π0 (pi0) - LIBERO finetuned v044 (GGUF for vla.cpp)

GGUF conversion of lerobot/pi0_libero_finetuned_v044 for inference with vla.cpp, a lightweight C++ inference engine for Vision-Language-Action models built on top of llama.cpp.

π0 pairs a PaliGemma vision-language backbone with a Gemma action expert running a flow-matching head. On LIBERO it matches the lerobot reference and runs roughly two orders of magnitude faster than the lerobot CPU path.

Usage

# Terminal 1 - serve (use the CUDA build for inference)
./build-cuda/vla-server --bind tcp://*:5566 \
    pi0-libero-finetuned-v044.gguf

# Terminal 2 - drive a LIBERO episode (inside the LIBERO uv venv)
python eval/client/run_sim_client_direct.py \
    --arch pi0 \
    --task libero_object --task-id 0 --n-episodes 10 \
    --n-action-steps 32 \
    --vla-addr tcp://localhost:5566

Notes:

π0 uses the gated google/paligemma-3b-pt-224 tokenizer. Run huggingface-cli login and accept the licence, or point --tokenizer at a local copy.
--n-action-steps is 32 for this checkpoint (50 was used for the Orin Nano sweep).

Benchmark

Full libero_object sweep (10 tasks × 20 episodes = 200 episodes):

Hardware	n_act	Success rate	client/step	client/call	Peak mem
RTX 3060 (sm_86)	32	87.5%	9.74 ms	312 ms	5548 MiB VRAM
Jetson AGX Orin (sm_87)	32	85.5%	27.90 ms	893 ms	640 MiB RAM
Jetson Orin Nano 8 GB (sm_87, split)	50	80.5%	39.10 ms	1955 ms	6068 MiB RAM

The Orin Nano run was split (server on the Nano, client + LIBERO sim on a second machine) because a co-resident client + sim + server exceeds 8 GB; its timings include cross-machine network latency.

License

Weights follow the upstream license of lerobot/pi0_libero_finetuned_v044 (Apache-2.0). The vla.cpp conversion tooling and inference engine are MIT-licensed.