Reinforcement learning-based adaptive dedicated lane planning for autonomous vehicles

As autonomous vehicles enter urban traffic systems, the allocation of Dedicated Autonomous Vehicle Lanes (DAVLs) becomes a strategic concern for transport operations and long-term infrastructure planning. This study develops a reinforcement learning framework that formulates DAVL management as a Markov Decision Process and applies Proximal Policy Optimisation to minimise congestion, energy consumption, and pollutant emissions on Seoul’s Naebu Ringway under varying AV market penetration rates. The results reveal non-linear and threshold-dependent responses: at low penetration (10%), premature DAVL allocation increases emissions and delays, whereas measurable benefits emerge only beyond 30–50%. At high penetration (≥50%), DAVL allocation improves traffic flow, increases average speed, reduces density, and lowers emissions by up to 33%. The study reframes DAVLs as both an operational tool and a planning instrument, providing insights for aligning automation readiness with sustainability goals, indicating that DAVL allocation is conditionally effective rather than universally beneficial.