This note is part of Quality testing.

This work presents a behavior planning algorithm for automated driving in urban environments with an uncertain and dynamic nature. The algorithm allows to consider the prediction uncertainty (e.g. different intentions), perception uncertainty (e.g. occlusions) as well as the uncertain interactive behavior of the other agents explicitly. Simulating the most likely future scenarios allows to find an optimal policy online that enables non-conservative planning under uncertainty.

This thesis considers risk aware planning and probabilistic prediction for autonomous systems under uncertain environments. Motion planning under uncertainty looks for trajectories with bounded probability of collision with uncertain obstacles. Existing methods to address motion planning problems under uncertainty are either limited to Gaussian uncertainties and convex linear obstacles, or rely on sampling based methods that need uncertainty samples. In this thesis, we consider non-convex uncertain obstacles, stochastic nonlinear systems, and non-Gaussian uncertainty. We utilize concentration inequalities, higher order moments, and risk contours to handle non-Gaussian uncertainties. Without considering dynamics, we use RRT to plan trajectories together with SOS programming to verify the safety of the trajectory. Considering stochastic nonlinear dynamics, we solve nonlinear programming problems in terms of moments of random variables and controls using off-the-self solvers to generate trajectories with guaranteed bounded risk. Then we consider trajectory prediction for autonomous vehicles. We propose a hierarchical end-to-end deep learning framework for autonomous driving trajectory prediction: Keyframe MultiPath (KEMP). Our model is not only more general but also simpler than previous methods. Our model achieves state-of-the-art performance in autonomous driving trajectory prediction tasks.

Steffen Heinrich describes a motion planning system for automated vehicles. The planning method is universally applicable to on-road scenarios and does not depend on a high-level maneuver selection automation for driving strategy guidance. The author presents a planning framework using graphics processing units (GPUs) for task parallelization. A method is introduced that solely uses a small set of rules and heuristics to generate driving strategies. It was possible to show that GPUs serve as an excellent enabler for real-time applications of trajectory planning methods. Like humans, computer-controlled vehicles have to be fully aware of their surroundings. Therefore, a contribution that maximizes scene knowledge through smart vehicle positioning is evaluated. A post-processing method for stochastic trajectory validation supports the search for longer-term trajectories which take ego-motion uncertainty into account. About the Author Steffen Heinrich has a strong background in robotics and artificial intelligence. Since 2009 he has been developing algorithms and software components for self-driving systems in research facilities and for automakers in Germany and the US.

In the past decade, the autonomous driving industry has seen tremendous advancements thanks to the progress in computation, artificial intelligence, sensing capabilities, and other technologies related to autonomous vehicles. Today, autonomous cars operate in dense urban traffic, compared to the last generation of robots that were confined to isolated workspaces. In these human-populated environments, autonomous cars need to understand their surroundings and behave in an interpretable, human-like manner. In addition, autonomous robots are engaged in more social interactions with other humans, which requires an understanding of how multiple reactive agents act. For example, during lane changes, most attentive drivers would slow down to give space if an adjacent car shows signs of executing a lane change. For an autonomous car, understanding the mutual dependence between its action and others' actions is essential for the safety and viability of the autonomous driving industry. However, most existing trajectory planning approaches ignore the coupling between all agents' behaviors and treat the decisions of other agents as immutable. As a result, the planned trajectories are conservative, less intuitive, and may lead to unsafe behaviors. To address these challenges, we present motion planning frameworks that maintain the coupling of prediction and planning by explicitly modeling their mutual dependency. In the first part, we examine reciprocal collision avoidance behaviors among a group of intelligent robots. We propose a distributed, real-time collision avoidance algorithm based on Voronoi diagrams that only requires relative position measurements from onboard sensors. When necessary, the proposed controller minimally modifies a nominal control input and provides collision avoidance behaviors even with noisy sensor measurements. In the second part, we introduce a nonlinear receding horizon game-theoretic planner that approximates a Nash equilibrium in competitive scenarios among multiple cars. The proposed planner uses a sensitivity-enhanced objective function and iteratively plans for the ego vehicle and the other vehicles to reach an equilibrium strategy. The resulting trajectories show that the ego vehicle can leverage its influence on other vehicles' decisions and intentionally change their courses. The resulting trajectories exhibit rich interactive behaviors, such as blocking and overtaking in competitive scenarios among multiple cars. In the last part, we propose a risk-aware game-theoretic planner that takes into account uncertainties of the future trajectories. We propose an iterative dynamic programming algorithm to solve a feedback equilibrium strategy set for interacting agents with different risk sensitivities. Through simulations, we show that risk-aware planners generate safer behaviors when facing uncertainties in safety-critical situations. We also present a solution for the "inverse" risk-sensitive planning algorithm. The goal of the inverse problem is to learn the cost function as well as risk sensitivity for each individual. The proposed algorithm learns the cost function parameters from datasets collected from demonstrations with various risk sensitivity. Using the learned cost function, the ego vehicle can estimate the risk profile of an interacting agent online to improve safety and efficiency.

This work develops a motion planner that compensates the deficiencies from perception modules by exploiting the reaction capabilities of a vehicle. The work analyzes present uncertainties and defines driving objectives together with constraints that ensure safety. The resulting problem is solved in real-time, in two distinct ways: first, with nonlinear optimization, and secondly, by framing it as a partially observable Markov decision process and approximating the solution with sampling.