ALMOND Targets Autonomous 6G Satellite Control

A research paper introduces ALMOND, a framework for autonomous control of 6G low Earth orbit satellite networks, where translating human requests such as “uninterrupted Arctic video” into reinforcement learning rewards can take days of manual trial and error. Existing large language model translators cannot learn from failures in the policies they generate.

ALMOND converts natural-language demands into multi-objective reward functions and analyzes failed reinforcement-learning trajectories to infer operational constraints, validating its findings counterfactually. It also uses a large language model to guide graph neuroevolution, which develops and critiques policy architectures. The ground-based, compute-frugal system sends satellites distilled policies below 50,000 parameters; inference takes under 15 ms, and satellites run no on-board large language model.

In a digital twin of 350 satellites using real ephemeris data, ALMOND raises polar handover success from 55% to over 95%, converges about three times faster than baseline reinforcement learning, and cuts reward engineering time from 48 hours to 4.5 hours.

A research paper introduces ALMOND, a framework for autonomous control of 6G low Earth orbit satellite networks, where translating human requests such as “uninterrupted Arctic video” into reinforcement learning rewards can take days of manual trial and error. Existing large language model translators cannot learn from failures in the policies they generate.

ALMOND converts natural-language demands into multi-objective reward functions and analyzes failed reinforcement-learning trajectories to infer operational constraints, validating its findings counterfactually. It also uses a large language model to guide graph neuroevolution, which develops and critiques policy architectures. The ground-based, compute-frugal system sends satellites distilled policies below 50,000 parameters; inference takes under 15 ms, and satellites run no on-board large language model.

In a digital twin of 350 satellites using real ephemeris data, ALMOND raises polar handover success from 55% to over 95%, converges about three times faster than baseline reinforcement learning, and cuts reward engineering time from 48 hours to 4.5 hours.