Wu, C., Kreidieh, A.R., Parvate, K., Vinitsky, E., Bayen, A.M.
This paper introduces Flow, a computational framework that uses deep reinforcement learning to optimize traffic control in mixed-autonomy environments where autonomous vehicles (AVs) share the road with human drivers. The key insight is that even a small percentage of intelligently controlled AVs can dramatically improve traffic flow for everyone. The framework enables researchers to train AI controllers that learn optimal driving behaviors through simulation, then deploy these in real-world scenarios.
Just 5-10% of vehicles need to be autonomous to see significant improvements.
Average traffic speed improvements in congested scenarios.
Uses deep reinforcement learning for optimal control policies.
Mixed autonomy works: You don't need 100% autonomous vehicles - a small percentage can stabilize traffic for everyone.
Scalable simulation: The Flow framework enables rapid testing of control strategies before real-world deployment.
Transferable results: Controllers trained in simulation successfully transfer to real traffic scenarios.
Open source: The framework is publicly available, enabling further research and development.
SmartRoads applies these same principles - we don't need every vehicle to be equipped. By coordinating a small percentage of vehicles with RTK precision, we can eliminate congestion for all road users. The Flow framework validates our core approach: intelligent coordination of the few benefits the many.