Motus
- hybrid
- 2025
- WAM + VLA
Model cardPaper ↗
Motus is a Mixture-of-Transformers model with understanding, video-generation and action experts, switched between VLA, world-model, inverse-dynamics and joint video-action modes.
Key facts
Rows below are the register record in lib/models.ts. A dash means the value is not confirmed from a primary source. Classification follows Reuss, NVIDIA Technical Blog (15 June 2026) and the model's own paper.
| Name | Motus |
|---|---|
| Lab | — |
| Year | 2025 |
| Paradigm | hybrid |
| Backbone | — |
| Source | https://arxiv.org/abs/2512.13030 |
What the paper claims
- Motus connects a vision-language understanding expert, a video-generation expert and an action expert through shared attention (tri-model joint attention), and uses a UniDiffuser-style scheduler so one set of weights can run as a world model, a VLA, an inverse-dynamics model, a video generator, or a joint video-action predictor. Bi et al., 2025
- It learns latent actions from optical flow as a pixel-level “delta action”, so the action expert can be pretrained on unlabeled videos as well as robot trajectories. Bi et al., 2025
- Training follows a three-phase pipeline (video pretraining, latent-action pretraining, embodiment-specific action fine-tuning) on a six-layer data pyramid from web video through target-robot data. Bi et al., 2025
- Reuss cites Motus, with BagelVLA, as an early hybrid that trains understanding, video generation and an action expert together rather than choosing one backbone family. Reuss, NVIDIA Technical Blog (15 June 2026)
What it evaluates
- The authors report a 15% improvement over X-VLA and a 45% improvement over π0.5 in simulation, and a 11–48% improvement in real-world scenarios, in the comparisons they publish. Bi et al., 2025
On this register
Hybrid. Mixes the vision-language-action (VLA) and world action model recipes. That axis is defined in what is a world action model? and in Reuss, NVIDIA Technical Blog (15 June 2026).
Other hybrid models on the register: BagelVLA, Cortex 2.0.
The backbone field is unconfirmed, so this card does not claim a video-model family.