Research map / Point clouds / Point cloud classification
Any3D-VLA: Enhancing VLA Robustness via Diverse Point Clouds
Xianzhe Fan, Shengliang Deng, Xiaoyang Wu, Yuxiang Lu, Zhuoling Li, Mi Yan, Yujia Zhang, Zhizheng Zhang, He Wang, Hengshuang Zhao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| lift depth to point clouds and fuse with 2D patches instead of 2D-only | VLA · Simulation pilot | Single-Trial SR: 45.3 -> 61.1 | Improved |
| point cloud fusion instead of RGBD image-plane depth | VLA · Simulation pilot | Single-Trial SR: 56.8 -> 61.1 | Improved |
| point cloud fusion instead of implicit-3D (VGGT) encoder | VLA · Simulation pilot | Single-Trial SR: 46.3 -> 61.1 | Improved |
| add depth-pretrained encoder branch (implicit depth) to 2D-only | VLA · Simulation pilot | Single-Trial SR: 45.3 -> 55.8 | Improved |
| add 2D branch to 3D-only (ablation) | Any3D-VLA · Simulation ablation | Single-Trial SR: 44.2 -> 61.1 | Improved |
| add 2D branch to 3D-only (ablation) | Any3D-VLA · Simulation ablation | Grasp SR: 91.6 -> 89.5 | Hurt |
| use DINOv2-L in addition to SigLIP instead of DINOv2-S (ablation) | Any3D-VLA · Simulation ablation | Test SR: 72.6 -> 80.0 | Improved |
| add DINOv2 to 3D + SigLIP (ablation) | Any3D-VLA · Simulation ablation | Single-Trial SR: 42.1 -> 61.1 | Improved |
| hybrid point cloud training instead of simulator-only (ablation) | Any3D-VLA · Real-world zero-shot (DA3) | Average SR: 60.0 -> 62.5 | Improved |
| hybrid point cloud training instead of simulator-only (ablation) | Any3D-VLA · Simulation (DA3 inference) | Test SR: 78.9 -> 82.1 | Improved |
| hybrid point cloud training instead of simulator-only (ablation) | Any3D-VLA · Simulation (Simulator inference) | Test SR: 80.0 -> 81.1 | No clear effect |
| hybrid point cloud training instead of sensor-only (ablation) | Any3D-VLA · Real-world Task 1 (DA3 inference) | SR: 80.0 -> 93.3 | Improved |
Closest papers by meaning
- Learning from Videos for 3D World: Enhancing MLLMs with 3D Vision Geometry Priors (NeurIPS 2025 · similarity 0.89)
- BridgeVLA: Input-Output Alignment for Efficient 3D Manipulation Learning with Vision-Language Models (NeurIPS 2025 · similarity 0.88)
- 3D-VLA: A 3D Vision-Language-Action Generative World Model (ICML 2024 · similarity 0.88)
- VLM-3R: Vision-Language Models Augmented with Instruction-Aligned 3D Reconstruction (CVPR 2026 · similarity 0.88)
- Geometrically-driven Aggregation for Zero-shot 3D Point Cloud Understanding (CVPR 2024 · similarity 0.87)
- Point2Real: Bridging the Gap between Point Cloud and Realistic Image for Open-World 3D Recognition (AAAI 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
