Deep Reinforcement Learning Based Autonomous Drift System for Abrupt Obstacle Avoidance

Abstract

Autonomous vehicles face significant challenges in executing emergency obstacle avoidance maneuvers beyond conventional driving limits. Previous approaches, relying on vehicle dynamics modeling or simplified learning methods, often struggle with generalization to diverse scenarios. This paper presents a novel deep reinforcement learning-based Drift Obstacle Avoiding (DOA) system for extreme obstacle avoidance. Our system integrates adrift path planner based on cubic spline curves with a controller trained in a high-fidelity simulator using expert demonstrations and a Refined-exploration Soft Actor Critic (RSAC) algorithm. Utilizing only real-world accessible vehicle states, the system exhibits robust performance across various scenarios, vehicle types, andchallenging road conditions. We validate our approach through extensive simulation testing and open-source our code and expert trajectories dataset at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/ustcly/DOA</uri>for further studies.