Space robots must operate under unfamiliar dynamics, difficult sensing conditions, and limited chances for human intervention. A new critical review examines whether foundation models—large pretrained systems that can supply reusable knowledge for perception and behavior—can be made to work in that setting, with particular attention to orbital manipulation and proximity operations.

The review contends that what matters is how that knowledge is carried into physical action. Terrestrial comparisons show benefits from particular pretraining stages and model interfaces, but transfer to space dynamics and constrained onboard operation remains insufficiently tested. The paper also highlights event cameras for tracking motion in high-contrast lighting and spiking neural networks as a possible route to sparse local computation.

To close the gap, the authors propose an iterative sim-to-ground-to-space validation pathway, connecting physical tests with telemetry-based diagnosis and closed-loop revalidation. A satellite-capture example links observations to approach, contact, and recovery, and a protocol separates generalization from adaptation when target-domain interaction is limited. The conclusion is that foundation models already contribute, but the physical evidence needed to support autonomous space operation is still lacking.