Network fetches are blocked in this session, so I’m restricting citations to long-standing institutional URLs I’m confident about and leaving out anything less certain.
Spacecraft guidance is divided by time scale, not by trust. Ground teams own the slow decisions: fitting weeks of radio tracking into an orbit estimate, designing correction burns, and judging whether a strange reading is physics or a broken sensor. Onboard software owns the fast ones: attitude control from star trackers and gyros, and the seconds-long phases such as landings, flybys and impacts, where the one-way light delay (4 to 24 minutes at Mars) makes a human in the loop physically impossible. The current shift is moving the middle ground onto the vehicle: optical navigation against known landmarks, satellite navigation received out at lunar distance, and onboard orbit determination. For each phase, the question is which side can act fast enough and which side can be wrong safely.
What does a spacecraft actually need to know?
Two separate problems hide under the word “navigation”. Orbit determination answers where the vehicle is and how fast it is moving, six numbers that must be known to metres and millimetres per second. Attitude determination answers which way it is pointing, and must be known to arcseconds so antennas, cameras and engines face the right way.
These use different sensors, different teams and different failure modes. Position usually comes from radio tracking on the ground. Attitude comes from star trackers and inertial units on board, and has been fully autonomous since the 1990s. NASA’s Basics of Space Flight remains the clearest free primer on both.
How does ground navigation still work in 2026?
NASA’s Deep Space Network runs three antenna complexes, in California, Spain and Australia, spaced roughly 120 degrees apart so a probe is always above one horizon. Each has a 70-metre dish plus several 34-metre dishes. You can watch which spacecraft each antenna is tracking in real time on DSN Now.
Three measurement types feed the orbit solution. Two-way Doppler gives line-of-sight velocity to a fraction of a millimetre per second. Ranging times a coded signal there and back, giving distance to about a metre. Delta-DOR compares the spacecraft’s signal against a quasar at two stations to pin down its angle in the sky. Ground software, still built on the extended Kalman filter that Stanley Schmidt’s NASA Ames group adapted for Apollo in the early 1960s, fuses these into an estimate with an uncertainty ellipsoid.
The human job is modelling the small forces that filter cannot see: solar radiation pressure, outgassing, thruster leaks and attitude-control nudges. The 1999 loss of Mars Climate Orbiter came from exactly this layer. Thruster impulse data was delivered in pound-force seconds where newton-seconds were expected, and the discrepancy that navigators noticed was never escalated. The case is still a cornerstone of the NASA Lessons Learned system, and the lesson is about process, not computing power.
What has moved onboard, and when?
Some popular timelines place the debut of inertial navigation in 1972. In fact Apollo flew a gyro-stabilised inertial platform and the Apollo Guidance Computer from 1966, and astronauts realigned it by sighting stars through a sextant. The “GPS analogs of 1997” you sometimes read about never existed. The real milestones are these.
- Attitude from stars, 1990s onward. Modern star trackers match a camera frame against a catalogue in under a second to arcsecond-class accuracy, with gyros propagating attitude between fixes. The reference catalogues now derive from ESA’s Gaia mission, which measured positions for roughly 1.8 billion stars.
- Autonomous optical navigation, 1998. Deep Space 1 was the first probe to determine its own trajectory by imaging asteroids against the star background.
- Autonomous terminal guidance, 2022. NASA’s DART handed control to its SMART Nav software about four hours before impact and steered itself onto Dimorphos, a target about 160 metres across, at roughly 6 kilometres per second. No ground team could have closed that loop.
- Terrain-relative landing, 2021 and 2024. Perseverance matched descent images to an orbital map and diverted around hazards in Jezero Crater. JAXA’s SLIM lander used crater matching to land about 55 metres from its target in January 2024, where earlier landers accepted ellipses of many kilometres.
- Satellite navigation beyond Earth orbit. Receivers now use the weak signals that spill past Earth from GPS and Galileo. NASA’s MMS tracked GPS at about 187,000 kilometres in 2019, and the LuGRE receiver on Firefly’s Blue Ghost lander obtained fixes on the lunar surface in March 2025. The policy framework for this “space service volume” is documented at GPS.gov.
- Crew backup, 2022. Orion’s optical navigation camera was tested on Artemis I, imaging Earth and the Moon so a crew could find its way home if communications were lost.
Who decides in each phase?
| Phase | Primary authority | Why |
|---|---|---|
| Cruise, weeks to years | Ground navigation team | Light delay is irrelevant; human judgment on force models and anomalies matters more than speed. |
| Attitude control, continuous | Onboard | Needs updates many times per second; fully mature and well understood. |
| Landing, flyby, impact, minutes | Onboard | Decisions must be made faster than a signal can travel to Earth and back. |
| Rendezvous with crew, hours | Shared | Automation flies the profile; crew and ground hold abort authority. |
| Anomaly recovery, days | Ground | Diagnosing a failed sensor requires engineering judgment the vehicle lacks. |
Where does it go wrong?
Almost never in the mathematics. Intuitive Machines’ Odysseus lander in February 2024 reached lunar orbit with its laser rangefinders disabled, because a safety switch had not been thrown before launch. Engineers rewrote the flight software in orbit to use NASA’s Navigation Doppler Lidar, a technology demonstration payload that happened to be aboard. The lander still touched down too fast and tipped over, but it survived and transmitted. Human improvisation rescued a mission that pure automation would have lost, and a human checklist error had caused the problem.
Mars helicopter Ingenuity’s final flight, in January 2024, ended when its vision-based navigation lost track over featureless sand and the rotor struck the ground. Optical navigation is only as good as the texture the camera can see. This is why planners keep radio tracking as the baseline and treat onboard vision as an addition, not a replacement.
Making explainer films for engineering clients, we see a version of the same mismatch. Clients ask for a glowing line and an “AI flying the ship”. The honest visual is an uncertainty ellipse shrinking as tracking accumulates, and a filter that is deterministic, not learned. We now have a flight dynamics reviewer check every trajectory shot before final render, because a reversed orbit direction found late means re-rendering the whole sequence, which costs days rather than hours.
What is changing before 2030?
Machine learning is entering perception, such as hazard detection and crater recognition, rather than the core estimator, because flight software must be certifiable and explainable. Pulsar navigation was demonstrated on the ISS in 2018, when the SEXTANT experiment fixed the station’s position to within about 16 kilometres from X-ray millisecond pulsars alone. The Deep Space Atomic Clock flew from 2019 to 2021 with stability good enough for one-way tracking, which would let a probe navigate from a single downlink. ESA’s Hera, arriving at the Didymos system in late 2026, is testing autonomous navigation close to an asteroid. Data exchange between all these systems follows the CCSDS navigation message standards, so an ESA estimate can feed a NASA filter without translation.
If you are starting in this field, separate the two problems first: read the attitude and orbit chapters of Basics of Space Flight, then spend an evening on DSN Now watching real missions being tracked. If you are building a system, keep radio tracking as the truth source, add onboard vision for the phases where light delay forbids a human, and write the units into every interface document.
