Abstract
According to recent statistics in the U.S., close to 6 million motor vehicle accidents were reported in 2022 with more than 42,000 fatalities and more than 1.6 million injuries. In 2021, vulnerable road users accounted for about one-third of all traffic fatalities including 7,388 pedestrians, 966 cyclists, and 5,932 motorcyclists. In 2021, about 116,000 pedestrians, 105,000 cyclists, and 201,000 motorcyclists were treated in hospitals after collisions. Motor vehicle accidents are the second leading cause of deaths resulting from unintentional injury in the U.S. Connected and autonomous driving with carefully developed crash mitigation methods will help in the reduction of road traffic accidents and the fatalities and injuries they cause, including the ones with vulnerable road users. Based on these facts, this proposal focuses on developing data driven collision avoidance algorithms that learn to avoid collisions with nearby vehicles and vulnerable road users, building upon our approach to pedestrian and bicyclist safety in our Year 1 and Year 2 Safety 21 projects. A deep reinforcement learning safety algorithm with safe exploration will be developed using realistic hardware-in-the-loop (HIL) and vehicle-in-virtual-environment (VVE) platforms.
The current state-of-the-art crash mitigation methods use a catalog of possible collision interaction scenarios using model-in-the-loop simulations for development and evaluation. While this approach will be successful in mitigating collisions for the scenarios in the catalog being used, it is not generic in nature and will not necessarily mitigate crashes in the many other possible dangerous interactions between the road actors. The motivation for the use of the data driven approach in this proposal is this limitation in scope and the difficulty of accurately modeling the interaction dynamics between vehicles and between vehicles and vulnerable road users in dangerous encounters. Our deep reinforcement learning based collision avoidance algorithm will learn from interactions in complex traffic environments, adapt to various traffic scenarios, and satisfy real-time performance requirements, making it a more robust and advanced solution. We will use a hybrid approach in that the trained deep reinforcement learning algorithm will only be activated and work during dangerous and close interactions with other road users. Training will also be based on such interactions using traffic microsimulation to generate realistic training interactions automatically as compared to the fixed nature of the scenario driven approach.
Since reinforcement learning algorithms suffer from the absence of safety guarantees during the learning process, we will integrate our deep reinforcement learning algorithm with control Lyapunov and control barrier functions to limit its training to feasible areas of interaction with other road actors within the road barriers, with realistic vehicle dynamic model constraints as compared to the oversimplified ones used in the literature. Vehicle-to-Vehicle and Vehicle-to-VRU connectivity will also be used as low latency, and high speed near field and over the cloud cellular and other means of communication are available and as they help significantly in non-line-of-sight situations in aiding the perception sensors of autonomous vehicles. Efficacy of the proposed method will be demonstrated using HIL simulations and with a real vehicle using VVE.
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
July – September 2025
1. SAE WCX 2025 abstract submission
2. Participation in Safety21 meetings
3. Inform project deployment partners of research progress of interest
October – December 2025
1. SAE WCX 2025 paper submission based on accepted abstract
2. Deployment Partner Consortium Symposium participation
3. Participation in Safety21 meetings
4. Inform project deployment partners of research progress of interest
January – March 2026
1. Submit journal paper
2. Participation in Safety21 meetings
3. Inform project deployment partners of research progress of interest
4. Prepare and submit progress report
April – June 2026
1. SAE WCX paper presentation
2. Participation in Safety21 meetings
3. Inform project deployment partners of research progress of interest
4. Prepare and submit final project report
Expected Outcomes/Impacts
Our deep reinforcement learning with control Lyapunov/barrier function exploration based algorithm will help avoid collisions of connected and autonomous vehicles with other vehicles and vulnerable road users and help automotive OEMs and suppliers in their development of advanced emergency braking algorithms for the FMVSS No. 127 standard by 2029. The developed collision avoidance algorithm will fuse vehicle-to-vehicle and vehicle-to-VRU communication with perception sensor data to improve accuracy including no-line-of-sight encounters. The VVE approach of our Year 1 and 2 projects will be further developed to inject real high collision risk vehicles and vulnerable road users into the development and evaluation environment to safely and realistically test extreme events naturally. This approach has the potential of being used in pre-deployment evaluation of connected and autonomous vehicles and the novel collision avoidance algorithm to be developed has the potential of improving road safety and eventually contributing to the goal of safer roads with less accidents especially those due to driver error. Connected and autonomous vehicle collision avoidance algorithms developed and evaluated in this approach with natural injection of unexpected real vehicle and vulnerable road user motion will also reduce cost and improve efficiency of development and deployment.
Expected Outputs
The research and technology outcome will be the development of the novel deep reinforcement learning algorithm with control Lyapunov/barrier function exploration using more realistic vehicle/VRU interaction models and vehicle dynamics constraints for collision avoidance with other nearby vehicles and vulnerable road users and corresponding conference and journal publications. The algorithm will be developed and demonstrated safely first using model-in-the-loop (MIL) and hardware-in-the loop (HIL) simulations and then with a real vehicle using the VVE method of our research group at the Ohio State University. The scenarios for training and evaluation will be created by injecting other vehicle(s) and vulnerable road users with unexpected motion within realistic traffic generated by a microscopic traffic simulator. The project will also contribute to graduate student theses and undergraduate student senior projects. Project results will be integrated into the relevant parts of the OSU courses ECE 5553 Autonomy in Vehicles (undergraduate and graduate) and ME 8322 Vehicle System Dynamics and Control (graduate). Exemplary project results will be shared using our lab’s Youtube site. Anticipated invention disclosure(s) will be followed, if applicable, with potential provisional patent applications and subsequent patent applications by OSU’s Technology Commercialization Office.
TRID
We searched TRID using the paper/article query “collision avoidance AND connected autonomous driving AND deep reinforcement learning” which resulted in four recent papers. The first paper is on cooperative adaptive cruise control (CACC) of a platoon with deep reinforcement learning (DRL) for absorbing speed oscillation and improving fuel economy and is not relevant. The second reference is on DRL based collision avoidance but uses an actor-critic network as opposed to our double deep reinforcement learning (DDQN) with control Lyapunov/barrier function exploration. Also, their application is autonomous ships which is not relevant for our proposal. The third paper uses multi-agent reinforcement learning-based connected and autonomous vehicle (CAV) control for traffic efficiency improvement while changing lanes and is not relevant for the CAV collision avoidance of our proposal. The fourth and last paper optimizes the cooperative decision-making of CAVs using DRL. Their approach will not apply to the sudden and evasive maneuvering of emergency collision avoidance and to avoiding collisions with vulnerable road users who are inherently non-cooperative. While these references and their use of DRL are interesting and useful, this search has shown the uniqueness of our approach for improving the safety of road traffic.
Individuals Involved
| Email |
Name |
Affiliation |
Role |
Position |
| aksunguvenc.1@osu.edu |
Aksun-Guvenc, Bilin |
Ohio State University |
PI |
Faculty - Research/Systems |
| guvenc.1@osu.edu |
Guvenc, Levent |
The Ohio State University |
Co-PI |
Faculty - Tenured |
| redmill.1@osu.edu |
Redmill, Keith |
The Ohio State University |
Other |
Faculty - Research/Systems |
Budget
Amount of UTC Funds Awarded
$81535.00
Total Project Budget (from all funding sources)
$187647.00
Documents
Match Sources
No match sources!
Partners
| Name |
Type |
| City of Marysville and Union County, Ohio |
Deployment Partner_ Deployment Partner_ |
| Transportation Research Center |
Deployment Partner Deployment Partner |
