To safely deploy autonomous vehicles on our roads, we first must ensure that they can accurately predict the movement of pedestrians, cyclists, and other fellow drivers. With so many people and vehicles on the roads today, though, behavior prediction can be a really daunting task. MIT researchers have developed a system to help in this challenge.
Dividing the problem into smaller parts
As part of their — deceptively simple — solution, the MIT team divided the problem of multiagent behavior prediction into smaller parts, enabling the computer to solve each one individually in real-time.
Using such a behavior prediction system, an autonomous car would first guess the relationship between two agents on the road, such as who has the right of way and who has to yield. The car’s computer then uses those relationships to anticipate the potential trajectories of multiple agents.
More accurate than existing models
The new technique has several advantages over other current machine learning models. According to the researchers, the new behavior prediction method was more accurate than that used by autonomous driving company Waymo. Plus, the MIT model used less memory since it broke the problem into simpler pieces.
“This is a very intuitive idea, but no one has fully explored it before, and it works quite well, says co-lead author Xin “Cyrus” Huang. “The simplicity is definitely a plus. We are comparing our model with other state-of-the-art models in the field, including the one from Waymo, the leading company in this area, and our model achieves top performance on this challenging benchmark. This has a lot of potential for the future.”
