No abstract

Autonomous navigation has recently gained great interest in the field of reinforcement learning. However, little attention was given to the time optimal velocity control problem, i.e. controlling a vehicle such that it travels at the maximal speed without becoming dynamically unstable (roll-over or sliding).Time optimal velocity control can be solved numerically using existing methods that are based on optimal control and vehicle dynamics. In this paper, we use deep reinforcement learning to generate the time optimal velocity control. Furthermore, we use the numerical solution to further improve the performance of the reinforcement learner. It is shown that the reinforcement learner outperforms the numerically derived solution, and that the hybrid approach (combining learning with the numerical solution) speeds up the training process.

This paper addresses the issue of autonomous competitive yet safe driving in the context of the Indy Autonomous Challenge (IAC) simulation race. The IAC is the first multi-vehicle autonomous head-tohead competition, reaching speeds of 300 km/h along an oval track modeled after the Indianapolis Motor Speedway (IMS). We present a racing controller that attempts to maximize progress along the track while avoiding collisions with opponent vehicles and obeying the race rules. To this end, the racing controller first computes a race line offline. During the race, it repeatedly computes a small set of dynamically feasible maneuver candidates, each tested for collision with the opponent vehicles. It then selects a collision-free maneuver that maximizes the progress along the track and obeys the race rules. Our controller was tested in a 6-vehicle simulation, managing to run competitively with no collision over 30 laps. In addition, it managed to drive within a close range of the leading vehicle during most of the IAC final simulation race.

The determination of the pose of the imaging camera is a fundamental problem in computer vision. In the monocular case, difficulties in determining the scene scale and the limitation to bearing-only measurements increase the difficulty in estimating camera pose accurately. Many mobile phones now contain inertial measurement devices, which may lend some aid to the task of determining camera pose. In this study, by means of simulation and real-world experimentation, we explore an approach to monocular camera localization that incorporates both observations of the environment and measurements from accelerometers and gyroscopes. The unscented Kalman filter was implemented for this task. Our main contribution is a novel approach to landmark initialization in a Kalman filter; we characterize the tolerance to noise that this approach allows.

The field-of-view of a wide-angle image is greater than (say) 90 degrees, and so contains more information than available in a standard image. A wide field-of-view is more advantageous than standard input for understanding the geometry of 3D scenes, and for estimating the poses of panoramic sensors within such scenes. Thus, wide-angle imaging sensors and methodologies are commonly used in various road-safety, street surveillance, street virtual touring, or street 3D modelling applications. The paper reviews related wide-angle vision technologies by focusing on mathematical issues rather than on hardware.

No abstract