pi05-libero-sft

pi0.5 LIBERO full-SFT (comparison baseline)

Full-parameter supervised fine-tune of π0.5 (pi05) on the LIBERO benchmark, trained with the openpi PyTorch backend. This is a full SFT checkpoint (not LoRA), intended as a comparison baseline.

Training

• Base: pi05_base (paligemma gemma_2b + action expert gemma_300m), action horizon 10.
• Config: pi05_libero, full fine-tune (no LoRA, no freeze).
• 60,000 steps, global batch 128 (32/GPU), 4× RTX A6000 DDP, bf16 + gradient checkpointing.
• Data: physical-intelligence/libero (LeRobot), 2 cameras (image + wrist_image).
• Final training loss ≈ 0.0036.

LIBERO eval (partial)

Evaluated on the 4 standard suites (openpi client/server, MuJoCo EGL). Numbers below are from a partial run (early-stopped), not the full 50 trials/task:

Reference (openpi pi0.5 @ 30k): spatial 98.8 / object 98.2 / goal 98.0 / libero_10 92.4 (avg 96.85).

Files

• model.safetensors — bf16 model weights (inference).
• optimizer.pt — AdamW optimizer state (for resuming training).
• metadata.pt — training metadata (global_step).
• assets/physical-intelligence/libero/norm_stats.json — normalization stats (required for inference).

Usage

Serve with openpi (auto-detects the PyTorch checkpoint via model.safetensors):

python scripts/serve_policy.py --env LIBERO \
  policy:checkpoint --policy.config pi05_libero --policy.dir <path-to-this-checkpoint>