A robotic system designed to prepare lunar soil for future oxygen-production plants has completed an integrated analogue trial at the European Space Agency’s LUNA facility, revealing that moving excavated material between machines could be one of the most important engineering problems to solve.
The Mobile Lunar Excavation and Size Separation System, known as MoLES3, combines a mobile excavation rover with a separate stationary unit that sorts regolith by particle size. During a 2.5-hour continuous test campaign, the system completed 37 excavation-and-deposit cycles while researchers tracked material losses, energy consumption, robotic-arm loading and separator performance.
The study does not demonstrate oxygen extraction itself. Instead, it examines the upstream process required to deliver appropriately prepared regolith to future oxygen-production equipment. That distinction is important because lunar resource plants will depend not only on chemical extraction technologies, but also on reliable excavation, transportation and feedstock conditioning.
Preparing Lunar Regolith for Oxygen Production
Lunar regolith contains substantial oxygen chemically bound within silicate and oxide minerals. Technologies including hydrogen reduction, molten-salt electrolysis and carbothermal reduction are being investigated as potential ways to release that oxygen for life support and rocket propellant.
Some of these processing routes can benefit from regolith with particular particle-size distributions or other material characteristics. This creates a logistical chain in which soil must first be excavated, transported and conditioned before it reaches the extraction reactor.
MoLES3 was developed to examine that chain as an integrated engineering system rather than treating excavation and particle separation as unrelated laboratory processes.
A Rover and a Separate Processing Unit
The architecture tested by the researchers physically separates the excavation system from the beneficiation equipment. A mobile rover carries a four-joint robotic arm and excavation scoop, while a stationary vibratory classifier sorts the collected material.
After digging, the rover carries its payload in a hopper to the stationary unit. The hopper engages a compliant rubber-membrane interface and releases the regolith into the classifier. The membrane was designed to tolerate small rover-positioning errors without requiring a mechanically complex precision docking system.
The stationary classifier uses vertically stacked screens to divide material into three particle-size groups: greater than 500 micrometres, between 100 and 500 micrometres, and less than 100 micrometres.
The smallest fraction was treated as the target material for the study because fine regolith can be relevant to hypothetical downstream oxygen-production processes.
37 Excavation Cycles Produced 2.1 Kilograms of Target Material
Across the 2.5-hour campaign, MoLES3 excavated 10.8 kilograms of regolith simulant. Approximately 10.0 kilograms reached the beneficiation system, and 2.1 kilograms emerged as the target fraction below 100 micrometres.
This corresponds to an excavation rate of 4.32 kilograms per hour and a target-fraction production rate of 0.84 kilograms per hour.
The researchers calculated a transfer efficiency of 93%, meaning about 7% of the excavated material failed to reach the classifier inlet. The loss was attributed primarily to residual material remaining inside the hopper during deposition.
The beneficiation yield ratio was 21%, while overall feedstock-preparation efficiency was 19% when transfer and separation were considered together.
The Separator Was Not the Main Source of Material Loss
At first glance, recovering only around one-fifth of the delivered material as sub-100-micrometre particles could appear inefficient. However, the EAC-1 lunar regolith simulant used in the integrated trial had previously been characterised as containing only about 18–24% of its mass in that fine-particle range.
The measured 21% beneficiation yield therefore closely matched the amount of target-size material physically available in the simulant. According to the researchers, this indicates that the screening system recovered most of the fine fraction that could realistically be recovered under the tested conditions.
The more significant recoverable loss occurred before separation, at the interface where the rover transferred regolith from its hopper to the stationary processor.
For this particular architecture, the result suggests that improving hopper geometry and material discharge could produce larger gains than attempting to extract substantially more performance from the classifier itself.
Beneficiation Used Only 11% of System Energy
The entire MoLES3 system consumed 265 watt-hours during the campaign. Only 30 watt-hours were attributed to beneficiation, representing about 11% of total system energy consumption.
The remaining energy demand was dominated by excavation and rover mobility. The measured excavation energy intensity was 24.5 watt-hours per kilogram of excavated material, while producing the target fine fraction required 126.2 watt-hours per kilogram when the entire system energy budget was considered.
These measurements provide an Earth-gravity experimental baseline that future systems can use when comparing different combinations of mobile excavation, material transfer and stationary processing hardware.
Separator Power Had a Threshold-Like Effect
Before the integrated LUNA trial, the researchers tested the vibratory classifier at two motor-power levels using the same EAC-1 simulant.
At 3.96 watts, the intermediate 100–500-micrometre material did not completely clear the screens during the available operating period. Increasing excitation to a resonant setting of 7.70 watts allowed that fraction to clear while also increasing recovery of the finest particles.
The result suggests that vibratory separation cannot always be treated simply as a process that becomes proportionally faster as power increases. Below a practical excitation threshold, material may remain trapped on the screens within the required operational window.
For future lunar processing equipment, this means separator motors may need to be selected according to whether they can complete separation for a specified material load and cycle time, rather than being sized only around average throughput.
Excavation Trajectory Mattered More Than Expected
The autonomous excavation sequence consisted of eight phases, including arm deployment, approach, alignment, regolith engagement, scooping, payload containment, delivery and deposition.
Analysis of robotic-arm loading showed that the highest actuator demand occurred as the scoop entered and moved through the regolith. However, the researchers found that elevated loading was associated more strongly with poor entry alignment and excavation trajectory than with differences between the simulants tested.
Misaligned scoop entry could increase torque without delivering a corresponding increase in collected material. It could also propagate positioning errors through later stages of the excavation cycle and reduce payload volume.
The findings indicate that trajectory planning may need to be treated as part of the mechanical design problem for long-duration lunar excavators, rather than merely as a software optimisation performed after the robotic arm has been designed.
Autonomous Digging Completed Every Trial Cycle
MoLES3 used an RGB-D camera and onboard processing to identify suitable excavation locations. Its perception pipeline acquired a three-dimensional point cloud, filtered and denoised the data, constructed a digital elevation model and evaluated candidate surface cells against slope and reach constraints.
A valid excavation coordinate was produced autonomously during all 37 cycles, with no manual override required for execution of the robotic arm.
A single operator remained responsible for rover navigation, monitoring and potential anomaly response. The passive transfer membrane also reduced the need for exact alignment during deposition, allowing the rover to approach the beneficiation unit without a complex precision-docking manoeuvre.
The test therefore demonstrated a supervised-autonomy model in which routine excavation could be automated while a human operator retained higher-level control.
Important Limits of the Trial
The researchers emphasise that the results should not be interpreted as direct predictions of lunar performance. The campaign was conducted under Earth gravity at ESA’s LUNA analogue facility using lunar regolith simulant.
The Moon’s roughly one-sixth Earth gravity could change excavation mechanics, regolith transfer, particle settling, screen residence times and vibratory separation behaviour. Motor settings and screen geometry that worked during the terrestrial campaign would therefore require renewed testing under reduced-gravity conditions.
The rover also operated only on the level Zone 1 terrain at LUNA. A second zone containing slopes of approximately 30–35 degrees was excluded after a stability assessment because it exceeded the safe tilt range of the rover chassis.
In addition, the RGB-D sensor and commercial Intel NUC computer were research components selected for rapid development rather than flight-qualified lunar hardware. A mission-ready system would need sensors and processors capable of tolerating lunar dust, extreme illumination, thermal-vacuum conditions, radiation and tight mass and power constraints.
What the Results Mean for Future Lunar Resource Plants
The study identifies several system-level engineering priorities for separated lunar feedstock architectures. A compliant transfer interface can reduce precision-docking requirements, excavation trajectories can influence actuator loads, separator power must be sufficient to complete processing within the available cycle time, and efficiency should be measured across the entire material-handling chain rather than at individual subsystems alone.
The researchers also point to a substantial scaling challenge. Previous modelling cited in the study suggests that a hydrogen-reduction plant producing roughly one tonne of oxygen per year could require approximately 10–15 kilograms per hour of sub-100-micrometre feedstock, depending on assumptions about reactor efficiency and operating duty cycle.
MoLES3 produced 0.84 kilograms per hour of that target fraction during the analogue campaign. The system was not designed as a production-scale oxygen plant, but the measured value gives engineers an experimentally grounded reference from which concepts such as multiple coordinated excavators, improved trajectories and redesigned hopper interfaces can be evaluated.
Next Steps Toward Integrated Lunar Oxygen Production
The researchers identify reduced-gravity testing as a major next step, particularly for the material-transfer interface and vibratory classifier. They also propose longer-duration testing to examine screen fouling and wear, refinement of the hopper outlet to reduce transfer losses, operation on steeper terrain and eventual integration with an actual oxygen-extraction reactor.
Such integration would allow researchers to measure efficiency from the first excavation of lunar material all the way through to oxygen production rather than evaluating only the upstream preparation stages.
MoLES3 therefore represents more than a rover excavation demonstration. The experiment provides a system-level baseline showing how seemingly small engineering details at the boundaries between machines can influence the performance of a future lunar resource plant. As lunar surface infrastructure develops, efficiently collecting, transferring and preparing regolith may prove just as important as the chemistry used to extract oxygen from it.