MORE | Spring 2024

Hardware-Net

Machine learning or gradient-based optimization with hardware in the loop typically requires differentiation through the hardware layer. For example, in the optimal control of a hardware system, finding an optimal system trajectory requires differentiation through unknown dynamics of the hardware, which is a challenge. In this project, differentiation through the hardware layer is possible using randomized smoothing. However, this leads to significant data inefficiency when incorporating into machine learning or optimization-based frameworks. Therefore, the research team proposes using a neural network gradient estimator as a proxy for such randomized smoothing for hardware layer differentiation.  Focus will be on the application of the proposed method in optimal control of a hardware robot.