Deep Reinforcement Learning based Planning for Urban Self-driving with Demonstration and Depth Completion

Research shows major interests in urban self-driving in recent years, both perception and motion planning considered to be significant topics. Current techniques of decision making for driving policy are modular and hand designed, which is expensive and inefficient. With the development of machine learning, learning-based approaches have become a mainstream research direction. However, the performance in urban driving scenarios is far from satisfaction due to the brittle convergence property of deep reinforcement learning and debased observation. To solve these problems, this paper proposed a learning-based method with deep reinforcement learning (DRL) and imitation learning (IL), and additionally a novel depth completion model for better perception. Our framework is built upon Soft Actor-Critic algorithm and introducing an update method that value function, Q-function and policy network all learn from the expert data. To tackle the observation problem, we proposed a reconstruction restraint deep fusion depth completion network which can predict the integrated and precise depth map of the environment with our own novel pre-processed datasets. In experiment, our autonomous driving agent transfer smooth from IL to DRL in training, and outperformed state-of-art methods in urban challenging scenes and still competing compared to our model with groundtruth input.