A Tübingen startup wants robots to learn consequences from video

Language models benefit from abundant digital text. Robots need information about how actions change physical situations, and collecting that experience can be expensive or risky.

Ontic Labs proposes using video and experience to train a world model that can predict consequences and support learning across tasks. The attraction is broader transfer: knowledge about motion and interaction might be reusable instead of requiring a new demonstration set for every narrow behavior.

The account comes from the founders’ research institution and describes an ambition under development. It does not establish a generally capable household robot or prove that video alone supplies all the needed causal knowledge. Its relevance is the proposed training strategy and the attempt to turn research into a German robotics company.