CRII: OAC: RUI: Real-Time, Mixed-Integer Model Predictive Control via Learned GPU-Acceleration

Barnard College, New York NY

Investigators

Abstract

From self-driving cars to robotic home-health aids, in order for autonomous systems to meet their potential, they must operate safely around humans in unstructured and dynamic environments. This requires these systems to quickly and accurately solve motion planning and control problems. Unfortunately, many state-of-the-art algorithms used to solve these problems today are too slow to run in real-time, limiting such systems. This project helps alleviate these issues by leveraging parallel computing and machine learning to develop new solvers that accelerate the computation of optimization-based algorithms used for planning and control. This project addresses critical scientific needs for practical online planning and control for field robots and results in open-source solver artifacts that can be used in wider scientific computing domains such as operations research. This project also directly feeds into the development of new open-source robotics courses and, as the project is located at an undergraduate women’s college, this research also provides opportunities for a number of women undergraduates to participate in research - many for the first time. This project addresses the computational challenges of mixed-integer trajectory optimization problems, which are crucial for motion planning and control in autonomous systems operating in unstructured environments. The project builds on recent research that has shown that these algorithms can be accelerated through parallelism using Graphics Processing Units (GPUs) and machine learning. The project develops an open-source GPU-accelerated mixed-integer solver architecture. A learned parallel search heuristic that accelerates the outer branch-and-bound layer of the overall solver is developed by leveraging domain knowledge and machine learning. A GPU-accelerated direct trajectory optimization solver is also developed for the underlying continuous problem. This underlying solver takes advantage of the block-tridiagonal structure of the Schur Complement of the trajectory optimization problem through a novel symmetric stair preconditioner, a preconditioned conjugate gradient solver, and a block-factorization-based solver. The overall solver is evaluated and compared to state-of-the-art approaches through simulation and on a physical quadruped robot to demonstrate its effectiveness in generating dynamic locomotion behaviors. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.