izlley/pi05-libero-vla

pi05_libero_izlley — π0.5 VLA fine-tuned on LIBERO (PyTorch, step 29999)

π0.5 (a flow-matching Vision-Language-Action model) full fine-tuned on the LIBERO benchmark, then converted from JAX to PyTorch with verified numerical parity. This is an independent study reproduction focused on the full training→evaluation→conversion pipeline for VLA models.

• Base model: π0.5 (physical-intelligence/pi05, via the openpi codebase)
• Architecture: VLM backbone + flow-matching action expert (π0 family). Hybrid FAST+Flow inference; action chunking.
• Task suite: LIBERO (Spatial / Object / Goal / Long, 4 suites)
• This checkpoint: PyTorch, training step 29999 (final). Format = model.safetensors + config.json + norm stats.

Headline results (closed-loop success rate, 100 episodes/suite)

Suite	Success
LIBERO-Spatial	98%
LIBERO-Object	99%
LIBERO-Goal	98%
LIBERO-Long	96%
Average	97.75%

• Step-10000 checkpoint already reached 97.5% avg → near-ceiling convergence by 10k (diminishing returns to 30k).
• Evaluated closed-loop with a custom harness (eval_rollout.py): success rate + rollout videos logged to W&B.

Training setup

• Full fine-tune (not LoRA) of π0.5, H100 × 4 with FSDP, 30k steps, ~2.9 s/it (22 h), loss 0.088 → 0.02.
• Data: LIBERO RLDS → LeRobot format conversion (izlley/libero, 1,693 episodes / ~33 GB), HF_LEROBOT_HOME pinned to persistent storage.
• Config: pi05_libero_izlley (repo_id = izlley/libero), keeping the upstream pi05_libero config intact.
• Logging hardened: update_param_ratio, per-module grad/param norms, histograms, open-loop val/action_mse.

JAX ↔ PyTorch parity

The original training was in JAX; this checkpoint is the PyTorch conversion of step 29999. Parity verified against the JAX weights: cosine similarity 0.99999, max abs diff 0.008 — i.e. the PyTorch export is numerically faithful for downstream inference.

Intended use & scope

• Use: research/education on VLA fine-tuning, LIBERO reproduction, JAX→PyTorch VLA deployment, inference experiments.
• Not: a production robot controller. LIBERO is a simulation benchmark; real-robot transfer is out of scope for this checkpoint.
• Independent study artifact; results are self-measured and reproducible via the openpi pipeline + the config above.

Provenance / license

Derived from π0.5 (physical-intelligence/pi05) under the openpi project (Apache-2.0) and the public LIBERO benchmark. Fine-tune weights released Apache-2.0. Base-model and dataset licenses of the respective upstreams apply.

Related

Part of a broader VLA/robotics self-study (VLA paper deep-dives, RL fine-tuning with RLinf, Isaac Lab sim platform). Companion analysis and lab notes are documented separately.