Every Embodied Pi0 Protected ROCm Checkpoint
This is the protected Pi0 fine-tuning checkpoint used in the Datawhale Every Embodied AMD ROCm tutorial.
Verified result
- Strict physical success:
12/14 - Unseen-position subset:
9/10 - Hard subset:
6/8 - Task: MuJoCo mug-to-plate pick and place
- Recipe: clean-success-only data, blue-task reweighting, protected continuation to the selected checkpoint
- Model SHA256:
44aa854a5f084d39ba375d1ffe951ca0c8b6d147bacdb1a9e637ed1725514eeb
This is a fine-tuned Pi0 checkpoint. The result depends on the matching eef_abs action bridge, 8-D state (6DoF + gripper + timestamp), action-chunk execution, normalization statistics, camera order, and strict physical-success predicate. It must not be described as raw Pi0 zero-shot success.
Tutorial
See the Every Embodied AMD ROCm topic and 15_pi0_end_to_end.ipynb.
Contents
The weights/ directory contains model and configuration files. Evaluation metadata is kept in evaluation/. Optimizer state and raw datasets are intentionally omitted.
License and attribution
Please follow the licenses of Pi0, PaliGemma, LeRobot, the base checkpoint, and the Every Embodied project. This model card describes a reproducible tutorial artifact, not a new foundation model.