Space Computing and AI
DraftOther sections cover the hardware. This one covers what runs on it and what runs alongside it on the ground: the software, autonomy and data infrastructure that determine what a space system can actually do.
Reading order
Autonomy comes first because its verification argument recurs throughout. The perception and data pages form a chain; the last three are infrastructure.
The constraint that shapes everything
Two numbers explain most of this section.
Light time and contact windows force autonomy. A Mars lander gets no ground help during its seven-minute descent, and a low Earth orbit satellite is out of contact for more than 95% of its life.
Downlink scarcity forces onboard processing. A high-resolution imager generates about 10 Gbit/s and can transmit roughly 30 GB per pass, a gap of more than an order of magnitude.
Everything else follows: planning must run onboard because conditions change between uplink and execution, vision must run onboard because landing cannot wait, and data must be triaged onboard because it cannot all be sent.
Four things worth knowing
Autonomy is limited by verification, not capability. A system that plans its own actions has a state space too large to test exhaustively. The practical answer is runtime assurance: let a complex planner propose, and interpose a simple verifiable monitor that rejects anything unsafe.
Machine learning is accepted for payload data and not for vehicle control. The dividing line is consequence. A misclassified cloud costs a downlink slot; a misperception during descent destroys the vehicle.
Onboard decisions are irreversible. A cloud classifier that wrongly rejects a scene has destroyed an observation nobody can retake. Prioritise rather than delete wherever storage allows.
Satellite network topology is deterministic. Unlike terrestrial networks, a constellation’s future connectivity is computable from orbital mechanics, which makes reactive routing protocols the wrong tool and contact graph routing the right one.
Pages
Why autonomy is forced by physics, the four autonomy levels and the verification burden each imposes, onboard planning with flight heritage, fault management proportional to urgency, and runtime assurance.
02AI for mission planningObservation scheduling as an NP-hard optimisation, where learned methods genuinely help and where they do not, cloud avoidance, downlink allocation optimised for data age, and the operator trust problem.
03Computer vision in spaceTerrain relative navigation and the accuracy improvement it delivered, relative pose for rendezvous, visual odometry and slip, star identification, and why safety-critical vision remains classical.
04Satellite image processingThe processing chain and where systematic errors originate, radiometric and geometric correction, orthorectification, atmospheric correction, cloud masking, and change detection.
05Edge computing in spaceThe downlink gap quantified, what onboard inference costs in power and thermal budget, radiation tolerance of commercial accelerators, what to run onboard, and the honest limits of the concept.
06Space roboticsManipulators and coupled base dynamics, surface rovers and the command cycle limit, free flyers, teleoperation under delay, and the unsolved problem of grappling an uncooperative target.
07Digital twinsWhat a twin is and is not, calibration against ground test and flight telemetry, anomaly detection through model residuals, operations rehearsal, and an honest assessment of the term.
08SimulationOrbit propagation and force models, six degree-of-freedom simulation, the four in-the-loop testing stages and what each catches, Monte Carlo dispersion, and the toolchains including SPICE.
09Distributed satellite systemsFormation flight and free-drift relative orbits, optical inter-satellite links and point-ahead, routing over a deterministic moving topology, delay tolerant networking, and why fractionated spacecraft did not happen.
10Space data platformsStandards including STAC, cloud-native formats and why partial reads matter, analysis-ready data, moving compute to the data, the Landsat open policy result, and the archive as a permanent obligation.