Elements: RUI: Accessible GPU-Accelerated Edge Optimal Control Library and Benchmarks

Nonlinear optimal control solvers are used in a diverse range of applications from robotics to manufacturing and utilities. Unfortunately, current software infrastructure fundamentally limits the performance of many of these systems due to its inability to effectively scale to large-scale problems. At the same time, advances in parallel computational hardware have shown promise for addressing these limitations. This project overcomes such computational challenges through acceleration on Graphics Processing Units (GPUs), and develops more general, accessible, and documented open-source GPU optimal control libraries that can support a broader range of scientific research, with a focus in robotics. These libraries are paired with open-source benchmark problems and datasets and integrated into machine learning (ML) frameworks to enable fair evaluations of new algorithms and to enable broader participation in this interdisciplinary field. Finally, this project develops an integrated educational curriculum that provides background knowledge, enabling researchers and practitioners worldwide to learn about these topics, leverage this cyberinfrastructure to improve their systems, and contribute to the current project. This project addresses fundamental shortcomings in current software libraries for optimal control, which do not consider the use of GPU-acceleration, and are often not compatible and comparable with each other. As such, this work addresses critical scientific needs for low-latency optimal control at the edge as well as unified APIs and benchmark problems and datasets. This project expands upon and generalizes existing proof-of-concept, open-source GPU-accelerated optimal control solvers for robotics enabling them to be broadly used across both the robotics domain, as well as for other optimization tasks. This not only includes support for general purpose constraints and supporting kernels commonly found in robotics and beyond, but also wrappers in high level languages and integration with popular machine learning (ML) frameworks. Open-source benchmark problems and datasets as well as unified APIs are provided to: enable the optimal control community to fully and fairly evaluate novel algorithms and implementations; enable low-barriers to entry for contributions from researchers and practitioners worldwide; and avoid the current unnecessary development of bespoke libraries by individual research groups and organizations. The team evaluates this approach by tracking the number of projects, researchers, educators, unique applications, and subfields using, citing, and contributing to this software, courseware, and benchmarks. 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.