Satellite InferenceAI Inference in OrbitGet in touch FORMATION AND CONTROL / CLUSTER GNC
Compute nodes fly a few hundred metres apart so their laser links stay short. In free fall the gap between two neighbours swings between its closest and about twice that, twice per orbit, and thrusters mostly cancel drag. Here every node is propagated numerically around the real Earth, with its bulge, its air and sunlight. Drag the view to look around, switch to fixed stars to see where the “dance” comes from, and make any node fail at any moment.
01 / WHY THE NODES SEEM TO DANCE
The default view flies with the cluster's centre and keeps the Earth below, so it turns once per orbit. Nothing circles the centre: every node is on its own orbit around the Earth, and their speeds differ by fractions of a metre per second. From the turning camera that difference looks like a 2:1 oval traced once per orbit. Switch to fixed stars and the same motion looks different, because only the camera changed.
George Hill described this motion in 1878 for the Moon; Clohessy and Wiltshire applied it to rendezvous in 1960, and it has guided every docking since. With no perturbations our numerical propagation reproduces their ellipses to within 0.5 m on a cluster 800 m across; the rest is the reference orbit's slight eccentricity and curvature.
Google's Project Suncatcher places 81 satellites within 1 km, all in the orbit plane, so the shape holds without propellant and, in a dawn–dusk orbit, no node shades another. Flat, two nodes can drift into each other's path. Our design turns the same lattice 18° out of the plane, so every pair stays apart without anyone steering; the simulator opens on it.
02 / THREE WAYS TO FLY A CLUSTER
Started from matched mean orbital elements, the 1 MW lattice drifts by 6 mm per orbit and holding its shape costs 0.006 m/s a year; cancelling drag costs 2.6 m/s a year on the solar-cycle average. When the centre node fails at the ascending node it passes 2 m from a neighbour after 5.3 hours: 10 neighbours step aside with 10 cm/s of burns in all.
Each node gets its own radial and cross-track amplitude, with the relative eccentricity and inclination vectors kept parallel (D'Amico and Montenbruck, 2006, the method TanDEM-X and TerraSAR-X fly). Any two nodes stay apart even if along-track control is lost, and a failed node passes 40 m away. The cost: the helix spans 5.4 km, and J2 turns the eccentricity vectors 3.2° a day, so keeping them parallel takes up to 13 m/s a year for the outermost node.
Giving every node a cross-track swing in step with its motion along the track makes each pair's relative eccentricity and inclination vectors parallel, as in the helix, while keeping the lattice and its four neighbours a node. In our design study, with 12 nodes of 100 kW, a dead node passes 3.2 m from a neighbour in the flat lattice and no closer than 89 m in the tilted one, for 12 m/s a year at the edge.
03 / CHOOSING THE DESIGN
Starlink is moving about 4,400 satellites from 550 to 480 km during 2026, and Amazon's Kuiper is authorised at 590–630 km. Below them, atomic oxygen erodes too much film in five years; higher up, every 50 km needs more shielding for the memory and a dead node stays up longer. 650 km is the lowest altitude above the constellations, in sunlight 96% of the year, and a tumbling dead node falls out in about 1.2 years on the solar-cycle average.
A 100 kW node is about 28 m across, so 142 m between neighbours leaves five node widths for navigation and control errors. At that range a laser link can carry far more than a node can power, so safety and propellant set the spacing, not the lasers.
Every node carries mass that does not shrink with its power, and a cluster needs spares that cannot be repaired. For one megawatt the balance in our design study is 100 kW nodes, about 3.5 t each: 10 to compute and 2 spare. The study compares 1,080 designs: see how we chose.
04 / GUIDANCE, NAVIGATION AND CONTROL
Earth's bulge (J2) speeds up and tilts every orbit slightly differently. Copying the ideal ellipses onto the nodes' instantaneous orbits lets the outer nodes slip 10 m per orbit in our propagation. Matching their orbits averaged over one revolution instead removes the slip to millimetres; Suncatcher's simpler fix, a 2:1.0037 axis ratio, keeps it under 3 m/s a year per kilometre.
Carrier-phase GPS between nearby spacecraft measures relative positions on board to a few centimetres; the PRISMA formation mission flew it at about 6 cm in 2010. Laser terminals on nodes holding their attitude turn at most 0.12° per second to follow a neighbour, and Suncatcher puts the pointing needed at about a microradian, within commercial optical terminals.
Healthy nodes cancel their own drag, which NRLMSIS puts at 2.6 m/s a year for a 20 kW node at 650 km on the solar-cycle average, plus allowances for drag differences between nodes and for debris avoidance. Nodes fly their arrays and radiators edge-on to the airflow; a dead node tumbles and feels about fourteen times more drag.
05 / FAILURE AND END OF LIFE
A node that loses power stops cancelling its drag while its neighbours keep cancelling theirs. It tumbles, sinks and pulls ahead. In a 1 MW flat lattice of 20 kW nodes the centre node leaves the cluster after about 7 hours; on the way it can pass within metres of a neighbour, depending on which node fails and when, which is why the simulator lets you choose both. Neighbours act after the failure is noticed, with single burns of a few centimetres per second. In our tilted design no neighbour has to act: the dead node passes at least 89 m away. A real cluster also needs collision checks on board and the ability to move together around debris.
At the end of the mission, the FCC requires satellites in low Earth orbit to re-enter within five years. Left in place, a healthy node at 650 km would take more than a decade, so it keeps about 100 m/s to lower its perigee to 300 km and re-enters within about five months. With electric thrusters, that burn and five years of station-keeping take a few kilograms of propellant, within the 5% the node model carries; chemical thrusters would need more.
06 / METHOD, CHECKS AND LIMITS
Every node and the cluster's reference orbit are propagated together in an Earth-centred inertial frame with a fourth-order Runge–Kutta integrator (20 s steps). Forces: Earth's gravity with the J2, J3 and J4 terms of EGM96; drag from NRLMSIS 2.1, tabulated along this dawn–dusk orbit by altitude and position so it keeps the day–night and latitude structure, in an atmosphere turning with the Earth (drag coefficient 2.2); and sunlight pressure with a conical Earth shadow. The reference orbit is sun-synchronous and frozen. Nodes start from relative orbital elements matched after averaging over one revolution. Station-keeping costs come from relative elements averaged over every revolution and the Gauss equations. A failed node stops thrusting and tumbles; it is noticed after 10 minutes and neighbours keep three node widths clear with single burns sized by the linear (Hill–Clohessy–Wiltshire) response.
Sun direction against JPL's DE421 ephemeris: within 0.0075° over 2026–2035. A day of flight against a separate Python implementation integrated with an eighth-order method: 0.37 m apart, and 34 µm for the relative position of two spacecraft 100 m apart. J2's turning of the pattern, -3.195° a day, against Brouwer–Kozai theory, -3.205°. Energy is conserved to one part in a billion over a day.
The Sun's and Moon's pull differ across a 1 km cluster by under 2×10⁻¹⁰ m/s², and Earth's non-zonal gravity by about 10⁻⁸ m/s² that averages out as the Earth turns. Navigation and thruster errors, attitude dynamics, geomagnetic storms that multiply drag for days, and collision probabilities are not modelled; drag differences between nodes (5%) and debris avoidance (1 m/s a year) are allowances. Node geometry is illustrative.
The density table, reference trajectories and these checks are generated by scripts in the site's repository and rerun as automated tests on every change. A communication tool, not mission analysis: flight design needs a full force model, attitude and navigation errors and certified tools.
Sources: G. W. Hill, Researches in the lunar theory (1878); W. H. Clohessy and R. S. Wiltshire, J. Aerospace Sciences 27 (1960); D. Brouwer and Y. Kozai, Astronomical Journal 64 (1959); S. D'Amico and O. Montenbruck, J. Guidance, Control and Dynamics 29 (2006); O. Montenbruck and E. Gill, Satellite Orbits (2000); D. Vallado, Fundamentals of Astrodynamics and Applications; EGM96; J. T. Emmert et al., NRLMSIS 2.0 (2021) and NRLMSIS 2.1 via pymsis; JPL DE421 via Skyfield; Google Research, Project Suncatcher, arXiv:2511.19468 (2025); Ardaens, D'Amico and Montenbruck, PRISMA GPS navigation flight results (2010); FCC Second Report and Order 22-74 (2022); FCC authorisation of the Kuiper system (2020); SpaceX on lowering Starlink satellites (January 2026); reported reusable launch capacities to low Earth orbit: Starship Block 3 about 100 t, Falcon 9 17.5 t. Checks generated 2026-10-05. The SI-P1 pathfinder is a hosted payload; formation flight applies to future clusters of SI Nodes.