Unified Prediction and Planning

Traditional autonomous driving pipelines treat trajectory prediction and motion planning as separate modules. Prediction forecasts what surrounding agents might do, while planning determines the ego vehicle’s path — but these systems rarely communicate effectively. My PhD research develops unified frameworks that jointly reason about prediction and planning, enabling more informed and safer driving decisions.

Key contribution: PlanTRansformer (PTR) — a Gaussian Mixture Transformer framework that integrates goal-conditioned prediction, dynamic feasibility, interaction awareness, and lane-level topology reasoning into a single architecture.

Aerial Imagery for Traffic Understanding

Bird’s-eye-view data from drones provides a unique perspective for understanding urban traffic dynamics. Unlike vehicle-mounted sensors, aerial imagery captures the full spatial context of an intersection, including occluded areas and complex multi-agent interactions. I develop datasets and methods that leverage this perspective for trajectory prediction and planning.

Key contribution: DeepUrban — a drone-based dataset for dense urban traffic scenarios with 3D traffic objects, map data, and scene context, demonstrating up to 44% improvement in prediction metrics when combined with existing benchmarks.

Interaction-aware Motion Forecasting

Predicting the future trajectories of traffic participants requires understanding how agents interact — vehicles yielding, pedestrians crossing, cyclists merging. My work models these multi-agent interactions explicitly, producing multimodal predictions that capture the inherent uncertainty of human behavior in traffic.