Research map / 3D vision / 3D generation
PhysMani: Physics-principled 3D World Model for Dynamic Object Manipulation
Peng Yun, Shouwang Huang, Hao Li, Jinxi Li, Jianan Wang, Bo Yang
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 |
|---|---|---|---|
| incorporate predicted 3D future dynamics into the 3DFA policy | 3DFA · PhysMani-Bench (16 tasks) | mean success rate (%): 37.8 -> 45.9 | Improved |
| remove velocity components D_t from future-aware tokens (ablation) | PhysMani · PhysMani-Bench ablation protocol | mean success rate (%): 45.4 -> 37.9 | Hurt |
| add learnable query token L instead of self-attention for dynamics fusion (ablation) | PhysMani · PhysMani-Bench ablation protocol | mean success rate (%): 40.8 -> 45.4 | Improved |
| concatenate 2D optical flow to RGB input of 3DFA | 3DFA · PhysMani-Bench (16 tasks) | mean success rate (%): 37.8 -> 37.5 | No clear effect |
| reduce world-model optimization iterations T from 50 to 40/30/20/10 (ablation) | PhysMani · PhysMani-Bench (16 tasks) | mean success rate (%) | No clear effect |
| use PhysMani online world model instead of FreeGave for future prediction (1st frame) | FreeGave · Future frame prediction | PSNR: 19.47 -> 26.90 | Improved |
| use PhysMani online world model instead of FreeGave for 10th future frame | FreeGave · Future frame prediction | PSNR: 17.89 -> 20.00 | Improved |
| add PhysMani future dynamics to 3DFA on real robot | 3DFA · Real-world dynamic tasks (4 tasks) | mean success rate (%): 45.3 -> 62.5 | Improved |
| add PhysMani future dynamics to 3DFA on Beat Buzz | 3DFA · PhysMani-Bench normal speed, task Beat Buzz | success rate (%): 82.5 -> 78.3 | No clear effect |
| use VLA pi0.5 instead of 3D policies on high-speed variants | pi0.5 · PhysMani-Bench (16 tasks) | mean success rate (%): 37.8 -> 8.3 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- FreeGave — FreeGave: 3D Physics Learning from Dynamic Videos by Gaussian Velocity (CVPR 2025) · better on 3 of 3 shared results
- ManiGaussian — ManiGaussian: Dynamic Gaussian Splatting for Multi-task Robotic Manipulation (ECCV 2024) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Dream to Manipulate: Compositional World Models Empowering Robot Imitation Learning with Imagination (ICLR 2025 · similarity 0.88)
- EnerVerse: Envisioning Embodied Future Space for Robotics Manipulation (NeurIPS 2025 · similarity 0.88)
- MolmoMotion: Forecasting Point Trajectories in 3D with Language Instruction (NeurIPS 2026 · similarity 0.87)
- Predictive Inverse Dynamics Models are Scalable Learners for Robotic Manipulation (ICLR 2025 · similarity 0.87)
- DreamVLA: A Vision-Language-Action Model Dreamed with Comprehensive World Knowledge (NeurIPS 2025 · similarity 0.87)
- DynaRend: Learning 3D Dynamics via Masked Future Rendering for Robotic Manipulation (NeurIPS 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.
