Datasets:

filesmuggler
/

experiment-cube-cup-100-l_diagonal-full

license: apache-2.0
task_categories:
  - robotics
pretty_name: Cube-to-cup pick-and-place (L-shape + diagonal)
tags:
  - LeRobot
  - so100
  - so101
  - pick-and-place
  - cube
  - cup
  - imitation-learning
  - compositional-generalization
configs:
  - config_name: default
    data_files: data/*/*.parquet

This dataset was created using LeRobot.

Cube → cup pick-and-place (SO-ARM, L-shape + diagonal)

Teleoperated demonstrations of picking a small dark cube and placing it into an orange cup on a white table. Recorded on an SO-ARM follower (robot_type: so_follower) with LeRobot v3.

Task: Grab the cube and put it in the cup

Episode 0, top camera. Cube in the back row, cup in front-center (the most common cup cell).

Collection strategy

Start-pose sampling follows the L in-domain collection idea from Gao, Xie, Xiao, Finn, and Sadigh, Efficient Data Collection for Robotic Manipulation via Compositional Generalization (RSS 2024; arXiv:2403.05110; Hugging Face paper).

That work treats environmental axes (object pose, container, texture, …) as a factor grid. Instead of recording every combination, an L covers each factor along one axis while holding the other at a base value, and relies on the policy to compose the missing pairs. This dataset applies that idea to two factors on a 3×3 table:

cube (target) cell
cup (goal) cell

Concretely:

L, cube arm — vary cube cell while the cup stays near a base place pose (here goal cell 8, 33 episodes).
L, cup arm — vary cup cell while the cube stays near a base pick pose (here target cell 5, 25 episodes).
Diagonal / extra pairs — additional cube→cup combinations off the L, so the dump is denser than a pure L (l_diagonal-full in the repo name).

A full 9×9 grid would be 81 pairs. This run has 35 / 81 observed pairs: the L plus a diagonal, not uniform coverage. That is intentional, not an accident.

Why this dataset

Single skill, repeated with varied start poses on a 3×3 table grid, collected to test composition rather than to fill every pair.
Two RGB views: overhead observation.images.top and observation.images.wrist (640×480, AV1, 30 FPS).
Joint-space action and observation.state (6-DoF + gripper, degrees).

Trained policies: smolvla_cube_cup, act_cube_cup, pi05_cube_cup.

Snapshot


Episodes	99 (0–98). Filename says 100; one episode was not recorded.
Frames	29,432 (~16.4 minutes at 30 FPS)
Episode length	min 6.5 s · median 9.7 s · max 19.5 s (episode 62)
Cameras	top, wrist
License	Apache-2.0

Spatial coverage (first-frame review)

Every episode has a visible cube (target) and cup (goal) on frame 0. Counts below are not uniform, and they match the L: mass on T5 and G8.

Cube (target) cell counts (1 = back-left, 9 = front-right):

	left	mid	right
back	3	7	8
mid	13	25	14
front	6	12	11

Cup (goal) cell counts:

	left	mid	right
back	12	8	8
mid	8	7	5
front	8	33	10

Unique cube→cup pairs: 35 / 81
Most common pair: cube cell 8 → cup cell 2 (7 episodes)
32 cube locations were boxed by hand (dark cube vs white table); cups were automatic
Same-cell starts (cube and cup share a cell): episodes 46, 47, 80, 81, 82, 87

Bar chart of target (green) and goal (blue) first-frame counts. Orange dots mark the annotated robot.

Intended use and limits

Use for: imitation learning / VLA fine-tuning on a short-horizon pick-and-place skill, and for studying whether a policy can compose unseen cube/cup cell pairs from L + diagonal data (the setting Gao et al. describe).

Limits:

One table, one lighting setup, one cube, one cup.
Pair coverage has large holes (especially picks from cell 1 and places away from cell 8). That is by design for an L, not a claim of full combinatorial coverage.
This is an independent SO-ARM dataset inspired by their collection strategy; it is not the WidowX / kitchen data from the paper, and we do not use BridgeData V2 here.
First-frame overlays are a review aid, not ground-truth object poses. A few automatic cube locks are ambiguous when the cube is near the gripper.
Joint values are in the LeRobot degree convention; absolute table XYZ from nominal FK is approximate.

Dataset Structure

See meta/info.json for the full LeRobot v3 schema (codebase_version: v3.0).

Citation

@misc{filesmuggler_cube_cup_l_diagonal_full,
  title  = {Cube-to-cup pick-and-place (L-shape + diagonal)},
  author = {filesmuggler},
  year   = {2026},
  url    = {https://huggingface.co/datasets/filesmuggler/experiment-cube-cup-100-l_diagonal-full},
  note   = {Collection layout follows the L strategy of Gao et al., RSS 2024}
}

@inproceedings{Gao-RSS-24,
  author    = {Jensen Gao and Annie Xie and Ted Xiao and Chelsea Finn and Dorsa Sadigh},
  title     = {Efficient Data Collection for Robotic Manipulation via Compositional Generalization},
  booktitle = {Proceedings of Robotics: Science and Systems},
  year      = {2024},
  address   = {Delft, Netherlands},
  month     = {July},
  doi       = {10.15607/RSS.2024.XX.013},
  url       = {https://iliad.stanford.edu/robot-data-comp/}
}