REFINE: REachability-based trajectory design using robust Feedback lInearization and zoNotopEs

REFINE is a safety real-time trajectory planning framework for autonomous driving. REFINE utilizes a parametrized robust controller that partially linearizes the vehicle dynamics in the presence of modeling error. Offline zonotope-based reachability analysis is performed on the closed-loop, full order vheicle dynamics to compute the corresponding control-parametrized, over-approximate Forward Reachable Sets (FRS). Real-time trajectory planning is achieved by solving an optimization framework in real-time with the pre-computed, control-parametrized FRS being used to ensure vehicle safety. The link to the project website can be found here.

Overview

0. Installation

REFINE is built on Ubuntu 20.04 with ROS Noetic Distribution, and the algorithms are implemented in MATLAB 2022a and C++17.

• Clone the git repos and CORA by running the following from the top-level:

./download-dependencies.sh

• Run install.m.
• In split.m, replace line 20 with cd(your_matlab_directory/toolbox/matlab/strfun) and line 22 with cd('your_CORA2018_directory/global functions/globOptimization').

1. Offline reachability analysis

Vehicle behavior is offline computed in REFINE via reachability analysis using zonotopes. See Offline_Reachability_Analysis for detail.

2. Simulation

REFINE is evaluated in simulation on a full-size Front-Wheel-Drive vehicle model. See Full_Size_Vehicle_Simulation for detail. Simulation video can be found here.

3. Hardware Implementation

REFINE is evaluated in real hardware testing on a All-Wheel-Drive 1/10th race car robot. See Rover_Robot_Implementation for detail. Hardware Demo can be found here.