Abstract
This project tackles the critical research question: *How can pre-trained autonomous driving models be seamlessly and dynamically adapted across diverse operational design domains (ODDs) to achieve safe and robust deployment?* ODDs define the specific conditions under which an autonomous driving system can operate safely—spanning factors like road types, traffic density, weather patterns, and geographic features. For instance, a model trained to navigate Pittsburgh’s urban intersections must also adapt to Pittsburgh’s rural routes with sparse signage and rolling hills, or even underdeveloped, unpaved, mixed-traffic roads. What is the right way for a robot to, in real time, smoothly switch between different planning trajectories to ensure safety and consistency amid changing driving conditions?
By focusing on adapting pre-trained models rather than retraining them from scratch, this project offers a scalable approach to ensure safe deployment. This is particularly critical in data-limited settings—collecting training data for every possible scenario can cost millions and take years. For example, equipping autonomous systems to operate in rural areas avoids costly new data collection while preserving high safety standards. A single deployment failure in an untested environment could cost lives and erode public trust, emphasizing the importance of rigorous adaptation. With an estimated 1.35 million road traffic deaths annually worldwide in drastically different driving conditions, this project aims to ensure that autonomous driving technologies can be deployed safely. In other words, this research algorithmically and proactively catalyzes translation of new technologies to data-limited areas, instead of passively relying on commercial technology spillover – the latter would take years or even decades.
We leverage diverse datasets to ensure robust testing. These include driving data from TIER IV in Japan, and over 10 million miles of publicly available driving data from Waymo, collected across 25 U.S. cities and capturing diverse urban environments. In Pittsburgh, we have access to truck-mounted camera data that covers mixed urban and rural routes, while stationary cameras funded through a PITA partnership with RIDC capture real-time trajectory data in high-traffic industrial areas. Simulation-based fine-tuning complements these datasets, addressing data gaps in under-documented scenarios, such as rural roads with limited signage. This rich dataset ecosystem provides the foundation for testing and validating model adaptations in diverse operational design domains.
The project employs open source, cutting-edge algorithms, including inverse reinforcement learning, guided diffusion, blueprints, and fine-tuning. Inverse RL provides an overall framework to complement rule-based modules to directly and flexibly extract algorithm parameters (sometimes implicitly) from expert (human or simulated) driving. Guided diffusion allows for precise and smooth generation and adjustments of such parameters in constrained domains, such as adapting autonomous driving algorithms from Pittsburgh's dense urban streets to its hilly rural roads. Blueprinting ensures scalable and structured design of template models that are transferable across diverse contexts, while fine-tuning efficiently adapts pre-trained models to novel environments, particularly in low-data settings. Open source models already implement these algorithms. We adapt these powerful emerging methods to address challenges in rural areas and infrastructure-limited regions.
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 project leverages simulation and targeted real-world testing to develop and validate autonomous driving models across diverse contexts.
Months 1-3: Focus on establishing robust simulation frameworks to adapt pre-trained models. Data from TIER IV (Japan), Waymo (U.S.), and Pittsburgh’s truck-mounted and stationary cameras are integrated into simulation platforms. Using the Autoware driving stack, three simulation contexts are created: urban Pittsburgh, rural Pittsburgh, and other areas. Baseline models are developed with guided diffusion, blueprinting, and fine-tuning, with key metrics defined, such as collision reduction, navigation accuracy, and scalability.
Months 4-6: Refine and validate models through extensive simulation. Urban models optimize motion planning for high-density intersections using RIDC camera data. Rural models address challenges like unpaved roads and sparse signage using truck-mounted camera data. Other simulations focus on mixed-traffic scenarios and low-infrastructure constraints. Collaboration with TIER IV supports testing under left-hand traffic scenarios in Japan. Iterative feedback ensures robust, adaptable algorithms.
Months 7-9: Transition from simulation to controlled real-world testing. Urban and rural models are tested in Pittsburgh's controlled environments such as RIDC locations to evaluate accuracy and scalability. Discrepancies between simulation and real-world performance are resolved.
Months 10-12: Collaborations with the City of Pittsburgh and other CMU departments. Results are documented in a comprehensive policy report and presented at the Safety21 Annual Deployment Partners’ Consortium Research Showcase. Follow-up funding proposals target NSF, the World Bank, and the Gates Foundation.
Expected Outcomes/Impacts
This project will produce adaptable autonomous driving models, open-source software modules integrated into the Autoware stack, and simulation-tested deployment frameworks for diverse operational design domains (ODDs). These outputs will improve safety by reducing collision rates in urban intersections, enhancing navigation accuracy on rural roads, and ensuring reliable operation in infrastructure-limited environments, such as mixed-traffic conditions. By leveraging pre-trained models and fine-tuning instead of retraining, the project also reduces costs.
The research will directly influence policy decisions by promoting CAV deployment practices, ensuring certain areas are prioritized for rigorous testing before rollout. It will provide key stakeholders, including the U.S. Department of Transportation (USDOT), state transportation agencies, and international partners, with actionable guidance to prevent unsafe deployment of autonomous vehicles, fostering safety in transportation systems.
The project’s findings will benefit the transportation system by enhancing reliability through robust real-time adaptability, durability by addressing edge cases and diverse road conditions, and cost efficiency through scalable, low-resource deployment strategies. These advancements support a safer transportation future while setting a foundation for global regulatory frameworks and commercialization opportunities.
Expected Outputs
This project will produce several outputs to advance autonomous driving technology. Such as adaptable autonomous trajectory generation and selection modules for the Autoware open source auto-driving stack, optimized for diverse operational design domains (ODDs), including urban, rural, and infrastructure-limited environments. These models will be developed and refined using cutting-edge methods such as inverse reinforcement learning, guided diffusion, and fine-tuning to ensure robust performance and transferability.
The project will release these open-source software modules into the Autoware driving stack on Github, featuring advanced motion planning and risk-aware optimization algorithms.
A comprehensive dataset of simulation and field-testing results will be shared with the research community, offering a valuable resource for studying the adaptation of autonomous systems across ODDs.
The research findings will also be prepared for publication in high-impact journals, such as Transportation Research Part A: Policy and Practice, INFORMS Journal on Applied Analytics, and IEEE journals on robotics and vehicular technologies. These publications will disseminate the project’s results to both academic and practitioner audiences, accelerating adoption of the methodologies and insights. Additionally, the project may generate invention disclosures or patent filings related to risk-aware optimization and adaptation processes. Together, these outputs will drive safety and scalability in autonomous vehicle deployment.
TRID
Operational Design Domains (ODDs), originally developed in aerospace engineering to define the boundaries of operational capabilities for systems like drones and aircraft, are now being adapted to Connected and Autonomous Vehicles (CAVs). Despite their growing relevance, the body of research explicitly applying ODDs to CAV deployment remains relatively sparse. They may focus on defining ODDs or using them for route planning, but do not address dynamic adaptation or seamless model transfer across diverse ODDs.
This project is unique in its emphasis on leveraging cutting-edge open-source algorithms—such as end-to-end learning, pre-training, and fine-tuning—which have matured significantly in recent years but have not yet been extensively applied to CAVs. By adapting and deploying these techniques, the project bridges the gap between advancements in machine learning and their practical application in autonomous vehicle safety.
Collaborative opportunities exist with researchers working on ODD definition and optimization, as well as open-source initiatives like the Autoware stack TIER IV is leading. Partnering with these efforts will enable the integration of dynamically adaptive models into existing frameworks, advancing both technical capabilities and real-world deployment.
Individuals Involved
| Email |
Name |
Affiliation |
Role |
Position |
| haohao@andrew.cmu.edu |
Hao, Hao |
Carnegie Mellon Heinz College |
Other |
Student - PhD |
| ejafari@andrew.cmu.edu |
Jafari, Esmatullah |
Carnegie Mellon Heinz College |
Other |
Staff - Business Manager |
| loril@andrew.cmu.edu |
Lawrence, Lori |
Carnegie Mellon Heinz College |
Other |
Staff - Business Manager |
| yidim@andrew.cmu.edu |
Miao, Yidi |
Carnegie Mellon Heinz College |
Other |
Student - PhD |
| ptang@andrew.cmu.edu |
Tang, Pingbo |
Carnegie Mellon University |
Co-PI |
Faculty - Tenured |
| pyzhang@cmu.edu |
Zhang, Peter |
Carnegie Mellon Heinz College |
PI |
Faculty - Untenured, Tenure Track |
Budget
Amount of UTC Funds Awarded
$98047.00
Total Project Budget (from all funding sources)
$196094.00
Documents
Match Sources
No match sources!
Partners
| Name |
Type |
| TIER IV |
Deployment Partner Deployment Partner |
| RIDC |
Deployment Partner_ Deployment Partner_ |
