Evanxw0715/BiCAT-SO101-Dataset

BiCAT SO-101 Dataset

Teleoperated bimanual SO-101 robot tasks in LeRobot v3.0 format. Each task folder is a complete dataset with its own meta/, data/, and videos/ directories. Episode indices start at 0 within each task.

Folder	Task instruction	Episodes	Frames	FPS	Size
block/	Put the block into the box.	100	54,412	30	0.55 GB

Total dataset size: approximately 0.55 GB, counting files under block/. GB uses decimal units (1 GB = 1,000,000,000 bytes). Total recording time is 30.2 minutes; episodes are 14–23 seconds long.

.
├── README.md
└── block/
    ├── data/
    ├── meta/
    └── videos/

Robot and recording setup

All episodes were recorded with two SO-101 follower arms (LeRobot robot type bi_so_follower), teleoperated with two SO-101 leader arms (bi_so_leader). Each arm has five joints and a gripper, driven by Feetech STS3215 servos.

Camera	Feature	Device
Top	observation.images.top	Intel RealSense D435, RGB stream only
Left wrist	observation.images.left_wrist	USB camera mounted on the left follower's wrist
Right wrist	observation.images.right_wrist	USB camera mounted on the right follower's wrist

Left and right are from the robot's point of view. Before and after every episode, all four arms were moved slowly to the same fixed home pose. Every episode therefore starts from the same arm configuration.

Features

Feature	Shape	Description
observation.images.top, observation.images.left_wrist, observation.images.right_wrist	480 × 640 × 3	RGB video, 30 FPS, AV1
observation.state	12	Measured follower joint positions
action	12	Leader joint positions, sent to the followers as position targets

observation.state and action share one joint order: left arm first, then right arm. Each arm lists shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper, named for example left_shoulder_pan.pos.

Joint values are in degrees, except the gripper. A joint's 0° is the middle of its calibrated range of motion, not a particular pose. The gripper is on a 0–100 scale, where 0 is fully closed and 100 is fully open.

Episode sources

The 100 block/ episodes come from three recording sessions of the same task:

Session	Episodes recorded	Episodes used	New episode indices
session_001	51	51	0–50
session_002	35	35	51–85
session_003	15	14	86–99

The sessions were merged into one dataset, and episode indices were renumbered from 0 to 99. The last episode of session_003 was left out to give exactly 100 episodes; it was not a failed demonstration. Any demonstration judged bad during recording was discarded and re-recorded then, so it never entered a session. session_002 stopped when a wrist camera disconnected. The episode in progress was not saved, and all completed episodes were kept.

Every episode was checked before merging, and the merged dataset was checked again:

• Timestamps are exactly 1/30 s apart, and each episode's row count matches its recorded length.
• There are no missing joint values, and both arms move in every episode.
• Each episode has the expected number of frames in all three camera videos.
• No camera video has black or frozen frames.
• Every merged episode's actions match its source episode exactly.

All episodes use the instruction "Put the block into the box."

Load a task locally

The LeRobot loader expects a dataset root containing meta/info.json. Point it at one task folder after downloading the repository:

from pathlib import Path
from huggingface_hub import snapshot_download
from lerobot.datasets.lerobot_dataset import LeRobotDataset

snapshot = Path(snapshot_download(
    repo_id="Evanxw0715/BiCAT_SO101",
    repo_type="dataset",
))
block = LeRobotDataset("local/block", root=snapshot / "block")

The videos are grouped into files. Episode video locations and timestamps are recorded in meta/episodes/. Recorded with LeRobot 0.6.1.