Evaluation of Domain Randomization Techniques for Transfer Learning

摘要

To address the challenge of resource-intensive data collection from real robotic environments, many deep learning applications use synthetic data to train their networks. This creates new problems when transferring the obtained knowledge from the simulated to the real world domain. Various aspects of the simulation, which do not influence the learning objective, can be randomized to enhance generalization to new domains. In this paper, we analyze the effect of these domain randomization techniques. To get an insight into their benefits, we apply them while training a grasp success classifier based on state-of-the-art CNN for an industrial robot as a showcase. We generated a large synthetic data set containing 1.44M RGB images with 48 permutations of 6 different randomizations and a base scenario as training data. The resulting networks, each trained on a different subset of this data set, are evaluated on 3k real world images of the robot performing grasps. We observed the effectiveness of randomization of perspective, distractors, lighting and the grasped box. Notably, we show that pretrained networks benefit from these techniques in particular.