Motubrain: A World Action Model

Built on the proprietary UniDiffuser framework and a three-stream MoT architecture, Motubrain unifies perception, planning, and execution within a single model. It integrates world prediction and action generation into a unified framework, supporting tasks such as VLA, world modeling, video generation, inverse dynamics modeling, and video-to-action prediction. With strong cross-embodiment adaptation, cross-task generalization, and long-horizon execution capabilities, Motubrain enables robots to develop human-like physical intuition and reliably accomplish complex end-to-end tasks across real-world environments, including homes, industrial facilities, and commercial settings.

Real-World Task Demonstrations

Motubrain can execute long-horizon task sequences end-to-end, involving more than ten atomic actions while maintaining stable subtask transitions and temporal consistency. The physical action dynamics it models exhibit strong generalization and cross-embodiment transferability, enabling unified generalization across diverse robotic embodiments.

One Brain. Many Skills.

One Brain. Any Robot.

One Brain. Directly Long-Horizon.

One Brain. Foresight.

World Action Model Motubrain: A New Paradigm for Multi-Task Generalization and Scalable Evolution in Embodied Intelligence

Motubrain is built on the UniDiffuser unified modeling framework. Through Cross-Modal Prior Fusion, it integrates vision-language knowledge (VLM), video dynamics knowledge (Video Generation Models), and action skill knowledge (Action Experts) within a single model, enabling unified representation and generation across language, video, and actions.

Number of Tasks
Number of Episodes

Top-Ranked on Two Leading International Benchmarks

Motubrain ranked first on two leading international benchmarks, demonstrating strong performance in the unified modeling of perception, planning, and action.

RoboTwin 2.0 Leaderboard

On RoboTwin 2.0, Motubrain achieved scores of 95.8 in the Clean setting and 96.1 in the Randomized setting, ranking #1 in both. It is the only model on the leaderboard to achieve an average score above 95 in randomized environments and attains perfect or near-perfect performance across most individual tasks.

WorldArena Leaderboard

On WorldArena, Motubrain ranks #1 with an overall EWM score of 63.77 and achieves leading results across key motion metrics, including Motion Quality, Flow Score, and Motion Smoothness, consistently setting the benchmark for performance.