Chang'e-6 Sampling System

Overview

Chang’e-6 returned the first samples from the far side of the Moon. The mission was built as the Chang’e-5 backup and reuses its sampling architecture: a rotary-percussive drill for a subsurface core and a four-joint mechanical arm carrying a scooping sampler for surface material [1], [2]. The camera configuration is likewise similar to Chang’e-5’s, including a binocular camera.

The sampling window was about 48 hours between landing on 2 June and ascent on 4 June [7], against Chang’e-5’s 3.12 h of drilling plus 15.82 h of surface sampling on the near side [6], and the far-side location imposed a relay link. Chang’e-5’s manual per-point confirmation and specialist pose fine-tuning were too slow for that budget [1]. Chang’e-6 replaced both with a vision pipeline, selecting its own sample points and correcting its own manipulator pose.

Autonomy is required rather than merely desirable when three conditions hold together: changes in the environment or spacecraft occur, those changes are not predictable, and the required response time is shorter than the next communication cycle [4]. The last of the three is the operative one here.

Specifications

Parameter	Value	Source
Sampling mechanisms	rotary-percussive coring drill and a four-joint scooping arm	[1], [2]
Arm joints	4, sampler attached at joint 4	[1]
Arm reachable region	sector of 2.8 m radius and 120 degrees
Sampler attitude control	pitch adjustable, roll not adjustable
Sampler sensing	touching disc with force sensor, front shovel, narrow-field close-up camera
Binocular camera image size	2352 x 1728 pixels
Binocular baseline and focal length	200 mm, 15.4 mm
Binocular working imaging distance	2 to 5 m
Expected 3D reconstruction error	5 mm at 2.5 m
Sampler pose measurement precision	1.1 mm and 0.4 degrees
Sampler pose measurement time	1.6 s average

Mission profile

Parameter	Value	Source
Relay satellite	Queqiao-2, launched 20 March 2024 on a Long March 8	[7]
Launch	3 May 2024, Long March 5 from Wenchang
Circumlunar orbit insertion	8 May 2024
Lander-ascender separation	30 May 2024
Landing	2 June 2024, South Pole-Aitken basin
Landing point	approximately 154.0 W, 41.6 S, southern Apollo basin	[1]
Sampling	2 to 3 June 2024, drill and robotic arm	[7]
Ascent	4 June 2024
Rendezvous and sample transfer	6 June 2024
Earth return	25 June 2024, Siziwang Banner, Inner Mongolia
Surface window available for sampling	about 48 hours
Returned sample mass	1935.3 g	[8]

The program timeline page places sampling on 2 and 3 June and ascent on 4 June [7], which brackets the roughly 48 hour surface window the autonomy was designed against [1]. The page states neither a sample mass nor landing coordinates; the 1935.3 g figure comes from the separate CNSA handover announcement [8] and the coordinates from the vision sampling paper.

Mobility

Not mobile. The mechanical arm has four joints with the sampler attached at joint 4, and its reachable region on the terrain is approximately a sector of radius 2.8 m and 120 degrees [1]. Scooping allows more flexible site selection than drilling because the sampler can reach any point within that sector, whereas the drill samples only the column beneath it, 1.68 m of travel for about 1 m of core on Chang’e-5 [6].

The sampler can adjust pitch but not roll [1]. A slope component in the roll direction under the shovel therefore cannot be corrected by the arm and directly degrades control of scooping depth, so it is avoided by choosing where to sample, which is what the roll term in the site selection loss function below does.

Power and energy

No power figures are published for the sampling system. The lander operated for about two days inside a single lunar day [7] and was never required to survive a lunar night. The Chang’e-5 drilling chain, on which the Chang’e-6 drill is based, has a maximum single-machine power above 800 W [6].

Thermal

No thermal design is published for the Chang’e-6 sampling system. The Chang’e-5 drill, on which it is based, uses the helical blades of its rotating outer pipe both to discharge cuttings and to dissipate drilling heat [2], there being no convective path out of a borehole in vacuum. Lunar surface temperature at 45 degrees latitude runs from a mean 350 K at local noon to a mean 89 K before sunrise [9].

Compute and avionics

The sampling system comprises the four-degree-of-freedom mechanical arm, the sampler, and a binocular camera [1]. The sampler itself consists of a touching disc, a shovel and a close-up camera, with a force sensor at the bottom of the touching disc indicating contact with the surface. The close-up camera has a narrow field of view and is used to monitor sampling and inspect the collected soil.

The binocular camera is the primary sensor for the autonomy. Its images are 2352 by 1728 pixels, the baseline between the two cameras is 200 mm, focal length is 15.4 mm, and the working imaging distance is 2 to 5 m [1]. That configuration was chosen against the arm’s 2.8 m by 120 degree reach: it covers the whole sampling sector while achieving the expected 3D reconstruction error of 5 mm at 2.5 m.

The processing chain is CREStereo for stereo matching and depth estimation, the Point Cloud Library for point cloud processing, YOLOv5 to extract the local region of the sampler and so reduce reconstruction cost, SIFT and FPFH as feature descriptors, TEASER for global point cloud registration and ICP for fine registration [1]. The source does not state where each stage executed, so the division between onboard and ground computation is not established here. Autonomy that must run inside a single communication cycle has to be resident on the spacecraft; where changes can be predicted and modeled, ground operators prescribe the actions instead and no autonomy is needed [4].

Autonomy

Every step of the sampling implementation is autonomous, with the operator only monitoring and making a simple decision if required [1]. Space autonomy is traded against non-autonomous approaches on risk and cost, and mission objectives are commonly scaled back toward non-autonomous practice where possible; autonomy becomes a requirement only where the response time is shorter than the next communication cycle [4].

Sample point selection

Terrain point clouds reconstructed from the binocular pairs are processed automatically to produce a slope map and to segment obstacles and pits, classifying points as collectable area, obstacle or unreachable, and pit [1]. Obstacles below a size threshold are classified instead as collectible rocks.

Candidate points are scored by a multi-objective loss function over four factors, weighted equally for Chang’e-6 [1]. The landing zone within which this operates was itself selected against an 8 degree slope limit [3]. The factors are local slope, proximity to obstacles and pits, distance from the limits of the arm’s reachable region, and specifically the slope component in the roll direction, which is penalized because the arm cannot compensate it. A minimum point-spacing constraint prevents selected points from crowding, and simulated annealing is used to select the set with the lowest loss.

For Chang’e-6 this produced ten sample points over a region whose average slope was 5.1 degrees, with a neighborhood radius parameter of 400 mm [1]. Points 1 to 3 and their neighboring areas were scooped. The wider landing zone has an average slope of about 5.74 degrees, with 76 percent of its area below the 8 degree maximum for safe landing, and contains 26,785 identified craters larger than 50 m, more than 96 percent of them between 100 m and 1 km across [3]. The three flat regions favored by engineering constraints have average slopes of 2.36 to 2.37 degrees at elevations of -5,197 and -5,277 m, with craters occupying 8.61 and 10.36 percent of their areas.

Pose measurement and fine-tuning

The arm cannot reach a commanded pose accurately. Flexible deformation and secondary effects produce end-localization errors of several centimeters, measured on Earth as about 30 mm and 2 degrees of absolute error even with a balloon used to compensate the gravity difference [1]. These errors are difficult to model and directly affect scooping depth.

Chang’e-6 measures the sampler’s actual pose visually instead. The sampler point cloud is reconstructed from a binocular pair, its region of interest located with YOLOv5, features matched against a template point cloud, and the transformation matrix estimated by TEASER followed by ICP, from which the sampler’s x, y, z and pitch angle follow [1]. Because occlusion and specular highlights make the reconstructed region much smaller than the template, the method replaces the template with the current reconstruction once a correct registration is confirmed, so subsequent registrations are made against like data.

Measured performance: precision of 1.1 mm and 0.4 degrees, average relative distance error of 3.4 mm over movements of 300 to 450 mm and 2.0 mm over the 100 to 200 mm fine-tuning range, average measurement time 1.6 s, and correct registration with less than 30 percent of the valid region visible [1].

Two systematic errors are then compensated explicitly [1]:

• Soil sinking. When the touching disc contacts the surface it squeezes a small pit, so the sampler sits lower than the intended touching pose. The offset is recovered as the difference between the calculated and the vision-measured touching height, and ranges from 5 to 20 mm depending on soil density in terrestrial experiments [1].
• Support force. The force sensor has an activation threshold, so a support force is already acting on the touching disc when motion stops, altering the arm’s deformation. This cannot be seen in the touching pose alone, so it is recovered by comparing the displacement the arm reports from its four joint angles against the displacement the vision system measures between a position above the sample point and the touching pose.

The final target height is the touching height plus a lifting height set by the target scooping depth and the sampler geometry, plus both offsets [1]. An alternative direct formulation is used for points with no available imagery or outside the binocular camera’s field of view.

Fine-tuning achieved an average scooping depth error of 3.8 mm in terrestrial experiments against a 22.0 mm target depth [1]. In-situ lunar soil relative density is about 65 percent in the top 15 cm [10], which is the range the touching disc squeezes into. Where terrain was rough and points widely spaced the per-point offsets differed from their mean by 2 to 5 mm; on flat terrain with closely spaced points the difference was under 3 mm.

The failure modes are asymmetric: too high and the shovel does not touch the soil and collects nothing; too low and it takes excessive soil and can become stuck, which is a hazard [1].

Communications

Far-side operation means no direct-to-Earth path. All Chang’e-6 surface communications were relayed by the Queqiao-2 satellite, launched on a Long March 8 from Wenchang on 20 March 2024 specifically to support Chang’e-6, Chang’e-7 and Chang’e-8 [7].

A relay adds latency and restricts communication to windows when the satellite is in view of both lander and Earth. Manual per-point analysis and manual pose fine-tuning both require several image-command round trips per sample point, and the long communication delay is identified as what makes manual operation time-consuming and drove the fine-tuning design [1]. Chang’e-4 had already demonstrated autonomous terrain-relative navigation with hazard assessment and avoidance for far-side landing without radiometric data [4]. Link parameters are not published.

Payload and instruments

The sampling hardware is the drill and the arm-mounted scooping sampler, the same two methods used on Chang’e-5: soil core drilling and soil and rock fragment scooping [1], [7]. Compared with drilling, scooping allows more flexible site selection because the sampler can reach any point within the arm’s sector.

The sampler assembly carries the touching disc with its force sensor, the front-end shovel that collects and transfers material to the container, and the close-up camera [1]. The Chang’e-5 arm from which it derives integrates shovelling, digging, suction, receiving and grasping, with repeated positioning accuracy better than 1 mm [6].

Instrument-level parameters for the Chang’e-6 drill are not published in the sources consulted here; the mechanism is stated to be the same approach as Chang’e-5’s dual-pipe soft-bag corer [1], [2].

A Chinese national flag carried by the lander was unfurled after sampling, the first flag deployed on the lunar far side [7].

Modes of operation

The sequence is fixed rather than mode-driven, and closely follows Chang’e-5’s: unlock the drill and scoop after landing, drill and core and transfer the core to its primary package device, scoop surface samples and transfer them to their own package device, then move the package to the sealing capsule on the ascender and seal it with drilled and surface samples kept separate [2], [7].

Within scooping, each sample point runs the same cycle [1]: determine the target sampling pose and the corresponding touching-moon pose from the terrain analysis; move the arm to an approximate position, deliberately holding a large height margin during the midway movement for safety; touch down and measure; compute the sink and support-force offsets; fine-tune; and scoop.

Ground operations

Operated by CNSA, with the sampling system developed and tested by the China Academy of Space Technology, whose testing ground supplied the terrestrial evaluation data alongside the lunar images captured during the mission [1], [7].

On Chang’e-5 an operator confirmed each sample point from imagery and a specialized operator performed pose fine-tuning, a procedure requiring complex training and carrying a high time cost [1]; that campaign ran 15.82 h for 12 surface sampling operations [6]. On Chang’e-6 the operator monitors and intervenes only when required. The reported result is a reduced time cost at maintained sampling quality, within a roughly 48 hour surface window [7].

Post-flight ground operations follow the Chang’e-5 pattern: the returner landed in Siziwang Banner and the sample container and certificate were handed over by CNSA to the Chinese Academy of Sciences on 28 June 2024, at which point the mass was announced as 1935.3 g [7], [8].

Technologies developed

Autonomous sample point selection from reconstructed terrain point clouds, with a loss function that encodes the manipulator’s own kinematic limitations, including the missing roll degree of freedom, as a terrain criterion [1].

Template-based 6D pose estimation of a manipulator’s end effector by point cloud registration, achieving 1.1 mm and 0.4 degrees with under 30 percent of the tool visible and no fiducial markers [1]. Insensitivity to occlusion, imaging distance and lighting direction reduces the complexity required of the sampler mechanism itself.

Explicit error decomposition for a scooping manipulator, separating arm flexible deformation, soil sinking under the contact disc, and force-sensor threshold support force, and measuring each with a different combination of arm kinematics and vision [1]. Soil sinking is 5 to 20 mm depending on density.

The method is extended by its authors to drilling and grabbing samplers, placing a drill away from obstacles and measuring pose for the vertical alignment drilling requires [1].

The returned material is farside mare basalt with an eruption age of about 2.8 Ga [5], which is the sample the site selection and the sampling autonomy were built to obtain [3].

References