Researchers at the University of Michigan have found that incorporating both near-miss incidents and crash scenarios into autonomous vehicle (AV) training can improve the safety performance of driving algorithms by 90%, offering a potential path to faster development and validation of self-driving technologies.
The findings, published in Nature Communications, suggest that expanding the types of safety-critical data used during simulation could improve the effectiveness of artificial intelligence models while reducing the time required to validate autonomous driving systems.
Near-Miss Events Strengthen AI Training
Autonomous driving systems rely on machine learning algorithms trained using a combination of real-world driving data and computer simulations. Developers continually refine these models by identifying failures, retraining the software and evaluating its performance under new traffic conditions.
Henry Liu, director of both Mcity and the University of Michigan Transportation Research Institute (UMTRI), said this iterative process often creates new challenges.
“With AI, we have something called the seesaw problem—you find a problem, and then you run simulation variations to train to solve the problem.”
“Then, unfortunately, sometime after the training, another unanticipated side of the same problem arises, or even completely new problems that have never appeared before.”
Rather than relying primarily on crash data to improve performance, the researchers found that training algorithms with both collisions and near-miss events produced significantly better results.
“What we’ve learned is that the training is more effective when you’re utilizing data from both crashes and near-misses.”
“They are both safety-critical scenarios, and a near miss means the vehicle was able to maneuver through a situation successfully without a crash.”
According to Liu, near-miss events occur far more frequently than collisions during simulated testing, providing substantially more data for model development.
“In simulated testing, near misses occur a thousand times more often than crashes. Bundling failures and near misses improves overall performance dramatically.”
Faster Validation of Autonomous Driving Systems
The research team evaluated the approach at the Mcity Test Facility, where the revised training method delivered a 90% improvement in safety performance compared with conventional simulation approaches.
The study also builds on earlier University of Michigan research addressing what researchers describe as the “curse of rarity” in autonomous vehicle testing, where serious safety events occur too infrequently during real-world driving to efficiently train AI systems.
Previous work by the research team showed that artificial intelligence-based simulation techniques could reduce the amount of on-road testing required by approximately 99.9%, helping accelerate the development of higher levels of vehicle automation.
The researchers said improved simulation methods could support the industry’s efforts to demonstrate the safety of Level 4 and Level 5 autonomous driving technologies while increasing public confidence in self-driving vehicles.