Research map / Video / Video generation
From Imagined Futures to Executable Actions: Mixture of Latent Actions for Robot Manipulation
Yajie Li, Bozhou Zhang, Chun Gu, Zipei Ma, Jiahui Zhang, Jiankang Deng, Xiatian Zhu, Li Zhang
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 MoIDM latent-action interface with flow IDM only (ablation) | MoLA · CALVIN ABC-D | average length: 4.24 -> 4.39 | Improved |
| add depth IDM to flow IDM (ablation) | MoLA · CALVIN ABC-D | average length: 4.39 -> 4.46 | Improved |
| add semantic IDM to flow + depth (ablation) | MoLA · CALVIN ABC-D | average length: 4.46 -> 4.55 | Improved |
| add semantic IDM only (ablation) | MoLA · CALVIN ABC-D | average length: 4.24 -> 4.31 | Improved |
| pretrain MoIDM on large robot datasets vs from scratch (ablation) | MoLA · CALVIN ABC-D and other benchmarks | success rate | Improved |
| fine-tune MoIDM jointly instead of freezing (ablation) | MoLA · CALVIN ABC-D | success rate | Improved |
| use DiT flow-matching action head instead of autoregressive Transformer (ablation) | MoLA · CALVIN ABC-D | success rate | Improved |
| replace SVD with Wan2.2 5B video backbone (ablation) | MoLA · CALVIN ABC-D | success rate | Improved |
| use three specialized IDMs instead of unified alternatives (single RGB IDM, shared multimodal IDM, direct multimodal input) (ablation) | MoLA · CALVIN ABC-D | success rate | Improved |
| feed identical or noisy current frames instead of imagined future frames at inference (ablation) | MoLA · CALVIN ABC-D | success rate | Hurt |
| use MoLA latent actions instead of OpenVLA-OFT+ | VLA policy · LIBERO-Plus | average success rate: 79.5 -> 92.7 | Improved |
| use MoLA instead of pi0.5 in real world | real-world manipulation · Real-world UR5e | average success: 77.5% -> 73.0% | Hurt |
Closest papers by meaning
- VideoVLA: Video Generators Can Be Generalizable Robot Manipulators (NeurIPS 2025 · similarity 0.91)
- DreamVLA: A Vision-Language-Action Model Dreamed with Comprehensive World Knowledge (NeurIPS 2025 · similarity 0.91)
- Predictive Inverse Dynamics Models are Scalable Learners for Robotic Manipulation (ICLR 2025 · similarity 0.89)
- FLIP: Flow-Centric Generative Planning as General-Purpose Manipulation World Model (ICLR 2025 · similarity 0.89)
- EA-WM: Event-Aware Generative World Model with Structured Kinematic-to-Visual Action Fields (NeurIPS 2026 · similarity 0.89)
- Generate Subgoal Images before Act: Unlocking the Chain-of-Thought Reasoning in Diffusion Model for Robot Manipulation with Multimodal Prompts (CVPR 2024 · 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.
