RAIDS HKPC robot policy checkpoints

pi0.5

  • pi05/full/stage01_17_h40/ contains the ready-to-load full inference parameters for the Stage 1-17 checkpoints at 80k, 90k, 100k and 110k. The formal deployment uses the matching task catalog at pi05/deployment/tasks.json.
  • pi05/delta/ contains 14 independent inference deltas. Each uses the same pi05/base/params/. Reconstruct a selected step with the project CPU exporter and its own normalization assets. Stage 1-8 H40 120k is distributed as base plus delta.
  • pi05/full/ also includes the existing Stage 1-8 full inference exports. pi05/historical/ contains the inference parameters and associated normalization from eight earlier training checkpoints.
  • The PaliGemma tokenizer is supplied under pi05/tokenizer/. The pi0.5 base was published by Physical Intelligence. See Gemma terms and redistribution notice.

ACT

Each policy directory under act/ contains its model weights and loading configuration. The Stage 9.5 and Stage 17 bundles include their normalization and task bindings. The Stage 17 dashboard defaults to act/stage17_h40/checkpoints/050000/ and provides six steps at 5k intervals from 25k. ACT refinement variants retain their learned feature estimator and warm-up policy where required.

Download the selected policy at a fixed model repository revision. The accompanying deployment software is distributed separately. The LeRobot training data is shared through a private handover repository.

Downloads last month

-

Downloads are not tracked for this model. How to track
Video Preview
loading

Collection including BBBXR/raids-hkpc-checkpoints