SeonghoonYu/RACE_gauss_libero

RACE gauss (pi0.5 + Gaussian transition-proximity) — LIBERO checkpoints

All 1k-interval checkpoints (1000–20000) of the gauss (gauss_timing) method for horizons h5 / h10 / h15 / h20, fine-tuned from official pi0.5 LIBERO on physical-intelligence/libero. Eval-only bundle: model.safetensors + checkpoint-embedded norm stats (assets/physical-intelligence/libero/norm_stats.json)

• metadata.pt. Optimizer states are NOT included.

Layout

pi05_libero_gauss_h{5,10,15,20}/gauss/{1000..20000}/
  model.safetensors
  metadata.pt
  assets/physical-intelligence/libero/norm_stats.json
assets/gauss_timing/libero_cps.npz   # serving-time load dependency (see below)

Serving (openpi fork with gauss_timing)

CUDA_VISIBLE_DEVICES=0 python scripts/serve_policy.py --port 8000 --env LIBERO \
  policy:checkpoint --policy.config pi05_libero_gauss_h10 \
  --policy.dir <download_root>/pi05_libero_gauss_h10/gauss/13000

Notes:

• Config names pi05_libero_gauss_h{H} are defined in the RACE repo's src/openpi/training/config.py; serving is observation-only (the timing head predicts b internally — no changepoint inputs needed at inference).
• assets/gauss_timing/libero_cps.npz must exist relative to the server's CWD (repo root). It is loaded at policy build time but unused for inference.
• Replan convention: replan = horizon (h5→replan5, h10→replan10, h15→replan15, h20→replan20).
• Clean-LIBERO best steps: h5 9000 (97.40), h10 13000 (98.65), h15 6000 (98.35), h20 18000 (96.55).