Fast-WAM revolutionizes the world model, improving robot reaction speed.

The StarMap team's new world model, Fast-WAM, significantly improves robot reaction speed by changing training and inference methods. While maintaining dynamic modeling of the world during training, Fast-WAM eliminates redundant prediction branches during inference, directly extracting the "world representation" containing implicit physical laws. This innovation reduces Fast-WAM's inference latency to approximately 190 milliseconds in multiple experimental tasks, achieving about a four-fold speed improvement compared to traditional WAM.