source release with its sensor data untouched — the image columns are byte-identical, verified by round-trip comparison. Only the three bookkeeping columns (episode_index, task_index, and the global row counter index) are rewritten, because episodes are renumbered and tasks are merged.

Three changes relative to the source's 49 tasks:

1. Merges — scene batches of one configuration become one task: pen_uncap_gray_left_b5 + _b9 → 143 episodes; pen_uncap_red_left_b9 + _b5 → 69; handover_b9 + handover_b5 → 104.
2. Drops — 15 tasks / 138 episodes absent from the paper in both training and testing: the four separate_cups tasks (100), pen_uncap_blue_left_b5 (25), and 10 single-demo pick-and-place tasks (13).
3. Renaming — task strings are natural-language instructions rather than folder names.

Hand-naming convention (important)

Pen-uncap task names here follow the paper's convention, where <left>/<right> names the hand that picks up the pen. The source release's folder names use the mirror convention, naming the hand that uncaps. So the hand label is flipped relative to aloha_data_unique:

This dataset	Source task	Episodes
...gray pen with the right hand...	pen_uncap_gray_left_b5 + _b9	143
...gray pen with the left hand...	pen_uncap_gray_right_b5	51
...red pen with the right hand...	pen_uncap_red_left_b9 + _b5	69
...blue pen with the left hand...	pen_uncap_blue_right_b5	25
...second blue pen with the right hand...	pen_uncap_blue2_left_b5	25
...second blue pen with the left hand...	pen_uncap_blue2_right_b5	25
...red pen with the left hand... (unseen)	pen_uncap_red_right_b5	22

Pick-and-place hand labels are not flipped — the two conventions agree there.

The four extra bimanual configurations (handover, cup_stack, stir, water_wipe) keep short folder-style names, as no instruction template for them appears in the paper.

Splits

The dataset ships as a single train split; the seen/unseen division is by task name. The 6 unseen configurations are:

Pick up the pear and place it in the basket with the left hand.
Pick up the orange juice and place it in the basket with the left hand.
Pick up the kiwi and place it in the basket with the right hand.
Pick up the banana and place it in the basket with the right hand.
Pick up the red pen with the left hand, grasp the cap with the other hand and uncap it.
Pick up the red egg with the right hand, place it in the box, and close the box.

Excluding those leaves exactly the 1,318 training episodes the paper reports.

Notes

• meta/stats.json numeric features are recomputed over these 1,349 episodes. The per-channel image statistics are carried over from the source release unchanged.
• Models trained on aloha_data_unique are not directly comparable: the training distribution and normalization statistics both differ.