Berkeley Researchers Deploy 100 RL-Controlled AVs to Smooth Highway Traffic
Researchers from the Berkeley Artificial Intelligence Research (BAIR) lab successfully deployed 100 autonomous vehicles (AVs) controlled by reinforcement learning (RL) algorithms into rush-hour highway traffic. The primary objective was to mitigate "stop-and-go" waves, also known as phantom jams, which are caused by amplified human driving fluctuations and lead to congestion, increased fuel consumption, and higher CO2 emissions. By utilizing fast, data-driven simulations trained on real-world data from Interstate 24, the team developed decentralized controllers that allow AVs to smooth traffic flow using standard radar sensors. The experiment demonstrated that a small proportion of intelligently controlled AVs can significantly improve energy efficiency and throughput for all drivers, including those in human-driven vehicles. This approach offers a scalable alternative to traditional infrastructure-heavy solutions like ramp metering. The findings, detailed in a recent paper, highlight the practical challenges and successes of transitioning RL controllers from simulation to large-scale field deployment, marking a significant step forward in mixed-autonomy traffic management.
Wire timeline
Berkeley Researchers Deploy 100 RL-Controlled AVs to Smooth Highway Traffic
Researchers from the Berkeley Artificial Intelligence Research (BAIR) lab successfully deployed 100 autonomous vehicles (AVs) controlled by reinforcement learning (RL) algorithms into rush-hour highway traffic. The primary objective was to mitigate "stop-and-go" waves, also known as phantom jams, which are caused by amplified human driving fluctuations and lead to congestion, increased fuel consumption, and higher CO2 emissions. By utilizing fast, data-driven simulations trained on real-world data from Interstate 24, the team developed decentralized controllers that allow AVs to smooth traffic flow using standard radar sensors. The experiment demonstrated that a small proportion of intelligently controlled AVs can significantly improve energy efficiency and throughput for all drivers, including those in human-driven vehicles. This approach offers a scalable alternative to traditional infrastructure-heavy solutions like ramp metering. The findings, detailed in a recent paper, highlight the practical challenges and successes of transitioning RL controllers from simulation to large-scale field deployment, marking a significant step forward in mixed-autonomy traffic management.
The Berkeley Artificial Intelligence Research Blog