Docking Uninvited at Twenty-Eight Thousand km/h

On-orbit servicing — refueling, repairing, or repositioning existing satellites — requires a servicer spacecraft to autonomously rendezvous and dock with a client satellite that was not designed for servicing. Unlike ISS docking (cooperative, with reflectors and transponders), most servicing targets are non-cooperative: they may be tumbling, have no docking interfaces, and present complex geometries that confuse visual navigation algorithms. The autonomous guidance, navigation, and control (GNC) system must handle approach from kilometers away to centimeters away, transitioning between sensor modalities (star tracker → lidar → camera) while managing collision risk with a high-value asset.

Over 3,000 satellites worth $300+ billion are in geostationary orbit, many nearing end of fuel life but otherwise functional. Extending their operational life by 5–15 years through refueling is worth $50–100M per satellite to operators. Active debris removal (ADR) to prevent Kessler syndrome also requires non-cooperative rendezvous. Several companies (Astroscale, Northrop Grumman MEV, ClearSpace) have demonstrated basic rendezvous, but only with cooperative or semi-cooperative targets. Scaling to true non-cooperative servicing could sustain the space economy while addressing the debris crisis.

Northrop Grumman's MEV-1 and MEV-2 successfully docked with cooperative GEO satellites (Intelsat) by grasping their engine nozzles — but these missions required months of planning, ground-in-the-loop operations, and a known, stable target geometry. Astroscale's ELSA-d demonstrated magnetic capture of a cooperatively designed target. DARPA's RSGS (now Robotic Servicing Vehicle) has been under development for years with limited flight demonstration. The core challenge is that visual navigation algorithms trained on synthetic renders of satellites fail when confronted with real lighting conditions (specular reflections, deep shadows, Earth albedo changes). Lidar provides range data but struggles with reflective surfaces (solar panels, MLI blankets). The last few meters of approach — where collision risk is highest — remain dependent on human oversight via ground link, which introduces unacceptable time delays for LEO operations.

Robust pose estimation algorithms that work across lighting conditions, surface materials, and unknown satellite geometries — likely requiring sim-to-real transfer learning with domain randomization. Sensor fusion architectures that gracefully hand off between modalities as range decreases. Safety-assured control systems that can guarantee collision avoidance even under sensor degradation. Standards for "servicing-friendly" satellite design (grapple fixtures, fiducial markers, refueling ports) would help for future satellites but don't address the existing fleet.

A team could develop a pose estimation pipeline for satellite rendezvous using synthetic training data (tools like Blender for rendering) and test it on publicly available satellite imagery. The SPEED+ dataset (Stanford/ESA) provides benchmark data for spacecraft pose estimation. A robotics-focused team could prototype a proximity operations controller on a planar air-bearing testbed simulating the final approach phase. The computer vision and control problems are accessible with standard academic tools.