Autonomous Driving Control Research

Advisor: Dr. Yiqiang Han, Dr. Umesh Vaidya

This project is intended to create a data-driven control framework for high speed racing style navigation for autonomous driving.

• Use Augmented Koopman Operator (polynomial + deep neural network) to achieve system identification based on recorded vehicle control and state data.

• Use both iterative update as well as dynamic programming approach to obtain global reference trajectory (Racing - line). This reference trajectory can be used to test the control performance between Deep Koopman MPC, Kinematic Nonlinear MPC and Pure Pursuit under high speed.

• The experiment is conducted in both F1tenth Simulation as well as Scaled RC car.

This project use Deep Neural Network as koopman operator to transform nonlinear vehicle dynamics into a LTI state-space system, so that the MPC on lifted state-space can be solved as a QP problem. '

• Use the collected data to train a neural network lifting function, then solve the linear invariant state-space matrices from the lifted states.

• Apply model predictive control on lifted state-space for control. Compare its navigation performance with adaptive pure pursuit controller, nonlinear model predictive controller and linearized model predictive controller.

• Language, package and platform used in this project: ROS, Python, C++, Pytorch, Numpy, CasADi, CVXPY

' This project use a novel method to assign velocity and look-ahead distance to conventional pure pursuit, hence improve pure pursuit's performance in fast navigation scenario. '

• Generate velocity profile based on discrete yaw change calculated from real-time local path and tire force data.

• Create nonlinear mapping between current vehicle velocity and look-ahead distance.

• Use iterative methods to improve lap time.

• Language, package and platform used in this project: ROS, Python, C++, Numpy

' Rebuild 1/10th scale RC Car into an autonomous driving research platform. '

• The sensors used for this project: Intel realsense tracking camera t265; Intel realsense depth camera D435i; SLAMTEC RPLIDAR A2

• The central computing unit onboard: Jetson Nano/ LattePanda

• Slam-gmapping and LiDAR were used to build 2D-map for Clemson Flour-Danial Engineering building's basement.

• Use particle-filter to fulfill LiDAR based localization

• Allow both manual and autonomous control mode.