twanghcmut/aloha_placing_drawer_kitchen_lerobot

ALOHA right-arm multitask — drawer + kitchen

Two single-task LeRobot v2.1 datasets merged into one multitask set, so a single policy can be trained on both behaviours with the language instruction selecting between them.

This is a derivative work. All 300 episodes come unchanged from @locht131:

source	episodes	frames	becomes
locht131/aloha_placing_drawer_lerobot	150	37,797	episodes 0–149, task_index 0
locht131/aloha_placing_kitchen_lerobot	150	42,060	episodes 150–299, task_index 1

No frames were added, dropped, resampled or re-encoded. The videos are the original files. Cite the sources, not this repo, for the data itself. The sources declare no license; the Apache-2.0 tag above covers only the merge scripting and metadata, and if @locht131 wants this taken down or relicensed, that governs.

Contents

Format	LeRobot v2.1
Robot	ALOHA, right arm only
Episodes / frames	300 / 79,857
fps	15
Tasks	2
Cameras	observation.images.color.high (overhead), observation.images.color.wrist_right — 480×640
Episode length	179–383 frames (11.9–25.5 s)

Tasks:

task_index	instruction
0	Use the right gripper to open the top drawer, pick up the red square block, place it inside the top drawer, and then close the drawer.
1	Use the right gripper to pick up the banana and place it into the pot. Then pick up the lid with the right gripper and place it on top of the pot to close it.

Features are unchanged from the sources and identical between them: action (13), observation.state (13), observation.velocity (7), observation.effort (7), plus the two video streams.

What the merge actually changed

Renaming episode files is not enough — three columns inside each parquet are global and had to be rewritten, or the dataset loads with silently wrong episode boundaries:

• episode_index — kitchen shifted by +150
• index — the global frame counter, kitchen shifted by +37,797, so the merged set runs contiguously 0…79,856
• task_index — kitchen set to 1

Metadata rebuilt accordingly: episodes.jsonl (300), episodes_stats.jsonl (300, per-episode stats carried over with shifted indices), tasks.jsonl (2), info.json totals.

Verified after merging: 300 parquet and 300 mp4 per camera; episode_index matches every filename; index strictly contiguous 0…79,856; frame_index restarts at 0 in every episode; 150 episodes per task. Zero discrepancies.

GR00T metadata

meta/modality.json, meta/stats.json and meta/relative_stats.json are included for Isaac GR00T finetuning. The modality map is the right-arm one:

{"state":   {"single_arm": {"start": 0, "end": 6}, "gripper": {"start": 6, "end": 7}},
 "action":  {"single_arm": {"start": 0, "end": 6}, "gripper": {"start": 6, "end": 7}},
 "video":   {"front": {"original_key": "observation.images.color.high"},
             "wrist": {"original_key": "observation.images.color.wrist_right"}},
 "annotation": {"human.task_description": {"original_key": "task_index"}}}

relative_stats.json (shape 16 × 6) matters if you keep single_arm as RELATIVE: that is the only representation use_relative_action acts on and the only one that reads this file, and the loader silently skips it when absent — so a missing file does not error, it just normalizes with the wrong statistics.

Sanity check on the merged statistics: the merged action mean lies between the two sources on every dimension, weighted toward kitchen as its larger frame count implies (dim 5: drawer 0.233, kitchen −0.188, merged 0.011).

Deployment note

15 fps. A client interpolating between model steps must derive its sub-step count from that ratio:

smooth_step = control_hz / 15

Reusing a value from a 25 or 30 fps robot stretches every trajectory by the ratio and the arm creeps through the task without finishing it.