pi0.5 — OpenMicrowave, cumulative step 20,000
Original RLinf RoboCasa365 continual-learning model weights, uploaded without conversion or parameter changes.
- Training sequence: OpenMicrowave.
- Cumulative checkpoint step: 20,000; each listed task stage lasts 20,000 steps.
- Latest training task:
OpenMicrowave. - Source run:
robocasa365_single_task_openmicrowave_fullbc_from_c0. - Initialization: the existing C0 calibration checkpoint.
- Model configuration:
pi05_robocasa365_pretrain_human300.
The repository contains the original PyTorch state dictionary at
model_state_dict/full_weights.pt, the associated action/input normalization
statistics at norm_stats.json, and RLinf inference settings in
rlinf_model.yaml. Optimizer state is not included.
These are native RLinf weights, not a Transformers from_pretrained export.
Download the repository with huggingface_hub.snapshot_download, set
ROBOCASA365_EVAL_MODEL_PATH to that downloaded directory and
ROBOCASA365_NORM_STATS_PATH to its norm_stats.json, and use the RLinf OpenPI
model loader with the supplied settings.
No performance or causal-analysis claim is attached to this upload.