Abstract
Hurricanes such as Helene and Milton, which struck only weeks apart in the fall of 2024, caused devastating damage and significant loss of life. Effective emergency response after a disaster like a hurricane begins with strategic planning. This critical first step, extensively studied in prior research, focuses on identifying optimal evacuation routes and ensuring emergency responders can reach affected areas quickly.
However, extreme weather events often render even the most carefully planned routes dangerous to navigate. For example, the heavy rainfall from Hurricane Milton triggered the formation of large sinkholes, columns of debris from collapsed buildings, and partially flooded areas along a large portion of Florida’s western coast, severely compromising roadways and creating dangerous conditions for emergency responders and motorists alike. This raises an important question: can we design systems to enhance driver safety in such hazardous conditions? In this proposal, we aim to study the feasibility of such systems and take the first steps toward deploying them in real-world, hazardous driving scenarios.
To achieve this goal, we will develop a control framework for vehicles operating in hazardous conditions. In such cases, the vehicle’s behavior is highly nonlinear and not constrained to 2D motions, assumptions that usually lead traditional control methods to fail. This is because the vehicle’s model at the core of these approaches is insufficient to capture the complexity of the dynamics encountered during extreme driving. Our key technical insight is that, while such models may be insufficient for precise planning and control, they can serve as abstract, imperfect simulators for training learning-based controllers. In our prior work on simulation-to-reality transfer for high-speed drone racing, published in Nature (2023), we addressed this challenge by augmenting imperfect models with stochastic, data-driven components. This hybrid model enabled us to train control policies using reinforcement learning, achieving robust performance despite the limitations of the underlying simulator.
In this proposal, we aim to extend these findings to ground vehicles, where frequent interaction with the terrain introduces dynamics more unpredictable than for drones. We plan to address this difficulty by developing a novel model-based reinforcement learning approach based on sampling multiple hypotheses and selecting, at test time, which hypothesis best fits the current state. Similarly to other model-based approaches, this control framework will be more data efficient than the model-free approaches used for drones, at the cost of a limited drop in performance. However, differently from other model-based approaches, it will be conditioned on the uncertainty and predicted accuracy of the model, which will make it less susceptible to unpredictable dynamics.
Description
Timeline
Strategic Description / RD&T
Section left blank until USDOT’s new priorities and RD&T strategic goals are available in Spring 2026.
Deployment Plan
The proposal directly addresses the challenge of enhancing driver safety and vehicle control in hazardous conditions. Our proposed approach focuses on the development of a hybrid simulation leveraging both first-principles based and data driven techniques to model the vehicle dynamics in extreme road conditions, e.g., unstable terrain or debris-laden environments. Despite these models being imperfect, they can be used in conjunction with our lab’s recent advances in simulation to reality transfer research to control physical vehicles in the real world.
The anticipated outcomes include the development of novel techniques for controlling vehicles, such as simulation-based control algorithms and real-world deployment strategies, alongside contributions in the form of patents and high-impact research publications. Overall, our work has the potential to augment existing driver-assistance technology by offering scalable, data-driven solutions capable of handling dynamic and unpredictable environments. By doing so, it could influence the regulatory standards of future advanced driver-assistance systems, ensuring their wide spread applicability in high-risk scenarios.
Expected Outcomes/Impacts
The proposed work will be implemented and piloted through strategic collaborations with disaster response agencies (DVRPC) and industry partners (The Autoware Foundation), who will provide critical support, including access to simulation platforms, testing facilities, and operational expertise.
The proposed work will demonstrate adaptive and superhuman driving capability on uneven off-road terrain with sand, gravel rocks, slush and other hazardous environment conditions. We will demonstrate safe planning and robust control of the vehicle in adverse driving conditions.
The proposed work will engage with DVRPC to identify hazard situations and the traversability requirements. We will work with Autoware to integrate the work in real vehicle platforms.
Expected Outputs
The proposed work’s key anticipated outcomes include:
Research Impact: This work will extend the frontier of what is possible with only an abstract and imperfect model of a real-world system. We will develop a fundamentally new approach to combine first-principle methods with data-driven techniques for grounded vehicle navigation in extreme conditions. Instead of designing our approach to be completely autonomous, we will focus on systems that support drivers in such extreme conditions. Indeed, while completely autonomous driving is still a far-fetched goal, assisted driving systems are already a reality today. This has important short-term implications for transportation and driver’s safety.
Publications: Our research will be disseminated through publications at major international venues including robotics, controls, and machine learning conferences.
Data: Prior art has focused on collect large datasets in normal driving conditions. While these datasets are large, they contain only very limited long-tail events. We will collect and open-source a new dataset that especially focuses in hazardous driving conditions, which could be a great complement to existing driving datasets.
Software: We will release the source code and associated software of our novel control framework to handle hazardous driving conditions.
Patent Filings: The proposed work will lead to patent filings to protect the intellectual property on the developed assisted driving technology.
Student Training: One PhD student will be trained during the course of the research. The student will collaborate with our stakeholders.
Pilots developments: The proposed work will lead to pilot developments coordinated with DVRPC's emergency response and evacuation project planning. We will locate off-road driving test sites in the Greater Philadelphia region (e.g. Philly Pump Track) to demonstrate the capabilities of the system in adverse environment and road conditions.
TRID
Increasing driver safety under extreme conditions is a critical priority for the U.S. DoT to achieve its goal of zero fatalities on the road.
The proposal addresses multiple key themes under the RD&T Plan. The proposal addresses safety, the top priority area of the US DOT according to the RD&T Plan (page III). Safety is listed as the first priority goal in the RD&T plan (page 5). The human element of the safety goal of the DOT is amplified in the safety grand challenge called out in the RD&T Plan: “Advance a future without transportation-related serious injuries and fatalities” (page 11). The design of data-driven safety systems is especially called out with the plan calling for systems that “Evaluate the safety performance of infrastructure design and develop and promote the use of effective safety countermeasures” – a research objective that our proposed work directly falls under (page 17).
The RD&T plan also calls for pedestrian safety systems through a more careful analysis of human-technology interaction. It specifically seeks research innovation to “Learn how people respond to the roadway environment, including signs, markings, and traffic control devices, emerging vehicle and roadway technology, innovative operational changes, and pedestrian and bicyclist safety” (page 18). The proposal designs roadway technology for pedestrian safety, which addresses this precise objective. More broadly, the proposal adheres to the “safe systems approach” called for in the RD&T plan, which calls for novel roadway infrastructure for safety, stating that “Human error is to be expected, so road infrastructure and vehicle technology must be designed and operated so that deaths and serious injuries are managed through system safety engineering.” (page 21).
Individuals Involved
| Email |
Name |
Affiliation |
Role |
Position |
| aloque@seas.upenn.edu |
Loquercio, Antonio |
University of Pennsylvania |
Co-PI |
Faculty - Untenured, Tenure Track |
| rahulm@seas.upenn.edu |
Mangharam, Rahul |
University of Pennsylvania |
PI |
Faculty - Tenured |
Budget
Amount of UTC Funds Awarded
$110000.00
Total Project Budget (from all funding sources)
$210000.00
Documents
Match Sources
No match sources!
Partners
| Name |
Type |
| Delaware Valley Regional Planning Commission |
Deployment Partner_ Deployment Partner_ |
| The Autoware Foundation |
Deployment Partner Deployment Partner |
