Domain Randomization for Sim2Real Transfer

A robot trained in one simulator can fail when real lighting, friction or object weight differs. Domain randomization changes those conditions during training, giving the robot practice across many simulated settings. Lilian Weng's 2019 survey explains the intended benefit: real conditions may fall within the variation the robot has already encountered.

Choosing that variation is part of the engineering problem. Ranges that are too broad can include unrealistic or infeasible situations and make learning harder. Guided randomization uses feedback from task performance or real observations to shape the training distribution, concentrating effort on variation that can help bridge the simulator-to-reality gap.

“Due to the sample inefficiency of deep RL algorithms and the cost of data collection on real robots, we often need to train models in a simulator which theoretically provides an infinite amount of data.”

“The gap is triggered by an inconsistency between physical parameters (i.e. friction, kp, damping, mass, density) and, more fatally, the incorrect physical modeling (i.e. collision between soft surfaces).”