Project-IRA/TPSoSe2026_Pi05_LeRobot_SO101_Finetuning_V4_Lego

Pi0.5 SO-101 — V4 Lego (single task)

Pi0.5 fine-tuned on the Lego colour-sorting task only. This was the first Pi0.5 run and it demonstrated the model was worth scaling to all four tasks. For multi-task use, prefer V7 Full V2.

Demo

Pi0.5 sorting lego bricks onto colour-matched plates on the physical SO-101.

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

Base model	lerobot/pi05_base
Robot	SO-101 follower (6-DOF)
Training data	Project-IRA/TPSoSe2026_Dataset_Lego_LeRobot_SO101
Recommended checkpoint	030000
Inputs	desk_view + wrist_left camera images, 6-dim joint state, English instruction
Outputs	6-dim continuous action chunks

Quality

Works really well on Lego colour sorting. Recommended checkpoint 030000.

This model only knows the Lego task — it has never seen desk cleanup, dice, or ball handover data, and will not do them.

Training

Single L40S GPU — no FSDP, no multi-GPU. Pi0.5 fits on one 48 GB L40S at batch 24 with gradient checkpointing and bfloat16, as long as it is not sharded.

Setting	Value
Base	lerobot/pi05_base
Dataset	TPSoSe2026_Dataset_Lego_LeRobot_SO101 (460 episodes, 46 prompts)
Steps	30000, completed (--save_freq=5000)
Recommended checkpoint	030000 (the final one)
Batch size	24
Precision	--policy.dtype=bfloat16
Gradient checkpointing	on
Vision encoder	unfrozen
Parameters	4,143,404,816 — all trainable (num_learnable_params == num_total_params)
Optimizer	AdamW, betas (0.9, 0.95), grad clip 1.0
LR schedule	peak 2.5e-05, 1000 warmup steps, cosine decay to 2.5e-06 over 30000 steps
Normalization	ACTION: MEAN_STD, STATE: MEAN_STD, VISUAL: IDENTITY
Camera keys	native wrist_left / desk_view (no rename_map)
SLURM	--gres=gpu:L40S:1 --cpus-per-task=16 --mem=92G --time=96:00:00

lerobot-train \
    --dataset.repo_id=TPSoSe2026_Dataset_Lego_LeRobot_SO101 \
    --policy.type=pi05 \
    --policy.pretrained_path=lerobot/pi05_base \
    --policy.normalization_mapping='{"ACTION": "MEAN_STD", "STATE": "MEAN_STD", "VISUAL": "IDENTITY"}' \
    --policy.compile_model=false \
    --policy.gradient_checkpointing=true \
    --policy.dtype=bfloat16 \
    --policy.freeze_vision_encoder=false \
    --policy.train_expert_only=false \
    --policy.device=cuda \
    --policy.push_to_hub=false \
    --batch_size=24 --steps=30000 --save_freq=5000 \
    --num_workers=16 --tolerance_s=0.01 \
    --output_dir=outputs/train/pi05_lego \
    --job_name=pi05_lego \
    --wandb.enable=false

Note the contrast with the multi-task V7 runs, where the best checkpoint is very early (8k–10k): this single-task run is best at its final checkpoint. One task means less overfitting pressure at equivalent step counts.

Pi0.5 requires lerobot[pi] installed from GitHub main, not the PyPI release.

Usage

Model files are nested under outputs_V4/, so from_pretrained("Project-IRA/...") on the repo ID will not work:

outputs_V4/train/pi05_lego/checkpoints/<step>/pretrained_model/   <- weights
outputs_V4/train/pi05_lego/checkpoints/<step>/training_state/     <- resume only

Checkpoints present: every 5000 steps from 005000 to 030000. Use 030000 (the final one). Repo total ~147 GB.

Fetch just that checkpoint's weights (~13 GB instead of 147 GB):

hf download Project-IRA/TPSoSe2026_Pi05_LeRobot_SO101_Finetuning_V4_Lego \
  --include 'outputs_V4/train/pi05_lego/checkpoints/030000/pretrained_model/*' \
  --local-dir ./pi05_lego

from lerobot.policies.pi0.modeling_pi0 import PI0Policy   # pi05 shares the PI0 module

policy = PI0Policy.from_pretrained("<local path to the checkpoint's pretrained_model/>")
policy = policy.to("cuda").eval()

Inference dtype: checkpoints are saved from a bfloat16 training run. If you hit GPU OOM at inference, confirm the loaded policy is in bfloat16 and not silently upcast to float32.

Pi0.5 is ~4B parameters. On-robot inference from the robot PC is impractical; we served it over the asynchronous gRPC inference server shipped in the code repo (https://github.com/Project-IRA/interactive-robotic-arm) and ran the robot as a thin client.

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.

Environment notes

All training ran on a SLURM cluster with L40S GPUs. Two environment details were required and are easy to miss when reproducing:

• ffmpeg libraries for torchcodec. A minimal conda env supplies the shared libraries that torchcodec discovers at runtime: export LD_LIBRARY_PATH=$CONDA_PREFIX/envs/ffmpeg_libs_v8/lib:<venv>/lib/python3.12/site-packages/nvidia/npp/lib:$LD_LIBRARY_PATH
• --tolerance_s=0.01 on every run, to accommodate timestamp jitter in the recorded episodes.

Multi-GPU runs additionally set PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True. Datasets and the virtualenv were copied to node-local /scratch before training rather than read from shared storage.

No Weights & Biases logging was enabled for any run (--wandb.enable=false), so there are no public training curves — the job.*.err SLURM logs are the record.

Tasks and prompts

The model is conditioned on English natural-language instructions. Prompt phrasing was varied roughly every 10 episodes during recording, giving 93 distinct prompts in the merged dataset. Use one of the training prompts verbatim for best results — the full lists are on the dataset card.

Limitations

• Behaviour cloning. The policy imitates teleoperated demonstrations and has no notion of recovery beyond what was demonstrated. It is susceptible to covariate shift and can fail to recover from states outside the demonstration distribution.
• Recovery data is incidental, not systematic. Recovery behaviour appears in the data only where the operator happened to make and correct a mistake during recording; no recovery episodes were scripted deliberately.
• Single environment. All data comes from one lab desk with one lighting setup, one camera geometry, and one set of physical objects. Expect degradation elsewhere.
• Prompt sensitivity. Language conditioning was trained on a fixed set of phrasings (listed in the dataset card). Prompts far from those phrasings may behave unpredictably.
• No formal evaluation. Quality assessments below are qualitative, from operators observing rollouts on the physical arm. There are no success-rate numbers.
• Not safety-rated. Supervise all physical execution and keep the workspace clear.

Upstream licensing & attribution

This model is a derivative work of Apache-2.0 licensed components:

Component	Upstream	License
LeRobot framework	https://github.com/huggingface/lerobot	Apache-2.0
lerobot/pi05_base (Physical Intelligence, openpi)	https://github.com/Physical-Intelligence/openpi	Apache-2.0

Apache-2.0 permits relicensing derivative works. We retain the upstream copyright notices, license text, and NOTICE files for the incorporated material, as Apache-2.0 Section 4 requires. The upstream components remain under Apache-2.0 — only this project's own contributions (the fine-tuned weights and training configuration) are offered under CC BY-SA 4.0.

CC BY-SA 4.0 was chosen because it is share-alike: derivatives must be released under the same licence, so this work cannot be taken closed-source. The project's source code lives in a separate repository under its own licence — see https://github.com/Project-IRA/interactive-robotic-arm.

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}}
}