Motus: A Unified Latent Action World Model
What it is
Motus is a unified world model for robotics that integrates video generation, vision-language understanding, and action prediction into a single framework. It uses a Mixture-of-Transformers (MoT) architecture and a UniDiffuser-style scheduler to model latent actions and joint video-action predictions.
The problem it solves
Decoupled models for video generation and robot action often struggle to share information effectively; Motus provides a unified latent action space to bridge this gap.
What you could use it for
- Robotic manipulation and motion planning
- Video generation and action joint prediction
- Fine-tuning for specific robotic tasks (e.g., RoboTwin2)
- Large-scale action pretraining using a data pyramid
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.