Vision-Language-Action policies and multi-task manipulation datasets for the SO-101 low-cost robot arm, built on LeRobot.
Code: https://github.com/Project-IRA/interactive-robotic-arm — the repositories here and that repo depend on each other; the code repo contains the recording, validation, merging, training and inference pipeline that produced everything in this organization.
Authors: Cleo Baten, Bela Keppler, Jonas Sapper
If you just want a policy that works on an SO-101:
Project-IRA/TPSoSe2026_Pi05_LeRobot_SO101_Finetuning_V7_Full_V2— checkpoint 008000
This is the strongest model in the org: Pi0.5 fine-tuned on all four tasks, with image augmentation. Note its unusual internal directory layout, documented on its card.
Pi0.5 sorting lego bricks onto colour-matched plates on the physical SO-101.
| Task | What the arm does |
|---|---|
| Sort Lego Color | Sort lego bricks onto colour-matched plates |
| Desk Cleanup | Clear objects from the desk into a bowl |
| Dice Throw | Pick up a dice cup and tip the dice onto the table ("true random") |
| Fetch Ball | Pick up a ball and place it into a waiting human hand |
Quality ratings are qualitative operator assessments from rollouts on the physical arm. There are no formal success-rate numbers.
| Model | Base | Trained on | Best checkpoint | Quality |
|---|---|---|---|---|
| Pi05 V7 Full V2 | Pi0.5 | all 4 tasks | 008000 | Best overall — works really well |
| Pi05 V7 Full | Pi0.5 | all 4 tasks | 010000 | Works really well |
| Pi05 V4 Lego | Pi0.5 | Lego only | 030000 | Works really well (Lego only) |
| SmolVLA V6 Full | SmolVLA | all 4 tasks | ~150000 (untested) | OK-ish, across most tasks |
| SmolVLA V5 Full | SmolVLA | all 4 tasks | 100000 | Works, but worse than V6 |
| SmolVLA V2 Lego | SmolVLA | Lego only | 030000 | OK-ish (Lego only) |
| SmolVLA V3 Desk Cleanup | SmolVLA | Desk Cleanup only | — | Does not work |
| SmolVLA V1 Misc | SmolVLA | misc early data | — | Does not work |
The headline result: Pi0.5 substantially outperforms SmolVLA on this setup. The ~450M SmolVLA appears capacity-limited across four tasks, while Pi0.5 handles the full multi-task set well — and needs far fewer steps to get there (8k–10k vs 100k–200k).
| Dataset | Episodes | Content |
|---|---|---|
| Full Merged Final V1 | 930 | The training set. All 4 tasks, 93 prompts, 844,208 frames |
| Lego | 460 | Colour sorting |
| Desk Cleanup | 350 | Desk clearing (250 of these went into the merge) |
| Dice Throw | 120 | Dice cup |
| Fetch Ball | 100 | Ball handover |
| Collection | — | Early miscellaneous data; superseded |
| 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.
Language conditioning was varied deliberately: the prompt phrasing was changed roughly every 10 episodes during recording. The merged dataset therefore contains 93 distinct prompts across 10 episodes each, all in English. Full prompt lists are on each dataset card.
All models and datasets here are CC BY-SA 4.0 — share-alike, so derivatives must stay open. Upstream LeRobot, SmolVLA and Pi0.5/openpi components are Apache-2.0; their copyright notices and licence text are retained as Apache-2.0 Section 4 requires.
@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}}
}