There is insufficient compute to operate these things locally in dynamic environments. The models for doing that kind of robotics inference are running on racks of H200’s.