pi0 — ManiGuard real-teleop `clutter` (DROID schema)

LoRA fine-tune of pi0 on 60 real Franka teleop trajectories. This is the pi0 half of a model comparison: the pi0.5 counterpart was trained on the same data with the same recipe, so the two differ only in the model.

Task prompt: "pick up the smallest mug and place it in the bowl"
Dataset: IDEAS-Lab-Northwestern/real-clutter-60-droid-refined (private) — 60 episodes, 13,171 frames, 15 fps
Warm start: gs://openpi-assets/checkpoints/pi0_droid/params
Norm stats: pi0_droid's bundled DROID statistics, reused verbatim — not recomputed on this dataset. Bundled per rung at <step>/assets/droid/norm_stats.json. ⚠️ They are pi0's own; the pi0.5 sibling ships pi05_droid's different file. Do not swap them.
LoRA: gemma_2b_lora (rank 16) + gemma_300m_lora (rank 32), batch 4, EMA off, 50,000 steps.

Controller — read this before running on a robot


Action	`actions[0:7]` = arm joint VELOCITY (rad/s) · `actions[7]` = next-frame gripper target (normalized 0=open, 1=closed)
Apply as	joint-velocity command to the arm + position target to the gripper
Chunk	`action_horizon = 10` (pi0.5 sibling uses 15)
⚠️ No delta reconstruction	the actions are already velocities. openpi's `LeRobotDROIDDataConfig` deliberately applies no delta transform — do not integrate or difference them, unlike the ManiGuard sim checkpoints which emit absolute joint targets
State	`joint_position` (7) + `gripper_position` (1)
Cameras	`exterior_image_1_left` ← exterior third-view · `wrist_image_left` ← wrist · `exterior_image_2_left` is a zero-filled placeholder (the rig has no second exterior camera) and is masked
Rate	data collected at 15 fps

Checkpoints

Ladder 10000/ 20000/ 30000/ 40000/ 50000/ (the final rung is openpi's 0-indexed 49999, relabeled). train_state/ is omitted — inference only.

The pi0.5 sibling also ships 10k–50k, so any rung compares like-for-like.

The best real-robot rung is not assumed to be the last one; sweep the ladder on the robot.

Serving

openpi policy server with config pi0-droid_real_clutter_60_refined_lora from the ManiGuard openpi fork (maniguard/openpi_sft/train_configs.py), pointed at a downloaded rung. Evaluated on a real Franka, not in simulation.

Paper & Citation

Part of ManiGuard: paper (arXiv:2608.17386) · code · docs

@misc{peng2026maniguard,
  title         = {{MANIGUARD}: A Benchmark and Data Suite for Specification-Grounded
                   Safety Evaluation and Improvement of Robotic Manipulation},
  author        = {Peng, Yiyan and Wang, Philip and Zhan, Simon Sinong and Lyu, Yiqi
                   and Ni, Zhenyang and Yan, Jixin and Wong, Fiorelli and Jiao, Ruochen
                   and Yin, Hang and Cao, Xinyu and Shao, Huajie and Li, Manling
                   and Zhang, Ruohan and Zhu, Qi},
  year          = {2026},
  eprint        = {2608.17386},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2608.17386},
}

License

The fine-tuned weights derive from a Physical Intelligence openpi base model whose VLM backbone is PaliGemma; use of these weights is therefore subject to the Gemma Terms of Use (including the Gemma Prohibited Use Policy), which downstream users must pass on. The openpi training code and ManiGuard's own contributions are Apache-2.0.