Project-IRA/TPSoSe2026_Dataset_Dice_Throw_LeRobot_SO101

SO-101 Dice Throw — 120 episodes

Teleoperated SO-101 demonstrations of picking up a dice cup and tipping the dice onto the table. A dynamic, non-quasi-static task: the arm must lift, rotate and release rather than simply pick and place.

Part of Project-IRA — Interactive Robotic Arm. Code: https://github.com/Project-IRA/interactive-robotic-arm

Episodes	120
Distinct task prompts	12
LeRobot codebase version	v3.0
Robot type	so_follower (SO-101, 6-DOF)
Control frequency	30 fps
Language	English

Composition

120 episodes across 12 prompt phrasings, 10 episodes each. All recorded as "Clean + Variation" blocks.

Recording scheme within each block:

• The cup is picked up from two different positions on the table under a single prompt.
• The dice are always thrown toward the middle / left side of the workspace.
• Between 1 and 6 dice are in the cup, varying across episodes.
• The arm rotates to the left to throw, then sets the cup back down close to the robot.
• Later blocks rotate slightly further after the dice leave the cup.

The task is framed as a physical true-random number generator — the robot rolls the dice, the result is read off the table.

Robot setup

Robot	SO-101 follower arm (6-DOF), robot_type: so_follower
Teleoperation	SO-101 leader arm
Control frequency	30 fps
State / action space	6-dim: shoulder_pan.pos, shoulder_lift.pos, elbow_flex.pos, wrist_flex.pos, wrist_roll.pos, gripper.pos
Camera observation.images.desk_view	800x600, h264 (recording)
Camera observation.images.wrist_left	640x480, h264 (recording)

Inference note: both cameras are run at 640x480 during inference, not at their recording resolutions, to reduce the payload sent to the inference server.

Schema

Feature	dtype	shape
observation.state	float32	(6,) — shoulder_pan.pos, shoulder_lift.pos, elbow_flex.pos, wrist_flex.pos, wrist_roll.pos, gripper.pos
observation.images.desk_view	video	(600, 800, 3), h264, 30 fps
observation.images.wrist_left	video	(480, 640, 3), h264, 30 fps
action	float32	(6,) — same joint layout as state
timestamp, frame_index, episode_index, index, task_index	—	bookkeeping

Recording protocol

• Recorded by teleoperating the SO-101 follower with an SO-101 leader arm.
• 10 episodes per prompt. The prompt phrasing was deliberately changed roughly every 10 episodes, so language conditioning sees many surface forms of the same intent.
• Object positions and scene difficulty were varied systematically within each block (e.g. early episodes with a single object, later ones with several).
• Recovery behaviour is incidental. Where the operator made a mistake mid-episode and corrected it, that correction stayed in the data. No recovery episodes were scripted deliberately, so recovery coverage is uneven.

Prompts

All 12 prompts, in English, 10 episodes each.

1. Pick up the cup and throw the dice onto the table
2. Roll the dice using the cup
3. Grasp the cup and lift it and tip it over to release the dice
4. Empty the dice cup onto the flat surface
5. Reach for the dice container and pick it up and toss the contents across the table
6. Take the cup with the dice inside and flip it to pour them out
7. Throw the dice on the table
8. Generate true random numbers by throwing the cup with dice
9. Pick up the cup and roll the dice into the center of the workspace
10. Grab the dice cup and cast the dice onto the open table area
11. Grab that cup and let the dice fly onto the desk
12. Perform a dice roll by picking up and turning the cup over

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("Project-IRA/TPSoSe2026_Dataset_Dice_Throw_LeRobot_SO101")
print(ds.meta.info)

Train a policy on it:

lerobot-train \
    --policy.path=lerobot/smolvla_base \
    --dataset.repo_id=Project-IRA/TPSoSe2026_Dataset_Dice_Throw_LeRobot_SO101 \
    --batch_size=64 --steps=200000 \
    --policy.device=cuda

Models trained on this dataset

No single-task model was trained on this dataset. Dice throwing is covered by the multi-task models trained on Dataset_Full_Merged_Final_V1 — best results from Pi05 V7 Full V2.

Limitations

• Single environment. One desk, one lighting setup, one camera geometry, one set of physical objects. Policies trained here should not be expected to transfer.
• Teleoperated demonstrations vary in quality and speed between operators and sessions.
• Recovery coverage is uneven — see the recording protocol note above.
• No held-out split. The dataset ships as a single train split; all episodes were used for training. Evaluation was done by running policies on the physical arm.

Licensing

Released under CC BY-SA 4.0. This is a share-alike licence: you may use, share and adapt this dataset, including commercially, provided you give attribution and release any derivative dataset under the same licence. It cannot be taken closed-source.

Recorded with LeRobot (Apache-2.0); the LeRobot dataset format and tooling remain under their original licence.

Citation

@misc{project_ira_2026,
  title        = {Project-IRA: Interactive Robotic Arm},
  author       = {Baten, Cleo and Keppler, Bela and Sapper, Jonas},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/Project-IRA}},
  note         = {Code: \url{https://github.com/Project-IRA/interactive-robotic-arm}}
}