BAAI Wujie·RoboBrain Orca Team technical report. Orca is a general world foundation model using Next-State-Prediction rather than next-token/frame/action prediction.
Pre-training: 125K hours video, 160M event annotations. Two paradigms: unconscious learning (continuous video) and conscious learning (language-described events, VQA). Backbone: frozen Qwen3.5 VLM. Downstream readouts: text, image (Stable Diffusion 3.5 adapters), embodied action (Action Expert).
Robotics: matches π0.5 across five manipulation tasks with only 200 real-world recordings per task for Action Expert; base never saw action labels during pre-training.