Abstract
Autonomous driving promises to be the main trend in the future intelligent transportation systems due to its potentiality for energy saving, and traffic and safety improvements. However, traditional autonomous vehicles’ behavioral decisions face consistency issues between behavioral decision and trajectory planning and shows a strong dependence on the human experience. In this paper, we present a planning-feature-based deep behavior decision method (PFBD) for autonomous driving in complex, dynamic traffic. We used a deep reinforcement learning (DRL) learning framework with the twin delayed deep deterministic policy gradient algorithm (TD3) to exploit the optimal policy. We took into account the features of topological routes in the decision making of autonomous vehicles, through which consistency between decision making and path planning layers can be guaranteed. Specifically, the features of a route extracted from path planning space are shared as the input states for the behavioral decision. The actor-network learns a near-optimal policy from the feasible and safe candidate emulated routes. Simulation tests on three typical scenarios have been performed to demonstrate the performance of the learning policy, including the comparison with a traditional rule-based expert algorithm and the comparison with the policy considering partial information of a contour. The results show that the proposed approach can achieve better decisions. Real-time test on an HQ3 (HongQi the third ) autonomous vehicle also validated the effectiveness of PFBD.
Highlights
Autonomous driving (AD) has been vastly investigated in different domains for several decades and has a wide variety of applications, especially in intelligent transportation systems
To learn an optimal policy, we creatively propose to use the deep reinforcement learning (DRL) algorithm to learn the aforementioned three key parameters instead
Since the proposed approach focuses on behavioral decision and planning with only abstract entities for autonomous driving, we tested the method with our vehicle simulation platform instead of real traffic
Summary
Autonomous driving (AD) has been vastly investigated in different domains for several decades and has a wide variety of applications, especially in intelligent transportation systems. (LIDAR (light detection and ranging) raw inputs, global direction and GPS), a full convolutional neural network [5] generated driving paths in a more explainable output Both rule-based and supervised learning based decision policies have the bottleneck of exploring unknown or complex scenarios. Aradi et al [9] used the REINFORCE algorithm to learn the driving policy mapping 16 continuous states as inputs to train the steer and acceleration demand These methods can hardly be practical, since modeling a vehicle’s lane change behavior with limited information is impossible in real traffic. The proposed PFBD method selects a route generated from the planning space with a safety guarantee It learns better policy through exploring the environment instead of struggling with the rules. The PFBD method achieved competitive performance compared to rule-based methods
Sign-in/Register to view highlights & summary 
Published Version (
Free)
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call