No abstract

In this paper we present a model predictive control algorithm designed for optimizing non-linear systems subject to complex cost criteria. The algorithm is based on a stochastic optimal control framework using a fundamental relationship between the information theoretic notions of free energy and relative entropy. The optimal controls in this setting take the form of a path integral, which we approximate using an efficient importance sampling scheme. We experimentally verify the algorithm by implementing it on a Graphics Processing Unit (GPU) and apply it to the problem of controlling a fifth-scale Auto-Rally vehicle in an aggressive driving task.

We present an information theoretic approach to stochastic optimal control problems that can be used to derive general sampling based optimization schemes. This new mathematical method is used to develop a sampling based model predictive control algorithm. We apply this information theoretic model predictive control (IT-MPC) scheme to the task of aggressive autonomous driving around a dirt test track, and compare its performance to a model predictive control version of the cross-entropy method.

Scaled platforms constructed from modified RC cars are popular in the academic and hobby communities. These platforms are typically 0.2 m to 1 m long and weigh between 1 kg and 25 kg. Costs range from a few hundred to tens of thousands of dollars, largely determined by the size, sensors, and computing. Construction, maintenance, and programming is typically handled by a small team of students or researchers. Recently, several open source projects released complete documentation and interface software, which is in contrast to the one-off nature of older work that often lacked enough information to replicate. Documentation for open source platforms normally includes parts lists, build instructions, 3 and interface software for the sensors and actuators. Availability of tutorials, simulation environments, and public datasets vary by project. Common sensors include wheel speed, inertial measurement unit (IMU), cameras, depth sensors, ultrasonic, and light detection and ranging (Lidar) units. The target environment for these platforms is typically indoors on a smooth surface. The Donkey Car [5] is an easy to build 1:16 scale autonomous platform for the DIY Roborace events targeted at hobbyists. Onboard computing and sensing are a Raspberry Pi 3 with a matching wide angle camera. The Berkeley Autonomous Race Car (BARC) [6] is a 1:10 scale vehicle designed as a simple and affordable research platform for self-driving vehicle technologies that has been successfully used to demonstrate various control algorithms. The onboard ODROID-XU4 is similar in computational performance to the Raspberry Pi 3, and the sensor suite includes a hobby grade camera, IMU, four ultrasonic range finders, and Hall effect wheel speed sensors. The F1/10 project [7] and accompanying Autonomous Racing Competition allows teams to race against one another using a common 1:10 scale platform developed at the University of Pennsylvania. Computing on the F1/10 platform is performed by an Nvidia Jetson. The sensor suite includes a hobby IMU, compact indoor Hokuyo 2D Lidar, and optional Structure and Zed depth and motion sensing cameras. The 1:10 scale Rapid Autonomous Complex-Environment Competing Ackermann-steering Robot (RACECAR) [8] from Massachusetts Institute of Technology is a platform for researchers creating applications for self driving cars. RACECAR also uses the Nvidia Jetson for computing, and includes the same Hokuyo Lidar and Zed stereo camera as the F1/10 platform. Table 1 provides a comparison of these open source scaled platforms.

Abstract-We present an algorithmic framework for stochastic model predictive control that is able to optimize non-linear systems with cost functions that have sparse, discontinuous gradient information. The proposed framework combines the benefits of sampling-based model predictive control with linearization-based trajectory optimization methods. The resulting algorithm consists of a novel utilization of Tube-based model predictive control. We demonstrate robust algorithmic performance on a variety of simulated tasks, and on a real-world fast autonomous driving task.