[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-92656":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":9,"language":10,"languages":9,"totalLinesOfCode":9,"stars":11,"forks":12,"watchers":13,"openIssues":14,"contributorsCount":15,"subscribersCount":15,"size":15,"stars1d":15,"stars7d":15,"stars30d":16,"stars90d":15,"forks30d":15,"starsTrendScore":15,"compositeScore":17,"rankGlobal":9,"rankLanguage":9,"license":18,"archived":19,"fork":19,"defaultBranch":20,"hasWiki":21,"hasPages":19,"topics":22,"createdAt":9,"pushedAt":9,"updatedAt":23,"readmeContent":24,"aiSummary":25,"trendingCount":15,"starSnapshotCount":15,"syncStatus":12,"lastSyncTime":26,"discoverSource":27},92656,"PhysisForcing","DAGroup-PKU\u002FPhysisForcing","DAGroup-PKU","PhysisForcing: Physics Reinforced World Simulator for Robotic Manipulation",null,"Python",87,2,4,1,0,3,41.73,"MIT License",false,"main",true,[],"2026-07-22 04:02:06","\u003Cdiv align=\"center\">\n\n\u003C!-- Optional: drop a logo at assets\u002Fphysisforcing_logo.png -->\n\u003C!-- \u003Cimg src=\"assets\u002Fphysisforcing_logo.png\" width=\"220\"\u002F> -->\n\n# PhysisForcing: Physics Reinforced World Simulator for Robotic Manipulation 🔥\n\nPeiwen Zhang\\*, Yufan Deng\\*, Shangkun Sun\\*, Juncheng Ma, Duomin Wang†, Jonas Du,\nZilin Pan, Ye Huang, Hao Liang, Songyan Huang, Ruihua Zhang, Enze Xie†, Ming-Yu Liu, Daquan Zhou†‡\n\n> Peking University · NVIDIA\n>\n> \\*Equal Contribution  †Co-Project Lead  ‡Corresponding Author\n\n\u003Cp align=\"center\">\n  \u003Ca href=\"https:\u002F\u002Fdagroup-pku.github.io\u002FPhysisForcing.github.io\u002F\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FProject-Page-1f8acb\">\u003C\u002Fa>\n  \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.28128\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-Paper-b31b1b\">\u003C\u002Fa>\n  \u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fpapers\u002F2606.28128\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FHugging%20Face-Daily%20Paper-ffcc4d\">\u003C\u002Fa>\n  \u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002FDAGroup-PKU\u002FPF_Cosmos\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F%F0%9F%A4%97%20Weights-PF__Cosmos-ffcc4d\">\u003C\u002Fa>\n  \u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002FDAGroup-PKU\u002FPF_Wan\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F%F0%9F%A4%97%20Weights-PF__Wan-ffcc4d\">\u003C\u002Fa>\n  \u003Ca href=\"LICENSE\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLicense-MIT-3fa34d\">\u003C\u002Fa>\n\u003C\u002Fp>\n\n\u003C\u002Fdiv>\n\n\u003Cp align=\"center\">\u003Cimg src=\"assets\u002Fteaser.png\" width=\"100%\"\u002F>\u003C\u002Fp>\n\n**PhysisForcing** is a **training-time, plug-and-play** framework that makes robotic video generation **physically plausible**. It focuses supervision on interaction-critical regions and aligns generation at two levels — a **pixel-level trajectory** loss and a **semantic-level relational** loss — on an intermediate DiT feature, dropping into existing video backbones (Wan, Cosmos) with **zero extra inference cost**. It ranks **first on R-Bench, PAI-Bench, and EZS-Bench**, and as a world model lifts the **WorldArena-action planner** closed-loop success rate from **16.0% → 24.0%**.\n\n`PF_Cosmos` = Cosmos3-Nano + PhysisForcing  |  `PF_Wan` = Wan + PhysisForcing\n\n\n## 🔥 News\n- `[2026.07.08]` 🎉 We release the **inference code & model weights** for **PF_Cosmos** and **PF_Wan**! Training code will be open-sourced **within this week**.\n- `[2026.06.29]` 🔥 We release the **[arXiv paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.28128)** and **[project page](https:\u002F\u002Fdagroup-pku.github.io\u002FPhysisForcing.github.io\u002F)** of PhysisForcing.\n\n## 📑 Todo List\n\n- [x] Inference code & checkpoints for **PF_Wan** & **PF_Cosmos**\n- [ ] Training code & auxiliary model checkpoints\n\n## 🎥 Demo\n\n**The video below is a compressed preview. Full HD demos and the complete set of side-by-side qualitative comparisons (playable videos across embodiments and tasks) are best viewed on the [Project Page](https:\u002F\u002Fdagroup-pku.github.io\u002FPhysisForcing.github.io\u002F).**\n\n\nhttps:\u002F\u002Fgithub.com\u002Fuser-attachments\u002Fassets\u002Fe6d5fefd-5f90-4610-a38f-112a151f48e4\n\n\n\n## 📈 Results\n\nAll scores are normalized percentages (higher is better, only the **Avg.** column shown). Bold = **PhysisForcing** variants (`PF_Wan14B`\u002F`PF_Wan5B` on Wan, `PF_Cosmos` on Cosmos3-Nano). The first three are embodied video-generation benchmarks; the rightmost is the WorldArena Action Planner world-model evaluation (IDM closed-loop success rate).\n\n\u003Ctable>\n\u003Ctr valign=\"top\">\n\u003Ctd>\n\u003Ctable>\n\u003Ctr>\u003Cth align=\"left\">R-Bench\u003C\u002Fth>\u003Cth>Avg.\u003C\u002Fth>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">Veo 3.1\u003C\u002Ftd>\u003Ctd align=\"center\">56.3\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">Hailuo v2\u003C\u002Ftd>\u003Ctd align=\"center\">56.5\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">Cosmos 3-super\u003C\u002Ftd>\u003Ctd align=\"center\">58.1\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">Seedance 1.5 Pro\u003C\u002Ftd>\u003Ctd align=\"center\">58.4\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">Wan 2.6\u003C\u002Ftd>\u003Ctd align=\"center\">60.7\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">Abot-PhysWorld\u003C\u002Ftd>\u003Ctd align=\"center\">52.9\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">Wan2.2-A14B\u003C\u002Ftd>\u003Ctd align=\"center\">50.7\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">Wan2.2-A14B (ft)\u003C\u002Ftd>\u003Ctd align=\"center\">57.9\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">\u003Cb>PF_Wan14B\u003C\u002Fb>\u003C\u002Ftd>\u003Ctd align=\"center\">\u003Cb>62.0\u003C\u002Fb>\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">Cosmos 3-nano (ft)\u003C\u002Ftd>\u003Ctd align=\"center\">61.5\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">\u003Cb>PF_Cosmos\u003C\u002Fb>\u003C\u002Ftd>\u003Ctd align=\"center\">\u003Cb>63.8\u003C\u002Fb>\u003C\u002Ftd>\u003C\u002Ftr>\n\u003C\u002Ftable>\n\u003C\u002Ftd>\n\u003Ctd>\n\u003Ctable>\n\u003Ctr>\u003Cth align=\"left\">PAI-Bench (robot)\u003C\u002Fth>\u003Cth>Avg.\u003C\u002Fth>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">Wan 2.5\u003C\u002Ftd>\u003Ctd align=\"center\">80.96\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">GigaWorld-0\u003C\u002Ftd>\u003Ctd align=\"center\">80.87\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">Veo 3.1\u003C\u002Ftd>\u003Ctd align=\"center\">80.45\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">WoW-Wan 14B\u003C\u002Ftd>\u003Ctd align=\"center\">79.53\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">Sora v2 Pro\u003C\u002Ftd>\u003Ctd align=\"center\">76.52\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">Abot-PhysWorld\u003C\u002Ftd>\u003Ctd align=\"center\">84.91\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">Wan2.2-A14B\u003C\u002Ftd>\u003Ctd align=\"center\">78.93\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">Wan2.2-A14B (ft)\u003C\u002Ftd>\u003Ctd align=\"center\">79.90\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">\u003Cb>PF_Wan14B\u003C\u002Fb>\u003C\u002Ftd>\u003Ctd align=\"center\">\u003Cb>81.73\u003C\u002Fb>\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">Cosmos 3-nano (ft)\u003C\u002Ftd>\u003Ctd align=\"center\">84.03\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">\u003Cb>PF_Cosmos\u003C\u002Fb>\u003C\u002Ftd>\u003Ctd align=\"center\">\u003Cb>85.17\u003C\u002Fb>\u003C\u002Ftd>\u003C\u002Ftr>\n\u003C\u002Ftable>\n\u003C\u002Ftd>\n\u003Ctd>\n\u003Ctable>\n\u003Ctr>\u003Cth align=\"left\">EZS-Bench\u003C\u002Fth>\u003Cth>Avg.\u003C\u002Fth>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">WoW-Wan 14B\u003C\u002Ftd>\u003Ctd align=\"center\">77.80\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">GigaWorld-0\u003C\u002Ftd>\u003Ctd align=\"center\">75.49\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">Cosmos-Predict 2.5\u003C\u002Ftd>\u003Ctd align=\"center\">73.94\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">UnifoLM-WMA-0\u003C\u002Ftd>\u003Ctd align=\"center\">62.94\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">Kling 2.6-Pro\u003C\u002Ftd>\u003Ctd align=\"center\">79.39\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">Abot-PhysWorld\u003C\u002Ftd>\u003Ctd align=\"center\">80.30\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">Wan2.2-A14B\u003C\u002Ftd>\u003Ctd align=\"center\">77.16\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">Wan2.2-A14B (ft)\u003C\u002Ftd>\u003Ctd align=\"center\">79.04\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">\u003Cb>PF_Wan14B\u003C\u002Fb>\u003C\u002Ftd>\u003Ctd align=\"center\">\u003Cb>80.54\u003C\u002Fb>\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">Cosmos 3-nano (ft)\u003C\u002Ftd>\u003Ctd align=\"center\">80.29\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">\u003Cb>PF_Cosmos\u003C\u002Fb>\u003C\u002Ftd>\u003Ctd align=\"center\">\u003Cb>81.08\u003C\u002Fb>\u003C\u002Ftd>\u003C\u002Ftr>\n\u003C\u002Ftable>\n\u003C\u002Ftd>\n\u003Ctd>\n\u003Ctable>\n\u003Ctr>\u003Cth align=\"left\">Action Planner\u003C\u002Fth>\u003Cth>Avg.\u003C\u002Fth>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">Genie Envisioner\u003C\u002Ftd>\u003Ctd align=\"center\">15.0\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">TesserAct\u003C\u002Ftd>\u003Ctd align=\"center\">18.0\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">RoboMaster\u003C\u002Ftd>\u003Ctd align=\"center\">14.0\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">Vidar\u003C\u002Ftd>\u003Ctd align=\"center\">10.5\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">WoW\u003C\u002Ftd>\u003Ctd align=\"center\">20.5\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">Wan2.2-5B (base)\u003C\u002Ftd>\u003Ctd align=\"center\">16.0\u003C\u002Ftd>\u003C\u002Ftr>\n\u003Ctr>\u003Ctd align=\"left\">\u003Cb>PF_Wan5B\u003C\u002Fb>\u003C\u002Ftd>\u003Ctd align=\"center\">\u003Cb>24.0\u003C\u002Fb>\u003C\u002Ftd>\u003C\u002Ftr>\n\u003C\u002Ftable>\n\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftable>\n\nFull per-metric tables and ablations are in the [paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.28128).\n\n## ⚙️ Usage\n\nThis repo hosts two **self-contained** inference bundles — each ships its own framework\ncode, environment, weights, and example inputs. Pick the one you need and follow its README:\n\n- **[`pf_cosmos\u002F`](pf_cosmos\u002FREADME.md)** — `PF_Cosmos` (Cosmos3-Nano + PhysisForcing), image-to-video.\n- **[`pf_wan\u002F`](pf_wan\u002FREADME.md)** — `PF_Wan` (Wan2.2-A14B + PhysisForcing), image-to-video.\n\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002FDAGroup-PKU\u002FPhysisForcing.git\ncd PhysisForcing\u002Fpf_cosmos   # or: cd PhysisForcing\u002Fpf_wan\n# then follow that bundle's README: environment -> weights -> run\n```\n\nThe two bundles use different PyTorch\u002FCUDA stacks, so set each up in its own environment.\n\n> 🚧 Training code & auxiliary checkpoints: **coming soon.**\n\n## 🙏 Acknowledgement\n\nOur work builds on many excellent projects: [Wan](https:\u002F\u002Fgithub.com\u002FWan-Video\u002FWan2.1), [Cosmos](https:\u002F\u002Fgithub.com\u002Fnvidia-cosmos), [CoTracker3](https:\u002F\u002Fgithub.com\u002Ffacebookresearch\u002Fco-tracker), [Depth-Anything-2](https:\u002F\u002Fgithub.com\u002FDepthAnything\u002FDepth-Anything-V2), [V-JEPA](https:\u002F\u002Fgithub.com\u002Ffacebookresearch\u002Fjepa), and [R-Bench \u002F RoVid-X](https:\u002F\u002Fgithub.com\u002FDAGroup-PKU\u002FReVidgen).\n\n## ✏️ Citation\n\nIf you find PhysisForcing useful, please consider giving a ⭐ and citing:\n\n```bibtex\n@article{zhang2026physisforcing,\n  title={PhysisForcing: Physics Reinforced World Simulator for Robotic Manipulation},\n  author={Zhang, Peiwen and Deng, Yufan and Sun, Shangkun and Ma, Juncheng and\n          Wang, Duomin and Du, Jonas and Pan, Zilin and Huang, Ye and Liang, Hao and\n          Huang, Songyan and Zhang, Ruihua and Xie, Enze and Liu, Ming-Yu and Zhou, Daquan},\n  journal={arXiv preprint arXiv:2606.28128},\n  year={2026}\n}\n```\n\n## License\n\nThis project is released under the [MIT License](LICENSE).\n","PhysisForcing 是一个面向机器人操作的物理增强型世界模拟器框架，旨在提升视频生成模型在机器人任务中的物理合理性。它采用训练时即插即用的设计，在像素轨迹和语义关系两个层次对扩散Transformer（DiT）中间特征施加物理一致性监督，支持无缝集成现有视频主干模型（如Cosmos、Wan），且不增加推理开销。项目已在R-Bench、PAI-Bench等机器人世界建模基准上取得SOTA性能，并显著提升WorldArena等闭环规划系统的成功率。适用于需要高物理保真度的机器人仿真、具身智能训练与世界模型构建场景。","2026-07-10 02:30:03","CREATED_QUERY"]