Research map / Vision-language and multimodal / Robot action and navigation
Moto: Latent Motion Token as the Bridging Language for Learning Robot Manipulation from Videos
Yi Chen, Yuying Ge, Weiliang Tang, Yizhuo Li, Yixiao Ge, Mingyu Ding, Ying Shan, Xihui Liu
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 latent motion token pre-training on video (ablation) | Moto-GPT · CALVIN (ABC->D) | average length: 2.14 -> 3.10 | Improved |
| add latent motion token pre-training on video (ablation) | Moto-GPT · SIMPLER overall | success rate: 0.480 -> 0.614 | Improved |
| add latent motion token pre-training on video (ablation) | Moto-GPT · Real-world tasks (avg) | success rate (%): 23.33% -> 60% | Improved |
| add motion token pre-training with only 1% of labeled data (ablation) | Moto-GPT · CALVIN (ABC->D) data efficiency | success rate (%): 0% -> 52.5% | Improved |
| add motion token pre-training under visual distractor and novel object (ablation) | Moto-GPT · Real-world generalization | success rate: +20% and +30% | Improved |
| add human video (SSV2) to OXE pre-training | Moto-GPT · SIMPLER Move Near | success rate | Improved |
| keep next-motion-token loss and motion tokens during fine-tuning (co-fine-tuning) (ablation) | Moto-GPT · CALVIN (ABC->D) | success rate | Improved |
| use latent motion tokens instead of frames as video representation for classification | token classifier · CALVIN video classification | semantic accuracy: 0.828 -> 0.797 | No clear effect |
| add 7 latent motion token chunks to initial frame | video classifier · CALVIN video classification | semantic accuracy: 0.292 -> 0.797 | Improved |
| compare Moto (static RGB only) to GR-1 (extra gripper RGB and proprio) | Moto · CALVIN (ABC->D) | average length: 3.06 -> 3.10 | No clear effect |
Later papers that compared with this method
2 later accepted papers report Moto as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- LatentVLA: Taming Latent Space for Generalizable and Long-Horizon Bimanual Manipulation (AAAI 2026 · better on 2 of 2 shared results)
- World Guidance: World Modeling in Condition Space for Action Generation (ICML 2026 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Robots Pre-train Robots: Manipulation-Centric Robotic Representation from Large-Scale Robot Datasets (ICLR 2025 · similarity 0.91)
- Latent Action Pretraining from Videos (ICLR 2025 · similarity 0.91)
- Vision-Language-Action Pretraining from Large-Scale Human Videos (ICML 2026 · similarity 0.91)
- Phoenix: A Motion-based Self-Reflection Framework for Fine-grained Robotic Action Correction (CVPR 2025 · similarity 0.90)
- Unleashing Large-Scale Video Generative Pre-training for Visual Robot Manipulation (ICLR 2024 · similarity 0.89)
- VidMan: Exploiting Implicit Dynamics from Video Diffusion Model for Effective Robot Manipulation (NeurIPS 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.
