基于3D CNN-DDPG端到端无人驾驶控制

End-to-End autonomous driving using deep deterministic policy gradient based on 3D convolutional neural network

文中基于希望直接应用低成本可见光摄像头解决无人驾驶中的刹车、油门和转向控制的问题为目的,采用了深度卷积神经网络和深度确定性策略梯度强化学习结合的方法。通过加入三维卷积神经网络,学习出连续的车辆摄像头视觉感知视频图像帧中的时序属性特征,使得智能体能够利用时序特性更平稳和安全得控制车辆。在开源无人驾驶仿真平台TORCS上进行实验,得出三维卷积神经网络和深度强化学习结合的方法对解决无人驾驶中的控制问题提供了可行性的方案结论。

In this paper,based on the purpose of using low-cost visible light cameras to address the issues of brake,throttle and steering control in driverless applications,a method that combines deep convolutional neural network and deep deterministic policy gradient(DDPG)is adopted.By adding three-dimensional convolutional neural network,the sequential attribute features of continuous visual perception of video frames of vehicle cameras are learned,which enables the agents to control the vehicles more smoothly and safely by using the temporal features.Experiments on TORCS,The Open Racing Car Simulator,show that the method of combining 3D convolution neural network and deep reinforcement learning provides a feasible solution to the control problem in autonomous vehicles.