OrbitTAMP: a language model translates instructions into physically valid space rendezvous maneuvers

Stanford researchers anchor the reasoning of an LLM in orbital planning modules: 98% of mission specifications correctly reconstructed with the best models.

Rendezvous and proximity operations between spacecraft are now planned by hand by engineers, who translate a high-level intention into safe and achievable trajectories. A work of the team of Marco Pavone and Simone D'Amico proposes OrbitTAMP, a framework that makes it possible to control this planning in natural language without losing the guarantee of physical validity.

The LLM proposes, planners check

The language model does not directly generate a trajectory: it converts the setpoint into a partial mission specification, chosen from a graph of reusable behaviors. Specialized planners then complete the decisions left open, before a trajectory optimization produces a plan compatible with orbital dynamics and operational constraints.

With state-of-the-art models, the architecture returns exactly to the expected specification in 98% of cases, significantly better than direct generation by the LLM. With a compact model of 9 billion parameters, a review step guided by an auditor increases this rate from 75% to 88%. The authors see this as an auditable basis for agents capable of planning maneuvers in orbit.