Grant: Research Tools and Instruments

As cyber-physical systems (CPS) become increasingly integrated into daily life-from drones to self-driving cars and healthcare robots-the imperative to guarantee their reliable and effective function in unpredictable settings has escalated. Extensive real-world trials by autonomous vehicle firms like Waymo highlight the complexity of this task. The addition of machine learning (ML) and artificial intelligence (AI) elements further complicates the matter due to their opaque nature and adaptability. The planned Research Tools and Instruments (RTI) grant aims to advance groundbreaking studies in formal control methods for safety-critical CPS incorporating ML/AI. Ongoing partnerships between the PI, global experts, and Canadian enterprises have been crucial in pinpointing major challenges. This project seeks to leverage the computational capabilities of GPU clusters and integrate cutting-edge robotic technologies, offering scalable solutions to high-dimensional systems, ensuring the satisfaction of complex requirements, and safeguarding the use of ML/AI in CPS control. Our initiative centres on the acquisition of a compact GPU cluster, designed to accelerate computational processes inherent in machine learning, formal methods, and control. The cluster will enable real-time processing of large datasets, essential for addressing the dynamism of practical scenarios, especially when coupled with ML/AI components. What sets our research apart from others, such as the repetitive training of generative, diffusion, or large language models, is our use of GPU-enabled computation to conduct real physical experiments with robots. The robotic platform in the applicant's lab and Waterloo's RoboHub will serve as physical embodiments of CPS, allowing for a tangible assessment of safety-critical control algorithms developed using formal methods. A range of aerial and ground vehicles are available for integration with GPU-enabled CPS control, including Craziflies, TurtleBots, and Huskies. As such, the proposed research will contribute not only to a better understanding of the fundamental computational challenges in CPS control but also to a future where CPS can seamlessly and safely integrate into human-centric environments. The main research objectives are: 1) to build a GPU-enabled modelling framework for "pixel-to-action" control of safety-critical CPS using Koopman-Zubov operator learning; 2) to develop GPU-accelerated computational physics-informed neural networks (PINNs) for CPS control; 3) to test the GPU-in-the-loop modelling and control framework with real robotic platforms involving both moderate-scale ground robots and teams of UAVs. It is expected that the outcomes of the proposed research will significantly advance the state-of-the-art in safety-critical control of CPS with formal guarantees and will train highly qualified personnel (HQP) for Canada's CPS sectors, including robotics, automotive, and autonomous systems.