Datasets:
pretty_name: Bimanual xArm Packing (Real World)
task_categories:
- robotics
tags:
- LeRobot
- bimanual
- xarm
- real-world
- packing
Bimanual xArm Packing (Real World)
Teleoperated demonstrations of a bimanual packing task on two real xArm robots. The arms pack
household objects (for example a sponge, a Cheez-It container and an Altoids container) into
small, medium and large boxes. Each episode is one subtask, such as "Pack Sponge into medium
box.", and the task field holds that instruction. Instruction wording varies across episodes
(for example with and without "the"), as recorded during collection.
Layout
Each round folder is a complete LeRobot dataset (codebase version 2.1).
| Folder | Episodes | Frames |
|---|---|---|
baseline/round1 |
164 | 97,522 |
baseline/round2 |
282 | 167,612 |
baseline/round3 |
283 | 147,111 |
baseline/round4 |
281 | 157,816 |
baseline/round5 |
280 | 154,808 |
baseline/round6 |
282 | 151,189 |
beta-data/round1 |
566 | 331,819 |
beta-data/round2 |
653 | 351,865 |
beta-data/round3 |
626 | 344,594 |
baseline holds standard demonstrations. beta-data holds demonstrations collected while a
person perturbs the scene during the task, so episodes include recovery from disturbed states,
and that person is often visible in the camera views.
Features
All rounds share one schema, which matches the simulation dataset
kshitiz25/sim_packing_task, and are
recorded at 60 Hz.
| Key | Shape | Description |
|---|---|---|
observation.state |
float32 (20,) | Per arm, left then right: end-effector position (x, y, z), rotation as the first two columns of the rotation matrix (r00, r10, r20, r01, r11, r21), gripper |
action |
float32 (20,) | Target pose in the same 20-d layout as the state |
observation.images.base |
video, 3x224x224 | Overhead camera |
observation.images.left_wrist |
video, 3x224x224 | Left wrist camera |
observation.images.right_wrist |
video, 3x224x224 | Right wrist camera |
task_index |
int64 | Index into meta/tasks.jsonl |
Videos are AV1-encoded MP4 files; decoding them needs an FFmpeg build with an AV1 decoder
(for example libdav1d). meta/high_level_prompts.json stores the goal-level prompt for each
episode.
Usage
Download one round and read its frames:
from huggingface_hub import snapshot_download
import pandas as pd
root = snapshot_download(
"kshitiz25/real_world_packing_task",
repo_type="dataset",
allow_patterns="baseline/round1/*",
)
ds_root = f"{root}/baseline/round1"
frames = pd.read_parquet(f"{ds_root}/data/chunk-000/episode_000000.parquet")
Point a LeRobot release that reads codebase version 2.1 at ds_root to load a round as a
LeRobotDataset.