Autonomous driving in complex urban environments requires a planner to reason about interactions with surrounding traffic while reacting at real-time rates. Large-scale imitation learning can generalize across diverse scenarios, but common generative planners face a difficult trade-off between behavioral diversity, inference speed, and trajectory quality.

Our research investigates how end-to-end learning can enable intelligent and adaptive

autonomous driving. We develop models that combine information about the road, nearby

vehicles, and the driving objective to predict future interactions and generate an appropriate vehicle trajectory.

A central focus of our work is bridging the gap between learned driving behavior and reliable physical execution. We explore approaches that improve responsiveness, trajectory quality, passenger comfort, and consistency with vehicle constraints while maintaining real-time performance. Ultimately, we aim to enable autonomous driving systems that can perceive, anticipate, plan, and act safely in complex real-world traffic environments.