Project-IRA/TPSoSe2026_Dataset_Lego_LeRobot_SO101

SO-101 Lego Colour Sorting — 460 episodes

Teleoperated SO-101 demonstrations of sorting lego bricks onto colour-matched plates. The largest single-task dataset in Project-IRA.

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

Episodes	460
Distinct task prompts	46
LeRobot codebase version	v3.0
Robot type	so_follower (SO-101, 6-DOF)
Control frequency	30 fps
Language	English

Composition

460 episodes across 46 prompt phrasings, 10 episodes each. Recorded in blocks with progressively varied conditions:

Block	Setup
Multi-colour sorting	2–4 different colours sorted onto their matching plates
Single-colour (red / green / yellow / blue)	Early episodes one brick, last 2–3 several bricks of that colour
Single-colour + distractors	Target colour plus another colour present; only the target is sorted
Full multi-colour	Early episodes few bricks of several colours, last 2–3 all colours
Plate-position variation	Plate moved to a different position every 5–10 episodes; bricks from varying start positions
Colour-ID with one plate	Only one plate on the desk; ~70% target-colour-only scenes, ~30% with one other brick colour present

All 460 episodes are included in the merged training dataset.

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 46 prompts, in English, 10 episodes each.

1. Sort the lego by color
2. Put each lego on the plate that's the same color as it
3. Pick up the legos and sort them onto the right color plate
4. Move all the legos to the plate that matches their color
5. Take each lego and set it on the plate with the matching color
6. Put red lego on red plate
7. Move the red lego brick to the red plate
8. Pick up the red lego and put it in the red area
9. Sort the red lego bricks onto the red plate
10. Take only all red bricks and place them on the red plate
11. Put the green lego onto the green plate
12. Move the green lego brick to the green plate
13. Pick up the green lego and put it in the green area
14. Sort the green lego bricks onto the green plate
15. Take only all green bricks and place them on the green plate
16. Put the yellow lego brick onto the yellow plate
17. Move the yellow lego to the yellow plate
18. Pick up the yellow lego and put it into the yellow area
19. Sort the yellow lego bricks onto the yellow plate
20. Take only all yellow bricks and place them on the yellow plate
21. Put the blue lego brick onto the blue plate
22. Move the blue lego to the blue plate
23. Pick up the blue lego and put it in the blue area
24. Sort the blue lego bricks onto the blue plate
25. Take only all blue bricks and place them on the blue plate
26. Place each lego brick on the plate that matches its color
27. Pick up the lego bricks and sort them onto the correct color plate
28. Move all lego bricks to their matching colored plate
29. Organize the lego bricks by placing each one on its corresponding color plate
30. Take each lego brick and put it on the plate with the same color
31. Put red lego in red plate
32. Sort the red lego brick in the red plate
33. Put blue lego in blue plate
34. Pick the blue brick up and put it in the blue plate
35. Put yellow lego in yellow plate
36. Take the yellow lego and put it into the yellow plate
37. Put green lego in green plate
38. Sort the green lego into the green plate
39. Sort the yellow brick onto the yellow plate
40. Put all yellow lego bricks onto their matching yellow bowl
41. Place the red lego brick onto the red plate
42. Grab the red block and drop it on the red tray
43. Transfer the green lego to the green plate
44. Lift the green block and place it in the green bowl
45. Deposit the blue lego brick onto the blue plate
46. Put the blue lego bricks into the blue bowl

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("Project-IRA/TPSoSe2026_Dataset_Lego_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_Lego_LeRobot_SO101 \
    --batch_size=64 --steps=200000 \
    --policy.device=cuda

Models trained on this dataset

Model	Best checkpoint	Quality
Pi05 V4 Lego	030000	Works really well
SmolVLA V2 Lego	030000	OK-ish

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}}
}