Autonomous vehicles powered by intelligent algorithms will be an important element of our nation’s
future mobile infrastructure. The vehicles will operate in unstructured environments and in collaboration
with people. This requires in-time decision-making, often with safety-critical components. Such fast
reactivity to events can only be processed “at the edge”, partly based on incomplete and uncertain
knowledge. As of today, it is unclear how autonomous vehicles must react under unusual conditions in
so-called edge-case scenarios, which present major safety challenges, to ensure safe operation. This
research project proposes a new and innovative approach to address this challenge: Can increasing
the agility of an autonomous vehicle improve safety?
This project aims to develop autonomous vehicle software along two dimensions and their active
interaction: agility and safety. The project focuses on (1) increasing the agility of autonomous vehicles
by developing new methods for safe and agile trajectory and behavior planning so that the vehicle
can maneuver in and above the driving dynamic limit, (2) automated reasoning about unsafe dynamic
situations that may occur while driving, and (3) developing new methods for generating test and
edge case scenarios to explore under what conditions an autonomous vehicle would fail. Finally,
photorealistic simulation platforms will evaluate the proposed methods and algorithms. Successful
completion of this project will be a major step towards adaptive and safe software for autonomous
vehicles. The capabilities of an autonomous vehicle will be massively extended so that it can drive
everywhere, at any time, and under all conditions.
Summary:
The project is divided into four work packages (WP) that aim to test and evaluate critical scenarios for autonomous vehicles, create new evaluation criteria and metrics, design behavior planning under uncertainty, and develop a trajectory planner for agile maneuvers. In WP 1, test scenarios are generated to place autonomous vehicles in extreme situations, initially using tools like CommonRoad. New criteria and metrics are identified to evaluate the performance of a trajectory planner in critical situations. WP 2 focuses on integrating and processing uncertainties using a behavioral planner based on Deep Reinforcement Learning. Latent variables are used to model a higher-level strategy for interacting with dynamic road users. WP 3 involves trajectory planning in uncertain environments, with a focus on generating drivable trajectories and optimal behavior. The trajectory planner is closely linked to the final control of the vehicle. WP 4 aims to demonstrate the capabilities of the developed algorithms through a project demonstration, specifically addressing emergency maneuvers in one of the critical scenarios.
Outcome:
Test scenarios generated for autonomous vehicles in extreme situations
New evaluation criteria and metrics for performance assessment of trajectory planners
Integration of uncertainties and stochastic properties into behavior planning using Deep Reinforcement Learning
Trajectory planner developed for agile maneuvers, including the ability to transition to unstable and aggressive driving states
Successful demonstration of an emergency maneuver at the handling limits based on the developed algorithms
