angkul07/so101-tray-sort

SO-101 tray sort (33 teleop episodes)

Real-robot teleoperation demonstrations on a single SO-101 arm (leader → follower), recorded with LeRobot on 2026-10-02.

Task (identical language instruction on every episode):

Put the vaseline, the tape and the marker into their compartments in the tray, then drag the tray to the marked area.

Every episode follows the same order: vaseline → tape → marker → drag the tray to the taped zone → arm back to rest.

At a glance

Robot	SO-101 follower (robot_type: so_follower), teleoperated with an SO-101 leader
Episodes	33
Frames	39,092 at 30 fps (21.7 min)
Episode length	30.3 – 48.0 s (mean 40.0 s)
Camera	1 wrist camera, observation.images.front, 640×480, AV1
Format	LeRobot dataset codebase_version: v3.0

Features

key	shape	notes
observation.state	6	follower joint positions
action	6	leader joint positions (the command sent to the follower)
observation.images.front	480×640×3 video (H×W×C)	wrist camera, despite the key name

Joint order: shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper. The five arm joints are in degrees, and the gripper is in 0–100 (percent of its calibrated range).

Scene layouts

The vaseline and tape start spots swap between two layouts. The marker, tray and drag zone never move.

episodes	layout
0–9	A (vaseline near)
10–19	B (tape near)
20–29	A
30–32	B

The layout is not stored as a feature, so use the table above. Layout-B episodes are about 6 s longer on average (43.2 ± 3.7 s vs 37.1 ± 3.6 s), which is consistent across all of them.

Quality checks (run before upload)

• The video frame count equals the parquet row count, frame indices are contiguous, and there are no NaNs.
• Every episode starts and ends at the rest pose. The spread of the start pose across episodes is about 3° on most joints (5° on wrist roll).
• The follower trails the leader by about 4 frames (133 ms), with a mean tracking error of 0.9°.
• The largest single-frame step is 8°. Only one pause longer than 1.5 s occurs mid-task across all 33 episodes.

Known caveats

• Idle frames are not trimmed. On average there is 2.0 s of no motion at the start of each episode (max 4.7 s) and 1.8 s at the end (max 3.1 s), about 10% of all frames. Trim them before training if they matter to you.
• The wrist hits its stop. wrist_flex sits at its mechanical stop (about 100°) for 22% of task frames. The leader reads the same value, so this is the true posture, not clipping.
• The camera is mounted sideways. The wrist image is rotated about 90°, and auto-exposure washes out a few end frames (episodes 21, 22, 30).
• Episode 0 may be a practice run. Its start scene differs from the other layout-A episodes. Consider dropping it.
• Gripper gap while holding. While an object is held, action (leader) closes further than observation.state (follower) can.

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("angkul07/so101-tray-sort")
print(ds.num_episodes, ds.num_frames)   # 33 39092
frame = ds[0]                            # dict with action, observation.state, observation.images.front, ...

Visualize in the browser with the LeRobot dataset visualizer.

Collection setup

• The leader drives the follower at 30 fps. Each episode was started and stopped by hand from the keyboard, and failed attempts (hesitation, a drop, a re-grasp, a tipped tray) were discarded and redone.
• The camera, lighting and table layout were fixed for the whole session, with nothing else on the table.