The most telling sessions at the 2025 SpaceOps conference were not about rockets but about software. Researchers presented an MLOps framework for machine learning aboard spacecraft — the deployment discipline of modern web engineering, rebuilt for machines that fly.
The problem it formalizes is new to spaceflight. Traditionally, a spacecraft runs the software it was launched with; updates are rare, audited events. Machine learning breaks that model. Models drift as conditions change, need retraining on ground data, and must be re-validated — compressed, quantized, regression-tested — before being pushed uphill to an embedded target that cannot be visited.
The framework borrows shamelessly from the cloud world: versioned training pipelines, shadow deployments, staged rollout across a fleet, telemetry on model performance itself. What it adds is the aerospace layer — radiation-tolerant targets, constrained power and memory, and certification.
The authors point toward architectures where fleets of satellites behave like distributed data centers: spacecraft that run analytics on their own sensor data and coordinate what they send home.
For ML engineers, this is the frontier where ordinary skills become scarce skills. Any team can train a model; almost none can fly one. The gap between the two is a career — and it is being formalized, conference paper by conference paper.
The presentation also proposed a future worth watching: spacecraft that archive their own observations with onboard triage, keeping raw data in orbit and downlinking only what downstream users actually request — an inversion of the current model that could rewrite how constellations are designed and how bandwidth is budgeted.