SeonghoonYu/RACE_gauss_liberoplus

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

Two eval-only checkpoints of the gauss (gauss_timing) method, selected as the best step on the full LIBERO-Plus benchmark (arXiv:2510.13626, 10,030 perturbed variants, 1 trial/variant, seed 7, leaked task.language prompt as on the leaderboard):

tag	checkpoint	training data	horizon / replan	LIBERO-Plus overall
zero-shot best	pi05_libero_gauss_h10/gauss/10000	plain LIBERO (physical-intelligence/libero), frozen VLM	h10 / replan 10	83.7% (8400/10030, 95% CI 83.0–84.5)
SFT best	pi05_libero_plus_gauss_h15/gauss/100000	LIBERO-Plus training set (Sylvest/libero_plus_lerobot), frozen VLM, init from pi05_libero_gauss_h15/gauss/15000	h15 / replan 15	84.2% (8442/10030, 95% CI 83.4–84.9)

"Zero-shot" = trained only on clean LIBERO and evaluated on LIBERO-Plus without seeing its perturbations (comparable with the leaderboard's pi0 53.6 / pi0-Fast 61.6 rows). "SFT" = post-trained on the benchmark's own training set (14,347 episodes, 100k steps at batch 128), i.e. an in-domain number comparable with OpenVLA-OFT+ (79.6) / SRPO (82.1).

Per-dimension success rate

dimension	n	zero-shot (h10@10k)	SFT (h15@100k)
Background Textures	1076	97.9	94.5
Light Conditions	1142	97.5	95.0
Sensor Noise	1601	89.6	86.4
Objects Layout	1525	87.7	83.5
Language Instructions	1537	87.0	82.7
Camera Viewpoints	1599	70.0	82.3
Robot Initial States	1550	64.8	70.6
Overall (micro)	10030	83.7	84.2
Overall (macro-dim)	—	84.9	85.0

Per suite (zero-shot / SFT): spatial 87.9 / 86.6, object 87.5 / 89.3, goal 80.5 / 79.8, libero_10 79.4 / 81.2.

Robot-init protocol: _initstate_N (N≠0) variants use reset() only (no set_init_state), so the perturbed initial pose is preserved. Sensor noise is applied to the agentview camera only (benchmark behaviour).

Layout

pi05_libero_gauss_h10/gauss/10000/                # zero-shot best
  model.safetensors
  metadata.pt
  assets/physical-intelligence/libero/norm_stats.json
pi05_libero_plus_gauss_h15/gauss/100000/          # SFT best
  model.safetensors
  metadata.pt
  assets/Sylvest/libero_plus_lerobot/norm_stats.json
assets/gauss_timing/libero_cps.npz                # serving-time load dependency (zero-shot config)
assets/gauss_timing/libero_plus_cps.npz           # serving-time load dependency (SFT config)

Optimizer states are NOT included.

Serving (openpi fork with gauss_timing)

# zero-shot best
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/10000

# SFT best
python scripts/serve_policy.py --port 8000 --env LIBERO \
  policy:checkpoint --policy.config pi05_libero_plus_gauss_h15 \
  --policy.dir <download_root>/pi05_libero_plus_gauss_h15/gauss/100000

Notes:

• Config names 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 at inference).
• assets/gauss_timing/libero_cps.npz / libero_plus_cps.npz must exist relative to the server's CWD (repo root). They are loaded at policy build time but unused for inference.
• Replan convention: replan = horizon (h10→replan10, h15→replan15).
• The full 1k-step clean-LIBERO sweeps are in SeonghoonYu/RACE_gauss_libero.