This paper proposes a trajectory planning method for obstacle avoidance in unmanned vehicles based on reinforcement learning considering front-wheel steering angle deviation fault, which affects the safety of the vehicle’s obstacle avoidance, and maintains a smoother body posture during obstacle avoidance. First, a reinforcement learning trajectory planning framework is established, and fuzzy rule processing is adopted to accelerate training for obstacle avoidance decision-making. Then, the Unscented Kalman Filter method is used to estimate the vehicle mass centroid sideslip angle, yaw rate, and lateral velocity. The differences between the estimated values and the actual state values of the vehicle are used to calculate the loss factor, analyzing the impact of front wheel steering angle deviation fault on the vehicle’s body posture. Next, the decision information, steering angle deviation fault information, and pitch/roll angle information are input into the TD3 algorithm for local trajectory planning, enabling the vehicle to safely avoid obstacles even with a certain degree of front wheel steering angle deviation, while minimizing body pitch/roll angles. Finally, through joint simulation and hardware-in-the-loop test, the results show that after adding fuzzy processing, the reinforcement learning for obstacle avoidance decision-making requires fewer training iterations; under normal driving conditions, vehicles using the proposed reinforcement learning trajectory planning method exhibit lower roll angle fluctuations under static and dynamic obstacle environments compared to that using dynamic programming methods; furthermore, when the vehicle experiences steering angle deviation faults, obstacles can be safely avoided by the proposed reinforcement learning trajectory planning method, and a smaller roll angle is also obtained.
Wang et al. (Fri,) studied this question.