loading page

Imitation Reinforcement Learning with Vision and Navigation for Autonomous Driving
  • +1
  • Lei He,
  • Mingyue Ba,
  • Yiren Wang,
  • Ling Han
Lei He
Changchun University of Technology
Author Profile
Mingyue Ba
Chongqing Changan Automobile Co Ltd
Author Profile
Yiren Wang
Changchun University of Technology
Author Profile
Ling Han
Changchun University of Technology

Corresponding Author:[email protected]

Author Profile

Abstract

Autonomous urban driving navigation remains an ongoing challenge, with ample scope for improvement, particularly in navigating through unfamiliar and complex environments. The images captured by cameras provide a wealth of environmental information; however, accurately determining the positions of obstacles within these images can be adversely affected by inclement weather conditions such as rain, snow, or haze. In response to these challenges, this paper presents a hierarchical framework named CNS-DDPG. CNS, which stands for Conditional Imitation Learning, involves the fusion of navigation state information with global path and vehicle state data. DDPG, or Deterministic Policy Gradient, is used for subsequent reinforcement learning. By carefully weighing the strengths and weaknesses of the image and perception module, our framework compensates for visual information captured by the camera by incorporating navigation state data. This design allows our model to perform effectively even in adverse weather conditions. However, the limitations of imitation learning, particularly the scarcity of diverse training data, prompted us to employ the reinforcement learning method DDPG in the second stage of training. This stage benefits from the learned weights of the pre-trained and optimal CNS model. This approach reduces the reliance on imitation learning data and mitigates the challenge of low exploration efficiency associated with randomly initialized weights in reinforcement learning. Additionally, we implement image enhancement techniques to mitigate overfitting associated with simple image types. To evaluate the effectiveness of our approach, we conducted experiments using the CARLA driving benchmark for urban driving. The car was controlled by a Raspberry Pi 4B, which was trained to navigate through an experimental area. Our experiments reveal that CNS-DDPG exhibits remarkable generalization capabilities, particularly in unfamiliar environments and challenging navigation tasks.