UW-trained robots complete dexterous tasks with zero real-world training data

UW-trained robot policies completed gear meshing, nut threading, and peg insertion without any real-world training data. They performed the tasks zero-shot from RGB input, and the same simulation-based recipe transferred to locomotion across difficult terrain.

That moves the bottleneck from collecting demonstrations to making simulations match reality and teaching reliable recovery. Simple, textureless parts transferred well, but a reflective board needed tape to reproduce its appearance, and recovery became unreliable when the board was not bolted down.