A thermodynamics-informed active learning approach to perception and reasoning about fluids

Abstract Learning and reasoning about physical phenomena is still a challenge in robotics development, computational sciences play capital role the search for accurate methods able to provide explanations past events rigorous forecasts of future situations. We propose thermodynamics-informed active learning strategy fluid perception from observations. As model problem, we take sloshing different fluids contained glass. Starting full-field high-resolution synthetic data particular fluid, develop method tracking (perception) simulation (reasoning) any previously unseen liquid whose free surface observed with commodity camera. This approach demonstrates importance physics knowledge not only data-driven (gray-box) modeling but also real-physics adaptation low-data regimes partial observations dynamics. The presented extensible other domains such as development cognitive digital twins learn observation which they have been trained explicitly.