buzinguyen/robotwin-cups-nesting

RoboTwin Cup-Nesting (Bimanual, Dual-Aloha)

500 successful expert demonstrations of bimanual cup nesting — a dual-arm robot nests three cups into one another — collected in the RoboTwin 2.0 simulator with strong domain randomization. Format: LeRobot v2.1.

What's in it

Episodes	500 (all success-filtered)
Frames	287,571 @ 30 fps
Embodiment	dual-arm Aloha (aloha-agilex), 14-DoF
Action / state	14-dim absolute joint positions
Cameras	head + left-wrist + right-wrist RGB, 240×320, h264
Domain randomization	background, cluttered table, lighting, table height

Instructions (a note on honesty)

Two instruction styles are present, both on cup-nesting episodes:

• 50 color-conditioned episodes: "stack the other cups into the {green|blue|red} cup" (17 / 17 / 16).
• 450 generic episodes: "stack the three cups by nesting them into one another".

The 450 generic episodes were originally captioned with a stale "stack the three bowls…" instruction inherited when this task was adapted from RoboTwin's built-in stack_bowls_three. The objects are cups in every episode (verified visually); only the captions were generic. They have been relabeled to the generic cup instruction above. If you need clean color-conditioning signal, use the 50 color-conditioned episodes; the other 450 are color-agnostic nesting demos.

Collection

Trajectories were generated by RoboTwin's motion-planning expert (curobo + mplib), success-gated (only episodes passing the nesting success check are kept), on a custom stack_cups_three task (our adaptation of RoboTwin's stack_bowls_three).

Known consumer gotcha: is_success

is_success is declared with shape (1,), so LeRobot stores each cell as list<bool> ([True]/[False]). In Python, bool([False]) is True — a non-empty list is truthy. Unwrap the scalar before use: bool(cell[0]), not bool(cell).

Loading

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("buzinguyen/robotwin-cups-nesting")

Attribution & Licensing

• This dataset (trajectories + renders) is released under the MIT License.
• Generated with RoboTwin 2.0; 3D assets (021_cup) and the aloha-agilex embodiment are from RoboTwin's repositories (MIT). The bowls task template is RoboTwin's stack_bowls_three; the cup task and all trajectories are ours.
• Full third-party notice: see THIRD_PARTY_LICENSES.md.

If you use this dataset, please cite RoboTwin 2.0:

@article{chen2025robotwin,
  title={RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain
         Randomization for Robust Bimanual Robotic Manipulation},
  author={Chen, Tianxing and others},
  journal={arXiv preprint arXiv:2506.18088},
  year={2025}
}