Research map / Video / Video generation
Learning 4D Embodied World Models
Haoyu Zhen, Qiao Sun, Hongxin Zhang, Junyan Li, Siyuan Zhou, Yilun Du, Chuang Gan
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 |
|---|---|---|---|
| predict depth and normal jointly with RGB instead of post-estimating them from generated video | CogVideoX · 4D scene generation (real) | Chamfer L1: 0.2191 -> 0.2030 | Improved |
| predict depth and normal jointly with RGB instead of post-estimating them | CogVideoX · 4D scene generation (synthetic) | Chamfer L1: 0.2884 -> 0.0811 | Improved |
| predict depth and normal jointly with RGB | CogVideoX · 4D scene generation (real) | depth AbsRel: 26.17 -> 22.07 | Improved |
| predict depth and normal jointly with RGB | CogVideoX · 4D scene generation (real) | FVD: 20.64 -> 21.59 | Hurt |
| predict depth and normal jointly with RGB | CogVideoX · 4D scene generation (synthetic) | normal mean angular error: 20.36 -> 14.75 | Improved |
| use 4D video world model over 4D point cloud diffusion (4D Point-E) | TesserAct · 4D scene generation (real) | Chamfer L1: 0.2211 -> 0.2030 | Improved |
| use 4D point clouds with inverse dynamics instead of 2D inverse dynamics (UniPi*) | 4DWM · RLBench close box | success rate: 81 -> 88 | Improved |
| use 4D point clouds with inverse dynamics instead of 2D inverse dynamics (UniPi*) | 4DWM · RLBench open drawer | success rate: 67 -> 80 | Improved |
| use 4D point clouds instead of 2D inverse dynamics (UniPi*) | 4DWM · RLBench open microwave | success rate: 72 -> 70 | Hurt |
| use 4D point clouds instead of 2D inverse dynamics (UniPi*) | 4DWM · RLBench weighing off | success rate: 68 -> 62 | Hurt |
| use 4D point clouds instead of 2D inverse dynamics (UniPi*) | 4DWM · RLBench sweep to dustpan | success rate: 49 -> 56 | Improved |
| use RGB-DN 4D prediction instead of Gaussian-splatting reconstruction (Shape of Motion) for novel views | TesserAct · RLBench novel view synthesis | PSNR: 10.94 -> 12.99 | Improved |
Closest papers by meaning
- Learning 3D Persistent Embodied World Models (NeurIPS 2025 · similarity 0.92)
- RoboScape: Physics-informed Embodied World Model (NeurIPS 2025 · similarity 0.90)
- DynamicVerse: A Physically-Aware Multimodal Framework for 4D World Modeling (NeurIPS 2025 · similarity 0.90)
- Learning Visual Feature-Based World Models via Residual Latent Action (NeurIPS 2026 · similarity 0.90)
- Empowering World Models with Reflection for Embodied Video Prediction (ICML 2025 · similarity 0.89)
- EnerVerse: Envisioning Embodied Future Space for Robotics Manipulation (NeurIPS 2025 · similarity 0.89)
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.
