[2Yin-B] Poster 4

[2Yin-B-23] Probabilistic Multimodal Spacecraft Trajectory Generation Using Transformer and Mixture Density Network

〇TOSHIHIKO YANASE¹, Akira Hatakeyama², Naoya Ozaki³ (1. Preferred Networks, Inc., 2. The Graduate University for Advanced Studies, 3. Institute of Space and Astronautical Science, JAXA)

Keywords:

Trajectory Generation,Transformer,Spacecraft,Machine Learning,Multimodality

This paper presents a preliminary study on multimodal time-series generation using a probabilistic model that combines a Mixture Density Network (MDN) with a Transformer. Specifically, rather than pursuing predictive accuracy, we conduct an empirical verification through numerical experiments to assess the modeling capability and applicability for explicitly handling the "multimodality of solutions" inherent in spacecraft trajectory data. In spacecraft trajectory generation, multiple valid trajectories may exist due to ambiguities in constraint conditions. However, conventional deterministic approaches tend to converge to the mean of the distribution or exhibit bias toward a specific solution. To address this issue, our proposed method adopts an architecture where the Transformer captures the time-series context, and the MDN outputs the probability distribution parameters (mixing weights, means, and variances) based on that context. In our experiments, we verified the behavior of the proposed model using simplified synthetic data designed to mimic the multimodal nature of spacecraft trajectories. The results confirmed that the proposed method, by introducing the MDN, can learn the underlying multimodal probability distribution and is capable of sampling diverse trajectory candidates, compared with deterministic methods.