Diffusion Policies for Generative Modeling of Spacecraft Trajectories

TL;DR

This paper tackles online, constraint-aware trajectory generation for spacecraft, focusing on 6-DoF powered descent. It introduces a compositional diffusion framework that can sample and plan trajectories while integrating diverse constraints and priors at inference time, including multi-landing-site risk maps, via energy-based diffusion and inpainting. The authors train a 6-DoF diffusion model from SCvx-derived data, then demonstrate flexible constraint enforcement through glideslope negation composition, state/control conditioning, and multi-landing-site sampling, achieving feasible trajectories faster and with lower constraint violations than unconstrained baselines. The approach offers a practical, modular toolkit for autonomous space missions, enabling efficient initial guesses and constraint-compliant planning without retraining for each new specification.

Abstract

Machine learning has demonstrated remarkable promise for solving the trajectory generation problem and in paving the way for online use of trajectory optimization for resource-constrained spacecraft. However, a key shortcoming in current machine learning-based methods for trajectory generation is that they require large datasets and even small changes to the original trajectory design requirements necessitate retraining new models to learn the parameter-to-solution mapping. In this work, we leverage compositional diffusion modeling to efficiently adapt out-of-distribution data and problem variations in a few-shot framework for 6 degree-of-freedom (DoF) powered descent trajectory generation. Unlike traditional deep learning methods that can only learn the underlying structure of one specific trajectory optimization problem, diffusion models are a powerful generative modeling framework that represents the solution as a probability density function (PDF) and this allows for the composition of PDFs encompassing a variety of trajectory design specifications and constraints. We demonstrate the capability of compositional diffusion models for inference-time 6 DoF minimum-fuel landing site selection and composable constraint representations. Using these samples as initial guesses for 6 DoF powered descent guidance enables dynamically feasible and computationally efficient trajectory generation.

Figures (19)

• Figure 1: Forward and reverse processes for a generative diffusion model.
• Figure 2: Sampled trajectories for 2D powered descent guidance around an obstacle, the gray circle. As this example illustrates, diffusion models are a powerful policy representation that can efficiently learn the inherently multi-model structure of the underlying task and thereby significantly reduce constraint violation compared to a "vanilla" dnn approach.
• Figure 3: By composing the trajectory dynamics distribution, shown in blue, with the Dirac delta perturbation function for satisfaction of the start and goal constraints, the compositional planning distribution, shown in red, allows for constraint-satisfying sampling without retraining.
• Figure 4: Diffusion model sampled trajectories and optimizer-generated trajectories for 6 DoF multi-landing site powered descent guidance.
• Figure 5: Comparison of the means and trajectory distributions for 1,000 diffusion model sampled trajectories and optimizer-generated trajectories for 6 DoF multi-landing site powered descent guidance. 50 out of the 1,000 sampled trajectories for each model are randomly selected and plotted with the means.