Research map / Video / Video generation
S-VAM: Shortcut Video-Action Model by Self-Distilling Geometric and Semantic Foresight
Haodong Yan, Zhide Zhong, Jiaguan Zhu, Junjie He, Weilin Yuan, Wenxuan Song, Xin Gong, Yingjie CAI, Guanyi Zhao, Xu Yan, Liu Bingbing, Yingcong Chen and 1 more
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 |
|---|---|---|---|
| add geometric distillation (ablation) | S-VAM · CALVIN ABC->D | average sequence length: 4.16 -> 4.01 | Improved |
| add semantic distillation (ablation) | S-VAM · CALVIN ABC->D | average sequence length: 4.16 -> 3.99 | Improved |
| add semantic distillation (5th task success) (ablation) | S-VAM · CALVIN ABC->D | 5th task success rate: 68.9% -> 64.1% | Improved |
| use self-distilled targets from own multi-step video instead of ground-truth future frame VFM targets (ablation) | S-VAM · CALVIN ABC->D | average sequence length: 4.16 -> 3.82 | Improved |
| add Uni-Perceiver (ablation) | S-VAM · CALVIN ABC->D | average sequence length: 4.16 -> 3.72 | Improved |
| keep original diffusion feature in condition (ablation) | S-VAM · CALVIN ABC->D | average sequence length: 4.16 -> 3.93 | Improved |
| use dense DINOv2 instead of global CLIP as semantic target (ablation) | S-VAM · CALVIN ABC->D (VFM teacher targets) | average sequence length: 4.01 vs. 3.72 | Improved |
| use DPAv3 instead of VGGT as geometric target (ablation) | S-VAM · CALVIN ABC->D (VFM teacher targets) | average sequence length: 3.99 vs. 3.74 | Improved |
| use video foundation model targets (VideoMAEv2, V-JEPA2) instead of DINOv2 (ablation) | S-VAM · CALVIN ABC->D (VFM teacher targets) | average sequence length: 3.90 and 3.74 vs. 4.01 | Hurt |
| use DPAv3 rather than VGGT in the combined target (ablation) | S-VAM · CALVIN ABC->D (VFM teacher targets) | average sequence length: 4.16 -> 4.04 | Improved |
| use DINOv2 rather than SigLIP in the combined target (ablation) | S-VAM · CALVIN ABC->D (VFM teacher targets) | average sequence length: 4.16 -> 4.06 | Improved |
| replace one-step features (VPP) with S-VAM geometric and semantic foresight | VPP · MetaWorld hard tasks | success rate: 52.6% -> 68.4% | Improved |
Closest papers by meaning
- DiT4DiT: Jointly Modeling Video Dynamics and Actions for Generalizable Robot Control (NeurIPS 2026 · similarity 0.88)
- VidMan: Exploiting Implicit Dynamics from Video Diffusion Model for Effective Robot Manipulation (NeurIPS 2024 · similarity 0.88)
- EA-WM: Event-Aware Generative World Model with Structured Kinematic-to-Visual Action Fields (NeurIPS 2026 · similarity 0.87)
- SAM-E: Leveraging Visual Foundation Model with Sequence Imitation for Embodied Manipulation (ICML 2024 · similarity 0.87)
- VideoVLA: Video Generators Can Be Generalizable Robot Manipulators (NeurIPS 2025 · similarity 0.87)
- Video Prediction Policy: A Generalist Robot Policy with Predictive Visual Representations (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.
