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

GeneralVLA-2 is a training-free framework that strengthens robot planning with two complementary components:

1. GeoFuse-MV3D produces conservative, mask-verified and appearance-preserving object geometry from calibrated multi-view observations.
2. Governed KnowledgeBank provides quality-, confidence-, lifecycle-, conflict- and geometry-aware experience retrieval.

GeneralVLA-2 does not introduce a newly trained monolithic checkpoint. This repository bundles selected runtime weights and representative GeoFuse-MV3D 3D Gaussian reconstruction outputs so that the commonly used vision and grasping components can be obtained from one repository.

Repository contents

• weights/sam/sam_vit_h_4b8939.pth: SAM ViT-H segmentation checkpoint.
• weights/clip-vit-large-patch14/: CLIP ViT-L/14 vision-language encoder and tokenizer files.
• weights/graspnet/checkpoint-rs.tar: GraspNet checkpoint used by the robot stack.
• geofuse_mv3d/examples/alarm_refined.ply: refined 3D Gaussian reconstruction for the alarm scene.
• geofuse_mv3d/examples/shoe_refined.ply: refined 3D Gaussian reconstruction for the shoe scene.
• generalvla2.yaml: compact deployment-oriented configuration.
• asset_manifest.json: machine-readable description of the bundled assets.

The PLY files are method outputs rather than neural-network checkpoints. They contain position, appearance, opacity, scale and rotation attributes for the reconstructed 3D Gaussians.

Download

hf download AIGeeksGroup/GeneralVLA-2 \
  --local-dir vendor/GeneralVLA-2

Code and project page

• Code: https://github.com/AIGeeksGroup/GeneralVLA-2
• Project page: https://aigeeksgroup.github.io/GeneralVLA-2
• Paper: https://arxiv.org/abs/2606.17480

Reported results

• On GSO-30, GeoFuse-MV3D reduces CD and LPIPS by 2.20% and 2.02%, while increasing PSNR and SSIM by 2.36% and 1.03% over the MV-SAM3D baseline under the same input-view protocol.
• Governed KnowledgeBank improves over ReasoningBank by 4.53% on Terminal-Bench 2.0 success rate and 3.73% on SWE-Bench Verified resolve rate, while reducing average steps by 4.95% and 5.65%.
• GeneralVLA-2 produces successful trajectories on all 14 evaluated RLBench tasks without parameter training and is best among the compared methods on 10 tasks.

License notice

The Apache-2.0 metadata applies to the GeneralVLA-2 configuration and documentation. Bundled third-party checkpoints retain their respective upstream licenses and terms.

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}
}