NASA KSC Surface Autonomy Test Site

NASA/Frank Micheaux. Public domain (NASA / US government work).

An enclosed, blacked-out analogue site at NASA KSC built by the ISRU Pilot Excavator project to test vision-based surface autonomy under lunar south pole illumination geometry. The article under test drives on prepared soil under six narrow beam stadium lights, navigates against a full-scale mock lander carrying fiducials, and is tracked by a 36 camera motion capture system whose pose solution is published as independent ground truth [1]. The site is listed by the KSC partnerships office as the Surface Autonomy Test Site [2].

Laboratory data

Parameter	Value
Operator	NASA KSC, ISRU Pilot Excavator project, STMD Game Changing Development
Location	NASA KSC, Florida, United States, 183 m east of the IPEx flight control building
Commissioned	Not published. First documented campaign August 2024
Type	Enclosed analogue surface site, blacked out, artificially lit
Floor area	39.6 x 21.3 m structure, 843 m2; practice area 33.5 x 21.3 m, 714 m2
Capabilities, area	Practice area, soil bed and rock field
Capabilities, fixtures	Lighting, mock lander
Simulant or terrain	Local Florida soil, tilled; albedo 0.26; baseline penetrometer resistance 0.5 kg/cm2
Instrumentation	36 OptiTrack Slim x13 motion capture cameras; 4 Axis situational cameras; Geotester penetrometer
Ground truth	OptiTrack beacon pose on a ROS topic, logged against vehicle telemetry
Fidelity limits	1 g, ambient atmosphere, terrestrial soil, no vacuum
Access	KSC partnerships office, Testing and Laboratories; no lead time or fee schedule published
Cited by	IPEx

Source: [1][2][4].

The site was built and is operated by the ISRU Pilot Excavator project under Space Technology Mission Directorate Game Changing Development [4]. The KSC partnerships office catalogs it as the Surface Autonomy Test Site, alongside the Granular Mechanics and Regolith Operations laboratory and its Artemis Arena and ASSIST chamber; the operators’ own engineering name for it is the IPEx Autonomy Test-Site [2]. Temperature and humidity sensors are distributed through the site and reported over the site network, but nothing sets them. A Ubiquiti AirFiber radio link connects the site to the control building, with a local server ingesting situational imagery and running the OptiTrack and network video recorder software so that only tracker pose and compressed imagery cross the link; article Wi-Fi is routed through an access point on the mock lander, and all network equipment is on uninterruptible supplies.

Capabilities

Enclosed practice area

Parameter	Value
Working volume	Structure 39.6 x 21.3 m, 1.8 m sidewall, 6.1 m ridge; practice area 33.5 x 21.3 m
Test article limits	Largest article run: the 30 kg-class IPEx alpha configuration
Vacuum	Not applicable. Ambient Florida pressure and humidity
Temperature	Not controlled. Measured and reported, not set
Illumination	See the lighting; blackout panels block 95 percent of daylight
Simulant or terrain	Local Florida soil, tilled level; see the bed
Slope	Not applicable. The bed is prepared level
Gravity offload	Not applicable. 1 g
Instrumentation	36 OptiTrack Slim x13 cameras on the perimeter; 4 Axis situational cameras

Source: [1][2][4].

The enclosure is a fabric-covered galvanized tube structure. With blackout end panels installed and the two staging foyers curtained off, the usable rover practice area is 33.5 x 21.3 m [1][2]. Slope work on IPEx wheels is run separately in the GMRO Regolith Test Bed [5].

Site selection ran as two trade studies. The first compared KSC, Cape Canaveral Space Force Station and off-center land, and selected KSC on logistics. The second compared available real estate at KSC against existing buildings and settled on open ground 183 m east of the IPEx offices, which sets the ground control separation used during operations [1].

The structure requirement was an interior dark enough to stand in for the lunar sky, so that artificial lighting is the only source, combined with a Florida climate envelope covering humidity, rainfall, moisture accumulation and lightning protection, a hurricane rating, and Authority Having Jurisdiction fire and exit compliance. Lean-to structures, brick and mortar buildings, portable structures and hybrids such as CONEX box walls under a rolled steel arch roof were considered before a fabric-covered galvanized tube tent was procured.

Blackout is achieved with tarps suspended from the tubular trusses that run the width of the structure, and with split curtains at each end that separate the practice area from the staging foyers and seal in the center with black Velcro. More opaque panel material was available but not bought.

Blackout parameter	Value
Truss spacing	Every 3 m across the structure width
Panel material	95 percent light-blocking polyester weave
Cost of more opaque material	Above 40 percent price premium
Panel drape	About 25 cm down from the ridge at the center of each panel

Lunar illumination system

Parameter	Value
Illumination	6 RuggedGrade StadiumPro IV fixtures, 1400 W each, 5000 K, 200,200 lm, 20 degree beam
Working volume	Fixtures at one end of the 33.5 x 21.3 m practice area

Measured illuminance falls steeply across the practice area. The narrowest available 20 degree beam was chosen to sharpen shadow edges [1]. Varying solar elevation and azimuth is listed as future work.

Measured illuminance	Horizontal	Vertical
30 cm from the lights	78,000 lux	192,000 lux
Sunward face of the largest rock	9,500 lux	n/a
Foot of the lander	360 lux	2,345 lux
Far end of the structure	94 lux	850 lux

The lighting configuration was selected by simulation before procurement. A model built in the Unity engine with the high definition rendering pipeline was used because HDRP accepts physical photometric units and reports resulting illuminance, so candidate layouts could be scored for uniformity and shadow sharpness against a simulated lunar sunlight baseline. The six fixture StadiumPro IV layout was adopted as the compromise.

Lighting alternative considered	Outcome
Single Clar Illumi Max 300 LED, 132,000 lux at 0.5 m, 5600 K	Adequate for short range work, did not cover the area
48 fixture layout	Improved uniformity, destroyed shadow quality
ARRI Daylight 18 kW metal halide	High intensity from few fixtures, costly and hard to integrate

Public domain (NASA / US government work).

Prepared soil bed and rock field

Parameter	Value
Simulant or terrain	Local Florida soil, tilled loose; albedo 0.26; baseline resistance 0.5 kg/cm2
Working volume	The 33.5 x 21.3 m practice area floor
Temperature	Not controlled
Instrumentation	Geotester pocket penetrometer, run as a survey after a campaign

The terrain is native Florida soil rather than a simulant. Preparation removes rocks above 7.6 cm and all vegetation and leaves a loose tilled surface resembling plowed earth, and penetrometer surveys run after the TRL 5 demonstration found that trafficked ground compacts to roughly twice the resistance of undisturbed ground, that the dig site compacted by as much as a factor of 11 from the 0.5 kg/cm2 baseline, and that the dump site loosened by a factor of 2.5 [1]. The operators read that across to the flight case: dig sites at the lunar south pole should be expected to compact progressively over a mission while dump sites become progressively looser.

Public domain (NASA / US government work).

The rock field is generated rather than arranged by eye: rocks are binned into discrete size ranges and placed at random positions at an areal density taken from a power law fit to Surveyor VI block counts, using the number of rocks per square meter between size bounds as N = a(d_min^b - d_max^b)/100 [1]. Cintala and McBride fitted power laws of the form N = a d^b to fragment size distributions measured from Surveyor surface photography [3].

Surveyor VI power law fit	Value
Rock size range placed	4 to 20 cm
a	0.154, errors +0.088/-0.019
b	-2.286, standard error 0.094
Correlation coefficient	0.983

Source: [3].

Mock lander

Parameter	Value
Working volume	2.75 m diameter, 1.2 m tall, standing in the practice area
Test article limits	Docking approach only; no published mass limit
Illumination	White top deck, gold metallic sides, silver metallic legs, lit by the stadium fixtures
Instrumentation	3 FLIR Blackfly S BFS-GE-88S6-M cameras, Kowa LM6JC 6 mm lenses; 16 AprilTag 36h11 fiducials

The lander is the navigation target and the site’s service node, 2.75 m in diameter and 1.2 m tall, finished with a white top deck, gold metallic sides and silver metallic legs, and carrying 16 AprilTag 36h11 fiducials, 15.24 x 15.24 cm overall with a 12.1 x 12.1 cm black-to-black edge, mounted along the top perimeter and at the mid-span of the legs and angle braces, and painted onto polycarbonate sheet, flat white vinyl for the light field and Musou black acrylic, quoted at up to 99.4 percent absorption, for the dark field, which holds tag contrast while suppressing specular return from the solar simulator [1].

Public domain (NASA / US government work).

As the service node it carries a network switch onto the flight control room network, the article’s Wi-Fi access point, the temperature and humidity sensors, three fixed cameras, and a wireless charger transmit coil suspended on cables and offset from the front face so that an off-nominal docking approach cannot damage the coil or the article.

Instrumentation

Pose ground truth comes from 36 OptiTrack Slim x13 cameras around the perimeter of the practice area, tracking a beacon on the article [1]. An independent characterization of the same class of hardware, a 16-camera Primex41 array at NASA Glenn’s lunar soil bins, checked the system’s self-reported mean wand error of 0.17 mm against a linear stage and found it honest as an aggregate figure, but found error in specific regions of the working volume running higher where test hardware occludes cameras [10]; the site’s own published pose accuracy is not broken out by region. The pose is published on a ROS topic, so it is logged alongside the article’s own telemetry in the same time series database and is available to the out-of-simulation operators live.

Situational coverage is four Axis cameras: two Q6128-E pan-tilt-zoom units on the ceiling in the front corners, a third Q6128-E on a movable ground tripod, and an M3058-PLVE fixed fisheye at the ceiling center giving a whole-field overview. Access to these views is restricted to the simulation coordinator, so that the in-simulation operators see only what a flight downlink would carry.

The lander’s three FLIR Blackfly S BFS-GE-88S6-M cameras are fixed focus and fixed focal length with Kowa LM6JC 6 mm lenses: two side-facing units look at the dig and dump areas and one front-facing unit is angled down for docking verification, and during the TRL 5 run only the docking camera was used, on the assumption that the side cameras might not be available for a full flight mission [1].

Soil state is measured with a Geotester pocket penetrometer, run as a survey after a campaign rather than continuously, comparing undisturbed ground, trafficked ground, the dig site and the dump site.

What it does not reproduce

Gravity and granular behavior. The site runs at Earth gravity (see Laboratory data above). The operators state directly that replicating the granular properties of the lunar south pole remains a challenge under that gravity. Bearing capacity, sinkage and excavation reaction forces measured here are not transferable to lunar conditions without a separate model.

Regolith. The bed is local Florida soil. Lunar simulants such as LHS-1 were considered and rejected because of the logistical, financial and safety burden of handling simulant in a facility of this size. The stated compensation is that Florida soil still provides a challenging traverse and excavation medium for autonomy work, not that it reproduces regolith. Simulant-based mobility and excavation testing for the same program is run separately in the KSC Regolith Test Bed with BP-1 [5].

Optical properties. Measured site albedo (see Laboratory data above) runs well above published lunar surface albedo. The site is brighter and less absorbing than the terrain it stands in for, which biases exposure and contrast in the direction of an easier problem.

Illumination uniformity and collimation. Six point sources at finite distance do not approximate a collimated beam from an effectively infinite source; measured illuminance falls by two orders of magnitude across the practice area (see the illuminance table above). The operators describe this non-uniformity as a challenge for consistent imaging but reuse it deliberately, as a range of conditions for testing adaptive exposure as the article crosses from the bright to the dim end. Solar elevation and azimuth are fixed by the physical fixture placement; varying them is listed as future work.

Blackout. The panel weave blocks most but not all light, and the panels drape short of the floor, leaving the enclosure short of a true lunar sky (see the blackout table above).

Craters. No craters are present. The paper states that their inclusion is an ongoing effort and lists more diverse terrain features, craters of various sizes among them, as future work.

Rock statistics. The Surveyor VI fit was derived from fragments typically below 1 m seen in surface photography [3]. Cintala and McBride found that extrapolating those fits to blocks above 2.5 m overestimates the areal density measured from orbital photography at every site but Surveyor III, and that the cumulative block population is therefore probably nonlinear with different slopes over different size intervals [3]. The same work found the areal density of blocks within 1.1 crater radii to be about a factor of ten higher than the intercrater density at all three mare sites, a clustering the site’s uniform random placement does not reproduce. The paper also notes that the Surveyor VII highlands site near Tycho carried a far higher large-block density than the mare sites, and that a polar site may differ again from the Surveyor VI model adopted [1][3].

Dust. Electrostatic lofting, adhesion to lenses and radiators, and abrasion are named as lunar south pole problems in the same paper but are outside what the site addresses; the stated focus is the visual effects seen by camera-based systems. Thermal vacuum and dust exposure for the same program’s cameras were run in the ASSIST chamber at the GMRO lab instead [4], a dirty thermal vacuum chamber run from 760 torr down to 3.5e-6 torr with a xenon arc solar simulator and a regolith-bed build plate for combined excavation and additive construction testing under simulated lunar or Martian pressure [9]. Airborne dust itself is instrumented separately again, in the GMRO lab’s regolith bins rather than at either the autonomy site or ASSIST: a laser and camera extinction sensor there found that spot measurements of dust density collapse toward zero above roughly 1e10 particles per cubic meter, so side-scatter measurement is needed once a cloud gets that dense, and that dust raised by a nine-laser test array took about an hour to clear from a test bed [11].

Campaigns run there

IPEx TRL 5 mission demonstration, August 2024. Five days of continuous operation beginning 25 August 2024, run as a mock mission against the IPEx concept of operations, which requires roughly 5.7 million actuator input revolutions and drives the actuator qualification program run in parallel at the GMRO lab [4]. The article was the alpha configuration of the TRL 5 hardware with larger external battery packs and surrogate avionics based on the final flight computer; all localization and autonomy software ran onboard. The site was configured with a statistically generated rock field unknown to the operators, low solar angle stadium lighting, reduced data transmission and a simulated 14 second round-trip delay standing in for a Commercial Lunar Payload Services lander relay, with ground control sitting in the IPEx flight control room 183 m from the site [1], with four telemetry displays on the flight control computer for in-simulation work and a separate console for the simulation coordinator carrying real-time, undelayed telemetry and imagery. Four shifts of four crew each ran the mission: primary operator, secondary operator, telemetry desk officer, and simulation coordinator, the last holding the only access to the situational cameras and performing out-of-simulation actions such as battery changes.

Measured result: continuous robot operation that, after subtracting thermal resets and wireless charging events, corresponded to a mock mission of several days of simulated time. Each repetition ran three drives, then excavation and unloading, then three more drives. The demonstration closed with operational margin in hand, and no equipment in the site lost power over the period.

Campaign result	Value
Continuous robot operation	77 h 46 min
Simulated mission time (after subtracting resets and charging)	205 h 18 min, about 8 days
Distance traversed	58.1 km across 334 mission repetitions
Maximum sustained drive speed	40 cm/s
Docking and recharge cycles	35, averaging 10 repetitions per battery charge
Drive length per leg	About 100 m, averaging nine minutes
Operational margin at close	59 h

Two datasets were released from the campaign. The instance segmentation set carries manually annotated instances across four classes: lander, rock, fiducial and fiducial inverse, with both bounding boxes and pixel-wise masks. Annotation was done in CVAT with Segment Anything as a serverless assist, with manual correction needed on heavily shadowed images. The imagery was collected not by IPEx but by a Clearpath Robotics Husky surrogate carrying prototype IPEx FLIR Blackfly S cameras with 6 mm Kowa lenses, driven in arcing and figure-of-eight patterns near the mock lander [1].

Instance segmentation dataset	Value
Images	2,343
Annotated instances	10,505
Mean mask area, lander	330,010 px
Mean mask area, rock	35,358 px
Mean mask area, fiducial	2,880 px
Mean mask area, inverse fiducial	2,361 px

The photogrammetry set was taken during the mapping phase of the mock mission from four cameras on synchronized intervals of linear driving. Stereo pairs were set to hyperfocal distance; the side cameras were manually focused to hold the lander at a fixed driving radius and ran a custom auto-exposure routine that exposed for the lander fiducials during arcing drives. The set has been processed into ground-based maps with COLMAP structure from motion and with neural radiance fields.

Photogrammetry dataset	Value
Images	1,514
Camera interval	Every 0.5 m of linear driving
Front and rear stereo pairs	IDS GV-51F1SE, Sony IMX547 sensor, binned to 1230 x 1028
Side cameras	Same sensor, binned to 2460 x 2056, monochrome
Side camera driving radius	7 m

Autonomous excavation algorithm development, 2022 onward. The autodig control work that the site’s excavation runs exercise was developed on the RASSOR 2 prototype in the GMRO lab regolith bin, not in the autonomy test site, because that bin holds BP-1 simulant [4]. RASSOR 2 is the counter-rotating bucket drum excavator IPEx descends from: two opposed drum pairs cancel the horizontal reaction force so the 66 kg vehicle can excavate without needing traction, at a design mass half that of its predecessor for double the payload capacity [7]. The force and torque a bucket drum of IPEx’s scale generates while cutting, which sizes the excavation rates the autonomy test site’s dig-and-dump cycle exercises, were measured separately on an instrumented gantry rather than at this site [8]. The autonomy test site is where the resulting stack is run end to end against a rock field and a lander, exercising the supervised autonomous navigation mode that the flight design assumes [4].

Wheel configuration testing, 2023 to 2024. Ten wheel configurations, drawn from 24 possible combinations of cleat shape and perforation, cleat spacing, and grouser shape, height and spacing, were driven on RASSOR 2.0 fitted with 30.5 cm IPEx-sized reconfigurable wheels as an IPEx surrogate in the KSC Regolith Test Bed, an 8 x 8 x 1.1 m bed holding 120 tons of BP-1 [5]. Tests covered circle driving, straight driving, a 20 degree prepared slope and drawbar pull, 83 tests in total, with slip computed from OptiTrack motion capture and power taken as summed motor current times pack voltage. Taller and squarer grousers slip less but draw more power, shorter and rounder grousers do the reverse, and solid versus perforated cleats made no discernible difference to either, so the lighter perforated cleats cost nothing; no single configuration won across all four test types, and the recommended baseline is a curved perforated cleat with short curved grousers [5]. This is the simulant-based counterpart to the autonomy site’s Florida soil traverses.

Navigation primitive development, 2022. The circular traverse pattern around the lander that the site’s layout is built to support was developed against dynamic movement primitives with an added obstacle avoidance term, modified to bound each obstacle’s influence by proximity rather than letting the sum of all obstacles push the trajectory outward, and evaluated first in the Simulated Excavation Environment for Lunar Operations, a Unity-based simulator with terrain deformation and modeled battery depletion, using a RASSOR 2.0 model in densely populated rock fields [6]. Across 10 rock distributions of 70 non-traversable rocks each at 0.15 rocks per square meter in a 22 x 22 m worksite, and a further 100 novel distributions, the modified method reached the goal with zero collisions in every run, against 426 collisions for the unmodified method and 425 for the original avoidance formulation, at a path-length cost of 44.11 plus or minus 0.65 m against a 43.45 m nominal circle [6]. The physical rock field in the test site is the corresponding hardware case.