Localization and mapping

A planetary vehicle has no GPS, no magnetic compass and no surveyed control network. Localization must produce a position and attitude good enough to point an antenna, place an arm on a target and aim a drive, and mapping must produce the terrain model the position is expressed in. Wheel odometry drifts with slip, inertial sensors drift with time, and the only external references are the sun, the stars, orbital imagery and what the cameras see. The MER requirements illustrate the scale: attitude within 1.5 degrees 3 sigma so the high gain antenna points within 2 degrees, and position error under 10 percent of distance traveled up to 100 m [1]. Both vehicles exceeded that budget on slopes, where slip reached 100 percent or more [1].

Design space

Four families of method cover what has been flown or tested, and they differ mainly in what reference they tie the estimate to.

Dead reckoning and absolute heading. Wheel odometry plus a gyro or IMU is the baseline on every rover. Field tests on Rocky 7 showed why heading dominates: with an absolute heading sensor (sun sensor, accelerometer tilt and clock) position error grew linearly at 6 percent of distance over 335 m, whereas integrated rate sensing grows faster than linearly [5]. MER propagates attitude at 8 Hz and reacquires it by sun search roughly every 20 sols [1]. The recursive filters behind these estimators descend from Kalman’s linear formulation [31], and the uncertainty-tracking frame structure that lets later observations be merged consistently goes back to early mobile robot world modeling [33].

Visual odometry. Tracking features between stereo pairs measures the slip that encoders cannot. On MER it converged on 97 percent of attempts on Spirit and 95 percent on Opportunity, but cost nearly three minutes per step on the 20 MHz RAD6000, so it was commanded only for specific motions [2]. On Curiosity it converged 20,588 times in 20,682 attempts and over 88 percent of driving used some form of it despite 47 seconds per step [3]. Running it while the vehicle moves removes much of that stop-and-process penalty [4]. Algorithms with no prior on motion, built on a separate disparity stage, exist for faster platforms but are demonstrated on terrestrial vehicles [35]. The robust-estimation and feature-tracking building blocks are older than any flight use [31][33].

Ground-based bundle adjustment and orbital map matching. Operations teams correct onboard dead reckoning on the ground. For Spirit, a pipeline of lander localization by Doppler and orbital triangulation, then incremental bundle adjustment of rover imagery, carried 3.76 km of traverse through sol 440 [7]. For Zhurong, error against a HiRISE-visible track was 3.11 percent of traverse for onboard dead reckoning, 1.75 percent for cross-site visual localization and 0.50 percent after matching site orthophotos to orbital imagery at four waypoints [8]. On Yutu-2, dead reckoning and ground visual positioning agreed to centimeters to decimeters between neighboring waypoints and differed by more than 1 m under heavy slip [9]. Combined descent, orbital and surface mapping was demonstrated in analog tests [24]. Onboard maximum likelihood matching of a local occupancy map to a global one is older still and was exercised on a rover testbed [37].

Absolute localization onboard. The first onboard global fix on a Mars rover came on sol 859, when Perseverance matched post-drive stereo orthomosaics to a HiRISE map, reaching 99 percent within 0.93 m in benchmarks; the 10 KB/s link to the co-processor made image transfer 75 percent of wall time [13]. Earlier surveys concluded that visual localization from the orbital data then available was not achievable at kilometer scale [6]; orbital resolution is what changed that. For the Moon, crater matching is the proposed route: crater detection from lidar, stereo or monocular images [16], particle-filter and Gaussian-mixture matching beating a 2 percent of distance baseline in simulation [15], and a stereo plus flood light variant for darkness reaching 1.9 to 4.8 m average error on analog field data [14]. None of the crater methods has flown. Night-time localization against a prior map has also been shown for ground vehicles outside the planetary field [32], and orbital peak bearings have been analyzed as a drift correction for Mars rovers by geometric bound only [34].

Terrain relative navigation

Absolute localization during descent is the one place a map-matching method has operated on a landing. The Mars 2020 Lander Vision System reduced a 3.2 km initial position error to under 40 m within 10 seconds from 4200 m altitude, matching landmarks coarse to fine against a 6 m per pixel onboard map; Perseverance touched down within 5 m of its selected target against a 60 m requirement [10]. The flight correction was a single sample and well inside the design case [11], and the pre-flight qualification used 656 helicopter runs, of which 7 exceeded 40 m [12]. The camera simulator used to verify it was compared against only two flight images [22]. The lighting limit is now the open issue: landmark matching degrades when the descent image and the reference map are lit from opposite azimuths, which a morning landing for a sample retrieval lander would force [22]. The enhanced design for that lander makes this system the primary EDL navigation sensor with a 60 m pinpoint target [23]. The history of the approach runs from 1950s radar map matching through JPL’s development to Mars 2020 [36], and the achievable accuracy has a Cramer-Rao bound set by map resolution and terrain content [21].

Small-body proximity operations use feature-based variants. OSIRIS-REx rendered catalogd features from a terrain model and correlated them with live images [17], and that Natural Feature Tracking drove the touch-and-go guidance update, cutting predicted position dispersion from 17 to 20 m to 5.7 to 8.1 m 3 sigma in a pre-flight analysis [18]. For lunar landing, ALHAT flash lidar hazard detection flew five times on the Morpheus lander: hazard detection worked on every flight, while hazard relative navigation largely did not [19]. A lidar-inertial factor graph replayed on those datasets and on the Mars 2020 descent reached horizontal error of 149.4 m at 6.4 km and 49.1 m at 15 m altitude, in post-processing only [20]. Photogrammetric descent reconstruction determined the Chang’e-4 landing site where no radio tracking existed [38].

Mapping and SLAM

Flight mapping is ground-processed stereo: site orthophotos and digital elevation models at 0.02 m grid spacing out to 15 m on Zhurong [8] and 0.02 m local models per waypoint on Yutu-2, supplemented by a 0.03 m orthophoto from descent images [9]. Orbital imagery also localizes after the fact: eight Pragyan turn locations were recovered by differencing two orbiter images taken 2.5 years apart [39]. Onboard SLAM has flown on the Astrobee free flyers on the ISS: a factor graph localizer cut average runtime from 1.13 s to 0.21 s and removed two lost-localization events [25], and its decoupled successor separates odometry from map-based absolute localization [26]. For aerial concepts, laser range updates fix scale drift in visual-inertial odometry in a 90 m outdoor flight [29], and SIFT matching to HiRISE-derived databases gave localization within 2 m on 99 percent or more of terrestrial flight frames [30]. Multi-robot loop closure is limited by the quadratic growth of candidates, addressed by prioritization [27] and by beacon-based loop detection [28]. These multi-robot and aerial methods are ground-tested, not flown on another planet.

What is not established

Most accuracy figures lack independent truth. Mars has no control network, so Spirit bundle adjustment accuracy is internal consistency [7], and the Zhurong figures are scored against a hand-digitized track from orbit [8]. The Lander Vision System flight is one descent, in benign light, with a 137 m initial correction against a 3.2 km design case [11]. The lunar crater localizers have no end-to-end result on real surface data, and the stereo detector failed on Chang’e imagery [15]; ShadowNav estimates 2D position only, with attitude assumed from a star tracker that was not in the loop [14]. The Morpheus tests were terrestrial with GPS available [19], and the factor graph descent results are post-processed with no runtime figures [20]. Onboard Mars absolute localization has one flight demonstration [13].