Research map / Video / Video generation
Geometry-aware 4D Video Generation for Robot Manipulation
Zeyi Liu, Shuang Li, Eric Cousineau, Siyuan Feng, Benjamin Burchfiel, Shuran Song
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 |
|---|---|---|---|
| remove multi-view cross-attention in the U-Net decoder (ablation) | OURS · four 4D generation settings | cross-view mIoU: 0.70 -> 0.41; 0.69 -> 0.44; 0.64 -> 0.26; 0.56 -> 0.32 | Hurt |
| remove multi-view cross-attention in the U-Net decoder (ablation) | OURS · four 4D generation settings | AbsRel-m: 0.11 -> 0.31; 0.07 -> 0.33; 0.07 -> 0.49; 0.20 -> 0.34 | Hurt |
| remove multi-view cross-attention in the U-Net decoder (ablation) | OURS · four 4D generation settings | FVD-n: 411.20 -> 497.43; 377.68 -> 451.54; 490.88 -> 597.05; 384.26 -> 694.23 | Hurt |
| add multi-view cross-attention over RGB features to SVD (ablation) | SVD · PutSpatulaOnTable, PlaceAppleFromBowlIntoBin, real world | FVD-n: 370.92 -> 536.02 (task 2); 659.52 -> 812.94 (task 3); 605.10 -> 1151.05 (real world) | Hurt |
| add multi-view cross-attention over RGB features to SVD (ablation) | SVD · StoreCerealBoxUnderShelf | FVD-n: 977.06 -> 941.73 | Improved |
| use generated RGB-D 4D video with pose tracking instead of baseline policies | OURS · three simulation manipulation tasks | task success rate: 0.12 / 0.12 / 0.25 -> 0.64 average | Improved |
| use global point clouds as input (DP3 vs DP) | diffusion policy · PutSpatulaOnTable, PlaceAppleFromBowlIntoBin | task success rate: 0.27 -> 0.27 (task 2); 0.00 -> 0.00 (task 3) | No clear effect |
| use point clouds as input (DP3 vs DP) | diffusion policy · StoreCerealBoxUnderShelf | task success rate: 0.10 -> 0.23 | Improved |
| add a third input view at inference without retraining (ablation) | OURS · StoreCerealBoxUnderShelf | FVD / mIoU: FVD 515.53 -> 566.86 -> 662.83; mIoU 0.62 -> 0.54 | Hurt |
| use OURS instead of SVD in inference cost | OURS · inference latency | inference time: 13.4 s -> 30.0 s | Hurt |
| use 4D Gaussian baseline instead of OURS | 4D Gaussian · novel-view multi-view generation | cross-view mIoU: 0.70 -> 0.39 (task 1); 0.56 -> 0.00 (real) | Hurt |
Closest papers by meaning
- RoMan-4D: Learning Robot Arm Manipulation from 4D World Models (ECCV 2026 · similarity 0.92)
- MVISTA-4D: View-Consistent 4D World Model with Test-Time Action Inference for Robotic Manipulation (ICML 2026 · similarity 0.91)
- Real2Edit2Real: Generating Robotic Demonstrations via a 3D Control Interface (CVPR 2026 · similarity 0.90)
- Choreographing a World of Dynamic Objects (CVPR 2026 · similarity 0.90)
- DynaRend: Learning 3D Dynamics via Masked Future Rendering for Robotic Manipulation (NeurIPS 2025 · similarity 0.89)
- EA-WM: Event-Aware Generative World Model with Structured Kinematic-to-Visual Action Fields (NeurIPS 2026 · similarity 0.88)
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.
