Orca models the world, not actions, to match top robotics bots
baai's orca matches specialized robotics systems across five tasks without using action labels. it tackles the chronic data shortage in robotics training.
baai introduced orca, a 4b parameter model that learns an abstract internal representation of world dynamics from 125,000 hours of video, rather than predicting specific tokens or actions. unlike conventional models, orca uses a frozen core with swappable 'heads' for text, images, and robot control to handle diverse outputs. it matches specialized robotics performance on five tasks despite training on only one-tenth of the video corpus and using zero action labels. this architecture signals a shift toward general world models that alleviate robotics' chronic data scarcity by learning physics and causality from raw video.