Release of Motus 2: A Self-Evolving World Model for Dexterous Robotic Manipulation
Motus 2 has been released as a novel, self-evolving general world model engineered specifically for dexterous robotic manipulation. It unifies joint video-and-action modeling to establish a continuous decision-and-learning loop for autonomous robotic policy self-improvement. Achieving human-like dexterous manipulation remains one of the greatest technical challenges in embodied AI and robotics. Motus 2 connects policy generation with environmental simulation, allowing robots to autonomously evaluate and refine complex hand-object interactions without relying solely on real-world data collection. Motus 2 employs UniDiffuser-style joint video-action modeling, functioning both as a policy that proposes executable actions and an action-conditioned world model that predicts visual consequences. This dual capability allows the system to run forward and counterfactual rollouts to continuously train and optimize its decision-making capabilities.
## BACKGROUND
In embodied AI, world models function as internal simulators of physical dynamics, helping artificial agents perceive environments, forecast outcomes, and plan complex tasks. Dexterous manipulation refers to human-like precision and object handling—such as dynamic re-grasping, rotation, and tool usage—which traditionally requires vast amounts of real-world training data.