(2024)3:36 | Adolfo Perrusquia, Weisi Guo, Benjamin Fraser, Zhuangkun Wei
The paper introduces a control-physics informed machine learning (CPhy-ML) framework to robustly infer drone intentions, which are often intangible and unobservable. The framework combines deep learning with aerospace model conservation laws to reduce bias and instability, achieving a 48.28% performance improvement over traditional trajectory prediction methods. The CPhy-ML framework is designed to infer drone intentions by evaluating the connection between the drone's purpose and its observed mission profile, enhancing the confidence of predictions. The framework includes a hybrid classifier for intention class prediction and anomaly detection, a deep mixture of experts for trajectory regression, and a reservoir computing network for trajectory prediction. The reward function inference is also addressed, using an off-policy model-based reward-shaping inverse reinforcement learning architecture to uncover the hidden reward function. The results demonstrate the effectiveness of the CPhy-ML framework in predicting drone intentions and improving the robustness and stability of the learning manifold. However, the framework is limited by the amount of data and its variability, and further research is needed to address these limitations.The paper introduces a control-physics informed machine learning (CPhy-ML) framework to robustly infer drone intentions, which are often intangible and unobservable. The framework combines deep learning with aerospace model conservation laws to reduce bias and instability, achieving a 48.28% performance improvement over traditional trajectory prediction methods. The CPhy-ML framework is designed to infer drone intentions by evaluating the connection between the drone's purpose and its observed mission profile, enhancing the confidence of predictions. The framework includes a hybrid classifier for intention class prediction and anomaly detection, a deep mixture of experts for trajectory regression, and a reservoir computing network for trajectory prediction. The reward function inference is also addressed, using an off-policy model-based reward-shaping inverse reinforcement learning architecture to uncover the hidden reward function. The results demonstrate the effectiveness of the CPhy-ML framework in predicting drone intentions and improving the robustness and stability of the learning manifold. However, the framework is limited by the amount of data and its variability, and further research is needed to address these limitations.