Absolute Localization

Absolute localization fixes a vehicle’s position and heading against something that does not move: the Sun, a prior map of the body, a lander, or a cached object. Every relative method a planetary vehicle carries accumulates error without bound. Visual odometry holds position growth to a few percent of distance driven and gyroscopes hold attitude to a rate, but neither ever returns to truth, and the growing uncertainty is what eventually stops the drive [3].

The drift each fix has to bound

Mars Exploration Rover flight software propagated position at 8 Hz from accelerometer, gyroscope and wheel odometry data while the rover drove at about 3.75 cm/s on flat ground, against a requirement that position error stay under 10 percent of distance traveled up to 100 m [1]. Attitude had to stay within 1.5 degrees, 3 sigma, because the high gain antenna pointing budget was 2 degrees, 3 sigma, in total. The Litton LN-200 gyroscope package accumulated enough error that attitude had to be reacquired after about 10,000 s of integration time, roughly every 20 sols [1].

Perseverance is configured to assume 5 percent of distance traveled as position uncertainty growth whenever visual odometry converges, against a demonstrated potential of 2 percent per 100 m, and 50 percent of distance traveled for any motion where it did not [3]. Uncertainty is a weighted sum over the two cases. A 655.8 m drive accumulated 32.92 m of position uncertainty on that accounting [3]. The consequence is geometric: operator-specified keep-in zones on the orbital map shrink by the uncertainty and keep-out zones grow by it, so a narrow passage closes and the drive stops even though the rover could physically pass.

Attitude from the Sun

Heading is the harder half. Roll and pitch are observable from a gravity vector, but yaw on a planetary surface has no magnetic reference on Mars and no useful one on the Moon, so it is taken from an image of the Sun against an ephemeris.

MER’s Surface Attitude Position and Pointing (SAPP) module implemented three primitives, all of which set an Attitude Quality Flag on completion [1]:

Primitive	Sensor	What it solves	Result
Nadir Update	accelerometers, averaged over 10 s	roll and pitch from the gravity vector, compared against Mars surface gravity	TiltOnly
Sunfind	pointable camera, 16 degree field of view	sweeps the camera in the plane perpendicular to nadir at the predicted solar elevation, then confirms with two further images checking that the Sun moved as predicted	Coarse
Sungaze	same camera, held fixed	images the Sun at intervals and feeds observed against predicted Sun vectors to the QUEST quaternion estimator until covariances converge	Fine

Sungaze ran on a ten minute timer and terminated early once the covariance reached 1.5 degrees, 3 sigma [1]. The Inertial Vector Propagator supplied predicted celestial body directions from conic approximations. Sunfind fails near local noon, when the Sun sits close to zenith and its azimuth is ill-conditioned. Operations converged on a shortcut: a Suncheck leaves the camera already pointed at the Sun, so a Sungaze can follow directly without repeating the Nadir Update, saving both time and IMU power.

Perseverance kept the architecture and changed the imaging. Its SAPP attitude update activity is a nadir update from the Rover Inertial Measurement Unit, an extended-verification-record-only gyrocompass run purely to trend gyro performance without touching onboard attitude, and a SunID update using the left NavCam [2]. Normal SunID, inherited from MSL, centroids a direct image of the Sun and is restricted to solar elevations between 30 and 60 degrees because centroiding degrades outside that band. Black SunID uses the NavCam’s own sun protection, which renders the Sun as a bright spot inside a black ring, gives better centroiding accuracy, and works between 15 and 87 degrees of elevation [2]. Both algorithms were imaged side by side on sol 169 during the Black SunID commissioning, after which Black SunID became the algorithm of choice.

Absolute heading changes the error law, not just its size. Rocky 7 carried a wide field of view sun sensor with accelerometer tilt and an onboard clock, and over 38 traverses totaling 335 m at Lavic Lake position error grew linearly at 6 percent of distance, a constant 3.4 degree heading error [16]. An integrated rate gyro instead grows relative position error with the square of distance at short range, which is why Sojourner, dead reckoning through wheel encoders with rate gyros, was an order of magnitude worse in relative heading error [16][27]. Much of the 6 percent traced to a calibration error whose sign flipped between morning and afternoon, so the vehicle drifted right before noon and left after. The sensor was a field-test article; the flown Sun-based methods are the MER and Perseverance image-based ones below.

Fixing position against an orbital map

Matching what the rover sees to what the orbiter saw is the only method that resets position error rather than slowing its growth.

Ground-in-the-loop matching is the flown baseline. Opportunity, Curiosity and Perseverance have all been localized by human mapping specialists orthorectifying NavCam imagery and correlating it to a HiRISE orbital map, at an expected accuracy of 50 cm, two pixels in the 25 cm HiRISE product [3]. The cost is a full Earth-Mars communication cycle per fix, which is what capped Perseverance autonomous driving at about three sols before a human had to re-localize it. The ground procedure itself is a feature correlation: an operator re-measures features in post-traverse imagery against their positions in the imagery that declared the current site frame, and takes a statistic over the deltas as the position correction. On MER a minimum of three point pairs was used and usually many more, with the median of the pairwise Euclidean distances as the error estimate, and the whole operation had to finish in no more than several minutes to fit the planning timeline [8].

Censible, onboard on Perseverance

The onboard replacement builds a top-down stereo orthomosaic from a post-drive NavCam panorama and matches it to the HiRISE orthomosaic at 25 cm and digital terrain model at 1 m using a modified census transform [3]. The census encoding compares each pixel with its neighbors and records whether the neighbor is larger or smaller, which survives the lighting difference between an orbital image and a ground-built orthomosaic, and it carries an explicit missing-data state because rover orthomosaics have holes where rocks and the rover body occlude the ground. A full 360 degree panorama is typically five NavCam stereo wedges at 96 degrees horizontal by 73 degrees vertical field of view, and sometimes only three or four when operational constraints intervene [3].

Accuracy over a benchmark of 264 post-drive panoramas spanning the first 2.5 years of the mission, against the human ground-based localization described above as truth, was 0.36 m mean, mode 0.25 m, with 99 percent inside 0.93 m and no outliers [3]. The benchmark deliberately included rocky terrain, sandy terrain, terrain covered in rover tracks, occlusions, all Mars seasons, and panoramas taken in late morning rather than the mid-afternoon lighting the orbital images were acquired in. Low outlier rate is the property that distinguishes it: normalized cross correlation and mutual information approaches reach comparable mean accuracy but produce outliers at a rate that blocked their transition to flight.

What it cost, and on which processor

Perseverance’s Rover Compute Element is a 133 MHz RAD750 pair and the algorithm was not run there [3]. It runs on the Helicopter Base Station, the Snapdragon 801 board flown to talk to Ingenuity.

	Rover Compute Element	Helicopter Base Station
CPU	RAD750, 133 MHz, x2	Snapdragon 801, 4 cores at 2.36 GHz
Memory	128 MiB ECC RAM	1.55 GiB non-ECC RAM
Storage	2 GB validated ECC	32 GB unvalidated ECC
Operating system	VxWorks 6.7	Linux 3.8
Architecture	PowerPC	ARM
Accelerators	FPGA in the Vision Compute Element	GPU, DSP
Radiation hardened	yes	no
Simulated upsets per device per day	13, corrected	4313, uncorrected

Source: [3], Table 1.

Phase 2 of the flight demonstration ran global localization in 32 s on the Snapdragon and reached 45 degrees Celsius; the same work on the RAD750 was estimated at an order of magnitude longer [3]. Most of that 32 s is not computation. The serial link between the Rover Compute Element and the base station is fixed at 10 KB/s, chosen originally to keep helicopter communication from disturbing the rover computer, and image and data transfer over it dominates about 75 percent of the elapsed time; a rover flight software update could raise the link to 40 KB/s. The RAD750 pair is the same processor that runs visual odometry and autonomous navigation, which is why an algorithm needing seconds of vision processing had to go somewhere else [1]. The radiation exposure is the price of the speed: the RAD750 datasheet expects 0.02 bits corrupted per day at 90 percent of worst-case geosynchronous environment, while a CREME-MC simulation of the Snapdragon 801 gives about 4314 bits per day, so the results have to be checked rather than trusted [3].

The operational effect is measured in meters of drive. Applying one fix at the end of each sol’s drive to the sol 717 to 719 three-sol plan removed the accumulated uncertainty and released 245.37 m of driving that the grown keep-out zones had blocked [3]. Perseverance’s record autonomous distance without human review is 699.9 m over three sols.

Landmarks instead of a map match

Bearing-only localization against topographic features is the cheaper alternative when a correlatable orthomosaic is not available. Measuring the azimuth to hills, peaked ridgelines and isolated boulders whose positions are known from an orbital digital elevation model gives a position fix whose accuracy is set by the geometric dilution of precision of the visible landmark set [4]. Using Perseverance’s NavCam as the sensor, at 0.33 mrad per pixel and an assumed landmark identification error of about 3.3 mrad, an analysis over a 30 km by 30 km Jezero elevation model resampled to about 50 m per pixel found 54 usable landmarks: 21 inside the crater, 13 on the rim and 20 outside it. Worst-case accuracy from a single set of measurements reaches about 10 m in favorable regions northwest of three prominent hills, degrades to 10 to 100 m over most of the mapped area, and becomes effectively unbounded in the riverbed, where the banks occlude every landmark [4]. At least two non-collinear visible landmarks are required, so the method is a sporadic drift correction on top of visual odometry rather than a navigation method in its own right. The expensive step is offline: computing the viewshed for every candidate landmark over a large operating area, which is done on the ground and delivered as a map overlay of terrain informativeness.

Craters are the lunar equivalent, and the reason is coverage: LRO Narrow Angle Camera imagery at up to 50 cm per pixel covers most of the Moon, and craters of 5 to 10 m diameter and larger are detectable from the surface [5]. The driving requirement is a proposed 1800 km traverse over four Earth years with multi-kilometer drives between stops and a 3 sigma position error under 5 m at all times. Simulated performance over 1792 stereo pairs, generated for crater diameters of 5, 7, 10, 12, 15, 17 and 20 m at ranges of 5 to 20 m and solar zenith angles of 0, 20, 40, 60 and 80 degrees, separates the two sensors sharply [5]:

Crater diameter	3 sigma position error, lidar	3 sigma position error, stereo
5 m	0.28 m	1.02 m
10 m	1.04 m	3.22 m
15 m	1.77 m	4.48 m
20 m	1.94 m	4.35 m

Source: [5], Tables 5 and 6. Simulated camera: 1024x1024 pixels, 90 degree field of view, 19 mm focal length, 1.5 m height, 30 cm stereo baseline. Simulated lidar: 75 degree vertical by 360 degree horizontal field of view, 0.333 degree angular resolution, 1.5 m height, tilted 23 degrees [5].

Stereo crater fitting degrades faster with range than lidar does [5]. A separate monocular detector built on YOLOv4 in TensorFlow, trained on 1000 hand-labeled Apollo surface images, reached 94 percent accuracy on its Apollo test set and detected craters larger than 10 m diameter at ranges of 15 m or more, with 5 m craters detectable to about 12 m, and showed no measurable sensitivity to solar zenith angle across the tested range. Transfer to Chang’e 3 and Chang’e 4 surface imagery, 168 and 1174 stereo pairs at 27 cm baseline, was evaluated qualitatively only [5]. The 5 m error budget is tighter than the 10 to 100 m a bearing-only landmark solution delivers on Mars [4]. No runtime or memory figure for the crater pipeline on a flight processor appears in the published work.

Crater matching in darkness and from the surface

Two JPL efforts extend crater matching to conditions where a sunlit stereo pair is not available. ShadowNav detects crater rim edges as stereo disparity discontinuities under an onboard lamp and matches them to an orbital crater map with a particle filter [10]. In Blender simulations it converged to 1.7 to 3.8 m final error across three trajectories, about 4 m better than the 2 percent odometry baseline, with edge detection best near 10 m range [10]. A field version on the Cinder Lakes crater field under a New Moon averaged 1.9 to 4.8 m error with maxima under 9 m, provided relative odometry stayed at 2 percent of distance or better, and ran in 16.47 s worst case on one Snapdragon 8155 core [9]. The light had to sit 0.2 m below the camera because of the opposition effect. Only 2D position was solved, with heading assumed from a star tracker that was not in the loop [9].

LunarNav built a particle filter and a Gaussian-mixture matcher for lidar or stereo crater detection, beat the 2 percent baseline in simulation with 0.75 to 2 m error, and raised the capability from TRL 2 to TRL 4 [11]. Its stereo detector did not generalize to real Chang’e 4 imagery, so no end-to-end result on real surface data exists [11]. None of the crater methods has flown.

A terrestrial analog from outside planetary work shows the structure without the lunar constraints: Night-Voyager localizes a monocular camera, IMU and odometer platform against a prior map of streetlight objects at night, and has no planetary test [30].

Ground-in-the-loop localization as flown

Every rover that has localized against orbital imagery in flight has done so with people in the loop. Spirit’s pipeline combined two-way Doppler and cartographic lander localization with onboard visual odometry and ground incremental bundle adjustment through sol 440 and 3.76 km, with accuracies that are internal-consistency estimates, since Mars has no surveyed control network [17]. Chang’e-3 compared dead reckoning, cross-site visual localization and DOM matching over the Yutu traverse [18]. The cross-site method matches ASIFT tie points inside search regions predicted by dead reckoning, rejects outliers by an inter-point distance test and bundle-adjusts the two-site network, on a ground workstation with manual intervention in difficult scenes [19].

Zhurong drove 2009 m true path in 206 site operations while the odometer read 1921 m. Localization error against a HiRISE-visible track was 3.11 percent of traverse length for onboard dead reckoning, 1.75 percent for cross-site visual localization and 0.50 percent after site DOM to orbital DOM matching at four waypoints [20]. On Yutu-2, dead reckoning and ground visual positioning agreed to centimeters or decimeters over up to 10 m between waypoints and diverged by more than 1 m when the rover slipped badly, where visual localization corrected the error [21]. Yutu-2 read 190.66 m on the odometer against 186.66 m from visual localization at the end of lunar day 5, about 2 percent slip [22]. Chang’e-4 had no radio tracking to the Queqiao relay, so its landing site was fixed from 180 descent images bundle-adjusted against a Chang’e-2 reference; the 415 m disagreement with the LRO-derived coordinate shows the terrain reference, not the photogrammetry, dominates the absolute error [23].

Orbit-side reconstruction needs no rover cooperation. Differencing two OHRC images 2.5 years apart located eight Pragyan in-place turns, though the validation reuses the commanded turn locations used to predict them [24]. A 0.30 m OHRC stereo DEM of the Chandrayaan-3 site, with a median triangulation error of 8.1 cm, supplies the kind of base map these fixes depend on [25].

Lander-relative fixes

Close to a lander, the lander itself is the reference. The Rocky 7 testbed was localized by lander-mounted stereo cameras that imaged the rover, detected it by differencing images across a commanded in-place turn of 0.3 radians, and triangulated its position [6]. Over 150 trials at 2 to 10 m range, the downrange estimate was biased 8.32 percent low with a standard deviation of 6.66 percent, and cross-range mean absolute error was 3.30 percent of downrange distance with a standard deviation of 4.07 percent [6]. Accuracy is proportional to range, so the method is useful only in the first few tens of meters. The same paper’s rover-mounted mast stereo localization [6] over 12 images taken at 1 m spacing along a straight line held 0.0317 m downrange and 0.0366 m cross-range average absolute error, two orders of magnitude better, which is why the lander-relative method did not survive into later missions.

The technique returned for surface construction robots, where a fixed structure is available and the working area is small. The ISRU Pilot Excavator uses a mock lunar lander carrying 16 AprilTags of the 36h11 family, each 15.24 cm square overall with a 12.1 cm black-to-black edge, mounted along the top perimeter and at the middle of the lander legs and support beams, to compute an absolute pose [7]. The tags are printed in Musou black acrylic, 99.4 percent light absorbing, on a white vinyl background, and detection covers both standard and inverted fiducials through a custom detector derived from the WhyCon system, because the lunar south pole lighting the site reproduces is a 6.5 degree solar elevation with long shadows and extreme contrast. Over eight days of testing on a 33.5 by 21.3 m area of Florida soil sieved to remove rocks above 7.6 cm, the vehicle ran 334 excavation and unload repetitions at a 9 minute average cycle, drove 58.1 km, sustained 40 cm/s, and completed 35 automated dockings, across 77 hours 46 minutes of hardware time [7]. The published report names ROS as the pose transport and describes the compute as a surrogate for the flight avionics, but gives no processor, clock, memory figure or detection runtime, and reports no pose accuracy or fiducial acquisition range.

Map-relative fixes for landers and aerial vehicles

The same correlation idea, at descent speeds, is the only absolute method that has flown autonomously on Mars. The Lander Vision System on Perseverance reduced an initial horizontal uncertainty of up to 3.2 km to 40 m in 10 s by matching five large patches in each of three descent images to a 12 m per pixel map, then up to 150 small patches per image to a 6 m per pixel map in an iterated extended Kalman filter fed by 200 Hz IMU data [13]. In flight it matched 15 of 15 coarse landmarks and found a 137 m position correction, from one descent in benign illumination [13]. Earlier survey work reduced the resolution requirement to a rule: about 20 by 20 map pixels are needed for a reliable fix, and the lidar simulations behind it retune the beam at every altitude in a way flight cannot [15]. A Cramer-Rao bound on correlation-based map-relative localization showed that for high signal-to-noise Mars imagery the error is dominated by estimator bias rather than noise [14]. Post-processed replays of ALHAT and Perseverance entry data put a LiDAR-inertial factor graph within 2 m on Morpheus flights, with no runtime reported [33].

For a Mars helicopter concept, SIFT features from a navigation image were matched to a database built from HiRISE and a DEM, with a pose prior from range-aided visual inertial odometry [12]. On terrestrial flights at 35, 50 and 100 m altitude it localized within 2 m on 99 percent or more of the frames that returned a result, at 0.3 s per image on a desktop core, and returned nothing usable at 12 m [12]. A single-point laser range update held range-VIO error under 0.5 percent of distance on a 90 m flight, while the sun sensor attitude contribution is simulation only [32]. Both are Earth demonstrations.

Ground-penetrating radar is a proposed sensor for texture-poor terrain; a two-week Mars Desert Research Station collection produced data but no accuracy figure [31].

Comparison

Method	Reference frame	Accuracy	Runtime and processor	Flight status
Sungaze, QUEST on Sun vectors	inertial, via ephemeris	1.5 deg, 3 sigma attitude	10 min timer, RAD6000 at 20 MHz	MER, flown, about every 20 sols [1]
Black SunID	inertial, via ephemeris	not published as a number	not published	Perseverance, flown from sol 169 [2]
Human orthomosaic to HiRISE match	HiRISE map	50 cm expected	one Earth-Mars cycle per fix	MER, MSL, M2020, operational [3]
Censible modified census transform	HiRISE map, 25 cm and 1 m DTM	0.36 m mean, 99 percent under 0.93 m	32 s, Snapdragon 801 at 2.36 GHz, 4 cores	Perseverance, flight demonstrated [3]
Bearing-only topographic landmarks	orbital DEM	10 m best, 10 to 100 m typical, unbounded in occlusion	offline viewshed on the ground	study only [4]
Crater landmarks, lidar	LRO NAC map	0.28 to 1.94 m, 3 sigma, by crater size	not published	simulation and Chang’e imagery [5]
Crater landmarks, stereo	LRO NAC map	1.02 to 4.48 m, 3 sigma	not published	simulation [5]
Lander stereo triangulation	lander	8.32 percent of range downrange	not published	Rocky 7 testbed, 150 trials [6]
AprilTag fiducials on a lander	lander structure	not published	not published	IPEx field campaign, 334 cycles [7]

Failure modes

Sun-based attitude acquisition fails on geometry. Sunfind is ill-conditioned near local noon [1], and SunID’s usable window is bounded by solar elevation, 30 to 60 degrees for the normal algorithm and 15 to 87 degrees for the black-sun variant [2]. Because the window depends on sol, latitude, season and rover tilt, planning has to schedule around it: Perseverance operations generate SunID time constraints from a script that combines the algorithm’s elevation limits, the Sun’s position and the terrain features that occlude it, and a separate occlusion report looks 10 sols ahead for high gain antenna blockage [2].

Orbital map matching fails where the terrain is not distinguishable at map resolution. The census transform approach recorded no failures across its 264-panorama benchmark [3], but the methods it displaced, normalized cross correlation and mutual information, failed by producing confident wrong answers rather than by refusing, and that outlier behavior is what kept them on the ground. Bearing-only landmark localization fails by occlusion, and does so predictably: the same drive campaign that gives 10 m accuracy on open ground gives worse than 85 m when the path drops into a riverbed [4].

Fiducial-based localization fails on lighting rather than on geometry: at 6.5 degree solar elevation, shadow masking, high contrast and non-uniform illumination across the work area were the recorded challenges of the IPEx campaign [7].

Operational coupling

Localization quality sets what can be commanded, not just what is known. Targeting accuracy requirements on MER ran from azimuth and elevation pointing for distant or atmospheric targets, through 3D coordinate targeting accurate enough for features within about 20 m, to 0.5 cm positioning for in-situ instrument placement with the arm [8]. Targets defined in imagery from an earlier rover position degrade as localization error accumulates, which is why the Maestro planning tool tracks target versions by planning position and labels a target created at the current position safe, one carried over from older data unsafe. The same argument drives single cycle instrument placement, which uses visual tracking from the mast cameras into the hazard cameras during the approach to place an instrument on a target 2 to 3 m away in one sol instead of three [8].

Attitude and position estimation are also what the ground team spends its shift on. Perseverance operations targeted a 5 hour daily downlink and uplink timeline against a heritage of 10 hours or more per shift on MER and MSL [2]. Automating the collection of attitude update, gyrocompass and SunID results into the operational report cut that report’s preparation from 16 minutes 41 seconds to 5 minutes 32 seconds in a sol 180 mock shift, measured end to end.

Perseverance’s ascent of the Jezero rim also records the first flight use of a non radiation hardened co-processor inside a rover onboard safety assessment, the same Snapdragon class used by the global localization demonstration [26].

What is not established

The flown onboard absolute fix is one demonstration on the Helicopter Base Station, not a routine capability on the rover’s own RAD750 pair [3]. No crater-based method has run on a rover with independent ground truth: the lunar results are simulation, a night field campaign, or small hand-labeled sets [5][9][10][11]. Heading from stars was assumed in the dark-crater work and never demonstrated [9]. Ground localization accuracies are measured against hand-digitized orbital tracks at a few waypoints, not continuously [20]. Sun-sensor accuracy for Black SunID is not published as a number [2]. Fiducial localization on the IPEx lander reports no pose accuracy [7]. Runtime figures on flight processors are absent for the crater pipelines [5].