SO101_PickAndPlace_Fruit

LeRobot dataset of an SO-101 follower arm performing a "pick banana and place" manipulation task, teleoperated and recorded via lerobot. Two RGB cameras (front + wrist) at 30 fps.

Dataset at a glance

Task	pick banana and place (single task)
Robot	so101_follower (6-DoF arm + gripper)
Episodes	102
Total frames	78,111 (~43.4 min @ 30 fps)
FPS	30
Episode length	min 536f (17.9 s) · median 724f (24.2 s) · mean 766f (25.5 s) · max 1791f (59.7 s)
Cameras	observation.images.front, observation.images.wrist — H.264, 640×480, yuv420p, 30 fps
LeRobot format	codebase_version v2.1 (per-episode files)

Format (LeRobot v2.1, per-episode)

SO101_PickAndPlace_Fruit/
├── data/chunk-000/episode_{000000..000101}.parquet     # one parquet per episode
├── videos/chunk-000/
│   ├── observation.images.front/episode_{000000..000101}.mp4
│   └── observation.images.wrist/episode_{000000..000101}.mp4
└── meta/
    ├── info.json            # schema, fps, totals, feature/video specs
    ├── episodes.jsonl       # per-episode {episode_index, tasks, length}
    ├── episodes_stats.jsonl # per-episode feature statistics
    └── tasks.jsonl          # {task_index: 0, task: "pick banana and place"}

Per-frame features (parquet columns)

Feature	dtype / shape	Description
action	float32 [6]	Target joint positions: shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper (.pos)
observation.state	float32 [6]	Measured joint positions (same 6 joints)
observation.images.front	video (480,640,3)	Front RGB camera frame (referenced from mp4)
observation.images.wrist	video (480,640,3)	Wrist RGB camera frame
timestamp	float32	Seconds from episode start
frame_index	int64	Frame index within the episode (0..length-1)
episode_index	int64	Episode id (0..101)
index	int64	Global frame index across the dataset (0..78110)
task_index	int64	Task id (always 0)

Loading

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("jinseonylee/SO101_PickAndPlace_Fruit")
print(ds.num_episodes, ds.num_frames)   # 102 78111
sample = ds[0]                           # dict with action / observation.state / images / ...

Data characteristics & quality notes

• Uniform encoding: every video is H.264 640×480@30 fps, ~8.3 KB/frame (front) / ~8.4 KB/frame (wrist) — consistent across all 102 episodes, no frozen/blank/oversized frames.
• Frame sync verified: for every episode, parquet rows == length == front mp4 frames == wrist mp4 frames; the global index is contiguous 0..78110 with no gaps or duplicates.
• Episode duration ranges ~18–60 s (60 s is the recording auto-timeout); short episodes are valid full demonstrations, not truncations.
• Assembly: merged from 5 collection sessions (10 + 10 + 29 + 50 + 3 episodes) into a single v2.1 dataset with episodes renumbered 0..101 and the global index made contiguous. All sessions share the identical schema (robot, fps, cameras, codec, resolution) and the same task.

Collection setup

• Robot: SO-101 leader→follower teleoperation (so101_leader / so101_follower).
• Cameras: two USB RGB cameras (front + wrist), 640×480 @ 30 fps.
• Pipeline: recorded on a Jetson edge device via AWS IoT Greengrass data-collection components, uploaded to S3, then exported to this LeRobot v2.1 dataset.
• Collection tooling: this data was collected using the open-source sample-lerobot-data-collection-on-aws-iot-greengrass project (AWS IoT Greengrass components for LeRobot data collection).

License

mit — see the collection tooling repository above for component licensing.