Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World

ABSTRACT

Bridging the 'reality gap' between simulated environments and the physical world is a major challenge in robotics. We introduce domain randomization, a simple yet powerful technique for training neural networks in simulated environments that transfer to the real world zero-shot. We train an object detector in a simulator that heavily randomizes rendering parameters, including lighting, object textures, background images, and camera positions. We hypothesize that with enough variability in the simulator, the real world may appear to the network as just another variation. We demonstrate this by successfully training a robotic arm to locate and grasp physical objects using a model trained entirely on heavily distorted, low-fidelity synthetic images.