Reproduce “Trajectory Planning for Autonomous Vehicles Using Hierarchical Reinforcement Learning” on OpenResearch

About this paper

Kaleb Ben Naveed, Zhiqian Qiao, John M. Dolan

Researchers from The Hong Kong Polytechnic University and Carnegie Mellon University developed a Hierarchical Reinforcement Learning framework for autonomous vehicle trajectory planning, integrating a PID controller for smooth execution and an LSTM layer for robustness against noisy observations. This combined approach achieved a 100% success rate without noise and 98% with noise in a lane change scenario, significantly reducing collisions and lane invasions compared to alternative methods.

Key insights

• Decomposing complex driving tasks into hierarchical sub-goals using HRL significantly accelerates learning convergence and improves sample efficiency for autonomous vehicle trajectory planning.
• Integrating a classical PID controller with a learning-based planner effectively ensures smooth trajectory tracking, leading to 0% lane invasion rates and reducing jerk for enhanced safety and comfort.
• The use of an LSTM layer in the HRL architecture provides robustness against noisy and incomplete observations, allowing the autonomous system to maintain high performance in dynamic and uncertain environments.

Summary by alphaXiv.