Datasets:
This dataset was created using LeRobot.
Dataset Structure
{
"codebase_version": "v2.1",
"robot_type": "tidybot_xarm7",
"total_episodes": 54,
"total_frames": 28748,
"total_tasks": 2,
"total_videos": 108,
"total_chunks": 1,
"chunks_size": 1000,
"fps": 30,
"splits": {
"train": "0:54"
},
"data_path": "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet",
"video_path": "videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4",
"features": {
"observation.images.overhead": {
"dtype": "video",
"shape": [
720,
1280,
3
],
"names": [
"height",
"width",
"channel"
],
"info": {
"video.height": 720,
"video.width": 1280,
"video.codec": "mjpeg",
"video.pix_fmt": "yuvj420p",
"video.is_depth_map": false,
"video.fps": 30,
"video.channels": 3,
"has_audio": false
}
},
"observation.images.wrist": {
"dtype": "video",
"shape": [
720,
1280,
3
],
"names": [
"height",
"width",
"channel"
],
"info": {
"video.height": 720,
"video.width": 1280,
"video.codec": "mjpeg",
"video.pix_fmt": "yuvj420p",
"video.is_depth_map": false,
"video.fps": 30,
"video.channels": 3,
"has_audio": false
}
},
"observation.state": {
"dtype": "float32",
"shape": [
8
],
"names": [
"joint1",
"joint2",
"joint3",
"joint4",
"joint5",
"joint6",
"joint7",
"gripper"
]
},
"observation.ee_pose": {
"dtype": "float32",
"shape": [
6
],
"names": [
"x",
"y",
"z",
"roll",
"pitch",
"yaw"
]
},
"observation.joint_torques": {
"dtype": "float32",
"shape": [
7
],
"names": [
"joint1",
"joint2",
"joint3",
"joint4",
"joint5",
"joint6",
"joint7"
]
},
"observation.base_pose": {
"dtype": "float32",
"shape": [
3
],
"names": [
"base_x",
"base_y",
"base_theta"
]
},
"observation.base_velocity": {
"dtype": "float32",
"shape": [
3
],
"names": [
"base_vx",
"base_vy",
"base_wz"
]
},
"action": {
"dtype": "float32",
"shape": [
11
],
"names": [
"joint1",
"joint2",
"joint3",
"joint4",
"joint5",
"joint6",
"joint7",
"gripper",
"base_vx",
"base_vy",
"base_wz"
]
},
"action.ee_pose": {
"dtype": "float32",
"shape": [
6
],
"names": [
"x",
"y",
"z",
"roll",
"pitch",
"yaw"
]
},
"action.arm_velocity": {
"dtype": "float32",
"shape": [
6
],
"names": [
"vx",
"vy",
"vz",
"wx",
"wy",
"wz"
]
},
"action.pose_target": {
"dtype": "float32",
"shape": [
6
],
"names": [
"x",
"y",
"z",
"roll",
"pitch",
"yaw"
]
},
"action.gripper_position": {
"dtype": "float32",
"shape": [
1
],
"names": [
"gripper_position"
]
},
"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:
[More Information Needed]
TidyBot top drawer with base motion (2026-10-23)
Teleoperated demos on a TidyBot (xArm7 + UFactory G2 gripper on a holonomic base), opening and closing the top kitchen drawer by driving the base. The arm reaches the handle and the jaws close on it. The arm then holds still while the base drives straight back about 0.45 m (open) or forward (close). The jaws then let go. Arm and base never move at the same time.
54 successful episodes: 27 Γ open the top drawer, 27 Γ close the top drawer, recorded
19:41β20:25. Failed attempts (operator-marked), takes with no base motion and one base test are
not included. Converted with remact/hw/tidybot_to_lerobot.py (converter v6). Per-episode
provenance (source session id, outcome, trim range, gripper taps, base availability) is in
meta/tidybot_episodes.jsonl.
How open and close were told apart: the episodes are not in task order. Each one was labelled by the direction the base drove (net odometry β0.32 to β0.50 m = open, +0.36 to +0.50 m = close). The label was then checked on the first frame of every episode: every open starts with the drawer shut, and every close starts with it pulled out.
| Key | Shape | Meaning |
|---|---|---|
observation.images.overhead |
720Γ1280Γ3 | overhead camera (MJPEG frames as captured) |
observation.images.wrist |
720Γ1280Γ3 | wrist camera |
observation.state |
8 | joint1β7 (rad), measured gripper position (0 open β¦ 1 closed) |
action |
11 | joint1β7 at the next frame (rad, absolute), commanded gripper position, base vx, vy (m/s), wz (rad/s) |
observation.ee_pose, action.ee_pose |
6 | TCP x, y, z (m), roll, pitch, yaw (rad), now / next frame |
action.pose_target |
6 | the teleop's commanded TCP pose (compliant pose control) |
action.gripper_position |
1 | measured gripper position at the next frame |
observation.joint_torques |
7 | controller-reported joint torques |
observation.base_pose, observation.base_velocity |
3 | wheel odometry (x, y m; heading rad) and measured base velocity |
action.arm_velocity |
6 | Cartesian velocity teleop command (zero here: these demos used pose control) |
Base
action[8:11] is the operator's base velocity command, [vx, vy, wz], in the robot's own
frame (x forward, y left, wz counter-clockwise). It is the command the base controller received,
held until the next one (~17 Hz). In these demos it is almost purely vx, with peaks of
0.08β0.15 m/s. vy stays below 0.03 m/s and wz is zero, so a policy trained on this data
will not learn to strafe or turn. The measured base velocity follows the command ~0.1β0.17 s
later (correlation 0.99). Base telemetry is present in every frame of every episode.
Gripper
Same as DarinDjapri/tidybot_drawer. Gripper positions use openpi's convention,
1 β opening / 84 mm: 0.0 = fully open, 1.0 = fully closed. A tap commands close (1.5 mm at
full force, action[7] = 0.982) or open (75 mm, action[7] = 0.107). action[7] switches
at the tap. The measured position in observation.state[7] follows ~0.1β0.2 s later. On the
handle the jaws stop at ~6 mm (0.93), not at the commanded 1.5 mm. At deployment, run action[7]
through hysteresis, e.g. close above 0.65 and open below 0.45. The hub logs gripper steps without
their direction, so each tap's direction is read from where the jaws end up afterwards.
Checks: the jaws follow all 108 command changes (half-way after a median 1.2 s). The headset's
own "Gripper: closing/opening" messages agree on 50 of 54 episodes. In 3 of the others, the
operator tapped close then open within 1 s and the jaws never moved, so the action follows the
jaws. In 20261023T201816-0ac9ddb5 (open), the jaws started half-open (50 mm), and a quick
open-then-close was merged into one tap. Its close command starts ~1 s early, during the reach.
Notes
- The layout follows openpi's 7-DoF joint-position recipe (
pi05_full_droid_finetune): absolute joint positions in the dataset, turned into deltas during training withmake_bool_mask(7, -4). Joints become deltas, and gripper and base stay absolute. The action key isaction, so setaction_sequence_keys=("action",)in the data config. Recorded at ~30 Hz. - It has the same 11-dim action and the same keys as
DarinDjapri/tidybot_drawer(arm-only drawer demos), so the two can be trained together. Note the prompts differ (open the drawerthere,open the top drawerhere). - The demos used the controller's compliant pose control, with F/T admittance while grasping.
- Images are timestamped on arrival at the hub, so the robot in an image is ~0.12β0.21 s older than the state paired with it. At deployment, pair the latest hub image with the current state in the same way.
- Downloads last month
- 52