Egocentric video shows what a person sees from their own perspective. Current video models can imagine such videos but do not understand the person's exact 3D body movements, like hand or upper-body motions.
EgoGeneralist takes a pretrained video generator and adds a special renderer that turns action data into video-like representations. This allows the model to use motion as either a control signal to generate video or as something to predict from video, without needing extra complex encoders.
The framework works on five different tasks using only lab-recorded training data, yet it generalizes to new scenes and interactions. It matches or beats task-specific models, bridging the gap between visual knowledge and precise 3D human motion.