This project explores a new vision of cyber-physical systems (CPSs) in which computing power and control methods are jointly considered. Our technical approach is in better understanding of computing hardware bounds, better data to build models of both computing platforms and CPS physics, new theory for feedback control algorithms that are aware of how computational limitations can be overcome with knowledge of the dynamical system, and experimental validation of the new theory on physical testbeds.
The novel computationally-aware results of this project can be made possible only by this significant research investment. Unless design changes are possible at the algorithm level, hardware designers must optimize for generic instructions. Unless computational performance is available as a model, control designers and CPS system integrators cannot trade off costs/benefits of new computing hardware. And unless open-experimental testbeds are available for data gathering and validation, theoretical results may be of limited impact, and translation to practice or to other CPS domains may be delayed.
Our results will permit advanced, verifiable control algorithms to run on less-sophisticated computing hardware, and with higher confidence than before. This is made possible through an integrated project in which understanding hardware performance is valued over generic hardware optimization, and in which CPS data are gathered at scale. These data are used to build models and controllers of computing platforms and CPS testbeds, allowing us to develop new methods of design-time and runtime switching.
If successful, the project opens new vistas in automotive, avionics, and other application domains in which computationally-intensive processes play a significant role at runtime. Automotive Original Equipment Manufacturers (OEMs) will be able to weigh whether updated software for advanced controllers or perception tasks can be deployed on vehicle models of the previous year. Regression analysis of candidate controllers can be concurrently tested in situ with deployed systems, without uploading significant amounts of data, if sufficient computing capacity is judged to be available based on runtime profiling. Algorithms that cannot currently be deployed due to lack of computing power, such as model predictive control, could be applied to provide robust autonomy solutions by improving the use of on-board energy sources and reducing the cost of the overall system, without compromising performance.