FORESEER: Recognize and utilize uncertainties by integrating data-based learning and symbolic feedback

Uncertainties resulting from intricate internal model uncertainties and external environmental disturbances significantly degrade robot planning and control performance. However, recognizing such persistently varying uncertainties in an explainable and lightweight manner is exceptionally challenging. We present two converged uncertainty prediction frameworks through the Fusion of Online Reactive Estimation and Sustained Experience Exploitation for Robots (FORESEER), enabling accurate prediction of two general kinds of uncertainties. Both frameworks feature properties of precision, lightweight, universality, and stability, in comparison with existing solutions. At first, a prediction algorithm for nonlinearly parametric uncertainties is developed by merging analytical basis learning with online symbolic adaptive estimation. Furthermore, an online prediction algorithm for more challenging composite uncertainties is proposed by seamlessly integrating learning-based feedforward and model-based/symbolic feedback observer. Benchmark comparisons on flying drones showcase the accuracy of the FORESEER on various real uncertainties including mass, aerodynamic drag, rain, and rope tension, leading to subsequent high-precision control. Moreover, an energy-saving and time-saving planning strategy is presented by utilizing the favorable wind. The developed algorithms hold the promising potential for direct combination with existing planning/control algorithms, promoting the environmental adaptability of robots.