CARLA Autonomous Driving Using Deep Reinforcement Learning

Engineering notes

This work employs state-of-the-art deep reinforcement learning (DRL) techniques for autonomous driving agent training. Using the open-source simulator Carla, we conduct experiments and construct a realistic urban simulation environment for training the model.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Artificial intelligence (AI) has made remarkable strides in many technological domains, and one of the most prominent areas of growth in AI is self-driving cars. Recent interest in deep learning, particularly in visual-based subsystems, has been very beneficial to this field. This work employs state-of-the-art deep reinforcement learning (DRL) techniques for autonomous driving agent training. Using the open-source simulator Carla, we conduct experiments and construct a realistic urban simulation environment for training the model. Because real-world legal frameworks cannot be applied due to the risks and ethical concerns involved, simulators are essential for testing. Many driving policies are now developed by hand, which often results in less-than-ideal solutions that are costly to develop, scale, and maintain. DRLs have shown promising results in learning driving behaviors. This work means the application of a DRL algorithm known as Proximal Policy Optimization (PPO) in a simulated driving environment, with a focus on route navigation. The primary objective is to investigate the training of agents in continuous state and action spaces using the DRL model. Our primary contribution is the development of PPO-based agents capable of reliably navigating Carla-based environments. We also employed a Variational Auto Encoder (VAE) to compress high-dimensional observations to accelerate learning in a low-dimensional latent space. The goal of this project is to create a completely autonomous driving system that steers the vehicle, ensures safe navigation, and lowers the likelihood of collisions. This work compiles our research, analysis, and discussions to address this complex issue.

Links and sources

• Official / arXiv page

Comments