Risk-based Motion Planning in Automated Vehicles
Speaker
Leon Tolksdorf
About this event
It is anticipated that there will be a long period where automated vehicles (AVs) and human road users share the same road network, with limited communication between them. As a result, AVs often operate with uncertain knowledge of their surroundings due to limitations in both sensing and estimation of the traffic scenario's current and future states. These limitations can be caused by factors such as noisy sensor data or lack of information about the intentions of other road users. Accounting for this uncertainty while designing motion planning and control algorithms for AVs is critical, as an AV may display hazardous behavior if it is overly confident in falsely estimated environment states or, contrarily, may act too cautiously when considering unnecessarily large safety margins in estimated states. Although safety regulations call for taking uncertainty into account, the right means to integrate uncertainty in motion planning has yet to be agreed upon. A promising way to account for uncertainty is through the concept of risk. Risk accounts for uncertainty in assessing safety by considering both the likelihood and severity of unsafe events, such as collisions. Research has shown that human drivers navigate traffic by trading off subjectively perceived risk against the potential rewards of reaching their destinations. Additionally, studies indicate that subjective risk tends to correlate with objective risk measures, while ethicists argue that risk can be used as a means to facilitate ethicality in road traffic by distributing risk fairly among road users. These relationships highlight the benefit of incorporating risk, as a proxy for uncertainty in safety, in motion planning algorithms. Incorporating risk in AV motion planning presents two main challenges. The first challenge is to develop a notion of risk that correctly incorporates an AV's uncertainty about the future motion of other road users, and that can be computed fast (and precisely) enough to be used in online (i.e., real-time) motion-planning algorithms. The second challenge is to develop an appropriate AV motion planning algorithm that incorporates risk and generates the desired AV behavior. This dissertation addresses both challenges by making five main contributions. First, we introduce an appropriate risk metric: the expected severity of a collision event. This metric, thus, is constituted of two terms, one accounting for the likelihood of a collision and another term considering the severity of a collision. Additionally, we investigate the AV behavior given optimization-based motion controllers that accommodate our risk metric. We find that a stochastic model predictive control (SMPC)-based motion planner adequately balances risk with travel efficiency in the presence of varying uncertainty. Second, to compute risk, we provide an computationally efficient (i.e., real-time computable) algorithm to estimate risk. Here, the focus lies on estimating the probability of collision (POC), as our risk metric reduces to a POC metric when the severity term is considered constant. Based on multi-circular shape approximations of the geometries of the AV and other road users, we present an algorithm that estimates the POC with good precision, and that is guaranteed not to under-approximate it. The algorithm is shown to execute significantly faster than standard Monte Carlo sampling. Further, we use the proposed algorithm within a SMPC-based motion planning strategy as a constraint and show that the SMPC generates smooth, reproducible trajectories while the controller deals with varying levels of uncertainty. Third, we extend our POC algorithm to risk estimation. Here, we propose a severity estimation method that allows us to consider different collision constellations, as literature shows that those are associated with different levels of harm to the passengers of vehicles. The resulting risk estimation algorithm leverages the advantages of the POC estimation algorithm, i.e., it is particularly suitable for optimization-based motion planning strategies, while allowing for the flexible integration of severity functions to represent different collision types. Fourth, we provide an extension of our risk metric to account for the risk from the individual perspective of each road user in the driving scene. Given a proposed SMPC-based motion planning strategy, our AVs behavior can be tuned such that it minimizes its own risk (egoistic perspective), the risk other road users face (altruistic perspective), and a balance of both (collective perspective). To evaluate safety and travel efficiency (i.e., reaching a desired destination in adequate time) for each of these risk metrics, we performed simulations on a database containing a wide variety of real-world and artificial traffic scenarios. The results show that a collective perspective balances the overall risk of all road users with the AV's travel efficiency in a favorable manner, and constitutes a suitable approach to facilitate ethicality in road traffic. Lastly, we use our proposed risk metric as part of the decision-making process on emergency maneuvers, i.e., emergency steering and emergency braking, for an AV designed to evade collisions with late-detected pedestrians. In simulations, we show that this metric outperforms deterministic risk estimation by choosing the appropriate evasive maneuver more frequently. In experimental testing, we verified the real-world applicability of this metric in an AV testbed designed for collision avoidance by having a pedestrian dummy walk out from a sight obstruction when the AV approaches. We find that risk is beneficial to deciding on accident avoidance maneuvers. The contributions introduced in this thesis show that risk, and thus uncertainty, can be systematically incorporated in AV algorithms for motion planning and decision-making. The proposed algorithms are computationally efficient for real-time optimization-based motion planning, navigating a fully automated vehicle smoothly through road traffic. Nevertheless, we also show that our risk metric is beneficial to the avoidance capabilities of low-automation safety systems. Furthermore, we show that risk-based AV behavior generation enables balancing the AV's interests in safety and travel efficiency for all road users, and thus, facilitates ethicality.
Host
