Duncan Eddy: Applying AI to Smarter Collision Avoidance and Space Traffic Management

As of May 2025, there are roughly 11,700 active satellites orbiting Earth, most of them crammed into low-Earth orbit where even a minor miscalculation can spell disaster. Individuals across industry, academia, and government are increasingly concerned about triggering Kessler syndrome—a runaway cascade of debris that could render swaths of orbit unusable for generations.

EcoAero had the opportunity to speak with Duncan Eddy, a research fellow in the Stanford Intelligent Systems Laboratory and in the Stanford Center for AI Safety. Eddy earned his PhD in Aerospace Engineering from Stanford in 2021, where he pioneered optimal decision-making frameworks for large-scale Earth-observing constellations. Prior to his fellowship, he led the Spacecraft Operations Group at Capella Space—fully automating the first U.S. commercial synthetic aperture radar constellation—and went on to head the constellation operations and space safety teams at Amazon’s Project Kuiper before joining AWS as a Principal Applied Scientist.

Driven by his industry experience at Capella and Amazon and the POMDP (Partially observable Markov decision process) expertise of his Stanford advisor, Eddy worked with a graduate student at Stanford, William Kuhl, to reframe collision avoidance as a Markov decision process rather than a static threshold rule. Traditional schemes instruct a burn whenever conjunction probability exceeds a fixed cutoff—often prompting early, fuel-heavy maneuvers. Their MDP-based policy instead weighs evolving uncertainties in orbital predictions, fuel budgets, and maneuver efficacy to decide whether to wait for clearer data or act immediately.

In practice, smaller operators tend to maneuver conservatively at the first warning, burning propellant days in advance when conjunction data messages still carry large covariance errors. Conversely, delaying too long can escalate delta-V requirements as miss-distance tolerances shrink. Eddy’s simulations reveal this trade-off: by optimizing expected cumulative propellant use against collision risk, his framework yields tactically timed maneuvers that can conserve significant fuel over a satellite’s lifetime.

To validate the approach, Eddy and colleagues built a sandbox using Space Track’s orbital catalog and synthetic conjunction scenarios. Through Monte Carlo simulations, they derived a decision policy mapping observed state information to maneuver actions. Early results indicate that, relative to rule-based cutoffs, the MDP policy can reduce fuel usage by up to 20% over a ten-year mission while maintaining equal or lower aggregate collision probabilities.

Despite promising proof-of-concept outcomes, Eddy cautions against full deployment without further real-world testing. His work relies on a simulated environment and simplified covariance models; integrating live conjunction data, operator-specific fuel-consumption profiles, and exhaustive on-orbit validation are essential next steps. Moreover, the “black-box” nature of MDP-derived policies may clash with industry preferences for transparent, predictable maneuver rules during joint operations.

Looking ahead, Eddy envisions research into how multiple operators’ decision policies interact—studying how heterogeneous maneuver rules influence overall congestion and collision rates. He also advocates for robust, open-access space situational awareness repositories to enable cross-validation of AI-driven strategies and foster collaborative innovation across academia and industry.

As LEO becomes ever more crowded, Eddy’s work illustrates the potential of AI-driven decision frameworks to enhance both safety and efficiency. At EcoAero, we’ll continue spotlighting such interdisciplinary breakthroughs, championing data transparency, and bridging academic innovation with operational practice—ensuring tomorrow’s constellations navigate our shared orbital commons both sustainably and prudently.