GeneralVLA-2: Geometry-Aware Reconstruction and Governed Memory for Robot Planning

This is the official repository for the paper:

GeneralVLA-2: Geometry-Aware Reconstruction and Governed Memory for Robot Planning

Haoyu Wang*, Guoqing Ma*, Zeyu Zhang*†, Yandong Guo, Boxin Shi, and Hao Tang^#

*Equal contribution. †Project lead. ^#Corresponding author.

Paper | Website | Models | HF Paper

This repository contains three code components used in the GeneralVLA project:

• GeneralVLA/: robot memory VLA runtime, model configuration, evaluation tools, and tests.
• GeoFuse-MV3D/: multi-view 3D geometry fusion and evaluation utilities.
• KnowledgeBank/: memory-augmented software-agent code and benchmark scripts.

Large checkpoints, datasets, generated outputs, robot logs, and benchmark trajectories are intentionally not included. Each component documents its own setup steps and expected external assets.

Citation

@article{wang2026generalvla2,
  title={GeneralVLA-2: Geometry-Aware Reconstruction and Governed Memory for Robot Planning},
  author={Wang, Haoyu and Ma, Guoqing and Zhang, Zeyu and Guo, Yandong and Shi, Boxin and Tang, Hao},
  journal={arXiv preprint arXiv:2606.17480},
  year={2026}
}

Repository Layout

.
├── GeneralVLA/
├── GeoFuse-MV3D/
├── KnowledgeBank/
├── THIRD_PARTY_NOTICES.md
└── README.md

Quick Start

Install and run each component from its own directory:

cd GeneralVLA
bash scripts/bootstrap.sh
pytest -q

cd GeoFuse-MV3D
pip install -r requirements.txt
python scripts/run_full_pipeline.py --config configs/paths.local.yaml

cd KnowledgeBank/third_party
pip install -e .
pytest tests -q

See the README inside each subdirectory for detailed asset paths, model configuration, and benchmark-specific instructions.

External Assets

The repository is code-only. Before running the full pipelines, prepare the external assets described by each component, including model checkpoints, benchmark datasets, WebArena services, GSO assets, and robot/runtime-specific configuration.

Component asset entry points:

• GeneralVLA/: project model assets are expected from https://huggingface.co/AIGeeksGroup/GeneralVLA.
• GeoFuse-MV3D/: use the official upstream assets documented in GeoFuse-MV3D/docs/external_assets.md.
• KnowledgeBank/: use the official benchmark/model-provider assets documented in KnowledgeBank/README.md.

Do not commit API keys, model checkpoints, local datasets, generated results, or private trajectories.