BAAI's Wang Pengwei on RoboBrain Orca: world models vs VLA for embodied generalization
Wang Pengwei, who leads the embodied model research center at China's BAAI, presented the RoboBrain Orca world model at the 2026 BAAI conference. He argues that today's four families of world models — language-, pixel-, 3D-structure- and visual-representation-centric approaches such as LLMs, video generators, World Labs' Marble and JEPA — each capture only a partial world and remain passively task-driven; Orca instead uses Next State Prediction, reasoning forward and backward about state transitions, and separates "unconscious" video learning from "conscious" event-level state-transition learning. He says Orca's pretraining used no action data, drawing on about 125,000 hours of internet video, roughly 130 million expert-model-labeled event-level state-transition samples and causal and general VQA data, and that BAAI is building a benchmark for state-transition world models because none exists; he expects some approach to break the 0% generalization barrier within a year, though 90%-plus would be needed for real usability.
If world models can crack the generalization bottleneck in embodied brains, they could reshape which technical route robots take from narrow tasks to general ones.