Calibrated Decision Models for Onboard Spacecraft Autonomy

Kurzbeschreibung

Beschreibung An autonomous spacecraft has to decide, minute by minute, whether to charge, observe, process data, downlink or stay safe. Generative LLM agents can take such decisions from a textual state description, but each answer is slow, costly and carries no trustworthy statement of its own reliability. System One decision models instead answer typed questions in a single forward pass and return a calibrated probability for every option. Trustworthy confidence could decide when the spacecraft acts on its own and when it defers to a safer controller or to the ground. Building on our work on autonomous satellite decision-making, the decision-model branch of our AUTOPS framework already connects the open-weight Kev model family (0.8B to 27B parameters) to an Orekit-based simulator of the EventSat 6U CubeSat. First zero-shot closed-loop runs fly complete simulated weeks and expose the open problems this thesis attacks: the energy cost of the decision cadence, a possible sensitivity to option order, confidence that is not yet calibrated for the mission, and a clear gap to a rule-based scheduler. The student works with the AUTOPS codebase, a data-centre GPU, an NVIDIA Jetson Orin platform and close supervision towards a joint publication. Voraussetzungen • Excellent academic record and the ambition to do publishable research • Solid machine learning background; hands-on experience with PyTorch and transformer • language models (e.g. LoRA fine-tuning) is a strong plus • Very good Python and software engineering skills (Git, testing, Linux, remote GPU servers) • Good command of probability and statistics (calibration, uncertainty, bootstrap) • Basic knowledge of spacecraft operations (power, attitude, communication); experience with imitation learning, reinforcement learning or planning is a plus