Car talk: will teaching vehicles to communicate with each other reduce accidents?

Kyunghyun Cho, Assistant Professor of Computer Science and Data Science, investigates vehicle-to-vehicle communication

While modern technology has helped us build robust accident-avoidance systems for cars, we’re still witnessing a high volume of highway related fatalities. Human error often plays a role in these accidents, which is why scientists are continually developing new solutions to help cars compensate for our blind spots.

To improve road safety, CDS’s Kyunghyun Cho, Assistant Professor of Computer Science and Data Science, along with researchers from NYU and FAIR (Facebook Artificial Intelligence Research), have recently been exploring the possibility of teaching vehicles how to communicate with each other.

Their research is timely since the US Department of Transportation (DOT) has recently proposed a rule that requires all new cars to have vehicle-to-vehicle communication capabilities by 2023.

“The approach,” the researchers explain, “is that vehicles will issue messages alerting each other of potential safety concerns so as to act upon these messages and avoid accidents.”

The details of the DOT’s proposal, however, remain unclear since we have not yet discovered exactly what type of information cars would need to send each other. Speed? Acceleration? Location? Moreover, while teaching vehicles to talk to each other seems to be a theoretically sound idea, would such a mechanism have a meaningful impact on road safety?

To begin answering these questions, the researchers used Box2D, a Python gaming framework, to simulate highway scenarios where cars are capable of direct vehicle-to-vehicle communication (V2V), and can transmit three specific types of information to each other: acceleration, steering wheel angle, and brake.

To gauge how effective V2V actually is, the researchers also used the simulation to construct baseline scenarios (where cars are not capable of communicating with each other), so that they could compare the success rate of cars with V2V and those without.

In the simulation, all the cars are aiming to drive as fast as possible to a highway exit without crashing into each other.

“Every episode consists of twelve cars,” state the researchers. “The vehicle’s size, exit, and starting position are selected randomly from five, two, and fifteen possibilities respectively. The length and angle of the highway also change randomly at each episode, and there are no lane markings, which makes driving even more challenging. These choices ensure diversity in driving scenarios.”

As the researchers were also working within the context of reinforcement learning (a machine learning technique that develops models by training them repeatedly on a scenario so that they improve with each iteration), they trained their cars under two different weather conditions — sunny and foggy.

“Our current results,” the researchers reported, “suggest that the proposed protocol will in fact make the road safer in adverse conditions and enable cars to go faster.”

Of course, the researchers note that a major drawback to their simulation is that it is “a limited interpretation of actual driving scenarios.” Yet, their work confirms is that V2V capabilities can become effective tools for improving road safety. After all, as the researchers suggest, developing V2V capabilities may lead cars to discover better solutions that scientists themselves have not even thought of.

“We have seen this recently with AlphaGo finding moves in Weiqi that human players had not discovered after many centuries,” the researchers point out. “In future research, we will compare the model’s learned policies to fixed safety responses.”

By Cherrie Kwok

Like what you read? Give NYU Center for Data Science a round of applause.

From a quick cheer to a standing ovation, clap to show how much you enjoyed this story.