Researchers have built a system that lets language models direct spacecraft maneuvers without letting them anywhere near the actual flying.
The framework, called OrbitTAMP, tackles a real bottleneck: planning spacecraft rendezvous and proximity operations currently requires engineers to manually convert high-level mission intent into trajectories that are both safe and physically feasible. OrbitTAMP instead has a pretrained LLM map a natural-language command into a partial mission specification, then hands that off to domain-specific planners that fill in the operational details, and finally a trajectory optimizer that produces a dynamically feasible path. In testing, this hierarchical approach hit 98% exact recovery of partial mission specifications when paired with frontier LLMs, far outpacing direct LLM generation. For a smaller 9B-parameter model, adding verifier-guided revision pushed accuracy from 75% to 88%.
The real story here isn't the LLM - it's everything bolted around it. The gains come almost entirely from constraining the model's output into a structure that planners and optimizers can check and correct, not from the model getting smarter about orbital mechanics. That's a pattern worth watching: as companies rush to put language models in charge of physical systems, the safe version looks less like "trust the model" and more like "cage it in verifiable scaffolding."
It's a sensible instinct. Nobody wants a chatbot improvising a burn sequence near another spacecraft, and this architecture reads as a template other safety-critical robotics domains could borrow rather than a one-off space hack.