Application of Reinforcement Learning in the Autonomous Driving Platform of the DeepRacer
Application of Reinforcement Learning in the Autonomous Driving Platform of the DeepRacer
复制标题
强化学习在DeepRacer自动驾驶平台中的应用
DOI：
10.23919/ccc55666.2022.9902325
复制
发表时间：
2022
期刊：
影响因子：
--
通讯作者：
Huaming Yan
中科院分区：
文献类型：
--
作者：
Wenjie Zhu;Haikuo Du;Moyan Zhu;Yanbo Liu;Chaoting Lin;Shaobo Wang;Weiqi Sun;Huaming Yan
本文围绕自动驾驶展开，主要介绍了Deep Racer上基于强化学习的自动驾驶云平台，利用AWS（Amazon Web Service）提高汽车的自动驾驶能力。应用DeepRacer平台提供的示例代码，赛车的训练和完成需要很长时间。因此，我们基于强化学习将路径规划应用到DeepRacer中。在以下几个方面取得了突破性进展：在DeepRacer上完成了强化学习的公式化和求解。将车辆模型简化为自行车，缩小了感知与关节之间的现实差距。介绍了一种新的系统框架VNARM（Vehicle Network Autonomous Racing Model）。在强化学习中，为了跟踪，适当地设置了奖励函数。车辆完成一圈的性能从接近30秒提升至不到9秒，同时保持较高的完成率。
This article revolves around autonomous driving, mainly introducing the autonomous driving cloud platform based on the reinforcement learning to improve the autonomous driving of the car on the Deep Racer using the AWS (Amazon Web Service). Applying the sample codes provided by the DeepRacer platform, the training and completion of the car requires a long time. Therefore, we applied path planning into DeepRacer based on reinforcement learning. Several breakthroughs have been shown as follows: The formulation and solution of RL are completed on DeepRacer. The model of vehicle is simplified as the bicycle and thus reality gap between perception and joints was narrowed. A novel system framework VNARM(Vehicle Network Autonomous Racing Model) is introduced. Reward functions were set properly for tracing in RL. The vehicle's performance of finishing one lap is increased from nearly 30 seconds to less than 9 seconds, while maintaining a high percentage of completion.