Abstract-Safety is of paramount importance in automated driving. One of the main challenges ensuring safety is the unknown future behavior of surrounding traffic participants. Previous works ignore this uncertainty or often address it by computing probability distributions of other traffic participants over time. Probabilistic approaches make it possible to predict the collision probability with other traffic participants, but cannot formally guarantee (i.e. cannot mathematically prove for given assumptions) whether a planned maneuver is collision-free. Our approach addresses exactly this problem: instead of computing probability distributions, we compute an over-approximation of all possible occupancies of surrounding traffic participants over time. This makes it possible to prove whether an automated vehicle can possibly collide with other traffic participants. The presented algorithm for occupancy prediction works on arbitrary road networks and produces results within a fraction of the prediction horizon. Experiments based on real-world data validate our approach and show that we could not find a behavior of a traffic participant that is not enclosed in our prediction.

Abstract-Formally verified methods for motion planning are required in order to guarantee safety for autonomous vehicles. In particular, we consider trajectory generation by considering the most probable trajectory of other traffic participants. However, if the surrounding vehicles perform unexpected maneuvers, a collision might be inevitable. In this paper, a fail-safe motion planner is developed, which generates optimal trajectories, yet guarantees safety at all times. Safety is achieved by maintaining an emergency maneuver which can safely bring the host vehicle to a stop while avoiding any collision. The emergency maneuver is computed by considering for a given time horizon the occupancy prediction which encloses all possible trajectories of the other traffic participants. The performance of the approach is evaluated through simulation against real traffic data.

This paper addresses the problem of following a vehicle with varying acceleration in a comfortable and safe manner. Our architecture consists of a nominal controller (here: model predictive control) and a safety controller. Although model predictive control attempts to keep a safe distance, it cannot formally guarantee it, due to the assumptions on the behavior of the leading vehicle. We address this problem by holding a formally verified safety controller available. Our novel mechanism gradually engages the safety maneuver since most critical situations resolve quickly. The overall approach is evaluated against real traffic data. The results show good position and velocity tracking performance, while safety and comfort are guaranteed.

Abstract-Dealing with the unknown future behavior of other traffic participants is one of the main challenges when generating safe trajectories for autonomous vehicles. When the ego vehicle (i.e., the vehicle to be controlled) follows a given trajectory, an emergency maneuver should be kept available for all times in order to avoid collisions. However, generating an emergency maneuver for each time step is computationally expensive and often not required. In this paper, we propose an algorithm for determining the maximum time horizon under which the ego vehicle can safely follow a given trajectory. First, an upper and a lower bound of this time horizon are computed. Then, binary search is used to find the maximum time horizon for which safety is still guaranteed. Our algorithm reduces the frequency of generating emergency maneuvers while still guaranteeing collision-free trajectories. The approach is tested on real traffic data, and it is shown that our algorithm indeed reduces the frequency of generating emergency maneuvers compared to previous work.

No abstract