نموذج المسيطر التنبؤي المستند على الشبكة العصبية لتتبع السيارة ذاتية القيادة.

Neural Network Based MPC for Tracking the Self Driving Car

Abstract

The road traffic environment is highly variable and unpredictable. Autonomous cars operating in such environments may face unexpected critical scenarios, where the risk of an accident rapidly increases compared to normal driving situations. These scenarios may arise due to unforeseen behavior from other road users or obstacles appearing on the road. In such crucial conditions, the primary objective of car motion control is to minimize the danger of an impending accident.
The purpose of this study is to develop a system that can help prevent accidents in unpredictable and variable road traffic environments by addressing the problem of motion planning and control in critical situations for autonomous cars. The system generates optimal paths and control inputs for the car to follow while avoiding obstacles and following the center of the track predictably. To achieve this objective, motion planning technique for self-driving cars are presented.
The motion planning technology utilized in this study is based on the A* and potential field algorithms, with an intelligent controller consisting of prediction supported by neural network technology. The model predictive controller predicts the car’s future for a finite time horizon using a mathematical model of the car. The controller utilizes a linearized and discretized kinematic bicycle model as an internal car model. The A* strategy path is used as it is a powerful tool for solving pathfinding problems due to its optimality, efficiency, admissibility, flexibility, and potential function for path planning in an environment with obstacles. The potential function is used due to its simplicity, safety, and low computational cost. The control inputs are the car’s steering angle and acceleration.
The model predictive controller solves the optimization problem as a quadratic programming problem that minimizes a cost function while satisfying a set of constraints. The neural network is used to adjust the rate of change of the steering angle value, adjusting the rate of steering angle is an important part of optimizing a self-driving car’s performance and ensuring its safety and reliability on the road. Without adjusting the rate of steering angle, the car’s ability to make precise turns and stop accurately may be compromised, which can lead to a higher risk of accidents and lower fuel efficiency. The cost function includes a set of objectives, including errors in desired and current states, inputs, and the rate of change of inputs, to guide the self-driving car away from high-cost regions. The decision-making module determines the next course of action for the car. After the subsequent maneuver is selected, a velocity profile and a lane center reference are generated for the self-driving car to track.
The quadratic programming problem is solved using convex optimization in the optimization tool of python, with a sample time of 0.1 seconds. The control parameters, including the cost function weights and the length of the horizon, are adjusted to make the lane safe and comfortable. The selected horizon length ranges from 8 m to 12 m, within which the control unit ensures that the car follows the intended path while avoiding obstacles.
All the results were achieved within the constraints of the specified car, with a maximum speed of 15 m/s, maximum reverse speed of 5 m/s, maximum steering angle of 45°, maximum steering rate of 30°, maximum deceleration of 6 m/s^2 and a maximum for acceleration 2.5 m/s^2