Research map / 3D vision / 3D generation
R3DP: Real-Time 3D-Aware Policy for Embodied Manipulation
Yuhao Zhang, Wanxi Dong, Yue Shi, Yi Liang, Jingnan Gao, Qiaochu Yang, Yaxing Lyu, Zhixuan Liang, Yibin Liu, Congsheng Xu, Xianda Guo, Wei Sui and 4 more
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 |
|---|---|---|---|
| add VGGT 3D features (ablation) | DP-multi · RoboTwin (4 tasks) | Success rate (%): 10.8% -> 39.8% | Improved |
| add Multi-View Feature Fuser (MVFF) (ablation) | DP+VGGT · RoboTwin (4 tasks) | Success rate (%): 39.8% -> 55.3% | Improved |
| add Asynchronous Fast-Slow Collaboration (AFSC) (ablation) | DP+VGGT+MVFF · RoboTwin (4 tasks) | Success rate (%): 55.3% -> 62.3% | Improved |
| add AFSC (TFPNet on non-key frames) (ablation) | DP+VGGT+MVFF · RoboTwin | Obs. encoder latency (ms): 78.3 -> 50.5 | Improved |
| increase tau from 4 to 8 (ablation) | R3DP · RoboTwin | Obs. encoder latency (ms): 50.5 -> 40.3 | Improved |
| increase tau from 4 to 8 (ablation) | R3DP · RoboTwin (10 tasks) | Success rate (%): 69.0% -> 65.7% | No clear effect |
| use multi-view input with naive feature concatenation | Diffusion Policy · RoboTwin (10 tasks) | Success rate (%): 36.1% -> 17.6% | Hurt |
| replace ground-truth depth with Depth Anything v2 estimated depth | DP3 · RoboTwin (10 tasks) | Success rate (%): 57.6% -> 28.2% | Hurt |
| use R3DP instead of pi0 fine-tuned with LoRA | pi0 · RoboTwin (10 tasks) | Success rate (%): 59.9% -> 69.0% | Improved |
| use R3DP instead of DP3 with real depth sensors | DP3 · Real-world (4 tasks) | Success rate (%): 48.4% -> 71.7% | Improved |
| add AFSC on top of VGGT and MVFF (ablation) | DP+V+M · Real-world (4 tasks) | Success rate (%): 66.7% -> 71.7% | Improved |
| add AFSC to reduce VGGT encoding cost (ablation) | DP+V+M · Real-world RTX 4090 | Obs. encoder latency (ms): 105.1 -> 62.18 | Improved |
Closest papers by meaning
- DiffuView: Multi-View Diffusion Pretraining for 3D Aware Robotic Manipulation (CVPR 2026 · similarity 0.89)
- Action-Geometry Prediction with 3D Geometric Prior for Bimanual Manipulation (CVPR 2026 · similarity 0.89)
- MVISTA-4D: View-Consistent 4D World Model with Test-Time Action Inference for Robotic Manipulation (ICML 2026 · similarity 0.89)
- DynaRend: Learning 3D Dynamics via Masked Future Rendering for Robotic Manipulation (NeurIPS 2025 · similarity 0.88)
- SAM-E: Leveraging Visual Foundation Model with Sequence Imitation for Embodied Manipulation (ICML 2024 · similarity 0.87)
- PDFactor: Learning Tri-Perspective View Policy Diffusion Field for Multi-Task Robotic Manipulation (CVPR 2025 · 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.
