Differential Model Predictive Control for Autonomous Driving in Under-structured Traffic Environment
In recent years, rapid growth in self-driving vehicles was largely enabled by breakthroughs in computer vision and reinforcement learning (RL). Current research focuses on applications of RL-based, model-free methods through simulation rather than classical model-based optimization. However, model-free algorithms often fail to generalize beyond the environment in which it was trained. For instance, self-driving cars trained in a regular urban environment will suffer to drive safely and efficiently in road environments in developing countries where traffic signals do not exist, roads are under-structured, and drivers drive under implicit rules. Subsequently, for autonomous vehicles, navigating safely but efficiently in under-structured, under-regulated road environment has remained as a major challenge. Thus, deployment of self-driving cars in developing countries is challenging, as it is hard for RL models to find the sweet spot between ensuring safety and avoiding congestion or delays, in such a traffic environment, with insufficient training data and poor generalization.
With a traffic trajectory dataset in under-regulated local traffic (intersections and highways) that my team collected, I aim to model human drivers behaviors and compute safe and efficient trajectory planner in such traffic environments by using differential Model Predictive Control (MPC), a state-of-the-art model-based controller, to both learn dynamics and plan trajectories by differentiating optimization problems in continuous space, which will mitigate problems in current model-based and model-free algorithms, given real-world trajectory data.
Message to Sponsor
- Major: Astrophysics
- Sponsor: Chen Fund
- Mentor: Eric Norman