NeuroVLA · LIBERO (all 4 suites, joint training)

Brain-inspired 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.

NeuroVLA couples a Qwen2.5-VL backbone with a layer-wise Q-Former that extracts action-relevant features from the VLM's hidden states, feeding a Spiking Neural Network (SNN) action head. The model was trained in a single supervised run (not continual learning) on a mixed stream of all 4 LIBERO suites, using the libero_all data mix.

Overview


Architecture	NeuroVLA (Qwen2.5-VL-3B + layer-wise Q-Former + SNN head)
Base VLM	`Qwen/Qwen2.5-VL-3B-Instruct`
Q-Former	Layers 36 → 37 · `num_query_tokens=8` · `output_dim=768`
Action head	DiT-based, `hidden_size=1024`, `action_dim=7`, `state_dim=7`, chunk 16
Training data	LIBERO · all 4 suites (Goal + Spatial + Object + Long) · `dataset_mix=libero_all`
Training type	Supervised fine-tuning (single run; not continual learning)
Attention	SDPA (not flash-attention, to avoid ABI pinning)
Optimiser	AdamW · `lr_base = 2.5e-5` · cosine-with-min-lr · 5 000 warmup
Step budget	50 000 (this release) · saved every 10 000 steps
Hardware / batch	2 × A800 80 GB · `per_device_batch_size = 16`

Files

├── README.md                   model card
├── framework_config.yaml       AlphaBrain framework configuration
├── dataset_statistics.json     action normalisation statistics (required for inference)
├── model.safetensors           full VLA weights (~7.7 GB)
├── resume_meta.json            training metadata (step count, GPU count)
└── qwen_pretrained/            Qwen2.5-VL tokenizer + preprocessor configs

Usage

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

export PRETRAINED_MODELS_DIR=/path/to/models   # must contain Qwen2.5-VL-3B-Instruct/

huggingface-cli download AlphaBrainGroup/neurovla-libero-all4suite \
    --local-dir ./neurovla_libero_all

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

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

Reproduction

# Framework's NeuroVLA pretraining entry
bash scripts/run_brain_inspired_scripts/run_neurovla_pretrain.sh \
    --yaml configs/neurovla_all4suite_libero.yaml   # (or equivalent config for 4-suite mix)

Expect multi-day training on 2 × A800 80 GB for the full 50 000-step schedule. The shipped framework_config.yaml is the exact training configuration used for this checkpoint.

Notes

Joint-training baseline, not continual learning. For the CL release of NeuroVLA (sequential training on LIBERO-Goal with Experience Replay), see AlphaBrainGroup/neurovla-cl-libero-goal and its LoRA variant.
Attention implementation is SDPA, chosen to avoid flash-attn ABI pinning across environments. Users who have a matching flash-attn wheel can override via --framework.qwenvl.attn_implementation=flash_attention_2.

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}
}