Issue A persistent simulation-to-reality gap limits what trained robots can do in the world

Summary

Deloitte lists training and learning as one of the barriers to scaling physical AI. Simulation offers speed, safety and scale, but approximate physics models leave a lasting gap between how robots perform virtually and how they perform in reality. Roboticist Ayanna Howard of The Ohio State University explains that simulated visuals are good, yet real settings contain nuances that look different, so a grasp learned virtually does not carry over exactly. She adds that robots adapt around specific tasks, such as gripping balls on surfaces with different friction, rather than learning holistic interaction with their environment, including how close to stand to people. Practising endlessly in the real world is not an option because things get broken. The report expects better physics engines, synthetic data and blended virtual and real training to help narrow the gap.