ShuaKang/robotwin_4tasks_2k_demos

RoboTwin 2.0 — four non-pick-and-place tasks, full domain randomisation

Simulated single-arm (right arm) manipulation demonstrations collected in RoboTwin 2.0 and exported to LeRobot v2.1, 1000 episodes / 175,112 frames over four tasks that are not pick-and-place: pressing, in-hand reorientation, an articulated object, and shaking. Rendering, camera, randomisation and export format are identical to the pre_full variant of the companion coverage dataset ShuaKang/my_roboTwin2.0_training, so the two can be mixed directly.

Collection code: https://github.com/Shua-Kang/my_roboTwin2.0

Tasks

folder	skill	episodes	frames	manipulated object
click_bell/lerobot_delta_hr	press	250	19,182	050_bell (2 ids)
rotate_qrcode/lerobot_delta_hr	in-hand reorientation	250	38,667	070_paymentsign (4 ids)
open_laptop/lerobot_delta_hr	articulated object	250	54,747	015_laptop (11 ids)
shake_bottle/lerobot_delta_hr	grasp + shake	250	62,516	001_bottle (20 ids)

• click_bell (press): approach the bell, push its top button down 4.5 cm, lift off (no grasp). Object: 050_bell (2 ids).
• rotate_qrcode (in-hand reorientation): grasp the QR sign, lift 7 cm, rotate it so the QR code faces the robot, put it down. Object: 070_paymentsign (4 ids).
• open_laptop (articulated object): push the laptop lid open to >= 50 % of its joint range (up to 15 re-grasps of the lid edge). Object: 015_laptop (11 ids).
• shake_bottle (grasp + shake): grasp the bottle, lift 10 cm, three +/-7pi/8 wrist twists while moving up/down 5 cm. Object: 001_bottle (20 ids).

Every episode is executed by the right arm (the scene generator spawns the manipulated object on the right half of the table, or for open_laptop samples the lid orientation until the right arm is chosen). Task success is verified by the task's own check_success at the end of each recorded trajectory.

Domain randomisation (all knobs on)

knob	setting
background / table textures	pool of 100 (same fixed subset as pre_full)
distractor objects	1–2 per episode from a 10-class pool
object placement	full task range
lighting	random
manipulated-object instance	random among the task's model ids (see table)

Camera setup (D4)

observation.images.head: RGB 320×240 at (0.10, −0.65, 1.50) m, pitched 52° down, vertical FOV 43°, so the whole right arm is in frame from the first step. Every robot link except the right arm is hidden (render only; physics unchanged). observation.images.right_wrist: RGB 320×240 D435 on the right wrist. The left-wrist and third-view streams exist in the raw HDF5 but are not exported here.

Format

LeRobot v2.1, 30 fps, one data/chunk-000/episode_XXXXXX.parquet + two mp4 per episode.

action — 14-D end-effector delta

[L: dpos(3) drotvec(3) grip(1) | R: dpos(3) drotvec(3) grip(1)], dpos = commanded_pos[t] − state_pos[t], drot = rotvec(R_cmd[t] · R_state[t]ᵀ) (world frame), gripper absolute in [0, 1]. Only the right-arm block moves; the left block is ~0 with a constant gripper.

observation.state — 20-D absolute

[L: pos(3) rot6d(6) grip(1) | R: pos(3) rot6d(6) grip(1)], rot6d = first two columns of the rotation matrix.

Language

task is one of 100 generated "seen" phrasings per episode (a further 100 "unseen" phrasings are in the raw collection). Placeholders name the object instance and the arm, e.g. "Press the center of the bell using the right arm."

Gotchas

• Integrate deltas on the previously commanded pose, not on the observed pose — the ~0.4 mm steady-state tracking error otherwise accumulates linearly and the gripper drifts by several cm over an episode.
• open_laptop episodes are long (150–230 frames) and contain repeated lid pushes; click_bell episodes are short (~75 frames).

Loading

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("ShuaKang/robotwin_4tasks_2k_demos", root=None, episodes=None)   # or point root at one task folder

dp_zarr_ws/: workspace-covering re-collection of the four skills (added 2026-10-05)

Diffusion-Policy zarrs of the four skills re-collected with all objects sampled in the common right-arm workspace x [0.05, 0.31] m, y [-0.20, 0.10] m (RoboTwin workspace_override), 250 episodes each, dataset v2 conventions (D4 head camera + right wrist, body/left arm hidden, right arm only, pre_full scene randomisation): ws_click_bell.zarr (19,839 frames), ws_open_laptop.zarr (55,927), ws_rotate_qrcode.zarr (40,265), ws_shake_bottle.zarr (62,019). Together with 1000 place_a2b_right episodes they form the 2000-episode pool pre_ws_2k in https://huggingface.co/datasets/ShuaKang/my_roboTwin2.0_training (dp_zarr_ws/), which also holds the retrieved 250-episode subsets. Format: data/{head_camera,right_camera,state,action}, meta/episode_ends.