Can one model handle robot vision, language, and control together?

A new foundation model class (WLA) unifies world modeling from egocentric video, language reasoning, and robot action prediction in a single autoregressive Transformer. It predicts text subtasks, subgoal images, and control actions jointly—using world prediction to guide action generation without slowing inference. On RoboTwin2.0, it reaches 92.94% success; on real-world RMBench, 56.5%. The 2B-parameter model runs at 40 ms per step and can learn from unlabeled cross-embodiment robot video.