AlphaBrainGroup/paligemma-pi0-libero-all4suite

PaliGemma-Pi0.5 · LIBERO (all 4 suites, joint training, openpi-aligned)

Vision-Language-Action (VLA) checkpoint released with the AlphaBrain framework. Trained jointly on all four LIBERO suites — Goal, Spatial, Object, and Long — for direct evaluation across the full LIBERO benchmark without retraining.

This is the openpi-aligned Pi0.5 training recipe: PaliGemma 3B VLM

• a 300 M flow-matching Gemma action expert, with the exact architecture layout and MEAN_STD action normalization used by OpenAI's openpi. Released as the steps = 15 000 checkpoint of a 60 000-step budget run, which is the best-performing openpi-aligned checkpoint in AlphaBrain's PaliGemma-Pi0.5 family on LIBERO.

Overview

Architecture	PaliGemmaPi0 (PaliGemma 3B + Pi0.5 flow-matching action expert)
Base VLM	google/paligemma-3b-pt-224 (bundled /datasets/peligemma equivalent)
Action expert	Gemma 300 M · depth=18, width=1024, mlp=4096, bfloat16
Action head	Flow matching, action_dim=7, horizon=10, 10 inference steps
Training data	LIBERO · all 4 suites (Goal + Spatial + Object + Long) · dataset_mix=libero_all
Training type	Supervised fine-tuning (single run; not continual learning)
Normalization	MEAN_STD (action & state mean/std hardcoded in framework_config.yaml)
Attention	SDPA (no flash-attn pinning)
Optimiser	AdamW · lr = 5e-5 · cosine-with-min-lr · 10 000 warmup
Step budget	15 000 (this release) / 60 000 planned
Hardware / batch	4 × A800 80 GB · per_device_batch = 32 · grad_accum = 2 · effective 256

Results

Evaluated on all 4 LIBERO suites, 50 rollouts per task × 10 tasks per suite = 500 episodes per suite.

Suite	Success Rate
LIBERO-Goal	96.7 %
LIBERO-Spatial	100.0 %
LIBERO-Object	100.0 %
LIBERO-10 (Long)	96.0 %
Avg (4-suite)	98.2 %

These numbers are very close to OpenPi's official Pi0.5 (98.0 / 98.8 / 98.2 / 92.4, avg 96.85) and establish a strong AlphaBrain Pi0.5 baseline at one quarter of the step budget (15 k vs 30 k).

Files

├── README.md                   model card
├── framework_config.yaml       AlphaBrain framework configuration (contains MEAN_STD norm stats inline)
├── model.safetensors           full VLA weights (~8.8 GB, includes VLM + action expert + flow-matching head)
├── resume_meta.json            training metadata (completed_steps=15000, effective_bs=256)
└── paligemma_pretrained/       PaliGemma tokenizer + preprocessor configs

Usage

git clone https://github.com/AlphaBrainGroup/AlphaBrain.git
cd AlphaBrain
pip install -e .

export PALIGEMMA_TOKENIZER_PATH=/path/to/paligemma_pretrained   # or bundled dir
export PI05_PRETRAINED_PATH=/path/to/this/download               # if you want to fine-tune from here

huggingface-cli download AlphaBrainGroup/paligemma-pi0-libero-all4suite \
    --local-dir ./paligemma_pi0_libero_all

python deployment/model_server/server_policy.py \
    --ckpt_path ./paligemma_pi0_libero_all --port 10093 --use_bf16

For evaluation on any of the 4 LIBERO suites, see the LIBERO eval pipeline.

Reproduction

bash scripts/run_base_vla/train.sh paligemma_pi0_openpi_aligned_v3

Expect multi-day training on 4 × A800 80 GB for the full 60 000-step schedule. The shipped framework_config.yaml captures the exact framework configuration used for this 15 000-step checkpoint.

Notes

• Joint-training baseline, not continual learning. For CL releases of AlphaBrain models see the qwengr00t-cl-* / neurovla-cl-* / paligemma-pi0-cl-* repos.
• Attention: SDPA — chosen so the checkpoint loads without a pinned flash-attn wheel. Users can override to flash_attention_2 via --framework.paligemma.attn_implementation=flash_attention_2 if available.
• MEAN_STD norm is baked into framework_config.yaml. A separate dataset_statistics.json is not required for inference.

License

MIT — see the parent repository.

Citation

@misc{alphabrain2026,
  title  = {AlphaBrain: A Modular Open-Source Framework for Embodied Intelligence Research},
  author = {AlphaBrain Team},
  year   = {2026},
  url    = {https://github.com/AlphaBrainGroup/AlphaBrain}
}