Inertial, stellar and optical navigation: locating yourself when Mars has no GPS

Getting lost on Mars is not a metaphor

On Earth a phone can obtain position in seconds from GNSS constellations, maps and networks. A Mars crew cannot assume that infrastructure exists. A rover going behind a ridge needs its position, attitude, energy reserve, return path and the uncertainty attached to all of them. Tens of metres may be irrelevant on a flat plain and critical near a cliff, exclusion zone or emergency cache.

Martian navigation is therefore a sensor-fusion problem. No instrument reports “truth” by itself. An inertial unit gives immediate motion information but drifts. Stars provide an absolute attitude reference but not complete surface position. Cameras can recognize terrain but suffer from dust, lighting and texture. Radio can measure range or relative velocity depending on geometry.

Robustness comes from combining imperfect measurements whose errors are not identical. Diversity reduces common failure, but fusion demands common reference frames, coherent clocks and honest uncertainty models.

The central question is not simply “where am I?” It is “where does the system believe it is, with what uncertainty, based on which measurements, and what can it still do without exceeding safety margins?”

Navigation error costs time and energy and can become a survival problem. In a landscape without continuous roads, a few kilometres of unexpected detour may consume a rover reserve or put an EVA crew too far from shelter. Navigation therefore needs more than a dot on a map: position, uncertainty, terrain state, retreat route and shared references with other vehicles.

The map itself is never perfectly true. Dunes move, dust obscures detail, shadows change and human operations modify the site. A settlement will have to continuously reconcile orbital mapping, local observations and surveyed reference points, much as a terrestrial city maintains its maps and buried networks.

Scenario: a rover loses the link and discovers that position is a probability

Imagine an autonomous rover 80 km from the base. The last orbital correction was three hours ago. The IMU continues measuring acceleration and rotation, but a small bias error accumulates. Visual odometry compares terrain between images, but a bland dusty area provides fewer features. An orbital map supplies a more absolute reference if the rover can recognize landmarks.

The system therefore does not have one perfect “position”; it has an estimate accompanied by covariance, a mathematical representation of uncertainty. As uncertainty grows, the rover can slow down, seek a beacon, climb to a location with better sky visibility or wait for an orbital pass. Intelligent autonomy begins by recognizing when navigation quality is degrading.

Safety depends on integrity as much as accuracy

A 20 m error may be acceptable on an open plain and catastrophic beside a cliff, power cable or habitat interface. Navigation therefore needs to report not only position but confidence and how quickly a bad solution can be detected. Integrity becomes essential when vehicles operate near people or share constrained corridors.

Different sensors fail in different reference frames

An inertial unit propagates motion continuously but accumulates bias and drift. A star tracker can provide an external attitude reference yet may be blinded by Sun geometry, plume light or contamination. Optical navigation can constrain line of sight to a planet, moon or terrain feature, but its accuracy depends on image geometry and feature knowledge. Combining these sensors works because their error mechanisms are different. The estimator should therefore preserve residuals and health information rather than blending every input into one opaque state. During a long Mars transfer, occasional absolute observations can bound inertial drift; near the surface, terrain or landmark measurements can update position while inertial data bridges periods when images are unavailable.

Calibration is part of navigation state knowledge

Bias estimates, alignment matrices and camera geometry change the meaning of raw measurements. A sensor that is electrically healthy can still degrade navigation if its calibration is stale after a thermal cycle or mechanical intervention. Long missions therefore preserve calibration history and test opportunities, allowing the estimator to separate a changed environment from a changed instrument.

Inertial navigation: knowing immediately how you move, then accepting drift

An IMU mainly measures acceleration and rotation. Integrating those measurements reconstructs velocity, attitude and position. The advantage is independence from external transmitters. The fundamental weakness is that small bias becomes large error when integrated.

Consider a pedagogical constant acceleration error of 0.001 m/s². After 100 s it creates 0.1 m/s of velocity error. Position error from a constant acceleration is roughly ½at²: 0.5 × 0.001 × 100² = 5 m. After 1,000 s the same uncorrected bias would imply about 500 m. Real systems are far more sophisticated, but the order of magnitude explains the need for external references.

Gyroscopes and accelerometers have noise, bias, scale factor, misalignment and thermal dependencies. Serious navigation does not hide them in one accuracy number. It propagates covariance or another uncertainty representation so guidance knows whether position is precise or merely plausible.

At Mars inertial navigation will remain a backbone for vehicles, aircraft, landers and robots. It is available everywhere, but it must be periodically constrained by observations of the outside world.

An inertial unit measures acceleration and rotation at high rate without any external infrastructure. It keeps working through radio outages, dust or tunnels. The price is drift: tiny biases are integrated into velocity and then position. A simplified constant acceleration bias of 0.001 m/s² produces roughly 5 m of position error after 100 s and 500 m after 1,000 s if nothing corrects it.

Real navigation therefore combines inertial continuity with periodic absolute corrections from stars, terrain, radio or beacons. Thermal calibration, mechanical alignment and bias modeling can matter as much as the headline sensor specification.

The Martian environment complicates calibration. Sensors do not live at constant temperature: a rover can move between very cold periods and locally warm electronics. Gyroscope and accelerometer biases can change with temperature. Serious designs therefore combine calibration tables, thermal measurements, re-alignment procedures and drift monitoring. A terrestrial data-sheet number is never a guaranteed performance across every Martian operating phase.

Inertial redundancy also raises the question of dissimilarity. Three identical units can vote through an isolated failure while remaining vulnerable to the same software defect, aging mechanism or environmental sensitivity. Critical functions can benefit from combining different sensing principles so a common cause cannot mislead every path in the same way.

Stars, Sun and horizon: recovering absolute references

A star tracker photographs the sky, recognizes stellar patterns and provides absolute attitude against a catalogue. In interplanetary space this is central because stars offer a highly stable reference. Surface use is different, but the principle remains valuable for selected platforms.

Optical navigation can also use the Sun, Phobos, Deimos, Mars itself or other known bodies. During cruise, line-of-sight measurements constrain spacecraft state. Near Mars, the planetary limb, stars and surface imagery add information.

A navigation camera is not simply “a camera with AI.” It becomes a geometric instrument when its calibration, focal length, orientation, timing and observed-object geometry are known. A beautiful image without calibration can be nearly useless for navigation.

A settlement therefore benefits from local calibration facilities, camera models and catalogues. Optical metrology becomes navigation infrastructure as well as scientific infrastructure.

A star tracker matches observed star patterns against a catalogue and can provide a highly accurate attitude reference. It is powerful in space or with clear sky but can be affected by the Sun, occultation, dust or field-of-view constraints. Sun sensors, horizon references and other cues can complement it.

Attitude directly affects surface operations. A few degrees of error can mispoint a high-gain antenna, distort map projection or change how a slope is interpreted. That is why position, attitude and time are normally estimated together rather than treated as unrelated instrument readings.

Recognizing the ground: visual odometry, orbital maps and terrain-relative navigation

Mars rovers have demonstrated visual odometry: comparing successive images estimates motion when wheels slip and mechanical odometry is uncertain. Perseverance added faster autonomous navigation, while Mars 2020 descent used terrain-relative navigation to compare surface imagery with an onboard map in real time.

The principle extends beyond landing. A base can maintain multi-layer maps containing terrain, hazards, roads, beacons, soil properties and radio coverage. Vehicles compare what they see with that reference while accepting that the map ages as construction moves regolith or new modules hide landmarks.

Navigation becomes configuration management. Which map version does the rover use? Does operations have the same one? Is a road still open? Was a moved beacon declared? Geospatial data deserve the same configuration discipline as software.

At city scale a living map becomes a shared asset connecting navigation, science, logistics, emergency response, urban planning and resource extraction.

Visual odometry tracks features between successive images to estimate motion and can correct the false distance suggested by wheels that slip in sand or over obstacles. At larger scale, comparison with orbital maps can re-anchor the route. Terrain-relative navigation during landing follows the same family of ideas: recognize what the camera sees and compare it with onboard knowledge of the world.

A city could extend this with ultra-high-resolution local maps, maintained landmarks, radio beacons and continuously updated 3-D models. Navigation becomes a shared asset: each rover improves the map used by the next. That collective memory is especially valuable where a localization mistake cannot be fixed by calling nearby roadside assistance.

Vision has its own failure regimes: uniform terrain with little texture, long shadows, dust on optics, low Sun, image compression or seasonal changes. Autonomous navigation should detect when images no longer contain enough trustworthy information rather than force a match. The camera must not only produce an answer; it must help quantify when that answer is weak.

Human operations also need richer maps than elevation alone. Traversable slopes, deep-dust zones, access to shelters, radio coverage, energy cost and EVA risk can become layers in the same route graph. The geometrically shortest path is not necessarily the safest path.

Filtering uncertainty: the Kalman idea without turning the chapter into a mathematics course

A Kalman filter and its many relatives implement an intuitive loop: predict state with a model, observe the world, compare prediction and measurement, then correct according to uncertainty. If inertial data are temporarily good but a camera is blinded by dust, the two sources should not receive equal weight. When the camera finds strong landmarks again, it can constrain drift.

The “state” contains quantities the system needs: position, velocity, attitude, sensor bias and sometimes other parameters. Covariance describes confidence and correlated errors. This is operationally important: a position known to ten metres is not the same object as a point whose uncertainty is half a kilometre.

Filters can fail when models are wrong or sensor faults are not represented. Innovation checks, anomaly detection and measurement rejection are therefore part of navigation. Sensor fusion is not a democracy where the majority is always correct.

Human interfaces should display uncertainty instead of a falsely precise dot. Ellipses, bands and confidence levels communicate what the system truly knows.

A Kalman filter does not magically discover the true position. It maintains an estimate and a description of uncertainty. Model propagation normally increases that uncertainty; a trusted measurement can reduce it. A noisy sensor receives less weight than a precise one. Covariance is the language used to describe this confidence and the relationships among errors.

Operationally, two vehicles can display the same estimated position but have very different navigation quality. One may be confident to metres and the other only to hundreds of metres. Speed limits, obstacle clearance and decisions to continue or stop should therefore depend on uncertainty, not only on the best estimate.

The Kalman idea: combine imperfect measurements because they are imperfect in different ways

An inertial sensor responds immediately but drifts. Terrain observations can be absolute but intermittent. A radio measurement may be strong in one geometry and weak in another. Kalman filters and related estimators exploit that complementarity: they propagate a model of vehicle state, then correct the estimate when new measurements arrive according to their uncertainty.

The important lesson is not matrix algebra. It is discipline: a measurement has both a value and a quality. If a radio position looks extremely precise but conflicts violently with inertial history, the system should question the sensor, timing or beacon identity before instantly jumping to the new solution.

Toward Martian PNT: from orbital relays to neighborhood beacons

The JPL PNT study argues that dedicated Mars services should be designed well ahead of large-scale human operations. Orbiters could distribute time and ephemeris information; surface beacons would strengthen accuracy around bases; vehicles would combine those signals with inertial and vision systems. The architecture can grow incrementally.

A first outpost may not need a planet-wide constellation. Carefully surveyed local control points, a few beacons and orbital relays can transform operations around one site. As settlements multiply, reference frames and timing services must be tied together.

Ownership and interoperability then matter. If each agency or company deploys incompatible beacons, users carry multiple receivers and emergency operations become brittle. International standards can prevent navigation infrastructure from fragmenting into closed islands.

Mars will therefore likely need a PNT system of systems rather than a copy of GPS. The goal is familiar — position, velocity and time — but distances, deployment history and user population are different.

A first Martian PNT service need not wait for a full constellation. A few orbiters with stable clocks and ranging capability, combined with local beacons and maps, would already improve autonomy. A permanent base could install surveyed geodetic points and reference stations around its operational zone.

As multiple settlements appear, compatible reference frames become strategic. Two cities should not describe the same location with incompatible coordinates or use time scales that cannot be converted reliably. Geodesy, timing and navigation-message standards eventually become institutional infrastructure as much as engineering detail.

A “Martian GPS” does not need to copy Earth’s satellite count

Terrestrial GPS is the result of a particular architecture: a large constellation, orbits selected for global coverage, atomic clocks, control stations, navigation messages and receivers that solve position from several signals. Mars has a different size, rotation period, user population and economy. Copying the number of satellites without copying the reasoning would be a design error.

An early base could use a regional system: a few orbital relays provide timing and radiometric measurements, local beacons cover work areas, rovers combine inertial sensing with vision, and orbital maps provide references. As settlements grow, the architecture can become planetary. The service definition matters more than the label: required accuracy, availability, integrity, coverage and time to alert when a solution becomes unsafe.

PNT can share telecommunications infrastructure

An orbiter already providing communications knows its orbit, carries a clock and transmits signals. It can contribute to navigation if messages, synchronization and geometry are designed for that function. The convergence is attractive because one spacecraft can provide communications, time distribution and ranging or Doppler observables. It also creates common-mode dependency: a constellation failure can remove several services at once.

A resilient Mars architecture therefore needs independent sources: onboard inertial navigation, stars, terrain, local beacons and perhaps vehicle-to-vehicle relative navigation. The correct choice is not “GPS or nothing,” but fusion of systems that fail differently.

What happens when sensors disagree?

A dangerous navigation failure is not always a black screen. Hard cases occur when several sensors still report plausible but incompatible numbers: an IMU drifts, a camera matches the wrong terrain, a radio beacon moves or time is wrong. The system must detect inconsistency before it becomes a bad trajectory.

Redundancy therefore needs diversity. Three identical sensors sharing one software or thermal failure may be less safe than independent inertial, optical and radio measurements. Cross-checks compare domains: visual velocity against inertial dynamics, radio range against the map and stellar attitude against vehicle motion.

In degraded mode the mission objective changes. A rover can slow down, avoid difficult terrain, return to a beacon or stop. A lander may not be able to stop and must choose a safe solution in the time remaining. Navigation must state what it can still guarantee.

A mature settlement preserves sensor logs and map versions so anomalies can be reconstructed. Navigation becomes a safety system and technical memory rather than just a number on a display.

A contradictory sensor is often harder than a dead sensor because it can continue producing plausible but wrong values. An IMU drifts, a camera matches the wrong feature or a radio beacon carries bias. FDIR must detect inconsistency, isolate the suspected source and reconfigure the estimate without triggering a cascade of bad decisions.

The vehicle must also know when to admit uncertainty. Slowing down, stopping, returning to known terrain or requesting local validation may be safer than continuing with exploding covariance. Safe navigation is not merely accurate; it has defined behavior when accuracy can no longer be demonstrated.

A decision tree can handle several degrees of disagreement. Small camera-versus-inertial divergence can increase uncertainty and trigger a search for another reference. Larger disagreement can reduce speed. Loss of several absolute references can cause a return toward a known beacon. Graduated behavior prevents a small discrepancy from causing a hard stop while also preventing silent drift.

After an incident, raw sensor data, software versions, temperatures, image quality and estimator decisions should remain available for diagnosis. In a settlement, that analysis can improve maps, procedures and maintenance criteria for every other vehicle.

Will Mars have GPS? The wrong question opens the right chapter

Terrestrial GPS works because satellites occupy known orbits, carry stable clocks and transmit synchronized signals. A receiver measures propagation times and solves for position using multiple sources. Copying the terrestrial constellation around Mars may be neither necessary nor economical. Settlement geography, user density, existing telecom relays and coverage requirements will be different.

An early base may combine local beacons, inertial navigation, terrain imagery and occasional orbital ranging. A region with hundreds of active vehicles may justify a broader PNT service. The same orbiters carrying communications can contribute time, Doppler and range measurements. The better question is therefore: what accuracy and availability does each operation require? A mining truck and a human lander do not have the same navigation risk.

JPL work on PNT architectures for the Moon and Mars emphasizes that Mars should not simply inherit a lunar design. Orbits, autonomy, coverage and clock technology shape the architecture. A future constellation could merge communications, positioning and timing rather than reproduce GPS as a separate system.

For society, position is a public-safety service. It enables rescue, convoy coordination, industrial asset tracking, grid synchronization and scientific georeferencing. PNT is therefore invisible infrastructure connecting mobility, safety, industry and science.

Inertial navigation is autonomous, but error integrates

An inertial measurement unit senses acceleration and rotation without relying on external signals. That makes it extraordinarily valuable during outages. But it integrates imperfect measurements. A small bias becomes a velocity error and then a growing position error, so inertial navigation is powerful in the short term and must be corrected over longer periods.

Consider a constant acceleration bias of only 0.001 m/s². In a simplified one-dimensional case, position error is approximately Δx = ½ a t². After 100 seconds the error is 5 metres. After 1,000 seconds it is 500 metres. Real systems are more complex, but the example explains immediately why external observations matter.

Star trackers provide a strong attitude reference. Cameras can recognize terrain or horizons. Radio provides range and radial velocity. Wheel odometry works until wheels slip. No sensor is universally best, so modern navigation fuses independent information and carries an explicit estimate of uncertainty.

Kalman filtering is one classic framework for that fusion. It combines a motion model with measurements according to uncertainty. Covariance tells the system how much confidence it should place in its estimate and which directions remain poorly observed. A rover that knows its uncertainty can slow down or request another observation before uncertainty becomes danger.

Optical navigation: Mars itself becomes a beacon

Cameras can turn the landscape into navigation infrastructure. Visual odometry estimates motion from successive images; terrain-relative navigation compares observed features with onboard maps; deep-space optical navigation can use planets, moons and stars.

The attraction is obvious: the environment itself supplies information. The limitations are equally important. Lighting changes, dust obscures landmarks, featureless terrain weakens matching, and maps can become stale. Software must know when images no longer constrain the state well enough.

For a city, maps become critical operational data. New roads, excavation, dune migration or a collapsed slope changes the world that autonomy expects. Mapping becomes a continuously updated civic service rather than a static science product.

Orbital PNT, local beacons, inertial sensors and optical methods can therefore reinforce one another. A robust settlement is unlikely to trust a single constellation; it will use an architecture whose source of truth changes with location, mission and failure state.