Semi-autonomous exploration of martian and lunar analogues with a legged robot using a Raman-equipped robotic arm and microscopic imager

1 Introduction

Future robotic surface exploration missions on Mars and the Moon require robust capabilities for in situ mineralogical and lithological characterisation, given their relevance for both In-Situ Resource Utilisation (ISRU) and astrobiological investigations. On the Moon, many key resources are located in terrain that is difficult to access, including volatile- and titanium-rich pyroclastic deposits, REE-bearing KREEP basalts, and water ice within permanently shadowed regions near the South Pole (Gaddis et al., 2003; Wieczorek and Phillips, 2000; Li et al., 2018). These environments are often associated with steep crater walls and rugged surfaces that challenge conventional wheeled robotic platforms.

Similarly, on Mars many of the most scientifically promising sites for biosignature exploration and ISRU prospecting occur in steep or otherwise hazardous terrain. These include layered sedimentary deposits in crater interiors, hydrothermal systems, and clay-rich outcrops exposed along crater and canyon walls (Farmer, 2007; Stack et al., 2020; Chojnacki and Hynek, 2008; Mandon et al., 2021). Water-ice exposures and metal-rich regolith have also been identified in high-latitude and highland regions, often within unstable slopes or fractured geological settings (Dundas et al., 2018; Badescu, 2009). Accessing and characterising such environments remains a major operational challenge for surface missions.

Past surface robotic missions to the Moon began with the Soviet Union’s Luna Program, which sent the Lunokhod 1 and Lunokhod 2 rovers in the 1970s (Witze, 2013). These missions conducted surface imaging, soil mechanics studies, and radiation measurements, contributing to early lunar surface resource studies. Around the same period, the United States launched the Surveyor Program (1966–1968), a series of robotic landers that provided crucial data on the lunar surface’s composition and mechanical properties. Some of these landers were equipped with robotic scoops, which performed in-situ regolith analysis, offering early insights into lunar soil characteristics and its potential for future resource utilization (Urey and Marti, 1968). More recently, China’s Chang’e Program (2007–present) deployed several rovers: Chang’e 3 conducted mineralogical analysis of regolith (Zhao et al., 2014). Chang’e 4 made the first-ever landing on the lunar far side, investigating its surface, subsurface, and water ice distribution (Wu et al., 2019) while Chang’e 5 returned lunar rock and regolith samples to Earth (Li et al., 2022). Additionally, India’s Chandrayaan-3 (Padma, 2023) with the Pragyan Rover, successfully landed near the South Pole, conducting in-situ analysis of lunar soil, and searching for water ice, highlighting its ISRU potential (Kanu et al., 2024). Also, the United States has recently resumed robotic lunar exploration. The Intuitive Machines Nova-C lander (Odysseus) landed in February 2024, marking the first U.S. lunar landing in over five decades and aiming to demonstrate ISRU technologies (Zimmerli et al., 2025). The most recent Firefly Aerospace Blue Ghost Mission 1 (2025) landed on the Moon in March 2025 and carried NASA-sponsored experiments to study lunar dust and geophysical properties (Witze, 2025).

Mars has also experienced extensive surface robotic exploration. Some of the past NASA missions include Viking missions (Soffen and Young, 1972) - the first rover to operate on Mars (Wilcox and Nguyen, 1998) - which explored Gusev Crater and found signs of past water (Haskin et al., 2005), Opportunity - which discovered evidence of ancient water in Meridiani Planum and operated for nearly 15 Earth years (Arvidson et al., 2015) and Phoenix, a stationary lander that studied water ice (Smith et al., 2009). More recent missions include China’s Zhurong rover (2021–2023), which conducted geological and climate exploration in Utopia Planitia (Tian et al., 2021) and NASA’s Ingenuity (Mars 2020 Helicopter), initially a technology demonstrator, which assisted with the aerial exploration of scouted areas of interest for the Perseverance rover (Balaram et al., 2021). The two currently active NASA rovers, Curiosity and Perseverance, are conducting scientific investigations: Curiosity has been exploring Gale Crater, discovering evidence of ancient freshwater lakes and possible habitable environments (Rampe et al., 2020) while Perseverance is operating in Jezero Crater, collecting rock samples for future return to Earth (Simon et al., 2023).

1.1 Challenges and operational strategies for lunar, martian, and deep space surface robotic mission

Operating robotic assets on planetary surfaces is a complex, multidisciplinary activity involving engineering, science, and mission-operations teams working remotely from Earth. On the Moon, short communication delays (∼1.3 s one way) permit relatively high operational tempo and limited near-real-time supervision, although most activities still rely on pre-planned, semi-autonomous execution (Alzaabi et al., 2024; Ghosh et al., 2024).

In contrast, Martian surface operations are strongly constrained by communication delays of 3–22 min one way. Daily operations typically involve analysing data from the previous sol, planning subsequent activities, and uplinking command sequences within restricted communication windows (Crisp et al., 2003; Milkovich et al., 2022). Although modern rovers possess substantial autonomous capabilities for navigation and hazard avoidance (Verma et al., 2023; Rankin et al., 2023), scientific progress remains relatively slow, with daily traverses generally limited to tens to hundreds of metres.

For missions to more distant planetary bodies, autonomy requirements become even more pronounced. For example, the Dragonfly mission to Titan will be required to operate largely independently due to one-way light-time delays exceeding 1 h (Barnes et al., 2021; Turtle and Lorenz, 2024). Similar constraints will apply to future missions to icy moons such as Europa (Cooper et al., 2002; Pappalardo et al., 2024) or Enceladus (Choblet et al., 2022; MacKenzie et al., 2022). These scenarios highlight the importance of advancing semi-autonomous, science-driven operational frameworks for planetary surface exploration.

1.2 Related work focused on lunar and martian analogue missions

In the context of space exploration on Mars and the Moon, optimizing operational concepts and enhancing sampling strategies are critical for maximizing scientific return. To address these challenges, numerous studies have been conducted in terrestrial analogue environments, investigating single-versus multi-robot exploration (Arm et al., 2023) as well as diverse operational frameworks and science-driven sampling methodologies. Single-rover science operations have been extensively tested in lunar and Martian analogue environments, including robotic lunar sample return deployments at the Sudbury impact structure, where instruments such as GigaPan (a high-resolution imaging system for panoramic photography), LiDAR (Light Detection and Ranging), C2SM (Canadian Camera for Surface and Subsurface Mapping), GPR (Ground-Penetrating Radar), a Raman spectrometer (a laser-based spectroscopic instrument for molecular analysis), and the CASS device (Camera for Autonomous Sample Selection) were used to assess operational strategies for remote sample selection and retrieval (Moores et al., 2012). Similarly, the Moon-Mars Analog Mission Activities at Mauna Kea in 2012 focused on optimizing geological mapping strategies using ground-penetrating radar (GPR) and science cameras (Yingst et al., 2015).

Beyond single-rover approaches, heterogeneous robotic teams have been explored to enhance autonomous exploration capabilities. The German Aerospace Center (DLR) developed a heterogeneous robotic team, integrating wheeled and legged robots for sample collection in a lunar analogue setting (Cordes et al., 2011). The ARCHES project further demonstrated a multi-robot system consisting of flying and driving robots at a Moon-analogue site on Mt. Etna, Sicily, in 2021. This mission focused on scientific sampling workflows, utilizing high-resolution RGB and spectral imaging alongside Laser-Induced Breakdown Spectroscopy (LIBS) for geological sample assessments (Schuster et al., 2020). More recently, as part of the ESA-ESRIC Space Resources Challenge, a team of legged robots was deployed to explore scientific applications in extreme planetary analogue environments, with emphasis on locomotion, mapping, task-sharing, and the deployment of scientific instruments on boulders and regolith to characterize lunar resources (Arm et al., 2023).

The most recent studies have compared operational strategies and the efficiency of multi-robot versus single-rover approaches for Martian cave exploration in lava tube analogue environments (Morrell et al., 2024). These investigations analysed the trade-offs between autonomous exploration and scientist-assisted operations, assessing mission efficiency and geological sampling performance. However, the science analysis focused only on images from stereo cameras and thermal cameras (Morrell et al., 2024) and no additional science payload for mineralogical or chemical analysis was included in these analogue missions for in-depth scientific assessment.

1.3 The scope of the study

This study builds upon previous work on legged robotic systems, specifically expanding on the MICRO and Raman robotic arm system developed for the ESA-ESRIC Space Resources Challenge (Arm et al., 2023). Whereas earlier studies primarily focused on advancing legged locomotion and multi-robot coordination, the present work concentrates on the scientific return achievable from a single robotic platform equipped with a simplified but complementary payload. In particular, we evaluate the extent to which panoramic context imaging, combined with microscopic imaging (MICRO) and Raman spectroscopy mounted on a robotic arm, can support semi-autonomous, science-driven sample selection and analysis in lunar and Martian analogue environments.

To conduct these tests, we deployed the legged ANYmal robot (ANYbotics, Switzerland), equipped with a microscopic imager (MICRO) and a Raman spectrometer (Metrohm, 2024) (MIRA XTR, Metrohm, Switzerland), mounted on an ETH in-house developed robotic arm (ETH Zurich Robotic Systems Lab, 2024) a key objective of this research is to assess the capability of integrating MICRO and Raman spectroscopy on a robotic arm for in situ characterization of sample morphologies and mineralogies in lunar and Martian analogue environments, particularly for applications in astrobiological investigations and in-situ resource utilization (ISRU) prospecting.

A central research gap addressed in this work is that, despite substantial progress in rover autonomy and analogue-field testing, there has been limited quantitative evaluation of whether a single robot with a compact instrument suite can conduct semi-autonomous scientific surveying at a level sufficient for geological interpretation, while also improving operational efficiency compared with fully human-supervised workflows. Existing surface exploration strategies remain strongly dependent on conservative, step-wise decision-making by mission operators, particularly in communication-limited environments.

To investigate this, we compare two operational concepts for robotic scientific surveying.

• Multi-target, semi-autonomous sampling with minimal human intervention.
• One-target-at-a-time exploration with human-assisted decision-making.

Ultimately, this research aims to demonstrate the feasibility of a semi-autonomous robotic system for planetary surface exploration and scientific surveying using a compact, operationally efficient payload. By optimizing sampling efficiency and data acquisition, this study contributes to the development of more autonomous and adaptive exploration strategies, facilitating faster, more effective missions on the Moon, Mars, and other celestial bodies while reducing reliance on human operators.

2 Materials and mission analogue preparation

2.1 Robotic platform and payload

Analog missions were conducted using a legged robot based on ANYmal D (ANYbotics, 2025), which we equipped with a robotic arm and scientific instruments for the ESA-ESRIC Space Resources Challenge (Arm et al., 2023). The robot weighs ∼60 kg (including the robotic arm) and has a nominal operation time of 90 min while continuously walking. The robot (Figure 1A) uses a DynaArm a custom 6-degree of freedom (DoF) robotic arm, to deploy the instruments for in-situ measurements. The DynaArm has a total weight of 10 kg, including the 2.3 kg of the scientific payloads. It has a reach of 0.9 m without any tools attached and a nominal payload capacity of 8 kg at half of the reach, with a maximum tool speed of up to 13.5 m/s (Arm et al., 2023). There were two instruments mounted on the robotic arm: a MICRO - microscopic imager mounted at the tip of the arm and a Raman spectrometer (MIRA XTR) mounted on the forearm (Figure 1A).

The primary objective of the first instrument, MICRO, was to capture close-up images of the samples, enabling the identification of rock textures and lithologies. This information is crucial for understanding the geological history of the site, providing context for further analysis, and selecting the most scientifically valuable samples for investigation. MICRO (Figure 1B) is a custom-built UV-VIS-NIR imager operating from UV to NIR at wavelengths of 395, 470, 525, 620, and 940 nm, with a field of view measuring 25.1 mm × 20.1 mm at a distance of 51.7 mm. It incorporates a USB microscope (Dinolite AD4113T-I2V) mounted on a linear actuator mechanism, a ring of 48 RGB LEDs, a time-of-flight (ToF) sensor, and control electronics. All parameters are listed in the Supplementary Table S1. The front of the casing features a foam ring to prevent stray light from entering when MICRO is in contact with a target. The ToF sensor determines the distance between the microscope and the target, enabling precise focusing. This distance feedback, combined with the linear actuator and an autofocus routine, ensures sharp image acquisition independent of the instrument’s placement.

The second instrument, Raman spectrometer (MIRA XTR) (Figure 1C), is a commercial device provided by Metrohm. It features an infra-red excitation laser with a wavelength of 785 nm and a maximum power of 100 mW. Additionally, it is equipped with a telescopic lens with autofocus, allowing remote measurements from a distance of 0.6–0.7 m, eliminating the need for direct contact with the sample. The spectrometer’s operational range spans from 400 to 2,300 cm⁻¹ with a resolution of 8–10 cm⁻¹. The main objective of Raman was to test its deployment from the robotic arm and assess the mineralogy of the samples to complement observations made by MICRO.

2.2 Laboratory test site and analogue rocks

All Martian and lunar analogue missions took place at the Marslabor (Figure 2) of the University of Basel, a laboratory facility designed to conduct simulations in preparation for missions such as ExoMars (Bontognali et al., 2021; Ligeza et al., 2025). The Marslabor facility encompasses an 80 m² room featuring a 40 m² test bed composed of Martian analogue materials. This includes basalt, specifically leucite-tephrite sourced from the Eifel region and mechanically fragmented into 8–10 mm diameter pieces, and a reddish sandstone rich in hematite, comprising a blend of finely shredded bricks with a particle size under 6 mm (Bontognali et al., 2021).

We conducted Martian analogue missions (Figure 2I) during daylight hours, adjusting illumination to reproduce representative Martian surface lighting conditions. These followed the laboratory illumination convention developed through studies conducted in Marslabor (Bontognali et al., 2021; Ligeza et al., 2025), in which the proportion of direct to diffuse light is set to approximately 5:1 (5000 and 1000 lux, respectively) when measured at the centre of the scene.

We selected rocks that are morphologically and mineralogically analogous to lithologies found on Mars. We selected a set of rock samples based on their relevance to Martian lithologies and their potential for biosignature preservation. Gypsum (A) has been previously detected on Mars and is considered a strong candidate for biosignature preservation due to its ability to encapsulate organic material (Vaniman et al., 2018). Carbonate rock (B1), composed primarily of calcite, has also been identified on Mars and is recognized as a potential biosignature reservoir due to its stability and formation in aqueous environments (Boynton et al., 2009). Additionally, we included siliciclastic carbonate rock (B2), which has the potential to preserve microbially induced sedimentary structures (Bosak et al., 2021) sandstone (B3), one of the most common lithologies on Mars, is often associated with past fluvial or aeolian processes (McSween, 2015). Finally, we selected sulphur-bearing basalt (C), which not only indicates volcanic activity but may also contain sulphur-bearing minerals such as jarosite, formed in the presence of water and considered a good candidate for biosignature preservation (Elwood Madden et al., 2004). Additional details on these samples, including size, rock textures, and known mineralogy, are documented in Table 1. Prior to the simulation experiments, we conducted handheld spectroscopic measurements of the samples (Supplementary Figure SA1) to determine their mineralogical composition and establish the ground truth for the simulations.

TABLE 1. A table summarizing the characteristics and composition of the Mars and Moon analogue samples and the data to be acquired with Raman spectrometer and MICRO during the analogue missions.

*Three different samples of the similar compositions, placed next to each other to increase the size of ROI

**Two different samples of the same compositions, placed next to each other to increase the size of ROI

***The known major mineralogy of the samples (ground truth) was recorded using handheld measurements with Raman prior to simulations, conducted in a laboratory setting at a 0.6 m distance to simulate data acquisition from the robotic arm.

****1 MICRO deployment = 1 RGB image, 1 UV, 1, IR, 1R, 1G and 1B images.

The lunar analogue mission (Figure 2II) was conducted at night to simulate the realistic lighting conditions of the lunar environment, particularly those near the lunar south pole. Performing the test in darkness minimized the influence of ambient light, ensuring controlled illumination conditions. To replicate the extreme contrast and long shadows characteristic of polar lunar regions, we used a 180-W light-emitting diode (LED) lamp (Aputure LS 120D II) positioned at a low angle of approximately 15° creating very long shadows emphasizing topography, following methods described by Arm et al., 2023. This setup allowed us to assess the effects of extreme lighting on robotic operations and target identification under conditions analogous to those expected in future lunar missions. Here, we selected one of the most common lunar lithologies and samples with high ISRU potential, such as: dunite (A), a rock type identified during the Apollo 15 and 17 missions (Shearer et al., 2015) which provides insights into the composition of the lunar mantle and could be a source of oxides, Anorthosite (B) a very common rock on the lunar surface rich in anorthite (Ryder, 1982) and rutile (C), a titanium dioxide-bearing mineral that may serve as an indicator of pyroclastic deposits on the Moon (Adams et al., 1974) and a potential resource for titanium extraction. Additional details on these samples are provided in Table 1, and their mineralogical composition is documented in the Supplementary Figure SA1.

3 Methods

3.1 Locomotion, target selection, and payload deployment

In all our analogue missions, the robot was remotely controlled by an operator through the mission control system. The mission control system contains a graphical user interface (GUI) that shows a digital elevation map and camera images where commands and tasks are transmitted. The mission control system shows a virtual representation of the environment with a digital elevation map (DEM) and camera images. It allows the operator to send navigation and measurement commands (Figure 3).

The robotic locomotion, navigation, and measurement approaches used in this study were originally developed for the ESA-ESRIC Space Resources Challenge (Arm et al., 2023). In this section, we detail the relevant interfaces and algorithms for this study. To enable mapping and localization, ANYmal is equipped with a VLP-16 Puck LITE LiDAR (Velodyne), six RealSense D435 active stereo sensors (Intel)—positioned at the front, left, right, and rear—for elevation mapping, and two FLIR Blackfly wide-angle cameras (front and rear) to provide RGB image streams.

The operator can select measurement targets directly in the infrared image stream of the RealSense cameras. When the operator clicks in the image interface, the respective depth reading allows a projection into 3D space. A marker at the respective location in the elevation map helps the operator to refine the location. The operator can mark multiple targets in the GUI and request measurements for each marked target at any time. Each measurement request consists of the target pose and the desired measurement (MICRO, Raman, or both).

After requesting a measurement, the robot approaches the target location. If the operator requested a Raman measurement, the arm moves into a target position such that the ray cast from the Raman spectrometer’s autofocus intersects with the target. If the operator requested a MICRO measurement, the arm approaches the target to a distance of 0.07 m and then uses the MICRO’s time of flight sensor to place the payload on the target surface and compensate for localization errors or inaccurate target marking. For locomotion and instrument deployment, we use a combination of a reinforcement-learning controller for the leg joints and a model predictive controller on the arm joints (Arm et al., 2023). This combination allows us to profit from the robustness of a reinforcement-learning controller for walking and balancing on uneven ground, and the accuracy of a model predictive controller for instrument deployment. Once positioned, the instruments complete the measurements autonomously and send the data back to the mission control station.

Two operational modes for scientific robotic surveys were tested, and their mission architectures are detailed in Supplementary Figure SA0. In the multi-target, semi-autonomous sampling mode with minimal human intervention, the operators selected multiple science targets, and the execution of the tasks was performed autonomously by the robot (Supplementary Figures SA0-A). In the one-target-at-a-time mode, operators took an active role in the mission (Supplementary Figures SA0-B) by selecting targets, reviewing science data on the go, and deciding how many measurements should be taken based on previous scientific findings. Below, we describe each of these concepts, along with the target selection criteria and trade-off decisions, in detail.

3.2 Operational concept: multi-target, semi-autonomous sampling with minimal human intervention

In this study, “semi-autonomous” operations refer to a mode in which humans retain responsibility for science decision-making (e.g., target selection and definition of the measurement sequence) based on panoramic imagery received from the rover. Autonomy is limited to the execution of predefined commands, including locomotion, waypoint navigation, instrument deployment, and data return.

This study tests a multi-target Martian exploration approach under these constrains, aiming to assess the feasibility and science return of semi-autonomous, systematic surveying. To ensure reproducibility and evaluate performance, we conducted four identical analogue missions, repeating the same sequence, to assess measurement consistency and target identification success rates across multiple trails. The mission timeline defined T₀ as the moment the rover transmitted the first panoramic overview image to mission control (Figure 3I, right panel). At the start of each mission, the robot performed a simulated post-landing deployment and initial scan using its navigation and front-facing cameras, allowing operators to identify potential targets visually through size, morphology, and albedo contrasts. Simultaneously, an elevation mesh generated from RealSense stereo data was transmitted to mission control to provide topographic context and support target recognition (Figure 3I, left panel). At T₀, three distinct targets (A–C) were clearly distinguishable within the mesh and corresponding infrared imagery, enabling the operations team to mark them for investigation.

Since this was a prospecting, multi-target exploration mission, the aim was to maximize lithological diversity in the sampling strategy. Three distinct targets were selected based on albedo contrast and morphological characteristics: Target A, a high-albedo rock; Target B, a rock at the boundary of two geological units; and Target C, a high-albedo regolith patch surrounded by dark-albedo rocks. In this operational concept, the operator initially identified a target in the infrared camera imagery and selected a waypoint near the target using the GUI. The operator then saved a sequence of pre-programmed science commands, enabling a fully autonomous execution of a systematic scientific survey at each site. Once uploaded, the robot executed all tasks autonomously, including locomotion, waypoint navigation, arm deployment, and science measurements. For each target (Table 1) the measurement sequence began with the MICRO imager, which was pre-programmed to be deployed first (Fig. III). Each MICRO deployment produced a standardized dataset including UV, IR, white light (RGB), and individual Red (R), Green (G), and Blue (B) bands. Following the MICRO deployment, three Raman spectra were acquired per sample to provide complementary spectroscopic analysis.

After completing the measurement sequence at each target, the robot autonomously continued its execution cycle, moving to the next target and saving data after each measurement. Once measurements were completed for all three targets, the robot transmitted the collected data to the base station.

3.3 Operational concept: one-target at a time exploration with human supported decision-making

The second operational concept we tested was a single-target lunar analogue mission, where decision-making was human-supported. In this traditional method, the operator determined the number of measurements per target before moving on to the next, allowing for real-time adjustments based on data quality and scientific priorities. This approach aimed to evaluate the differences in operational efficiency and scientific outcomes, specifically testing whether this approach would allow for more precise arm deployment on the target, and higher science quality compared to previously tested semi-autonomous, no human intervention multi-target strategy.

As in the Martian simulations, we defined T0 as the moment when the robot transmitted the first panoramic/overview image of the landscape to mission control (Figure 3II, right panel). The mission began with the robot standing up, simulating a landing sequence, and performing an initial scan of the terrain using its RealSense cameras. However, unlike in the Martian scenario, challenging lighting conditions were introduced to simulate strong contrast at the lunar South Pole, making it difficult to identify rocks in the infrared image stream. Additionally, the flatness of the terrain (Figure 3II, right panel) meant that rock features did not create significant topographic variations in the elevation map, making them harder to detect. Due to these challenges, the operator requested the robot to move closer to a potential target area for better identification. This led to the detection of Target A, a massive rock, which was selected for scientific analysis.

The operator identified a target in the camera image and selected a navigation waypoint in the graphical GUI. Once the robot reached the waypoint, the operator refined the measurement point using the depth camera and adjusted the final location in the map representation within the GUI. The operator also selected the instruments for data collection and sent the measurement task to the robot.

A key difference in this human-supported method was the ability to assess the data quality after each measurement. Unlike in the multi-target semi-autonomous approach, where all measurements were pre-planned, the operator could immediately review the incoming data and decide whether additional measurements were needed. For Target A, the operator deployed MICRO first, obtaining a close-up image of the rock. After confirming that the image quality was sufficient to assess texture, the operator requested three Raman spectra measurements from different angles to capture diverse mineralogical data. If the rock exhibited diverse mineralogy from the MICRO image, then operator requested 2 or 3 Raman spectra, if the rock or soil was homogenous then the operator requested only one Raman spectra.

Following this, the operator moved the robot to Target B, which consisted of two smaller rocks. Here, the same process was applied: one MICRO image was taken to confirm the rock texture, followed by two Raman spectra, one for each rock, to identify mineralogy. Lastly, Target C (regolith) was analysed. The MICRO image revealed a homogeneous target, and therefore, only one Raman spectrum was requested for analysis.

Similar to the Mars analogue mission, payload deployment followed the same data acquisition sequence. MICRO captured UV, IR, white light (RGB), and individual Red (R), Green (G), and Blue (B) bands, followed by Raman deployments. However, unlike in the multi-target approach, where all pre-programmed measurements were executed semi-autonomously, here data was reviewed after each measurement, allowing for adaptive decision-making based on measurement quality. Additionally, the operator chose how many Raman measurements were deployed and determined their specific locations on the rock.

4 Results

4.1 Science observations for martian and lunar analogue missions

The MICRO and Raman instruments provided detailed insights into lithology, texture, and mineralogy across the Martian and Lunar analogue missions. To present our findings, we selected the highest-quality, raw MICRO images, focusing on white light (RGB), UV, and IR, while omitting the separate R, G, and B channels for clarity. These omitted channels are available in the Supplementary Figures SA2–A6.

For the Martian analogue mission, we present the best-quality RGB, UV, and IR images from each mission cycle (Figure 4), while for the Lunar analogue mission, all three spectral images are shown for each target (Figure 5). Additional data from all mission cycles is available in the Supplementary Figures SA2–A6. For Raman spectroscopy, we include all spectra acquired during the four Martian analogue missions (Figure 6) and the Lunar analogue mission (Figure 7).

4.1.1 Identification of martian analogue rock textures with MICRO

Figure 4 presents MICRO images of three targets (A, B, and C) along with D, which represents an additional image of regolith from one of the mission cycles (M1, M2, M3, or M4). These images demonstrate MICRO’s capability in identifying rock textures and morphologies. For Sample A (gypsum), MICRO successfully captured its fine-grained, crystalline texture, with veins and cracks visible in both IR and RGB images. Sample B consists of three rock types: B1 (carbonate rock), B2 (siliciclastic carbonate rock), and B3 (sandstone/mudstone). These samples exhibited a fine-grained texture but lacked significant detail, as they were partially covered by regolith, making it challenging to identify lithological features solely through close-up imaging. Sample C displays a fine-grained powder on the rock surface. The IR image reveals that the powder exhibits a crystal-like or needle-like structure. The host rock appears gray, without visible phenocrysts or laminations, making it difficult to determine its exact texture. Finally, Sample D represents regolith from the Martian analogue test bed, which is composed of angular grains.

4.1.2 Mineralogical analysis with Raman spectrometer for martian analogue missions

To complement the morphological observations obtained with MICRO, we analysed 12 Raman spectra to determine the mineralogical composition of the rocks. The mineral phases were verified by comparison with reference spectra from the RRUFF database. Figure 5 presents Raman spectra for each target (A, B, and C), with numbers next to each letter (e.g., A1, A2, A3, A4) indicating spectra from four different mission cycles. For Sample A, all four mission cycles (A1–A4) show a clear peak at 1010 cm⁻¹, with minor peaks at ∼1138 cm⁻¹ and ∼415–419 cm⁻¹, corresponding to the Raman spectrum of gypsum. For Sample B, only mission cycle 3 (B3) detected the calcite peak at ∼1089 cm^-1. Although the spectra exhibit strong calcite peaks, they also contain several secondary peaks due to regolith covering the rock (Figure 4B), resulting in significant spectral noise. This was also the only successful on-target measurement for this sample (see Table 2). The spectra from B1, B2, and B4 also show multiple peaks, but they were not analysed for composition, as they were off target (see Table 2). For Sample C, all four mission cycles (C1–C4) successfully identified a sulphur peak at 473 cm⁻¹. These measurements are consistent with ground truth data obtained prior to the analogue missions using a handheld Raman spectrometer to establish sample mineralogy (Supplementary Figure SA1).

TABLE 2. A summary table for Martian and Lunar analogue missions - operations, payload deployment and target identification.

*1 MICRO deployment = 6 pictures (white, IR, UV, R, G, B)

**3 different samples of the similar compositions, placed next to each other to increase the size of ROI

***2 different samples of the same compositions, placed next to each other to increase the size of ROI

4.1.3 Identification of lunar analogue rock textures with MICRO

The mission images (Figure 6) demonstrate MICRO’s capability in identifying rock textures and morphologies. For Sample A (dunite), MICRO successfully captured olivine phenocrysts with accessory black oxide minerals in the matrix. For Sample B (anorthosite), the images reveal a binary distribution of mafic (black, pyroxene) minerals and felsic (anorthite) components, visible across RGB, UV, and IR channels.

Target C exhibits a very fine-grained brown powder within the regolith. In contrast to the MICRO images of Martian analogue rocks (Figure 4), these lunar analogue images provide a richer source of scientific information due to the greater heterogeneity in rock textures. This includes the porphyritic texture of Sample B (anorthosite) and the coarse-grained equigranular texture of Sample A (dunite) with accessory oxides.

4.1.4 Mineralogical analysis with Raman spectrometer for lunar analogue missions

To complement the morphological observations obtained with MICRO, we analysed six Raman spectra to determine the mineralogical composition of the rocks. Figure 7 presents Raman spectra for each target (A, B, and C). For Sample A, we acquired three Raman spectra (A1, A2, A3) from different areas of the rock across four mission cycles. However, the rock appears to be compositionally homogeneous, with two distinct peaks at ∼825 cm⁻¹ and ∼858 cm⁻¹, corresponding to olivine. Each measurement also exhibits a prominent peak at ∼472 cm⁻¹, which is unlikely to be related to the sample itself, as dunite is composed of >90% olivine (Table 1). This peak is likely an artefact or contamination from a previous sulphur measurement, as sulphur is highly Raman-sensitive and exhibits a strong peak at ∼473 cm⁻¹. Since this peak is not characteristic of for olivine, it was likely caused by residual sulphur powder adhering to the MICRO sponge during prior analyses. When MICRO was deployed on the dunite sample, this residual sulphur may have been transferred to the surface, leading to the detected Raman signal.

For Sample B (anorthosite), only one spectrum (B1) showed a weak anorthite signal at ∼517 cm⁻¹. However, this peak was weak, and the spectrum contained significant noise, with no detectable pyroxene peak at ∼650 cm⁻¹. Another measurement, B2, was also conducted on anorthosite, but no dominant peaks associated with anorthite, or pyroxene were observed, rendering the measurement unsuccessful. For Sample C, only one Raman spectrum was obtained, which corresponds to rutile, identified by its diagnostic peak at 615 cm⁻¹.

These measurements are consistent with ground truth data obtained prior to the analogue missions using a handheld Raman spectrometer to establish sample mineralogy (Supplementary Figure SA1). In particular, the olivine peaks at ∼825 cm⁻¹ and ∼858 cm⁻¹ observed in Sample A match the reference data, which does not exhibit the ∼470 cm⁻¹ peak, supporting the hypothesis that it is an artifact. Similarly, for rutile, the strongest peak in the reference data is also at 615 cm⁻¹, confirming its identification. In contrast, for anorthosite, the ∼517 cm⁻¹ anorthite peak is more prominent in the (Supplementary Figure SA1) compared to the analogue mission spectra. This difference is likely due to the handheld Raman measurements being conducted on clean samples, whereas in the analogue mission, dust covering the rock surface may have introduced additional spectral noise.

4.2 Summary of operational results

Table 2 summarizes the operational success and target identification for all four Martian analogue missions (Mars Mission 1, Mars Mission 2, Mars Mission 3, Mars Mission 4) and one Lunar analogue mission. The duration of the multi-target, semi-autonomous, Martian analogue missions (M1, M2, M3, and M4) was 21, 15, 12, and 23 min, respectively, while the human-supported, single-target Lunar analogue mission took 41 min (Table 2). The average time for instrument deployment and data acquisition in the Martian analogue missions was 1 min per sample for MICRO and 1.5 min per sample for the Raman spectrometer. For the Lunar analogue mission, both MICRO and Raman deployment and data acquisition took 1.5 min per sample. It is worth clarifying that the distance between Martian and lunar samples was different, on Martian ones it was on average 1.5 m, while for lunar samples it was 80 cm. However, it should also be clarified that locomotion time is almost negligible compared to the time required for analysis.

In the Martian analogue missions, we acquired one MICRO image per cycle, including white light (RGB), IR, UV, R, G, and B channels, and one Raman spectrum per cycle. In the Lunar analogue mission, we acquired one MICRO image, while three Raman spectra were acquired for Target A, two for Target B, and one for Target C.

4.2.1 Metrics used to evaluate mission performance

To evaluate the success of the missions, we introduced the following metrics (Table 2), consisting of three success factors: (A) positioning of the arm (MICRO and Raman) during deployment, (B) data quality from MICRO and Raman, and (C) scientific information from MICRO and Raman.

The first success factor (A) evaluates the positioning of the robotic arm for MICRO and Raman deployment. It is measured by the accuracy of on-target deployment.

• • A green rating indicates that the arm was on target, successfully aligning with the selected sample.
• • A yellow rating means the arm was partially on target, capturing only part of the intended target.
• • A red rating signifies that the arm was off target, resulting in an unsuccessful deployment.

The second success factor (B) assesses the quality of the acquired data, divided into two subcategories: B1 (MICRO image quality) and B2 (Raman spectrum quality).

• B1: MICRO image quality
• ○ A green rating represents an ideal image, where the texture and structural details of the rock are clearly visible.
• ○ A yellow rating indicates a sufficient image, where the image is slightly blurred but still usable for analysis.
• ○ A red rating corresponds to a poor image, where blurring prevents any meaningful scientific observations.
• B2: Raman spectrum quality
• ○ A green rating indicates an ideal Raman spectrum, with strong, clear peaks representing the mineralogy of the sample.
• ○ A yellow rating means the spectrum is sufficient, with the main peaks visible but affected by noise.
• ○ A red rating signifies a poor spectrum, where noise dominates, making mineral identification not possible.

The third success factor (C) evaluates the scientific information obtained from MICRO and Raman data, divided into two subcategories: C1 (scientific value of MICRO images) and C2 (mineralogical information from Raman spectra).

• C1: Scientific value of MICRO images
• ○ A green rating indicates that the rock texture is well-defined, with both main and minor minerals identifiable, allowing for rock classification.
• ○ A yellow rating means that some texture details are visible, but additional information would be required for a conclusive rock classification.
• ○ A red rating signifies that the texture is indistinct, preventing any meaningful identification.
• C2: Mineralogical information from Raman spectra
• ○ A green rating is given when the main peaks correspond clearly to the primary mineralogy of the sample.
• ○ A yellow rating signifies partial success, where the main peaks are detected but accompanied by multiple unidentified peaks.
• ○ A red rating indicates that the spectrum does not contain any major mineral-related peaks, rendering the analysis unsuccessful.

The information from the third success factor was used to evaluate the usefulness of the acquired data. These assessments were translated into a usefulness index, where 0 represented non-useful MICRO or Raman data, 0.5 indicated partially useful data, and 1 corresponded to fully useful data (see Table 2).

The usefulness index is a semi-quantitative heuristic framework intended for internal comparison. The numerical values represent ordered categories rather than continuous measurements, and the use of single-rate scoring remains a limitation, even though the assessment criteria are objective. However, overall this classification system allows for a structured evaluation of mission success, ensuring that target identification is assessed based on both technical and scientific criteria.

4.2.2 Multi-target, semi-autonomous exploration with minimal human intervention

In the first Martian analogue mission, 2 out of 3 targets were successfully identified, resulting in an overall success rate of 66.7%. The identified targets were A (gypsum) and C (sulphur with basalt). For Target A, arm positioning was partial, meaning only part of the rock was captured in the MICRO image with sufficient quality but minimal scientific information. However, Raman analysis provided a definitive identification by exhibiting a strong peak at ∼1010 cm⁻¹, allowing the sample to be classified as gypsum based on both MICRO and Raman data. For Target C, the deployment of MICRO was partially successful, as it captured the rock sample instead of the regolith with sulphur. However, sulphur patches were still visible on the rock, and the Raman spectrum exhibited a peak at ∼473 cm⁻¹, confirming its classification as sulphur. Target B was not identified because the arm was off target, capturing both MICRO and Raman data from regolith instead of the intended rock sample.

In the second Martian analogue mission, 2 out of 3 targets were again identified, maintaining a 66.7% success rate. As in Mission 1, Targets A and C were successfully identified, while Target B was missed due to off-target arm positioning. However, the key difference in this mission was that the arm was placed precisely on Target A, resulting in ideal MICRO image quality and Raman spectra, making the gypsum identification conclusive. For Target C, although the arm was correctly positioned, MICRO image quality was poor due to blurring caused by arm shaking, limiting the scientific information. Although the autofocus system controls the working distance along the optical (out-of-plane) axis, it does not compensate for lateral (in-plane) motion of the sensor head. During some deployments, high-frequency vibration of the robotic arm resulted in small in-plane displacements, which produced blurred images even when the optical focus distance was correctly maintained. In a few instances, image degradation occurred when the initial placement of the microscope was outside the autofocus compensation range.

Nevertheless, the Raman spectrum was ideal, producing a sharp peak at 473 cm⁻¹, which confirmed the target’s composition as sulphur.

The third Martian analogue mission was the most successful, achieving a 100% success rate, with all 3 targets correctly identified. The key factor in this mission was that all arm deployments were on target, enabling successful identification of all rock samples. While MICRO image quality was ideal for all three targets, rock texture identification remained partially successful, as the analogue rocks lacked distinct characteristic features—all were fine-grained, making visual identification difficult despite high image quality. However, Raman spectroscopy confirmed the composition of each sample. Notably, this was the only mission where Target B was successfully identified. For a successful identification, we considered one of three rock samples placed in this region: B1 (carbonate rock), B2 (siliciclastic carbonate rock), and B3 (sandstone), which were positioned to increase the target area for better localization (Table 1). In this case, the arm was deployed on B2 (siliciclastic carbonate rock). Although the MICRO image quality was ideal, the rock was covered by regolith (Figure 4B), making it difficult to derive scientific information from imaging alone. The Raman spectrum was also noisy (Figure 5B3) due to regolith coverage. However, the presence of a peak at ∼1089 cm⁻¹, characteristic of calcite, allowed for the successful identification of the target.

In the fourth Martian analogue mission, 2 out of 3 targets were identified, again resulting in a 66.7% success rate. Targets A and C were correctly identified, while Target B remained off-target due to misalignment of the robotic arm. Unlike the first and second analogue missions, in this mission, Target A was successfully aligned with the arm, but Target C was only partially on target. However, Raman data successfully confirmed the mineralogy of both samples, even though MICRO image information was insufficient for conclusive identification.

4.2.3 One-target at a time exploration with human-supported decision making

In the Lunar analogue mission, where one-at-a-time exploration with human-in-the-loop decision-making was employed, all three targets were successfully identified, resulting in a 100% success rate. Since the robot was commanded by an operator, it had greater freedom to approach the sample closely, ensuring precise arm deployment for all Targets A–C. Only one image per sample was needed to obtain ideal quality, containing the maximum amount of scientific information (Figure 6). For Raman analysis, three spectra from Target A (dunite) confirmed that the rock is predominantly composed of olivine. A single Raman spectrum for Target C was sufficient to classify it as rutile, with a distinct peak at ∼615 cm^-1. For Target B (anorthosite), the returned MICRO image was ideal (Figure 6B), enabling clear identification of anorthosite texture. However, only 50% of the returned Raman spectra were successful. One spectrum displayed a weak peak at ∼513 cm⁻¹ along with numerous secondary peaks, which may have resulted from either regolith interference affecting the signal or the complex mineral composition of the sample. Since this rock contains more than two mineral phases, it poses challenges for accurate identification using commercial Raman spectrometers (see Discussion section). Despite these limitations, the complementary use of MICRO and Raman still enabled the successful identification of this rock.

5 Discussion

5.1 Evaluating multi-target and single-target approaches for planetary surface exploration

We acknowledge that the two approaches, multi-target semi-autonomous and single target, were used to simulate distinct missions: one focused on identifying Mars analogue samples and the other on identifying Moon analogue samples. When comparing the results obtained with the two methods, it is important to consider the potential bias introduced by the intrinsic differences between the two sample sets, as one may be inherently more challenging to identify than the other. However, since mission success and execution speed are primarily governed by technical factors—such as the robot’s movements, instrument positioning, and the execution of analyses—we believe that comparing the time required to complete the Martian and Lunar missions using the two different approaches still provides a meaningful estimate of the time savings achievable with a semi-autonomous approach, as well as the associated trade-off in terms of scientific return (expressed as the total number of successfully identified samples).

The results from our analogue missions demonstrate the high potential of semi-autonomous, multi-target sampling with minimal human intervention as an efficient method for planetary surface exploration. Over an average mission duration of 16.5 min, the robot successfully identified at least 66.7% of targets per cycle, achieving 100% target identification in one out of four analogue missions. Despite the small target sizes (≤30 cm, Table 1), the robot successfully located and deployed the arm for scientific measurements, confirming the reproducibility of data across four identical missions. For example, Target A was consistently identified as gypsum, and Target C as sulphur throughout all mission cycles.

Even partially on-target deployments often enabled successful classification, albeit with reduced imaging quality (Table 2, eg., Mars Mission 2, Target C). When the MICRO image was suboptimal due to partial coverage or blurring, Raman spectroscopy compensated for this limitation, allowing definitive mineral identification. This highlights the complementary nature of MICRO and Raman spectroscopy in detecting and characterizing Martian lithologies and mineralogies, demonstrating its potential as a promising tool for an initial astrobiological science survey on Mars.

5.1.1 Mission performance - evaluation

The performance of the Mars and Lunar missions was further evaluated based on mission duration, sample identification success rates (Figures 8M1–M4,L), data collection efficiency (Figure 8A), and the usefulness of the collected data (Figure 8B). The results highlight distinct advantages associated with each mission type, depending on whether operational efficiency or scientific output is prioritized.

The Lunar mission, which lasted 41 min (Figure 8L), achieved 100% sample identification success, identifying all three samples while collecting nine data points, 7.5 of which were considered useful. This represents the highest data yield and quality among all missions, making the Lunar mission the most successful for science return. In contrast, the shorter Mars Mission 3 (15 min) also achieved 100% identification success but with a lower total data collection (six data points) and fewer useful data points (four). Despite this, Mars Mission 3 demonstrated high operational efficiency, achieving the same identification success as the Lunar mission in just 36.5% of the time.

The other Mars missions (1, 2, and 4) had only a 67% success rate, meaning that one out of three samples remained unidentified. Each of these missions collected six data points, yet their total useful data ranged between 2.0 and 3.0 points, significantly lower than both the Lunar mission and Mars Mission 3. The Lunar mission collected 1.5 times more total data than any Mars mission and nearly 1.88 times more useful data than Mars Mission 3. When compared to the average useful data across all Mars missions (≈2.88 useful points per mission), the Lunar mission produced nearly 2.6 times more valuable data per mission.

In terms of usefulness per sample, the Lunar mission led with 2.5 useful data points per sample, followed by Mars Mission 3 with 1.33 useful data per sample. The other Mars missions ranged from 0.67 to 1.0 useful data per sample, indicating greater variability and reduced efficiency in their data acquisition processes. This suggests that while shorter Mars missions can be highly efficient, they may require optimized sampling strategies to improve data usefulness.

From an operational standpoint, Mars Mission 3 demonstrated the best balance between speed and success, achieving full sample identification and relatively high data usefulness in a significantly shorter time. This efficiency is particularly valuable for time-constrained missions or operations where minimizing power and resource consumption is critical. However, when maximizing total scientific return is the priority, the Lunar mission remains the most effective, as it outperformed all Mars missions in both total and useful data collection, albeit requiring significantly longer operational time.

5.1.2 Data collection rate across analogue missions

The data collected per minute metric (Figure 8A) provides insight into how efficiently each mission gathered information relative to its duration. The results highlight key differences between Mars and Lunar missions, reflecting their distinct operational approaches and trade-offs between efficiency and total data yield.

The Mars missions consistently demonstrated higher data collection efficiency than the Lunar mission. The most efficient Mars missions, Missions 2 and 3, collected 0.40 data points per minute, while the least efficient Mars mission, Mission 4, collected 0.26 per minute. In contrast, the Lunar mission had the lowest collection rate of 0.22 data points per minute, making it the least time-efficient mission in terms of data acquisition.

Among all Mars missions, Mission 3 stood out as the most efficient, achieving 0.40 data points per minute and 100% sample identification success. This suggests that short-duration, well-optimized, semi-autonomous operations can maximize efficiency without sacrificing scientific return. The success of Mission 3 highlights the potential of automated, multi-target sampling strategies for rapid planetary exploration.

Despite its lower data collection rate, the Lunar mission collected the highest total data volume, with 9 data points, compared to the 6 data points collected in each Mars mission. This suggests that while longer missions allow for more total data acquisition, they do not necessarily maximize time efficiency. The lower collection rate of the Lunar mission could be attributed to longer instrument deployment times, higher precision operations, or increased human oversight.

5.1.3 Usefulness of data collection across all analogue missions

The usefulness of collected data (Figure 8B) varied significantly between Mars and Lunar missions, reflecting differences in operational efficiency and scientific output. Usefulness was evaluated based on the science information obtained from MICRO and Raman (see Section 4.2.1), where an index of 0 indicated not useful data, 0.5 represented partially useful data, and 1 signified useful data. Mars missions demonstrated higher data efficiency per minute, with Mars Mission 3 achieving the highest useful data collection rate (∼0.27 useful data points per minute), aligning with its 100% sample identification success rate. Other Mars missions had lower efficiency (∼0.10–0.17 useful data per minute), suggesting that instrument deployment precision and operational variability influenced data quality. In contrast, the Lunar mission collected the highest total useful data (7.5 points) but at a lower efficiency (∼0.18 useful data per minute). This indicates that longer missions yield more absolute data but may not optimize data acquisition efficiency, likely due to higher precision operations, increased human oversight, or more frequent instrument deployments. These findings highlight a trade-off between data yield and time efficiency, where short Mars missions prioritize rapid, autonomous sampling, making them more effective for time-sensitive planetary exploration, while longer Lunar missions maximize total data collection and measurement precision, benefiting from real-time human oversight.

5.2 Multi-target, semi-autonomous exploration and one-target at a time exploration - lessons learned

However, semi-autonomous target selection has limitations. Namely, selecting the a priori from far away led to off-target arm placement and payload deployment for Sample B, resulting in only 25% success across all missions. In the semi-autonomous scenario, the operator could not check the data after each measurement to request additional data. This contrasts with human-in-the-loop operations, where the target is selected from a closer position, which ensures precise arm placement (Table 2 – Lunar Mission). The reason for this discrepancy is that the targeting system used in this study was originally designed for a human-in-the-loop approach with big rocks as targets. A potential solution to this issue is using visual servoing close to the target to close the feedback loop during the arm positioning (Kragic and Christensen, 2002). Another recurring challenge was blurred MICRO images. Potential reasons for blurred images are that the arm moved slightly during measurements or that the microscope was not focused. A feasible mitigation strategy would be to autonomously check the images for blur (Chiang and Boult, 1997) and repeat the measurement if the image is too blurred.

To move to an even higher level of autonomy, the robots could detect targets of interest autonomously based on shape, colour, and texture. In scenarios where the data transmission is very slow (e.g., in the outer solar system), the robot could then autonomously take measurements of these targets. In other cases, the robot could send the target candidates back to the operator and allow the operator to prioritize the targets.

5.3 Complementarity of MICRO and Raman spectroscopy

These missions provide a clear demonstration of the synergy between MICRO imaging and Raman spectroscopy for the identification of Martian and Lunar lithologies. In most cases, a single successful Raman measurement per target was sufficient for definitive mineral identification, particularly in Martian analogue missions, where Raman spectroscopy proved highly effective. However, we acknowledge the limitations of the commercial Raman spectrometer used, which, due to budget constraints, operated at 785 nm with a detection range of 400–2300 cm^-1. This restricted the identification of complex mineralogies, such as anorthosite in the lunar analogue mission, which contains both anorthite and pyroxene (Figures 7B1,B2).

Our analogue missions demonstrated Raman’s effectiveness in identifying homogeneous mineralogies, such as: Martian targets A and C (gypsum and sulphur) and Lunar targets A and C (dunite and rutile). However, Raman was less effective for complex mineral assemblages, such as Lunar Target B (anorthosite), which consists of multiple mineral phases. Additionally, Raman struggled to detect oxides, such as spinel in dunite, which was visible in MICRO images (Figure 6A) but did not produce a clear Raman signal due to the low concentration of the minerals.

An important implication of this spectral limitation concerns lunar ISRU applications. Water ice exhibits Raman peaks primarily between ∼3000 and 3500 cm^-1 (Xue et al., 2013), which lie outside the 400–2300 cm^-1 range of the instrument used here. Consequently, while our system is well suited to the identification of many rock-forming minerals, it is not optimal for direct detection of water ice in permanently shadowed regions. Future mission concepts may therefore benefit from Raman instruments employing alternative laser wavelengths (e.g., 532 nm) and extended spectral ranges up to at least 3500 cm^-1, as being explored in studies such as LunaR (Cloutis et al., 2022; 2023). Such configurations would enable the detection of key vibrational features associated with feldspars, pyroxenes, olivines (Huang et al., 2000), and volatile-bearing phases including water ice. This is consistent with current planetary Raman implementations, including SHERLOC (248.6 nm) (Corpolongo et al., 2023) and SuperCam (532 nm) on Perseverance (Bhartia et al., 2021; Maurice et al., 2021), and the instrument planned for ExoMars (Rull et al., 2017).

Thus, while the Raman configuration deployed in this study was highly effective for many analogue lithologies, extending the spectral range would broaden mineral detectability and enhance the relevance of Raman spectroscopy for future lunar ISRU-focused missions.

5.4 Locomotion and manipulation control

The terrain conditions on the Moon and on mars are highly uncertain and challenging for walking robots. The focus of this study was on the operational implications rather than the control performance, Nevertheless, the granular terrain in the lab led to slippage and sinkage, and therefore tested the locomotion controller under harsh conditions.

All experiments were conducted in Earth gravity. As other planetary bodies of interest have lower gravity, the manipulation control of the robotic arm actually becomes easier, since the gravity compensation for the arm is less demanding.

5.5 Future studies–outlook

A key direction for future study will be the optimisation of science data products obtained from arm-deployed MICRO and Raman instruments. Although the present work concentrated primarily on demonstrating the feasibility of semi-autonomous multi-target operations, further development of instrument-level data processing could substantially enhance scientific interpretation. For example, the multispectral capability of the MICRO camera could be exploited to generate spectral band-index maps, similar to ones on another Martian missions, e.g., PanCam camera on ExoMars (Coates et al., 2017). Similarly, because the Raman sampling footprint lies within the MICRO field of view, the two datasets could be co-registered and jointly visualised to provide spatially resolved, multi-instrument context at each sampling site. A further avenue for development concerns automated quality-assessment of acquired data, such as deploying onboard sharpness or signal-to-noise metrics to trigger automatic re-acquisition when measurement quality falls below a defined threshold (Her and Yang, 2019). Together, these enhancements would improve the scientific robustness and reliability of semi-autonomous surface investigations, and represent natural extensions of the operational concepts explored in this study.

5.6 Implications for future planetary exploration

Our findings indicate that the multi-target, semi-autonomous exploration concept tested in this study represents a potential advance over current rover operational practices. Although existing Mars missions such as Curiosity and Perseverance already employ autonomous capabilities—for example, autonomous navigation and automated target detection through systems such as AEGIS (Francis et al., 2017) — these systems primarily support the execution of carefully pre-planned activities rather than high-level scientific decision-making. In the current state-of-the-art, Curiosity and Perseverance possess autonomy mainly for navigation, target detection, and execution of pre-approved actions; however, they do not autonomously plan multi-target sampling campaigns, determine the sequencing of investigations, or coordinate repeated instrument deployments across several targets within a single operational cycle. As a result, rover operations remain predominantly single-target, ground-controlled, and sequential over multiple sols, meaning that rovers rarely—if ever—conduct an entire “parcours” of scientific targets autonomously within one planning cycle. This conservative operational architecture reflects the risk of Mars missions, where sampling activities are high-value and fault-sensitive, and where chaining multiple sampling events without intermediate human review would significantly increase mission risk. In contrast, the approach demonstrated in this study explicitly tests whether semi-autonomous systems can coordinate the exploration of multiple scientific targets within a single operational cycle, with autonomy contributing not only to locomotion but also to the science of pre-programmed activities. Notably, this capability was demonstrated using a relatively simplistic payload, showing that meaningful multi-target scientific assessment can be achieved efficiently and without continuous human supervision. This suggests that such multi-target workflows could increase mission efficiency by reducing latency associated with ground-in-the-loop decision-making and by enabling rapid site-scale characterisation prior to committing to detailed sampling This would enable faster data collection, optimizing instrument utilization and increasing scientific return. The ability to semi-autonomously navigate, select, and analyse targets would be particularly valuable for rapid site characterization, resource prospecting, and biosignature searches, reducing the dependency on mission control decision-making and increasing mission efficiency in challenging terrains.

A practical consideration for future implementation relates to power and endurance constraints. The robotic platform used in this study has a nominal operational endurance of approximately 90 min under continuous terrestrial locomotion, and our analogue missions ranged from 12 to 41 min. In a planetary mission context, energy availability will be more restricted, and recharging opportunities may be infrequent or time-limited. On the Moon, such constraints may be mitigated through the integration of solar arrays, enabling periodic recharging during surface operations. On Mars, where solar input and environmental conditions vary substantially, endurance limitations would likely necessitate carefully optimised duty-cycling, staged traverses, and strategic scheduling of energy-intensive scientific activities. These considerations emphasise that semi-autonomous, multi-target surveying must operate within a constrained energy budget; however, the increased science efficiency demonstrated here suggests that greater scientific return may still be achievable per operational cycle compared with traditional single-target workflows.

While this study focused on validating the concept and showing a potential of multi-target exploration with a single robotic system, the approach could be extended to other robotic platforms, such as wheeled rovers, aerial drones, and swarm robotics. The incoming swarm and aerial robotic systems (Sand et al., 2013; Lorenz et al., 2018; Rabideau et al., 2025) could distribute exploration tasks, enabling more efficient resource mapping, increased scientific return, and improved mission resilience in extreme planetary environments.

6 Conclusion

Our study demonstrates that a multi-target semi-autonomous exploration approach is a viable option for geological investigations in planetary surface missions where the inability to control a robot in real-time significantly slows down exploration times and, consequently, the scientific return of the mission.

The integration of MICRO imaging and Raman spectroscopy on a robotic arm proved to be highly complementary, enabling efficient lithological and mineralogical characterization. This combination is particularly valuable for astrobiological and ISRU prospecting, as Raman identifies key mineralogy while MICRO provides textural context, ensuring a more robust interpretation of potential scientific targets.

Our results show that multi-target sampling optimizes instrument utilization and accelerates data collection. Mars Mission 3, operating in a shorter time frame (15 min), achieved a useful data collection rate of 0.27 useful data points per minute—50% higher than the Lunar mission (0.18 per minute). Despite the Lunar mission collecting 1.88× more total useful data, it operated for 2.7× longer (41 min vs. 15 min for Mars Mission 3), highlighting the trade-off between data yield and operational efficiency. Mars missions collected data 22% faster than the Lunar mission, reinforcing the suitability of multi-target, semi-autonomous strategies for time-restricted exploration where real-time decision-making is limited, such as Mars.

Conversely, the Lunar mission’s longer duration allowed for higher overall data yield and precision, demonstrating that single-target, human-in-the-loop approaches may be preferable when high-resolution, detailed measurements are required. These findings emphasize the need to balance mission automation, efficiency, and scientific return based on operational constraints and planetary environments.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/frspt.2026.1741757/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Morad Nazari, Embry–Riddle Aeronautical University, United States

Reviewed by: Cagri Kilic, Embry–Riddle Aeronautical University, United States

Sandra Lorenz, Helmholtz Center Dresden-Rossendorf, Helmholtz Association of German Research Centres (HZ), Germany