Anagrafe della ricerca

Trustworthy and long-term operations of autonomous robots in populated environments require the design of robust and interpretable control strategies. Two main kinds of controllers currently exist, on which the state of the art largely rests. On the one hand, data-driven controllers use machine learning in a black-box fashion, to provide robustness and generalization capabilities, thanks to the availability of large empirical data sets and the potentiality of Artificial Neural Networks (ANNs). At the other side of the spectrum, model-based controllers leverage the physical knowledge of the system, offering a more interpretable control action at the expense of typically high-dimensional models with limited granularity and adaptation capabilities. In this proposal, we devise a novel, robust and trustworthy control framework for autonomous robots, which seamlessly integrates the benefits of both data-driven and model-based approaches. Inspired to the emerging paradigm of Physics-Informed Machine Learning (PIML), the proposed control framework aims at attaining system robustness, as well as generalization capability and interpretability of learning-based models, with mathematically proved performance guarantees. The framework consists of three components: (i) a physics-informed surrogate model trained with an ad-hoc loss function to ensure physical consistency in the learning of the dynamical model of the robot, (ii) a residual Deep Neural Network (DNN) to provide real-time robustness and adaptability to unknown dynamics, and (iii) a nonlinear predictive controller with performance guarantees, which uses the predictions offered by the previous two models. The successful realization of this project will not only ensure trustworthy operations for safe-critical robotic systems (such as UAVs operating in densely populated areas), but will also attain the broader goal of advancing the state of the art in robust control of highly nonlinear and uncertain dynamical systems.