Sim-to-real

The practice of developing or training robot control policies in physics simulation and then transferring them to physical hardware. Because simulators imperfectly model friction, contact dynamics, sensor noise, and latency, transferred policies face a reality gap; techniques such as domain randomization and system identification are used to narrow it. The approach is central to learning-based control for legged robots and manipulation, where collecting equivalent training data on hardware would be slow or damaging.