Rsearchers at NYU Tandon have found a way for self-driving cars to share what they know about roads. This helps each car learn from others.

The researchers have described this study in an arXiv preprint and presented it at the AAAI Conference on Artificial Intelligence.

Their idea solves a tricky problem in artificial intelligence (AI) for self-driving cars. Usually, cars only share information when they pass close by, but that doesn’t happen much, so learning stays slow.

The researchers built a web of shared car experiences. A car from Manhattan could learn about Brooklyn roads without going there. This makes all cars smarter and ready for new situations.

Cached-DFL lets cars train their own AI models on board

The new method leverages Cached Decentralized Federated Learning (Cached-DFL). Federated Learning, a way machines learn together, often uses a central computer to manage updates. But Cached-DFL lets cars train their own AI models right on board. When cars get within 100 meters, they swap these models using fast device-to-device communication, not raw data. They can even share models from cars they met earlier, spreading info further. Each car keeps up to 10 models in a cache and refreshes its AI every two minutes.

To keep info fresh, the system drops old models when they pass a staleness threshold, a limit on how outdated something can be. The team tested this in a computer simulation using Manhattan’s streets. Virtual cars drove at 14 meters per second, turning at crossings with set chances. Unlike older methods that need cars to meet a lot, Cached-DFL spreads models through the group like messages in delay-tolerant networks (systems that save and send data when connections pop up).

Like in social media, where info jumps from friend to friend, cars act as messengers, sharing knowledge across the fleet. The method helps cars learn about roads, signs, and obstacles while keeping data private, especially in busy cities.

As AI shifts to edge devices, Cached-DFL offers a safe, fast way for cars to learn together. It could also help drones, robots, or satellites team up smarter. The researchers have shared their code for others to use.

Let us know your thoughts! Sign up for a Mindplex account now, join our Telegram, or follow us on Twitter.