Automotive Research Projects

Adversarial Scene Generation

Autonomous vehicles (AVs) are becoming increasingly more advanced due to the incorporation of machine learning components, they have complex vehicle dynamics, and they must appropriately handle a great deal of uncertainty due to environment complexity and variability. To date, most developments are focused on improving the robustness of on-road AVs, with comparably little work focusing on the off-road setting. Off-road vehicle dynamics are more complex due to tire-soil interactions and such AVs face uncertainty from the wide variability of natural, unstructured environments. 

A core research question is how can we efficiently evaluate the robustness of an off-road AV due to the many different natural operating conditions? A primary component of this question is simply in identifying reasonable scenarios (in which a human could identify and follow a good path quite easily) that confuse the AV’s autonomy stack inducing erroneous or undesirable behavior (such as abrupt stopping, colliding into obstacles, etc.). Our aim is to develop algorithms that can efficiently identify such scenarios, as well as learn how to perturb an existing scenario (in which the vehicle can navigate without problem) into one that confuses the vehicle. 

Data Driven Forecasting of Traffic Jams

Dramatic changes in the dynamics of complex systems occur with significant consequences. Such an unexpected change is usually undesirable and notably difficult to predict since models of complex systems are usually not accurate enough to predict reliably where and when critical thresholds may occur. One example of these systems is ground vehicle traffic jams, which remain a serious issue in today’s society. Despite advances in traffic flow management in recent years, predicting traffic jams is still a challenge. In this research, we have developed tools for model and predicting these complex systems, serving as early warning indicators and bifurcation forecasting methods, and investigate their application to predict traffic jams on roads.The physics-based methods our group has previously developed for linear and nonlinear systems, as well as the data-driven reduced-order models we have developed pave the way not only for predicting mistuned blisk dynamics, but more generally to model structural systems as a whole. 

These data-driven algorithms and machine learning techniques are rooted in dynamical systems theory and evaluate the stability and resilience of dynamical systems to forecast the risk of critical transitions using a limited number of measurements. These methodologies can be used to analyze stability of traffic models and address challenges related to the complexity of traffic dynamics. 

Optimal Distribution of Tasks in Human Autonomy Teams

There is a need for generating effective, adaptive, and risk-aware coordination strategies for Human-Autonomy Teaming (HAT), where AI-based autonomy is equipped with equal responsibilities as humans in strategic planning and execution, instead of replacing humans or being subordinate team members. Our research aims to prepare autonomy to become decision-makers and collaborate with humans in dynamic, strategic, adversarial operations, considering unforeseen events, human characteristics, workload and resource reduction, and adaptive team design under evolving situations. Currently, we have created a deep reinforcement learning framework to train autonomy using synthetic data and artificial agents with human characteristics. This includes developing an Unreal Engine game platform for actual human decision-makers to interact with autonomy in Virtual Reality, with the help of a machine learning-based workload adaptive user interface. Another line of work tends to tackle the challenges in the quantification and explainability of AI coordination strategies and adaptation for open novelty. 

The developed techniques could potentially benefit a wide range of applications including disaster-relief operations, battlefields, task scheduling for smart manufacturing, and inventory optimization. The research provides insights and opportunities for further developments and design of multi-agent systems that always consider humans as part of the system. 

Reachability-Based Trajectory Design

The research aims to enable the vehicle to safely traverse the 3D terrains with extreme mobility. Unlike on-road navigation, where the environment is well-structured and usually 2D planning is enough for solving the navigation problem, off-road navigation needs to take a 3D terrain geometry into consideration. And due to the 3D topology, a complex coordination of speed and steering is required to fully exploit the vehicle dynamics to navigate successfully. Without such extreme maneuvers, the terrain cannot be navigated. And the planning challenge is exacerbated by the need to do it rapidly without compromising mobility. The focus of this research is on trajectory planning to enable such maneuvers. 

In this context, trajectory planning refers to planning the entire set of vehicle states and controls as a function of time. It involves two components: a global plan that seeks a navigation solution from the current position to the goal while accounting for the prior knowledge of the environment; and a local plan that seeks a navigation solution for a short horizon while accounting for the online information.