Research map / 3D vision / 3D generation
Structured 4D Latent Predictive Model for Robot Planning
Zhiyi Li, Peilin Wu, Xiaoshen Han, Ruojin Cai, Yilun Du
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 |
|---|---|---|---|
| use decoded point clouds instead of voxels as inverse-dynamics input (ablation) | Structured 4D latent model · StackCube, 4 views | success rate: closed-loop 73% -> 84%; open-loop 57% -> 57% | Improved |
| use decoded point clouds instead of 3D latents as inverse-dynamics input (ablation) | Structured 4D latent model · StackCube, 4 views | success rate: closed-loop 80% -> 84%; open-loop 66% -> 57% | No clear effect |
| use closed-loop replanning instead of open-loop with 4 training views (ablation) | Structured 4D latent model · StackCube | success rate: 57% -> 84% | Improved |
| increase training camera views from 4 to 40 (ablation) | Structured 4D latent model · StackCube | Chamfer distance: 7.10 -> 6.81 | Improved |
| increase training camera views from 4 to 40 (ablation) | Structured 4D latent model · StackCube | planning success rate (open-loop): 57% -> 84% | Improved |
| increase training views from 4 to 40 (ablation) | Structured 4D latent model · StackCube | cPSNR: 28.89 -> 26.75 | Hurt |
| replace video generative planners with 4D latent predictive model | Structured 4D latent model · ManiSkill3 average | success rate: 24.7% -> 61.3% | Improved |
| replace TesserAct with structured 4D latents | Structured 4D latent model · RLBench average | success rate: 61.7% -> 75.3% | Improved |
| use 3D predictive rollout instead of DP3 under visual perturbation (noise) | Structured 4D latent model · StackCube zero-shot | success rate: 47% -> 80% | Improved |
| use predictive rollout instead of 3D Diffuser Actor at noise 0.10 | Ours-CL · StackCube zero-shot | success rate: 7% -> 86% | Improved |
| use 4D latent model instead of Wan-2.1 for generation | Structured 4D latent model · 4D generation | Chamfer distance: 43.09 -> 5.95 | Improved |
| use 4D latent model instead of TesserAct for image quality | Structured 4D latent model · 4D generation | SSIM: 0.86 -> 0.79 | Hurt |
Closest papers by meaning
- Geometry-aware 4D Video Generation for Robot Manipulation (ICLR 2026 · similarity 0.90)
- PlaySlot: Learning Inverse Latent Dynamics for Controllable Object-Centric Video Prediction and Planning (ICML 2025 · similarity 0.88)
- DynaRend: Learning 3D Dynamics via Masked Future Rendering for Robotic Manipulation (NeurIPS 2025 · similarity 0.88)
- Pre-training Auto-regressive Robotic Models with 4D Representations (ICML 2025 · similarity 0.87)
- Predictive Inverse Dynamics Models are Scalable Learners for Robotic Manipulation (ICLR 2025 · similarity 0.87)
- Efficient Robotic Policy Learning via Latent Space Backward Planning (ICML 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.
