PhyRC 2027 Phase 1 — pi0.5 LoRA wiring test
Participant project: https://github.com/leonyoon-3dai/PhyRC_2027_phase1
This first checkpoint tests the pipeline, not successful dressing. It uses the public pi0.5 base (gs://openpi-assets/checkpoints/pi05_base), openpi at 215abfb217dbac7d5f1273282331b9b1866c0479, and JAX LoRA for the language and action experts. The full inference parameters and optimizer/train state are retained. No conversion or merge was performed: inference uses the same LoRA architecture.
Training: 30 steps, batch size 1, 63 actual simulator transitions from 3 episodes, 30 processed samples (0.476 effective epochs). These are Codex-authored bounded scripted collection tests, labeled integration_test; they are not human demonstrations, live GPT API demonstrations, or successful dressing examples. Dataset revision: 057b59b1099407b20fe25dafa080285b34ab3aaa at https://huggingface.co/datasets/leonayoon/phyrc-2027-phase1-data.
Inputs: designated overview and two wrist RGB views plus public 32D robot state. Output: both robots' 18 normalized velocity/intent actions. Gripper +1 closes, -1 opens, 0 preserves intent. This checkpoint has not established overfitting, sleeve insertion, dressing success, RTX 3090 compatibility, or organizer scores.
The parameter directory, tokenizer, normalization and training request are stored under runs/20261022T0420Z-overfit/. Check SHA256.json. Upstream checkpoint folder name 29 corresponds to completed train state step 30. RNG basis is the upstream deterministic training seed and train-state step. Dataloader cursor and bitwise simulator state are not restored.
Use upstream openpi and component-specific pretrained-model/tokenizer terms. The project does not claim ownership of the base model or simulator assets. See the pinned openpi source and project THIRD_PARTY.md for attribution.