Domain Adaptation In Reinforcement Learning Via Latent Unified State Representation

Despite the recent success of deep reinforcement learning (RL), domain adaptation remains an open problem. Although generalization ability RL agents is critical for real-world applicability Deep RL, zero-shot policy transfer still a challenging problem since even minor visual changes could make trained agent completely fail in new task. To address this issue, we propose two-stage that first learns latent unified state representation (LUSR) which consistent across multiple domains stage, and then do training one source based on LUSR second stage. The cross-domain consistency allows acquired from to generalize other target without extra training. We demonstrate our approach variants CarRacing games with customized manipulations, verify it CARLA, autonomous driving simulator more complex realistic observations. Our results show can achieve state-of-the-art performance related tasks outperforms prior approaches latent-representation image-to-image translation.