Fast Plan & Furious Track
Oct. 2017Â - Dec. 2017
With the rapid development of autonomous systems such as UGVs and UAVs, maneuvering these vehicles remains a challenging problem, not only due to the inherent complexity of those nonlinear systems but also their need of real-time implementations.
​
Used a simple double integrator model in the higher level, in order to achieve a high computational speed.
Split the general frame into a hierarchical structure and perform the tasks of path planning and trajectory tracking respectively, for maximize the performance of our controller measured by optimality, feasibility and runtime.
Designed a path planner and tracking controller that combined MPC with probabilistic roadmaps.
Achieved obstacle avoidance using laser sensors and a near-field dynamic MPC approach.
This project discusses a real-time implementation of a hierarchical model predictive control (MPC) framework for high-speed autonomous vehicles, the Fast Plan and Furious Track (FPFT) framework. The higher-level path planner generates a collision-free and distance-optimal reference path by modeling the obstacles as convex sets and iteratively solving a nonlinear programming problem. The lower-level MPC-based tracking controller follows the reference trajectory using the real-time iteration scheme. The controllers are implemented using the automatic code generation capabilities of ACADO Toolkit. Numerical experiments carried out in MATLAB show that the path planner and tracking controller can achieve run-times in half second and milliseconds range, respectively.
​
The dominant issue of controlling an autonomous vehicle is to solve the collision avoidance problem. The problem is NP-hard in general. However, we find a paper that proposes a novel method for generating smooth nonlinear constraints using strong duality of convex optimization, which has been proved to be efficient in generic navigation and trajectory planning tasks for autonomous vehicles.
​
To satisfy the specifications of real-time performance, we used the framework introduced in another paper, where a real-time implementation of model predictive controllers for a 10-state nonlinear input-affine system has been introduced, and is able to achieve run-times in the microsecond range, using the code generation tools introduced in that paper.
​
The project combines the aforementioned techniques and forms a hierarchical MPC-based controller for optimal maneuvering of a four-wheel autonomous racing car. The controller is able to perform planning and tracking tasks simultaneously and steer the car to take over multiple opponents without collision. The results show that the controller works well in straight lane scenarios and fulfills the real-time requirements.
