Simulation to reality and back: A robot's guide to crossing the reality gap

Abstract

Simulation is an indispensable technology within robotics; however, the reality gap prevents many simulated solutions from transferring perfectly to reality. This thesis investigates the reality gap within the context of robotic manipulation. We present studies that first quantify and then benchmark the reality gap when comparing popular robotic simulators to a real-world ground truth collected using motion capture. We then present a promising new method for overcoming the reality gap that employs an online sim-to-real approach that utilises differentiable physics to iteratively narrow the gap and improve the simulation environment using data collected from the real system.