ESA GNC Conference Papers Repository
Enhancing the Guidance, Navigation and Control of Autonomous Parafoils using Machine Learning Methods
Artificial Intelligence techniques have developed into a transformative force across many industries. Their industrial adaption in aerospace Guidance, Navigation and Control (GNC) systems, however, has been rather limited to date. The Artificial Intelligence for Guidance, Navigation and Control (AI4GNC) project led by SENER Aeroespacial and funded by the European Space Agency (ESA) investigates the potential of several machine learning methods to enhance the performance and robustness of an aerospace GNC design. As a specific use case, we consider the descent and landing phase of an autonomous parafoil-guided return vehicle, inspired by the ESA Space Rider whose GNC software is developed by SENER. On this benchmark scenario, we demonstrate how a combination of machine learning methods can be used to significantly improve the performance of a baseline GNC design and gain insight into the system behaviour and its sensitivities. We investigate several complementing technologies at different hierarchical levels in the GNC and its design process and demonstrate the gained advantages on a full-complexity functional simulator, representative of industrial practice. At the intersection of guidance and controls, the project employs data-driven system identification to capture closed-loop system behaviour to serve as a basis for higher-level planning and guidance algorithms. Such models are typically cumbersome to derive from first principles since flight software, including lower-level controllers, actuator saturations and similar effects are part of the loop. In particular, neural networks which have been trained with an efficient deep-learning-based system identification method are used to augment an idealized baseline model which assumes perfect lower-level control. This is shown to effectively reduce residual errors while extending the region of validity compared to alternative linear variants and thereby provide an accurate system description to higher-level planning algorithms. Within the guidance layer, a robust trajectory planning technique is developed based on onboard optimization which can take numerous sources of uncertainty into account, such as wind conditions or uncertain system dynamics. The planning method relies on a novel extension of differential dynamic programming using results from robust control to formulate a sequence of semidefinite programs to find feedback & feedforward policies that efficiently steer the system despite the adversarial action of uncertainties. Extensive evaluations on the functional simulator show a clear hierarchy of achieved performances: (nominal) optimization-based guidance outperforms the baseline solution, while the novel robust variant shows the strongest performance. We furthermore present a developed GNC auto-tuner tool that utilizes Bayesian optimization (BO) to efficiently tune high-level GNC parameters for complex natural language constraints or objectives formulated in terms of temporal logic expressions. The use of Bayesian optimization enables a data-efficient stochastic black-box optimization of several key GNC parameters using a small number of (simulation) experiments. We further demonstrate that it is straightforward to employ the techniques in an antagonistic fashion leading to an effective worst-case-analysis tool. Our results show how such temporal logic-constrained BO can be efficiently used to improve system performance, explore parameter interdependencies and provide valuable insights to support the tuning of complex GNC systems. Finally, all developments are presented in a unified perspective highlighting synergies and sketching a general framework in which AI and data-driven techniques can contribute to the GNC discipline. This particularly highlights the increasingly central position of simulations, not only as a verification and validation tool but rather as an integral part of the GNC design process itself. With this, we envision a viable path forward towards the integration of AI techniques towards industrial practice, and towards realizing its considerable potential for the field.