PROXIMAL POLICY OPTIMIZATION FOR COLLISION AVOIDANCE AND MOTION PLANNING IN AUTONOMOUS VEHICLES: A MATHEMATICAL MODELING PERSPECTIVE

Motion planning and collision avoidance (MPCA) are critical in autonomous driving systems (ADS). Current deep reinforcement learning (DRL) ADS utilize discrete action space algorithms for MPCA, leading to inefficient and inaccurate MPCA in complex driving environments. This paper focuses on advanced mathematical modeling techniques in the context of autonomous vehicle (AV) dynamics and proposes an end-to-end proximal policy optimization (PPO)-based DRL framework for MPCA in complex environments. The proposed framework comprises three main components. First, a 3D virtual environment and a model car are designed, equipped with multimodality sensors including Cameras, LIDAR, and GPS. Second, an immediate reward system is designed with varied control bonds and a convolutional neural network (CNN)-based model for feature extraction and curve learning. The model is optimized using PPO in a Markov decision style of direct perception and driving. Finally, the network is embedded in the 3D model car for a real test drive. The framework is evaluated through various experiments, including a detailed ablation of driving in real scenarios. The results demonstrate that our framework allows AVs to perform efficient and accurate MPCA in a complex virtual environment, and it can generalize to deploy in a physical environment.