Affiliation:
1. University of California, Berkeley, Berkeley, CA
Abstract
Safety and efficiency are two key elements for planning and control in autonomous driving. Theoretically, model-based optimization methods, such as Model Predictive Control (MPC), can provide such optimal driving policies. Their computational complexity, however, grows exponentially with horizon length and number of surrounding vehicles. This makes them impractical for real-time implementation, particularly when nonlinear models are considered. To enable a fast and approximately optimal driving policy, we propose a safe imitation framework, which contains two hierarchical layers. The first layer, defined as the policy layer, is represented by a neural network that imitates a long-term expert driving policy via imitation learning. The second layer, called the execution layer, is a short-term model-based optimal controller that tracks and further fine-tunes the reference trajectories proposed by the policy layer with guaranteed short-term collision avoidance. Moreover, to reduce the distribution mismatch between the training set and the real world, Dataset Aggregation is utilized so that the performance of the policy layer can be improved from iteration to iteration. Several highway driving scenarios are demonstrated in simulations, and the results show that the proposed framework can achieve similar performance as sophisticated long-term optimization approaches but with significantly improved computational efficiency.
Publisher
American Society of Mechanical Engineers
Cited by
18 articles.
订阅此论文施引文献
订阅此论文施引文献,注册后可以免费订阅5篇论文的施引文献,订阅后可以查看论文全部施引文献