Comprehensive Analysis of Adaptive Soft Actor-Critic Reinforcement Learning-Based Control Framework for Autonomous Driving in Varied Scenarios

Authors:
Publisher website
Google Scholar
Abstract
In this paper, an energy-efficient autonomous driving framework combining adaptive soft actor-critic (ASAC) evaluation reinforcement learning (RL), and model predictive control (MPC) strategies is proposed. It enables autonomous driving vehicles (AVs) to obstacle-freely track the desired path that the improved A*-based path planner generated. As the core work is to maintain energy-efficient performance, the motor efficiency map is employed to help reduce energy consumption quickly and accurately in an MPC-based predictive path generator and the ASAC-based path-tracking modules. At the same time, the final control input is adapted to the vehicle kinematics, and the generalization performance of the frame is improved. In this work, the CarMaker tool trains and tests on a map of multiple scenarios. Experimental results show that the proposed method is superior to other control methods in path tracking performance, running time, and energy efficiency. Compared to human driving data, this ASAC method can reduce the energy consumption of autonomous vehicles by 9.432% without any energy feedback mechanism. The method can also deal with various disturbances, such as complex road conditions and vehicle mass changes, and track the path accurately.
Funding Information
  • Innovation and Technology (Fund ITP/047/21AP)
  • Power Electronics Research Centre, Department of Electrical and Elctronic Engineering, The Hong Kong Polytechnic University

This publication has 25 references indexed in Scilit: