mohsinarf/so_arm101_pick_and_place_tray

Green-cube-pick

A LeRobot imitation-learning dataset of an SO-101 (arm 101) robot arm picking up a green cube and placing it in a tray.

The dataset contains 85 human-teleoperated demonstrations recorded at 30 fps with two RGB cameras (a fixed external view and a wrist/arm-mounted view). It is stored in the LeRobot v3.0 format and can be loaded directly with the lerobot library to train visuomotor policies (ACT, Diffusion Policy, π0, etc.).

Dataset at a glance

Property	Value
Task	green cube pick (pick green cube → place in tray)
Robot	SO-101 (arm 101), 6-DOF
Episodes	85 (all in train split)
Total frames	22,308
Frame rate	30 fps
Episode length	201–327 frames (~6.7–10.9 s, mean ~8.7 s)
Cameras	2 × RGB, 640×480
Codebase version	LeRobot v3.0

Observations & actions

Both action (commanded joint positions) and observation.state (measured joint positions) are 6-D vectors over the same joints:

Index	Joint	Unit
0	shoulder_pan.pos	degrees
1	shoulder_lift.pos	degrees
2	elbow_flex.pos	degrees
3	wrist_flex.pos	degrees
4	wrist_roll.pos	degrees
5	gripper.pos	0–50 (open→closed)

Camera streams:

Key	Resolution	Codec	Description
observation.images.fixed camera	640×480×3	AV1 (yuv420p)	External fixed view
observation.images.arm camera	640×480×3	AV1 (yuv420p)	Wrist/arm-mounted view

Repository structure

.
├── data/
│   └── chunk-000/
│       └── file-000.parquet … file-005.parquet   # per-frame state + action (episodes packed per file)
├── videos/
│   ├── observation.images.fixed camera/chunk-000/file-000.mp4 …
│   └── observation.images.arm camera/chunk-000/file-000.mp4 …
└── meta/
    ├── info.json        # dataset schema, features, fps, paths
    ├── stats.json       # dataset-wide normalization statistics
    ├── tasks.parquet    # task_index → task string
    └── episodes/        # per-episode metadata + statistics

Each per-frame Parquet row contains: action, observation.state, timestamp, frame_index, episode_index, index, task_index.

Usage

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("mohsinarf/so_arm101_pick_and_place_tray")

print(dataset)
print(dataset[0])  # first frame: images, observation.state, action, ...

How to upload to the Hugging Face Hub

From the folder containing this README:

pip install -U huggingface_hub
huggingface-cli login

# Upload the full folder (creates the repo if it does not exist)
hf upload mohsinarf/so_arm101_pick_and_place_tray . --repo-type=dataset