BYLINE: Sydney Barrilleaux
Self-driving cars rely on artificial intelligence to predict where nearby cars will go. But when those predictions don’t match reality, that discrepancy can potentially lead to crashes and less safe roadways.
That’s why a recent study from the University of Georgia developed a new AI model to make self-driving cars safer.
This study introduces an AI model for self-driving cars, designed to predict the movement of nearby traffic and incorporate innovative features for planning safe vehicle movements.
"The planned trajectory of the self-driving car may turn out to collide with the actual trajectory of another vehicle.” —Qianwen Li, College of Engineering
The study used data from the I-75 freeway in Florida to predict other cars’ paths and determine the motion of the self-driving car when following another vehicle.
Previous research mostly predicts surrounding traffic movements and then plans a self-driving car’s motion. This separate approach, however, makes crashes and near-misses more likely.
“That’s why we wanted to consolidate those two steps — to make the autonomous vehicle operation safer,” said Qianwen Li, lead author of the study and an assistant professor in UGA’s College of Engineering. “And as illustrated by our experiments, that approach does help with safety performance.”
AI needs to do more than predict traffic
To keep drivers safe, self-driving cars have to be able to accurately anticipate the movements of surrounding traffic. However, it’s difficult to know what other drivers will do on the road.
“There are always differences between your prediction and the reality,” said Li. “The planned trajectory of the self-driving car may turn out to collide with the actual trajectory of another vehicle.”
The new model was designed to take prediction errors into account, as eliminating them isn’t possible.
Li’s group is also working on developing more complex AI models for self-driving car operations, such as large learning models like ChatGPT. Traffic scenarios could be fed to these models, and they would determine the best course of action.
However, large language models have limits. While they’re effective at making high-level decisions related to how to respond to different situations, planning the movements of a car isn’t what they’re built for.
“How do we make a perfect lane change that is safe and also efficient?” said Li. “How do we come to a smooth stop for pedestrians without inducing any riding discomfort? Basically, how do we design the specific trajectories? That part we do not ask ChatGPT or large language models to do because they do not have the capability to do so. Traditional trajectory optimization models can do a much better job based on our experiments so far.”
Balancing safety and mobility in vehicle artificial intelligence
Designing AI for self-driving cars is a balancing act. Maximizing safety often comes at the cost of mobility.
If a self-driving car is taught to drive as safely as possible, for example, it will stay far behind the car in front of it. While a safer option, that distance would likely reduce the number of cars that could fit on the roadway at a given time.
Similarly, focusing too much on mobility could result in cars driving too aggressively, increasing the risk of crashes.
“We’re still working on how we train the model in a way that can balance the safety and mobility performance,” said Li.
The study was published in Transportation Research Part E. Co-authors include Handong Yao of UGA’s College of Engineering, Xiaopeng Li of University of Wisconsin-Madison’s Department of Civil and Environmental Engineering, and Chenyang Yu of McGill University’s Mathematics and Computer Science Department.