Scholarlink

<p>Deep learning-based visual pose estimation systems have demonstrated strong performance for spacecraft rendezvous and proximity operations but remain challenging to deploy onboard due to their computational demands. This paper introduces a recursive six-degree-of-freedom dual-quaternion neural network–augmented estimator (DQ-MEKF–NN) that integrates physics-informed learning with a dual-quaternion multiplicative extended Kalman filter (DQ-MEKF). The proposed architecture employs two lightweight neural networks operating in parallel to estimate relative translation and attitude, each recursively ingesting pose measurements and reformulated dual-quaternion states propagated by the DQ-MEKF. Online learning is achieved through composite data-driven and kinematic consistency losses minimized using backpropagation and the AMSGrad adaptive optimizer. Monte Carlo simulations representative of real-world chaser–target rendezvous scenarios demonstrate that the proposed DQ-MEKF–NN significantly improves estimation accuracy over a baseline dual-quaternion filter, achieving up to 97% reduction in relative position root-mean-square error and up to 30% reduction in relative attitude error while maintaining estimator stability and computational tractability.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Neural Net Based Dual Quaternion Stochastic Filtering for Relative Spacecraft Motion Estimation

  • 期刊论文

  • 发表日期：

    2026年5月12日

查看全文
  • Warren Soh,
  • Regina Lee

摘要

Deep learning-based visual pose estimation systems have demonstrated strong performance for spacecraft rendezvous and proximity operations but remain challenging to deploy onboard due to their computational demands. This paper introduces a recursive six-degree-of-freedom dual-quaternion neural network–augmented estimator (DQ-MEKF–NN) that integrates physics-informed learning with a dual-quaternion multiplicative extended Kalman filter (DQ-MEKF). The proposed architecture employs two lightweight neural networks operating in parallel to estimate relative translation and attitude, each recursively ingesting pose measurements and reformulated dual-quaternion states propagated by the DQ-MEKF. Online learning is achieved through composite data-driven and kinematic consistency losses minimized using backpropagation and the AMSGrad adaptive optimizer. Monte Carlo simulations representative of real-world chaser–target rendezvous scenarios demonstrate that the proposed DQ-MEKF–NN significantly improves estimation accuracy over a baseline dual-quaternion filter, achieving up to 97% reduction in relative position root-mean-square error and up to 30% reduction in relative attitude error while maintaining estimator stability and computational tractability.