Glossary
Sim-to-real
The practice of training a robot policy in simulation and deploying it on a real robot.
Sim-to-real (sometimes "sim2real") is the dominant training paradigm for modern robots: train in a fast, cheap, parallel simulator, then move the policy to a real robot. The gap between simulation physics and the real world is what makes this hard.
The main techniques for closing the gap are domain randomization (randomize sim parameters so the policy is robust to many possible realities), system identification (measure the real robot and make sim match), and real-world fine-tuning (collect a small amount of real-world data after sim training).
As of 2026, sim-to-real is mostly solved for legged locomotion, partially solved for coarse manipulation, and still an open problem for contact-rich tasks.