pi0-umilingo-pickplace-v1

Pi0 model fine-tuned on the in-house Umilingo dataset for UR5e pick-and-place style manipulation.

Model Summary

• Base model: lerobot/pi0_base
• Fine-tuning framework: LeRobot
• Robot embodiment: UR5e + gripper
• Policy contract: 7D state and 7D action
• Relative actions: enabled
• Relative exclude joints: gripper

Dataset

• Hugging Face dataset: eurecat-robotics/umilingo_lerobot
• Local training dataset id: local/umi_ur5e
• Data root used for training: data/processed/umi_ur5e_lerobot

Training Snapshot

• Training run: 20260706_135737
• Uploaded checkpoint: 050000
• Checkpoint path in this repo: checkpoints/20260706_135737/050000

Intended Use

This model is intended for research and internal development of language-conditioned robotic manipulation policies on UR5e-like setups.

Limitations

• Trained on in-house data distribution; out-of-distribution performance is not guaranteed.
• Safety checks and execution constraints are required before any real-robot deployment.
• No formal benchmark claims are provided in this card.

Reproducibility Notes

Training and data processing setup are maintained in the SESAM policy workspace. Key config references:

• config/policy/pi0_umi_ur5e.yaml
• config/dataset/local_umi.yaml

Citation

If you use this model, please cite your project/repository and this Hugging Face model page.