pi05-droid-jointpos-pytorch

pi05-droid-jointpos-pytorch

PyTorch conversion of Physical Intelligence's pi05_droid_jointpos checkpoint (gs://openpi-assets/checkpoints/pi05_droid_jointpos, present in the official bucket though not listed in the openpi README), converted via openpi examples/convert_jax_model_to_pytorch.py with --config_name pi05_libero (matching the validator serving convention; the π₀.₅ weight set is identical across configs — verified key-for-key against the pi05_libero conversion), precision float32. Uploaded for RobotTrain subnet validator pipeline testing — a DROID/joint-position model is expected to score low on LIBERO, serving as a format-valid negative control (zero-score anchor) for benchmark discrimination tests.

Note: besides the checkpoint's original assets, assets/physical-intelligence/libero/norm_stats.json from the official pi05_libero checkpoint was added so the model can be served with the pi05_libero config (the validator's submission convention).

Important: the LIBERO score of this repo does NOT measure the original checkpoint's capability. The injected LIBERO norm_stats differ from the statistics this model was trained with, and its native state/action spaces do not match LIBERO's (8-dim EEF state / 7-dim delta-EEF actions), so under this pipeline its inputs and outputs are semantically mismatched by construction. Treat it strictly as a format-valid negative control for validator pipeline testing — never as an evaluation of the underlying model.

Layout: model.safetensors (fp32) + assets/<asset_id>/norm_stats.json. Load with openpi: create_trained_policy(get_config("pi05_libero"), <this dir>).