Chang'e-6 Sampling System

The problem

A far-side sample return has no direct-to-Earth link and a surface stay measured in hours. A sampler for that case has to pick where to dig and confirm that its own arm did what it was told, with the commanding delayed by a relay hop. Autonomy is required, not merely convenient, when the environment or spacecraft changes, the change is not predictable in advance, and the needed response is faster than the next communication cycle [2]. Chang’e-6 had about 48 hours between landing on 2 June 2024 and ascent on 4 June [11]. Chang’e-5 spent 3.12 h drilling and 15.82 h on 12 surface sampling operations on the near side [7].

The surface adds three difficulties. Regolith in place is denser than loose terrestrial simulant, with a relative density of about 65 percent in the top 15 cm at six Apollo and three Luna equatorial sites [3]. The returned Chang’e-6 material is unusually cohesive: a static angle of repose of 52.9 degrees and a dynamic angle of 70.4 degrees from a 5 g sample, with D60 of 48.4 um, above every Chang’e-5 and Apollo simulant tested [13]. And the landing zone, 150 to 158 W and 41 to 45 S in the southern Apollo basin, was chosen on orbital geology, with no published slope or roughness thresholds in the site paper [9]. A scooping arm with a flexible structure and no roll degree of freedom has to work inside those conditions without a human fine-tuning each point.

What has flown

Chang’e-6. The lander reused the Chang’e-5 architecture: a rotary-percussive drill for a subsurface core and a four-joint arm carrying a scooping sampler [1], [7]. It launched 3 May 2024 on a Long March 5, separated lander and ascender on 30 May, landed 2 June in the South Pole-Aitken basin, sampled on 2 and 3 June with the drill and arm, ascended 4 June, and the capsule landed in Inner Mongolia on 25 June [11]. All surface communication went through Queqiao-2, launched 20 March 2024 on a Long March 8 [6], [11]. The returned mass was 1935.3 g, a preliminary CNSA figure, and the samples were described as more viscous with clumps than Chang’e-5 material [12]. The scooping autonomy selected ten sample points over terrain with an average slope of 5.1 degrees, and points 1 to 3 and their neighborhoods were scooped [1]. The material is mare basalt with an eruption age of about 2.8 Ga [10], [5].

Chang’e-5. The precedent. An operator confirmed each sample point from imagery and a specialist fine-tuned arm pose, a procedure with high training and time cost [1]. Its drill traveled 1.68 m for about 1 m of core, its drilling chain drew above 800 W, and Chang’e-4 had already flown autonomous terrain-relative navigation with hazard avoidance for far-side landing [7].

Lunar PlanetVac. Acquired regolith on Firefly Blue Ghost Mission 1 at Mare Crisium on 3 March 2025 [32] in a five-second gas-driven sequence [16]. An earlier team abstract quotes 13 g and the later one 6.5 cc, and the two are not reconciled [15]. The head was off nominal, so this is one event at one site [32].

PRIME-1. The TRIDENT 1 m rotary-percussive drill was activated on the surface in March 2025 [20]. Rotation, extension and percussion were demonstrated, but the drill was oriented horizontally and buried in landing ejecta, so no drilling into undisturbed regolith occurred [17], [20].

Perseverance sampling. The nearest flown arm-and-corer record for another planet: through about sol 1000 the rover ran 7 paired core sampling sol paths, 1 paired regolith path and 7 single sampling paths, of which 6 yielded a sample [24]. Planned sol counts were exceeded, often by a wide margin, because of rover runout and drilling faults from unpredictable rock properties, and the first attempt returned an empty tube [24].

Yutu-2. Not a sampler, but the only in-situ far-side regolith mechanics: wheel sinkage of about 8 mm, bearing capacity near 4 kPa, friction angle 21.5 to 42.0 degrees and cohesion 520 to 3154 Pa, inferred by model inversion at the Longji site [14].

What is being tried now

Vision-based sample selection and pose correction

This family has flown once, on Chang’e-6. Terrain point clouds from a binocular camera give a slope map and segment obstacles and pits. Candidates are scored by a four-term loss over local slope, proximity to obstacles and pits, distance from the reach limit, and the roll-direction slope component the arm cannot correct. Simulated annealing selects the lowest-loss set with a 400 mm minimum spacing [1]. The sampler pose is then measured by registering its point cloud to a template, with TEASER followed by ICP, which replaces Chang’e-5’s manual fine-tuning [1]. The published accuracy is terrestrial only, as listed under the numbers below.

Learned scooping and perception

None of this has flown. Few-shot scooping with a deep kernel Gaussian process, meta-trained on 5,100 scoops over 51 terrains, was tested on 16 novel terrains and then unchanged on the JPL OWLAT testbed, in 1 g with household and landscaping materials, not regolith [23]. A DreamerV3 excavation agent trained in particle simulation has been evaluated only in simulation, with no hardware or sim-to-real test [28]. On the perception side, a semi-global matching pipeline for rover navigation maps was evaluated on Devon Island images without ground truth [25]. LunarStereo, over 50,000 synthetic south-pole stereo pairs, was used to fine-tune MASt3R and tested on synthetic renderings and real Chang’e-3 descent images [26]. For flown stereo, an analysis of 16,768 Perseverance NavCam pairs found holes large enough to stop autonomous navigation at solar phase angles under 20 degrees and on low-texture terrain [27]. Chang’e-6 uses a learned matcher (CREStereo) among classical point cloud tools [1].

Pneumatic acquisition

PlanetVac removes the arm and the pose-error problem, at the cost of a gas supply and a short acquisition event. It has flown once, as described above [16], [32].

Rotary-percussive coring and volatile drilling

Drilling preserves stratigraphy at the price of a heavy, power-hungry mechanism and loses volatiles and bedding in regular coring [8]. TRIDENT descends from a twenty-year lineage and exists in two flight units, for PRIME-1 and VIPER [17]. It has been run in two 2023 field campaigns at Haughton Crater and the Bishop Tuff [30], in a vacuum chamber where four 1 m holes in stratified simulant at 2.5 and 5 wt percent water were monitored by mass spectrometer [31], and in Antarctica, where drill telemetry identified subsurface horizons (not peer reviewed at the version held) [18]. The pre-flight program requirements for PRIME-1 are set out in [19]. ICP-Coring, a sealed liner to preserve volatiles in a returned core, is a design study with no integrated coring test [8].

Arm-based autonomy for ocean and icy worlds

A JPL Europa Lander prototype ran 58 field tests on the Matanuska Glacier in 2022 and recovered from 12 mid-level faults while excavating a 27 cm trench with no operator input; most system-level autonomy was exercised in software simulation [22]. It is a ground demonstration.

Crewed and long-range concepts

SAMPLERS, a suit-mounted scoop and hammer tool, is a mockup tested with four subjects in analog suits [21]. Endurance, a proposed 2000 km autonomous rover collecting up to 100 kg of regolith at 12 sites, is a concept briefing with requirements, not measurements [29].

The numbers

Parameter	Value	Source
Sampling mechanisms	rotary-percussive drill and four-joint scooping arm	[1], [7]
Arm reachable region	sector of 2.8 m radius and 120 degrees	[1]
Absolute arm error, Earth, balloon offload	about 30 mm and 2 degrees	[1]
Sampler pose precision, Earth	1.1 mm and 0.4 degrees, 1.6 s average	[1]
Landing	2 June 2024, approximately 154.0 W, 41.6 S	[11], [1]
Returned sample mass	1935.3 g	[12]
Landing zone average slope	about 5.74 degrees, 76 percent of area below 8 degrees	[9]
Craters over 50 m in the landing zone	26,785	[9]
Returned basalt age	about 2.8 Ga	[10]
Surface temperature near 45 degrees latitude	about 350 K at noon, about 89 K before sunrise	[4]
Chang’e-5 drilling and sampling time	3.12 h and 15.82 h, 12 surface operations	[7]
PlanetVac acquisition	five-second sequence, 6.5 cc reported later, 13 g reported earlier	[16], [15]
Perseverance sampling through about sol 1000	7 paired cores, 1 paired regolith, 6 of 7 single samples yielded	[24]
Europa Lander field prototype	58 tests, 27 cm trench, 12 autonomous fault recoveries	[22]

The pose pipeline uses CREStereo for matching, the Point Cloud Library, YOLOv5 to isolate the sampler region, SIFT and FPFH descriptors, and TEASER with ICP for registration, and registers correctly with under 30 percent of the sampler visible [1]. Two systematic errors are compensated explicitly. Soil sinking is the difference between calculated and vision-measured touching height. Support force arises because the force sensor has an activation threshold, and is recovered by comparing the displacement the four joint angles report with the vision-measured displacement [1]. Too high a final height collects nothing, and too low takes excess soil and can stick the shovel.

The Chang’e-6 sampler has no published power or thermal design. The returned scoop sample bulk density and particle size are broadly consistent with Chang’e-5 and Apollo values, each resting on a single sample or small aliquots [5]. A NASA design envelope is 391 K to 96 K at the equator [33].

What is not established

Every vision accuracy number [1], including the 1.1 mm and 0.4 degree pose precision and the 3.8 mm scooping error, is from an Earth testbed with a balloon offload. No on-orbit verification, achieved lunar scooping depths or flight failure and retry statistics are published, and individual fine-tuning results span 19.0 to 26.7 mm against a 22.0 mm target [1]. The claimed time saving over Chang’e-5 has no before-and-after operator timings, and the split between onboard and ground computation is not stated. The soil sink offset was measured in simulant, while the returned soil is far more cohesive than nearside material [1], [13]. Regolith density context comes from equatorial Apollo and Luna sites, not the Chang’e-6 site [3]. Queqiao-2 link parameters, drill instrument parameters and per-point masses are not published in the sources held [6], [12]. PlanetVac and PRIME-1 are single events, and TRIDENT has not drilled undisturbed lunar regolith [16], [17]. The learned scooping and stereo methods are untested on the Moon [23], [28].

References