Prince EDORH, Bruno HERISSE
DOI Number: N/A
Conference number: HiSST-2025-191
Hypersonic re-entry guidance poses significant challenges due to the sensitivity of optimal trajectories to dispersions in initial conditions and vehicle dynamics. Traditionally, guidance systems rely on precomputed optimal reference trajectories to mitigate these dispersions, employing on-line tracking algorithms to track to the nominal path. However, for large dispersions in initial conditions, this approach necessitates to embark either extensive databases of representative trajectories or robust online trajectory re-planning capabilities, both of which entail high computational or storage demands. Recent approaches have explored data compression techniques—such as Bézier curves or optimal
shooting points — to reduce the storage required for representing optimal trajectories and commands. In this paper, we demonstrate that the relationship between stored representative trajectory data and initial condition dispersions can be effectively learned and subsequently leveraged onboard. Artificial neural networks have been trained offline using a limited number of optimal trajectories within an initial dispersion box. The trained model is then used online to quickly recompute an initial reference trajectory suitable for guidance algorithms such as Bézier curves-based guidance (BCBG) or Proportional Navigation (PN). This approach enables coverage of a large dispersion box in Monte Carlo simulations while satisfying precision requirements and various path constraints significantly reducing computational and storage demands while maintaining robustness to dispersions.