Cross-domain imitation learning with feature synthesis

Abstract

This study suggested a cross-domain imitation learning method where the agent in one domain follows the behavior of an expert in another domain without accessing any reward function. One common method is to extract domain-independent feature vectors from inputs and determine how well the agent is trained. However, due to the difference between two domains, the behavior of the agent cannot be compared appropriately with that of the expert regardless of its degree of training. To overcome this problem, we extracted two feature vectors from each input

one contained only domain information and the other contained only policy expertness information. The expert data in target-domain can be produced by combining the target-domain feature vector and the expert-policy feature vector. The proposed method showed improved performance in several environments with domain discrepancies. From this study, we can expect to alleviate the difficulties of collecting demonstrations and increase the efficiency of training agents using imitation learning.