Neural Operators Plan Collision-Free Trajectories for Spacecraft Swarms

Autonomous spacecraft swarms face significant challenges in planning efficient, collision-free paths, especially as orbital congestion increases. Traditional optimization methods struggle with scalability due to the exponential growth of safety constraints. This paper proposes a novel permutation-equivariant neural operator designed to address these issues. The neural operator maps the distribution of spacecraft, targets, and debris directly to collision-aware trajectories for an entire swarm in a single forward pass. It is trained using self-supervised physics objectives and adversarial threats, without requiring optimal-trajectory labels. This approach allows it to generalize effectively from training on ten spacecraft to managing swarms of up to 1,000 spacecraft amidst over 11,000 cataloged objects, demonstrating superior accuracy and collision avoidance compared to existing methods.