Neural Projection Operators for Real-Time 6-DoF Powered Descent Guidance

This paper introduces a real-time trajectory optimization framework for six-degree-of-freedom (6-DoF) powered descent guidance using a neural network-based projection operator integrated within an alternating direction method of multipliers (ADMM) scheme. Conventional methods such as sequential convex programming (SCP) rely on iterative linearization and solving a sequence of convex programs, which can result in high computational cost and sensitivity to tuning parameters. These limitations pose challenges for onboard implementation in time-critical entry, descent, and landing (EDL) scenarios. To overcome this, we propose a learned projection operator that directly maps off-manifold points onto the surface defined by the nonlinear system dynamics. The projection network is trained using a geometric consistency loss that enforces both proximity and first-order stationarity on the dynamics manifold. Embedded into an Anderson-accelerated ADMM framework, this neural projection enables efficient enforcement of nonlinear dynamics constraints through a single forward-pass inference, significantly reducing per-iteration computation. Combined with a GPU-based implementation that exploits massively parallel projection and linear-algebra operations, our method achieves over an order of magnitude speed-up in computation time on a 6-DoF powered descent task, while maintaining accurate trajectory prediction and tight satisfaction of terminal and path constraints when compared against a well-tuned SCP baseline. Our GPU-based implementation achieves over an order of magnitude speed-up in computation time while maintaining accuracy in trajectory prediction and tight satisfaction of terminal and path constraints. The results highlight the method’s potential for deployment in onboard guidance systems requiring fast and reliable trajectory generation.