An Autonomous Vehicle with Deep Reinforcement for collision avoidance
Keywords:
Autonomous driving, Deep Reinforcement Learning, Neural Networks, Collision Avoidance, Policy Gradients, proximal policy optimization (PPO).Abstract
To save training and reinforcement learning efforts the approach is applied to the autonomous vehicle in search of an obstacle to avoid. Therefore, this study aims to allow an autonomous means to learn from errors and reprocess their precision movement to avoid collision in the work environment. The enhanced learning method uses a policy hierarchy to analyze state information and provide a distribution of all possible actions it might take. A random sample of proxy actions is selected from this distribution. Finally, we get some new features and states. This is repeated throughout the loop. An autonomous vehicle can use it to determine how and where it is moving. During a deep learning neural network that may encounter countless instances of different environments and actions performed by the autonomous vehicle is learned. In experiments, the deep camera and Lidar adopted as the input device are not affected by light intensity and the color of the road. The policy gradient method was trained with PPO about 21 days in 8 different cities, any different environments, and the results were ok.
