Moghadam, M. (CE) – Constraint-Aware Scene Understanding and Trajectory Generation Using Deep Reinforcement Learning for Autonomous Vehicles

Advanced driver-assistance systems (ADAS) are commonly organized as modular pipelines that transform raw sensor measurements into low-level actuation commands through perception, planning, and control. While learning-based methods have achieved state-of-the-art performance in perception and environment modeling, the planning layer remains a key bottleneck for reliable autonomy. Highway driving in particular requires long-horizon reasoning and socially aware interaction with multiple actors, while also producing smooth and dynamically feasible motion that can be tracked by classical controllers.

This thesis focuses on scene understanding and planning for highway driving. We study the problem through two complementary simulation environments: the high-fidelity CARLA simulator for motion planning and continuous trajectory generation under realistic vehicle dynamics and road geometry, and the lightweight HighwayEnv simulator for interaction-rich behavior planning at high episode throughput.

We present three planning contributions that increase autonomy. First, we introduce a modular hierarchical planning framework in Frenet space that combines long-term decision-making with short-term trajectory optimization. The approach includes a corridor-based dynamic obstacle avoidance strategy that generates spatiotemporal polynomial trajectories and supports diverse driving styles through interpretable parameter tuning. Second, we propose an end-to-end continuous deep reinforcement learning approach that unifies decision-making and motion planning into a single policy that outputs continuous polynomial trajectories in the Frenet frame. A spatiotemporal observation tensor and a temporal convolutional backbone enable the learned planner to exploit interaction history and outperform optimization-based and discrete RL baselines in CARLA. Third, we develop an interaction-aware behavior planning neural network architecture that couples trajectory prediction with high-level decision-making via a social pooling scene encoder built on actor histories and an ego-centered BEV representation. This unified design improves RL social awareness, safety, and overall driving performance in multi-agent highway scenarios in HighwayEnv.

Across extensive simulation studies, the results show that constraint-aware representations and learning-based policies can improve planning quality beyond hand-crafted objectives, especially when the policy is equipped with spatiotemporal social context while retaining classical feedback control for stable trajectory tracking. Finally, we provide supporting simulation and evaluation infrastructure, including observation tensor and neural network designs, BEV utilities, and scalable training and testing pipelines, to enable reproducible research on learning-based planning in interactive traffic.

Event Host: Majid Moghadam, Ph.D. Candidate, Computer Engineering 

Advisor: Gabriel Elkaim

Zoom- https://ucsc.zoom.us/j/95848602314?pwd=2jlktZ6BChlXcyqT3anX4ZuKrYV4wE.1

Passcode- 325939