The RoboScience research team has unveiled WorldFlow, the world's first natively 3D dynamics prediction world model, tackling the disconnect between visual representations and 3D motion. The model treats 3D motion as an independent generative modality on par with video, and leverages a complementary encoding mechanism and MoT architecture to simultaneously predict future frames and the 3D motion of objects. The team assembled 1.44 million paired samples from multi-source scenes for pretraining and built a dedicated benchmark to validate performance. Experiments show that its prediction accuracy significantly outperforms both single motion prediction and the two-stage approach of first generating video and then extracting trajectories. In real-robot tests, the probability of the robot generating correct trajectories rose to 97.2%, and the operation success rate climbed to 66.7%. By using object 3D motion trajectories as a universal interface, the model opens a new path for embodied intelligence.
