This paper describes an entire automated vehicle platform for software prototyping and demonstration. The project is build up with the cooperation of the Czech Technical University and Porsche Engineering Services company. The platform is based on a serial production car Porsche Panamera. The aim of this prototype is to close perception - planning - control loop and demonstrate it at a fully automated slalom ride. Besides the sensors already available in the stock Porsche Panamera car, our system is equipped with differential GPS, mono camera, and 16-layers LiDAR. The algorithms run on the NVIDIA TX2 platform, are integrated using Robotic Operating System (ROS) middleware and interface with the car ECUs via CAN and FlexRay buses. A Support Vector Machine with Histogram of Oriented Gradients (HoG) feature vectors from the camera image is used for cones classification. Cone localization is done relative to the car. It uses data from the camera and LiDAR. A Kalman filter derives accurate car odometry. It takes data from the GPS, car builtin accelerometer and actual speed and steers commands. The bicycle kinematic model is used to interconnect all of these pieces of information and for Kalman filter implementation. Clustering and filtering of all detected cones generate the cone map. Optimal trajectory planning respecting the car kinematics is crucial for the smooth slalom ride between the cones. A nonlinear quadratic optimization problem for this purpose is formulated and solved based on the bicycle model. Finally, the trajectory tracking control keeps the car on the planned slalom trajectory during the ride. Substantial work was devoted to experiments with a real vehicle and the fine - tuning of the system parameters. Validation of the system reveals interesting observations related to the precision, frequency and sensitivity of the system components and usability of ROS middleware for the autonomous driving tasks.

Autonomous cars represent the future of human transportation. We contribute to this course by cooperation with the automotive industry on making a Porsche Panamera car drive autonomously. Unfortunately, performing experiments with real cars is expensive, potentially dangerous, and sometimes not convenient. For this reason, we prototype many of the algorithms for the Panamera on an RC-car (a scaled-down model of autonomous car), which makes many experiments easier to conduct. To validate our work and to compare it with others, we participate in the International F1/10 Autonomous Racing Competition. In this racing event, student teams from all over the world compete in a so-called battle of algorithms. When developing these algorithms or combining a few of them in more complex control approaches, it is not always easy to decide which approach is better. In this paper, we present a comprehensive methodology for comparing the control approaches. We selected the algorithms for the comparison from the list of state-of-the-art approaches as well as approaches used by the teams within the competition. The compared algorithms include: - Iterative Follow The Gap algorithm, - Iterative Follow The Gap algorithm extended with location-based information (Map-based Follow The Gap), and - Static Optimal Trajectory Tracking algorithm. We compare the approaches above directly on the platform used in the competition rather than in simulations. Our evaluation is supported by a high-precision external localization system. We use the comparison criteria, selected based on our years-long experience, and which have proven themselves essential for successful participation in the competition. The criteria are: - racing performance (i.e., how fast is the car able to drive through the track), - obstacle avoidance slowdown (i.e., how much is the performance of the approach penalized by the presence of static obstacles), - hardware load (computing resources on the platform are limited; as the platform is selected by the organizers of the event), and - configuration effort (i.e., how much effort does it take to tune the parameters of the approach). The results of this paper will serve as a baseline for assessing the performance of control approaches developed in the future.