CADRE: A Cascade Deep Reinforcement Learning Framework for Vision-Based Autonomous Urban Driving

• DOI: 10.1609/aaai.v36i3.20259
• Publication Date: 2022-07-04T09:14:31Z
• Authors: Yinuo Zhao; Kun Wu; Zhiyuan Xu; Zhengping Che; Qi Lu; Jian Tang; Chi Harold Liu
• Keywords: Benchmark (surveying); Margin (machine learning); Obstacle avoidance

Vision-based autonomous urban driving in dense traffic is quite challenging due to the complicated environment and dynamics of behaviors. Widely-applied methods either heavily rely on hand-crafted rules or learn from limited human experience, which makes them hard generalize rare but critical scenarios. In this paper, we present a novel CAscade Deep REinforcement learning framework, CADRE, achieve model-free vision-based driving. derive representative latent features raw observations, first offline train Co-attention Perception Module (CoPM) that leverages co-attention mechanism inter-relationships between visual control information pre-collected dataset. Cascaded by frozen CoPM, then an efficient distributed proximal policy optimization framework online under guidance particularly designed reward functions. We perform comprehensive empirical study with CARLA NoCrash benchmark as well specific obstacle avoidance scenarios tasks. The experimental results justify effectiveness CADRE its superiority over state-of-the-art wide margin.