Add GeneralVLA-2 training-free deployment manifest

README.md

@@ -0,0 +1,90 @@
+---
+license: apache-2.0
+language:
+- en
+- zh
+library_name: pytorch
+tags:
+- robotics
+- vision-language-action
+- robot-planning
+- 3d-reconstruction
+- long-term-memory
+- training-free
+---
+
+# GeneralVLA-2: Geometry-Aware Reconstruction and Governed Memory for Robot Planning
+
+GeneralVLA-2 is a training-free, planner-facing extension of GeneralVLA. It improves the two evidence sources used by the 3D planning agent:
+
+1. **GeoFuse-MV3D** supplies conservative, mask-verified, appearance-preserving object geometry from calibrated multi-view observations.
+2. **Governed KnowledgeBank** supplies quality-, confidence-, lifecycle-, conflict-, and geometry-aware experience retrieval.
+
+The project keeps the original hierarchical interface: affordance perception, 3D trajectory planning, and low-level robot execution.
+
+## Important: checkpoint scope
+
+GeneralVLA-2 does **not** introduce a newly trained monolithic checkpoint. Its contributions operate at reconstruction and planning time without updating model parameters. This repository therefore provides the official deployment manifest and configuration instead of duplicating approximately 40 GB of base and third-party checkpoints.
+
+The GeneralVLA project assets are hosted at:
+
+- [AIGeeksGroup/GeneralVLA](https://huggingface.co/AIGeeksGroup/GeneralVLA)
+
+Additional upstream assets used by GeoFuse-MV3D, such as SAM 3D Objects and VGGT, must be obtained from their official releases. See `asset_manifest.json` for the complete provenance and purpose of each dependency.
+
+## 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
+
+## Setup
+
+```bash
+git clone https://github.com/AIGeeksGroup/GeneralVLA-2.git
+cd GeneralVLA-2/GeneralVLA
+
+hf download AIGeeksGroup/GeneralVLA \
+  --local-dir vendor/GeneralVLA/pretrain_model
+
+bash scripts/bootstrap.sh
+PYTHONPATH=src python -m robot_memory_vla.app.main \
+  --preflight \
+  --config-dir ./configs
+```
+
+The checked-in GitHub configuration already uses repository-relative paths. `generalvla2.yaml` in this model repository records the matching asset layout in a compact deployment-oriented form.
+
+## Base assets
+
+The GeneralVLA asset repository contains:
+
+- `LISA-7B-v1-explanatory`: language and segmentation reasoning checkpoint.
+- `segagent/zzzmmz/SegAgent-Model`: visual grounding model.
+- `sam_vit_h_4b8939.pth`: SAM ViT-H segmentation checkpoint.
+- `clip-vit-large-patch14`: CLIP vision-language encoder.
+- `checkpoints/v1/checkpoint-rs.tar`: GraspNet checkpoint used by the robot stack.
+
+These assets retain their own upstream licenses and terms. The Apache-2.0 metadata on this repository applies to the GeneralVLA-2 configuration and documentation, not to independently distributed third-party checkpoints.
+
+## 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.
+
+## 中文说明
+
+GeneralVLA-2 的新增贡献是训练无关的多视图几何增强与长期记忆治理，不对应一套重新训练的大模型权重。本仓库提供官方模型入口、部署配置和资产来源清单；实际运行时请从 `AIGeeksGroup/GeneralVLA` 下载基础权重，并按照 GitHub 仓库中各组件的说明准备 GeoFuse-MV3D 所需的官方上游资产。
+
+## Citation
+
+```bibtex
+@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}
+}
+```
+

asset_manifest.json

@@ -0,0 +1,47 @@
+{
+  "project": "GeneralVLA-2",
+  "checkpoint_scope": "training-free planner and reconstruction extensions; no newly trained monolithic checkpoint",
+  "code_repository": "https://github.com/AIGeeksGroup/GeneralVLA-2",
+  "base_asset_repository": "AIGeeksGroup/GeneralVLA",
+  "base_assets": [
+    {
+      "path": "LISA-7B-v1-explanatory",
+      "purpose": "language and segmentation reasoning",
+      "source": "AIGeeksGroup/GeneralVLA"
+    },
+    {
+      "path": "segagent/zzzmmz/SegAgent-Model",
+      "purpose": "visual grounding",
+      "source": "AIGeeksGroup/GeneralVLA"
+    },
+    {
+      "path": "sam_vit_h_4b8939.pth",
+      "purpose": "SAM ViT-H segmentation",
+      "source": "AIGeeksGroup/GeneralVLA"
+    },
+    {
+      "path": "clip-vit-large-patch14",
+      "purpose": "vision-language encoding",
+      "source": "AIGeeksGroup/GeneralVLA"
+    },
+    {
+      "path": "checkpoints/v1/checkpoint-rs.tar",
+      "purpose": "GraspNet robot grasping checkpoint",
+      "source": "AIGeeksGroup/GeneralVLA"
+    }
+  ],
+  "geofuse_upstream_assets": [
+    {
+      "repository": "facebook/sam-3d-objects",
+      "purpose": "SAM 3D Objects checkpoint and reconstruction runtime"
+    },
+    {
+      "repository": "facebook/VGGT-1B",
+      "purpose": "optional geometry-prior provider"
+    }
+  ],
+  "notes": [
+    "Third-party checkpoints retain their own licenses and access terms.",
+    "Generated PLY files and benchmark outputs are evaluation artifacts, not model weights."
+  ]
+}

generalvla2.yaml

@@ -0,0 +1,26 @@
+project:
+  name: GeneralVLA-2
+  code_repository: https://github.com/AIGeeksGroup/GeneralVLA-2
+  base_asset_repository: AIGeeksGroup/GeneralVLA
+  parameter_update_required: false
+
+runtime:
+  precision: fp16
+  load_in_4bit: true
+  device: cuda:0
+  retrieval_backend: simple
+  retrieval_top_k: 3
+  require_operator_confirmation: true
+
+assets:
+  lisa: vendor/GeneralVLA/pretrain_model/LISA-7B-v1-explanatory
+  segagent: vendor/GeneralVLA/pretrain_model/segagent/zzzmmz/SegAgent-Model
+  sam_vit_h: vendor/GeneralVLA/pretrain_model/sam_vit_h_4b8939.pth
+  clip: vendor/GeneralVLA/pretrain_model/clip-vit-large-patch14
+  graspnet: vendor/GeneralVLA/pretrain_model/checkpoints/v1/checkpoint-rs.tar
+
+components:
+  generalvla_runtime: GeneralVLA
+  geometry_reconstruction: GeoFuse-MV3D
+  governed_memory: KnowledgeBank
+