Transfer of Robot Perception Module With Adversarial Learning

Abstract

It is a recent trend in the field of robotic control to collect large amount of data from simulated environments and then use it to train deep neural networks. However, there are many essential differences between simulated data and data from the real world. Thus, models trained naively on simulated data often fail to generalize to reality. To address that problem, we propose two approaches to transferring robot perception module, which are based on domain adversarial neural networks (DANN) and generative adversarial networks (GAN), respectively. The former approach tries to extract domain-invariant features by a shared feature extractor and use the domain-invariant features to train a transferrable target localization model (TLM). Meanwhile, the latter approach attempts to learn a transformation from the source domain to the target domain and use the transformation to generate realistic synthetic samples. Then, the synthetic samples are exploited as training data for the TLM. The experiments show that given enough simulated data and only a small amount of real world data, the TLM adapted by our methods could generalize well to real-world environments without drastic performance decline.

Keywords

• Robotic grasping
• skill transfer
• domain adversarial neural networks (DANN)
• generative adversarial networks (GAN)