Datasets:
This dataset was created using LeRobot.
Dataset Description
Robot arm playing chess: teleoperated demos of an SO-101 arm picking chess pieces off a real board.
The current training set. 247 episodes, 35,982 frames, wrist camera only. Trained Dev-devadath/so101-chess-act-pick-wrist-v4.
Part of so101-chess: an SO-101 arm that plays chess on a real board. Everything except the grasp is inverse kinematics; a learned ACT policy does only the pick. Every episode starts with the arm hovering 8 cm above one piece, jaws open and pointing down, and ends with the piece lifted. Demos were teleoperated with an SO-101 leader arm.
- 30 fps; state and action are 6-D (5 joints + gripper, lerobot degrees,
so_follower) observation.images.wrist: gripper camera, 640x480observation.images.top(where present): overhead camera warped to a top-down board view, 480x480, a1 bottom-left- Gripper: red TPU V-groove pads glued to both jaws (except
so101-chess-pick-v0). The pads are in the wrist view, so a policy trained here expects them. - Board: 49 mm squares, ranks 1-5 reachable. White is on the arm's side.
- Recorded with lerobot at commit
5c43fa1c(0.5.2 dev). Load withvideo_backend="pyav".
It is all 271 episodes of Dev-devadath/so101-chess-pick-raw minus the 24 bad ones listed there, with the overhead
camera removed (remove_feature). The overhead camera was dropped because the first policy used it to
aim and broke when the far side of the board filled up.
The earlier wrist-only training sets are prefixes of this one: the first 122 episodes (first round
- d1/h1), 147 (+ the starting position) and 205 (+ full board ranks 1-2) trained the wrist v1, v2 and v3 policies.
Known issues
The demos teach a pose per square more than they teach aiming. A policy trained here returns to nearly the same joint pose on a square whatever the piece's exact position, so squares where the piece is often a few mm off (e2, d2, e4 in our games) are hit or miss.
Only ranks 1-5, and only the arm's own colour on ranks 1-2 in the full-board block.
b1's grasps are 10-16 mm sideways, as in the raw set.
Paper: none
License: apache-2.0
Dataset Structure
{
"codebase_version": "v3.0",
"robot_type": "so_follower",
"total_episodes": 247,
"total_frames": 35982,
"total_tasks": 1,
"chunks_size": 1000,
"data_files_size_in_mb": 100,
"video_files_size_in_mb": 200,
"fps": 30,
"splits": {
"train": "0:247"
},
"data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
"video_path": "videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4",
"features": {
"action": {
"dtype": "float32",
"names": [
"shoulder_pan.pos",
"shoulder_lift.pos",
"elbow_flex.pos",
"wrist_flex.pos",
"wrist_roll.pos",
"gripper.pos"
],
"shape": [
6
]
},
"observation.state": {
"dtype": "float32",
"names": [
"shoulder_pan.pos",
"shoulder_lift.pos",
"elbow_flex.pos",
"wrist_flex.pos",
"wrist_roll.pos",
"gripper.pos"
],
"shape": [
6
]
},
"observation.images.wrist": {
"dtype": "video",
"shape": [
480,
640,
3
],
"names": [
"height",
"width",
"channels"
],
"info": {
"video.height": 480,
"video.width": 640,
"video.codec": "h264",
"video.pix_fmt": "yuv420p",
"video.is_depth_map": false,
"video.fps": 30,
"video.channels": 3,
"has_audio": false
}
},
"timestamp": {
"dtype": "float32",
"shape": [
1
],
"names": null
},
"frame_index": {
"dtype": "int64",
"shape": [
1
],
"names": null
},
"episode_index": {
"dtype": "int64",
"shape": [
1
],
"names": null
},
"index": {
"dtype": "int64",
"shape": [
1
],
"names": null
},
"task_index": {
"dtype": "int64",
"shape": [
1
],
"names": null
}
}
}
Citation
BibTeX:
none
- Downloads last month
- 53