Quickly Adaptive Automated Vehicle’s Highway Merging Policy Synthesized by Meta Reinforcement Learning with Latent Context Imagination

Under a wide range of traffic cultures and driving conditions, it is essential that an automated vehicle performs highway merging with appropriate driving styles - driving safely and efficiently without annoying or endangering other road users. Despite the extensive exploration of Meta Reinforcement Learning (Meta RL) for quick adaptation to different environments and its application to automated vehicle driving policies, most state-of-the-art algorithms require a dense coverage of the task distribution and extensive data for each of the meta-training tasks, which is extremely expensive for the automotive industry. Our paper proposes IAMRL, a context-based Meta RL algorithm in which meta-imagination reduces real-world training tasks and data requirements. By interpolating the learned latent context space with disentangled properties, we perform meta-imagination. As a result of our autonomous highway merging experiments, IAMRL outperforms existing approaches in terms of generalization and data efficiency.