Yutu

CNSA / China Lunar Exploration Program [14].

The problem

A lunar rover has to land on terrain it cannot yet see at rover scale, locate itself on that terrain without GPS, drive without slipping into a hazard, and survive a lunar night. The Chinese landing program constrains the candidate area to an average slope of at most 8 degrees and a solar elevation of 30 to 50 degrees at landing, a limit that follows from spacecraft thermal control; latitudes below 30 degrees see large temperature swings and latitudes above 60 degrees violate the elevation window [8]. Soil strength at the site is not known in advance, so wheel design starts from Apollo and Luna priors [9]. Dead reckoning from a wheel odometer and an inertial unit accumulates error from wheel slip and drift, so a second position source is required [3]. At a landing site on the near side, commands and telemetry go through ground stations in view of the site; on the far side they go through a relay satellite [8], [23].

What has flown

Yutu, Chang’e-3 (2013). Chang’e-3 launched on a Long March 3B on 2 December 2013 and landed on 14 December 2013 at 19.51 W, 44.12 N in northern Mare Imbrium; the lander and rover separated and imaged one another on 15 December [6], [13], [1]. Powered descent ran from about 15 km altitude to touchdown with hazard avoidance by 3-D laser imaging from a 100 m hover, rejecting slopes steeper than 8 degrees over a 10 m baseline; the chosen 20 by 20 m area had slopes under 2 degrees and obstacles under 20 cm [16]. The rover carried four instruments: a Panoramic Camera, a Visible and Near-Infrared Imaging Spectrometer (VNIS), an Active Particle-induced X-ray Spectrometer (APXS) and a Lunar Penetrating Radar (LPR) [6]. It drove 114.8 m by cross-site visual localization, 111.2 m by DOM matching and 117.4 m by dead reckoning to waypoint N0209, each total including 4.4 m during separation [18]. The traverse crossed the ejecta of a young crater of about 450 m, and radar profiles resolved more than nine subsurface layers to about 400 m [15]. The CNSA pages give no cause for the end of mobility [1], [14], [2].

Lunokhod 1 (1970), the precedent. Lunokhod 1 is the earlier flown lunar wheeled vehicle, documented by a participant’s account of its chassis design and testing [26]. Its soil was characterized by a cone penetrometer, whose force against depth data for four terrain units were reduced to unit penetration resistance and compared with the Apollo 14 Simple Penetrometer [28]. Its traverse was later re-reconstructed on a 0.5 m per pixel LRO Narrow Angle Camera DEM, correcting the 1970 to 1971 ground-track record [27].

Yutu-2, Chang’e-4. Chang’e-4 went to the South Pole-Aitken basin with a relay satellite at Earth-Moon L2, and Yutu-2 carried a panoramic camera, a penetrating radar and a VIS-NIR spectrometer [6]. Yutu-2 reused the Chang’e-3 localization chain on the far side. At the end of lunar day 5 its odometer read 190.66 m against 186.66 m from visual localization, about 2 percent accumulated slip [3]. Onboard dead reckoning and ground visual positioning agreed to the centimeter or decimeter level over the up to 10 m between waypoints and diverged by more than 1 m where the rover slipped badly [11]. The command loop is relay-limited: a rover responds in about 10 s, the first telemetry frame arrives 80 to 100 s later and refreshes every 24 s, so confirming execution takes 3 to 4 minutes, and the rover moves at about 3.3 m/min [19]. Queqiao carries a 4.2 m parabolic relay antenna in a halo orbit at Earth-Moon L2 [23]. The LPR 60 MHz channel reached about 330 m penetration but its near-surface returns are saturated by coupling with the rover structure, so the 500 MHz channel with about 35 m penetration fills the shallow profile [21].

Zhurong, Tianwen-1 (2021), same method on Mars. Zhurong drove 2009 m of true path in 206 site operations over 360 sols while its odometer read 1921 m; onboard dead reckoning erred by 3.11 percent of traverse length [31]. Its suspension is active, with a main and auxiliary rocker arm and a clutch at the main rocker joint giving body lifting and single-wheel lifting for self-extraction [35].

What is being tried now

Visual localization and mapping. Flown. Cross-site visual localization matches ASIFT tie points between stereo pairs at adjacent waypoints inside search regions predicted from dead reckoning, rejects outliers by an inter-point distance test and closes the pose with a two-site bundle adjustment. Against indoor truth its error was 1 to 4 percent of traverse with a 3.3 percent mean, against 15 percent for dead reckoning alone; in flight it ran on a 32-core ground workstation in under 2 minutes per site including data transfer [10]. It ran at every waypoint within minutes of downlink; DOM matching, which matches orthophotos from Navcam stereo against orbital or descent basemaps, ran only where small craters or large rocks were present and was judged the most accurate [3]. The ranking rests on assertion rather than residuals against independent truth [3].

Descent-image and orbital mapping. Flown. For Chang’e-3, 180 descent images from 2.35 km to 10 m altitude with 26 ground control and 18 check points gave DEM and DOM products down to 0.05 m [3]; bundle adjustment residuals were 0.724 at control points and 0.908 at check points in the units reported [18]. The Chang’e-5 landing region near Mons Rumker has a prelanding geologic map of 14 units spanning about 3.7 to 1.2 Ga [20]. A two-stage procedure produces a continuous meter-scale orthophoto from LROC NAC images, with a planar block adjustment per subarea and heights interpolated from SLDEM2015 [25]. Rover Pancam stereo then refines the map: Yutu-2 5 cm DEMs showed the Chang’e-4 site sits on overlapping ejecta from Finsen and Alder craters [22].

Computer vision in teleoperation. Flown on Yutu-2, with the localization chain above and a Beijing Aerospace Control Center team doing the operation [11]. Commands carry a time parameter as well as a distance because a drive can fail unnoticed for minutes [19].

Wheel-soil modeling. Three families exist, none verified against Yutu flight data. A smoothed particle hydrodynamics continuum model run in Chrono, with a 137 kg six-wheel Yutu model, reproduces discrete element and single-wheel tractive force and torque curves at lower cost; it is a simulation [12]. Discrete element models resolve grains and selected lug geometry for a lunar micro rover, a 250 mm rim and 100 mm wide wheel with 18 lugs, on a JAXA slope bed in FJS-1 simulant, in two dimensions only [30]. Flight-derived estimation inverts wheel tracks: Yutu-2 track imagery gave cohesive and frictional moduli at the Chang’e-4 site [9], and a slip event on lunar day 41 gave a farside bearing capacity near 4 kPa against 2 to 3 kPa along the Lunokhod 1 traverse, with sinkage about 8 mm and never above 15 mm [4]. The continuum solver behind [12], Chrono::CRM, is validated against sphere cratering and cone penetration experiments at mean relative errors of 3.5 and 7.4 percent [34].

Regolith properties. Returned Chang’e-6 farside samples have a static angle of repose of 52.9 degrees, above any Chang’e-5 or Apollo simulant, from a D60 of 48.4 micrometers [24]. Chang’e-5 sampling used a drill with a goal of about 2 m of core, of which about 1 m was recovered [32]. At the Chang’e-3 site, Navcam stereo from 7 waypoints gave the rock size-frequency distribution on a crater ejecta blanket [17].

Thermal. The only flown vehicle-level thermal detail is a radioisotope heater unit on the APXS sensor head [5]. For the environment, Diviner nighttime cooling gives a thermal inertia of 55 J m^-2 K^-1 s^-1/2 at 273 K for the upper 4 to 7 cm [29], and the NASA specification for natural environments collects the lunar thermal and dust environments in one reference [33].

Space weathering and subsurface sounding. VNIS spectra of freshly excavated soil against weathered soil at the landing site measured the optical effect of space weathering in situ, with an optical maturity parameter of 0.34 at Site 5 and 0.18 at Site 8 [7]. LPR on Yutu was the first ground-penetrating radar operated on the lunar surface [6], [15].

The numbers

Parameter	Value	Source
Wheels	six, three per side, two-arm linkage per wheel	[12]
Wheel radius, width	0.13 m, 0.16 m	[12]
Mass in the published multibody model	137.0 kg, not an agency specification	[12]
Lander dry mass limit	1.22 t	[16]
Pancam, Navcam stereo base	270 mm on a shared mast bar; Hazcam body-mounted	[3]
Pancam image	2352 by 1728 pixels, 22.9 by 16.9 degrees	[3]
Navcam image	1024 by 1024 pixels, 46.6 degrees	[3]
Distance, cross-site visual localization	114.8 m	[18], [3]
Distance, DOM matching	111.2 m	[18], [3]
Distance, dead reckoning	117.4 m	[18]
Cross-site localization error, indoor truth	1 to 4 percent, mean 3.3 percent	[10]
Dead reckoning error, indoor truth	15 percent	[10]
VNIS spectral range	450 to 950 nm (100 channels), 900 to 2395 nm (300 channels), 5 nm sampling	[7]
APXS detector	silicon drift, 7 square millimeters, better than 140 eV at 5.9 keV, 0.5 to 20 keV	[5]
APXS sources	4 x 70 mCi 55Fe and 4 x 5 mCi 109Cd	[5]
APXS sensor head mass	754 g	[5]
APXS accuracy, ground test	major elements under 15 weight percent relative at 30 mm and 30 minutes	[5]
APXS recommended standoff	under 50 mm	[5]
Regolith conductivity, surface to 1 m	7.4e-4 to 3.4e-3 W m^-1 K^-1	[29]
Regolith density, surface to 1 m	1100 to 1800 kg m^-3	[29]

The APXS ground verification used two flat prepared samples and does not establish performance on irregular natural surfaces [5]. The simulation in [12] swept slip ratios from 0 to 0.8 at 0.2 m/s on soil parameter sets of 1627 to 1839 kg per cubic meter bulk density and 37.8 to 47.8 degree friction angle; tractive force and its oscillation amplitude rise with friction angle and density, and a larger rim diameter or wider wheel lowers the slip needed to climb a slope [12].

What is not established

No CNSA page gives Yutu’s mass, dimensions, speed, gradeability, power or thermal limits [1], [14]. No processor, memory, link frequency or data rate for the rover is published in the held sources. The cause of the end of mobility is not recorded in them [1], [14]. The localization ranking has no independent truth, and traverse totals are partly quoted from press releases [3]. The APXS test was not run on regolith at lunar temperature in vacuum [5]. The wheel-soil models are simulation or lunar-general and were not validated against Yutu flight data [12], [30]. The only soil strength estimates near the Chang’e-3 site come from later Chang’e-4 imagery analysis [9]. The 8 degree slope limit is quoted from another paper and its derivation from rover capability is not shown [8].