Multi-vehicle Collaborative Navigation using Machine Learning and Model Predictive Control
Francesco Borrelli, Professor
Mechanical Engineering
Applications for Fall 2024 are closed for this project.
In this project, we aim to achieve decentralized multi-vehicle collaborative navigation in an unsignalized intersection scenario. We will utilize an existing (custom) simulation environment to collect data. The data come from the optimal solution of the centralized optimal control problem or the nash equilibrium solution of the potential game (two are equivalent in our formulation). We will use various machine learning techniques to train a neural network that captures the multi-vehicle collaborative strategies.
In deployment, the neural network outputs the parameters of the cost function of the model predictive controller to guide each vehicle to follow the collaborative strategies in a decentralized manner.
Role: Students will work on collecting data, building and tuning neural network model architectures and parameters using Pytorch. Further, students will evaluate the model and co-analyze the model's open-loop performance to the closed-loop performance in simulation.
Qualifications: Python knowledge is mandatory. Experience with Pytorch or machine learning is preferred.
Day-to-day supervisor for this project: Hansung Kim, Graduate Student
Hours: 9-11 hrs
Mathematical and Physical Sciences Engineering, Design & Technologies