Project-IRA/TPSoSe2026_Dataset_Desk_Cleanup_LeRobot_SO101

SO-101 Desk Cleanup — 350 episodes

Teleoperated SO-101 demonstrations of clearing objects from a desk into a bowl. The most open-ended task in Project-IRA: a variable number of arbitrary objects in arbitrary positions.

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

Episodes	350 (250 in the merged training set)
Distinct task prompts	35
LeRobot codebase version	v3.0
Robot type	so_follower (SO-101, 6-DOF)
Control frequency	30 fps
Language	English

Composition

350 episodes across 35 prompt phrasings, 10 episodes each.

Only 250 of these 350 episodes are in the merged training dataset. Ten early sessions (100 episodes, prompts listed under "Excluded from final training" below) used a different coloured bowl and a messier scene, and were dropped for consistency.

Block	Setup	In merge?
V2/V3 early sessions	Different coloured bowl, messier scene	No
One-or-multiple objects	7–8 episodes with a single object, last 2–3 with several	Yes
Two objects	Episodes 1–6 two objects, 7–8 one, 9–10 several	Yes
Mixed counts	Episodes 1–3 one object, 4–7 two objects, 8–10 three to five	Yes
Single-item variation	One specific item (marker / keys / lego brick / ball / dice) into a bowl whose position changes every 5 episodes	Yes

Objects used include markers, keys, lego bricks, a ball, and dice, cleared into a green bowl beside the arm.

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 in English, 10 episodes each.

Included in the merged training dataset (25 prompts, 250 episodes)

1. Put all objects from the desk into the bowl
2. Clear the desk by placing all items into the green bowl
3. Move everything on the desk into the bowl next to the arm
4. Pick up all objects on the desk and drop them into the bowl
5. Tidy up the desk by putting all items into the bowl to the right
6. Place all items from the desk into the green bowl on the right side
7. Remove all objects from the desk and put them in the bowl
8. Clear the workspace by moving all objects into the nearby bowl
9. Pick up each object on the desk and place it into the green bowl
10. Put everything from the desk into the bowl directly to the right of the arm
11. Sweep the desk clean by placing all items into the bowl
12. Move all objects on the desk into the green bowl beside the robot
13. Collect all items from the desk and deposit them into the bowl to the right
14. Take each object off the desk and drop it into the green bowl next to the arm
15. Clear the desk by picking up each object and placing it into the bowl nearby
16. Grab every item on the desk and toss it into the green bowl
17. The desk needs to be cleared, move all objects into the bowl
18. Sort all desk items into the bowl sitting beside the robot arm
19. Make the desk empty by relocating all items into the bowl next to you
20. Transfer the objects from the desk surface into the green bowl nearby
21. Take the marker and put it in the bowl
22. Pick up the keys and put them in the bowl
23. Grab a lego brick and drop it into the bowl
24. Place the ball into the bowl to clear the desk.
25. Clean the desk by putting the dice into the bowl

Excluded from final training (10 prompts, 100 episodes)

These are present in this dataset but not in Dataset_Full_Merged_Final_V1.

1. Clean up the desk
2. Get the desk cleared off and everything into the bowl
3. Round up all the objects on the desk and put them in the bowl
4. Empty the desk completely by putting items in the green bowl
5. Pick everything up off the desk and drop it in the bowl
6. Clean the desk surface by moving items into the bowl next to it
7. Put each item into the bowl until the desk is cleared
8. Gather all the desk objects and place them in the bowl beside the arm
9. Drop every object from the desk into the green bowl
10. Place all desk items in the bowl so the desk is empty

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("Project-IRA/TPSoSe2026_Dataset_Desk_Cleanup_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_Desk_Cleanup_LeRobot_SO101 \
    --batch_size=64 --steps=200000 \
    --policy.device=cuda

Models trained on this dataset

Model	Quality
SmolVLA V3 Desk Cleanup	Does not work

Single-task SmolVLA fine-tuning failed on this task. Desk cleanup only became reliable via multi-task Pi0.5 training on the merged dataset — see Pi05 V7 Full V2.

This task is a strong instance of the multimodal action distribution problem: with several objects on the desk, any of them is a valid next target, and a policy that averages across those modes reaches for the empty space between them. Flow-matching action heads (used by both SmolVLA and Pi0.5) are chosen partly to address this.

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}}
}