Self-driving cars, like the Volkswagen Herbie and KITT from Knight Rider, are gradually transitioning from science fiction to reality. The benefits are clear: passengers can use their commute to relax, work, or be entertained while reducing accidents caused by human error. Additionally, autonomous vehicles offer increased mobility for individuals unable to drive themselves.
However, giving up control in the complex environment of road traffic requires highly advanced technology. Ongoing developments aim to bring fully autonomous vehicles to our roads, and one key area of focus is how these cars can communicate effectively — such as sharing updates on road conditions — to enhance safety and efficiency.
A research team from New York University (NYU) Tandon School of Engineering has developed a system to improve communication between autonomous vehicles, similar to how people interact on social networks. Their advancements were presented at the Association for the Advancement of Artificial Intelligence Conference on February 27, 2025.
The Current State of Autonomous Cars
Self-driving vehicles rely on sensors, cameras, and artificial intelligence (AI) to make informed decisions and navigate roads with minimal human input. The Society of Automotive Engineers classifies vehicle automation into six levels, from 0 (fully manual) to 5 (fully autonomous, meaning the vehicle can drive itself in all conditions without human intervention).
So far, no self-driving car has achieved full autonomy. The most advanced models, such as self-driving taxi services in California and Arizona, currently operate at Level 4. However, widespread adoption faces challenges, including the potential for accidents and concerns over data privacy.
AI enables autonomous vehicles to exchange knowledge when they interact directly, improving road navigation instantly. However, conventional model-sharing methods rely on immediate, one-on-one exchanges, which slows adaptation to new conditions. This is similar to how humans would struggle to spread information efficiently if they had to meet every recipient in person rather than passing messages through others.
Read More: Driverless Cars Still Have Blind Spots. How Can Experts Fix Them?
Making Car-to-Car Communication More Efficient
To overcome this limitation, researchers have introduced a new approach called Cached Decentralized Federated Learning (Cached-DFL). This method enhances how vehicles learn from one another, even if they seldom cross paths. Unlike traditional Federated Learning, which depends on a central server for updates, Cached-DFL allows vehicles to train AI models independently and exchange them directly.
When two vehicles come within close range — about 100 meters — they use high-speed communication to share trained models rather than transmitting raw data. This significantly accelerates adaptation and enhances learning efficiency compared to earlier decentralized approaches.
"It's a bit like how information spreads in social networks," explained Yong Liu, professor at NYU Tandon’s Electrical and Computer Engineering Department and supervisor of the project in a press release. "Devices can now pass along knowledge from others they've met, even if those devices never directly encounter each other."
Better Communication Enhances Safety
Cached-DFL addresses the challenge of enabling autonomous vehicles to learn from one another while maintaining data security. With this technology, self-driving cars can share crucial information about road conditions, signals, and obstacles — especially beneficial in urban areas where vehicles experience diverse conditions but seldom interact long enough for conventional learning methods to be effective.
"A car that has only driven in Manhattan could now learn about road conditions in Brooklyn from other vehicles, even if it never drives there itself. This would make every vehicle smarter and better prepared for situations it hasn't personally encountered,” Liu added.
As AI shifts from centralized servers to edge devices, Cached-DFL offers a secure and efficient way for self-driving cars to evolve collectively, improving their intelligence and adaptability. Moreover, this technology extends beyond autonomous vehicles; it can be applied to other networked systems of smart mobile agents — such as drones, robots, and satellites — to achieve decentralized learning and swarm intelligence. With researchers making their code publicly available, these advancements have the potential to accelerate innovation across multiple industries.
Article Sources
Our writers at Discovermagazine.com use peer-reviewed studies and high-quality sources for our articles, and our editors review for scientific accuracy and editorial standards. Review the sources used below for this article:
Cornell University. Decentralized Federated Learning with Model Caching on Mobile Agents
Having worked as a biomedical research assistant in labs across three countries, Jenny excels at translating complex scientific concepts – ranging from medical breakthroughs and pharmacological discoveries to the latest in nutrition – into engaging, accessible content. Her interests extend to topics such as human evolution, psychology, and quirky animal stories. When she’s not immersed in a popular science book, you’ll find her catching waves or cruising around Vancouver Island on her longboard.
