ten thousand satellites, no traffic control

Problem statement

The number of active satellites has increased from ~2,000 in 2019 to over 10,000 in 2025, driven primarily by megaconstellations (Starlink, OneWeb, Kuiper). The conjunction assessment system — which identifies potential collisions and triggers avoidance maneuvers — was designed for hundreds of tracked objects and generates thousands of false-alarm alerts per genuine risk. As constellation sizes grow toward 50,000+ planned satellites, the number of conjunction alerts scales quadratically with object count, and the current system cannot distinguish actionable threats from noise without human analysts who cannot keep pace.

Why this matters

A single collision in low Earth orbit generates thousands of debris fragments, each capable of destroying another satellite (the Kessler syndrome). The 2009 Iridium-Cosmos collision created over 2,000 trackable fragments still in orbit. With $300+ billion of space infrastructure at risk and the prospect of cascading collisions that could render entire orbital shells unusable for decades, the conjunction assessment problem is existential for the commercial space industry. Current operators report receiving hundreds to thousands of conjunction warnings per satellite per year, with false alarm rates exceeding 99.9%.

What’s been tried and why it hasn’t worked

The U.S. Space Command's conjunction data messages (CDMs) provide basic screening, but positional uncertainties (especially for objects without GPS) produce enormous alert volumes. SpaceX's autonomous collision avoidance system on Starlink maneuvers satellites based on onboard assessment, but this creates coordination problems with other operators who cannot predict Starlink's maneuvers. ESA's collision avoidance system uses a multi-step screening process (from thousands of events to a handful of actionable cases), but each step requires updated tracking data that arrives asynchronously. Machine learning approaches to filter alerts have shown promise in retrospective analysis but struggle with the rarity of actual collision events (class imbalance) and the catastrophic cost of false negatives.

What would unlock progress

Better orbital tracking data (especially for sub-10-cm objects) would reduce positional uncertainty and shrink the false-alarm volume. Standardized inter-operator data sharing — so constellation operators can share planned maneuvers in near-real-time — would prevent the current problem of cascading avoidance maneuvers. A probabilistic framework that quantifies collision risk continuously (rather than per-conjunction) could shift from event-based alerting to risk-budget management. This is fundamentally a data fusion and decision-under-uncertainty problem.

Entry points for student teams

A team could build a simulation of conjunction assessment scaling using publicly available TLE (two-line element) data from Space-Track.org (free registration), modeling how alert volume and false alarm rate scale with constellation size. Alternatively, a team could prototype a machine learning classifier for conjunction screening using the ESA Kelvins Collision Avoidance Challenge dataset — over 180,000 real conjunction data messages covering roughly 15,000 close-approach events for ESA satellites, prepared by ESA and openly downloadable — focusing on reducing false positives while maintaining zero false negatives. (Operator CDM archives, such as the Space Data Association's, are member-only.) The algorithmic and data science aspects are accessible; the space domain knowledge can be acquired.