Human-in-the-loop proof of concept for lunar teleoperation: performance, workload, and neuro-physiological insights from the LUNA Analog Facility

As lunar missions progress towards a sustained human presence, teleoperated robotic systems will play a critical role in surface exploration, construction, and logistics. Humans will teleoperate systems in diverse settings, ranging from suited astronauts on the lunar surface, to mission control operators on Earth. Each of these contexts presents distinct cognitive and physical challenges to the human operator. This proof-of-concept study examined teleoperation performance, workload, and neuro-physiological responses in two lunar analog operational contexts: a constrained ‘Lunar’ condition in which participants wore full protective equipment under simulated lunar lighting, and an unconstrained ‘Control Center’ condition simulating remote, non-analogue lunar tele-operation. Eighteen participants teleoperated a lunar rover in the LUNA Analog Facility, completing both a flat-ground navigation task and an obstacle course task in a high-fidelity analog environment. Performance was measured via completion time, path length, and steering reversal rate; subjective workload via NASA-TLX scores; and neuro-physiological response via electrocardiography (ECG) and functional near-infrared spectroscopy (fNIRS). Objective performance did not differ significantly between conditions on any metric, based on the 13 of 18 participants for whom complete performance data were available. Subjective workload was consistently and substantially higher in the Lunar condition, particularly during the obstacle course task. Heart rate was elevated in the Lunar condition in the full sample, while heart rate variability showed no significant differences between conditions. Neuro-physiological oxygenation data, available for 8 fNIRS-instrumented participants, showed no significant effect of Condition, Task, or their interaction on either oxy- or deoxyhemoglobin concentration change. Exploratory analyses suggested that the largest observed differences were more consistent with systemic (respiratory and thermoregulatory) confounds than with prefrontal neural activation. These findings emphasize the importance of designing human-robot systems for lunar missions with careful consideration of operator state and environmental constraints, independent of objective task success.

Robotic teleoperation is vital for lunar missions, enabling surface operations while minimizing astronaut risk (Cooper et al., 2005). Rovers support site exploration (Maurette et al., 1997), sample collection (Baker et al., 2024), infrastructure deployment (Seedhouse and Llanos, 2021), and astronaut assistance (Burridge et al., 2003). However, the harsh lunar environment, characterized by extreme lighting, abrasive regolith, and limited visibility (Harrell et al., 2021), significantly challenges teleoperation performance.

Future systems will involve near-field operations from the lunar surface (Panzirsch et al., 2022) and far-field control from Earth or the Gateway station (Burns et al., 2019). Teleoperation must remain robust across diverse conditions, from suited astronauts on rugged terrain (Schreiner et al., 2015; Bagdigian and Stambaugh, 2010) to operators in pressurized habitats (Timman et al., 2023). Each context imposes unique demands; bulky suits impair mobility, while remote operation introduces latency and reduced sensory feedback.

Mission success depends on both robotic technical specs and human operator limits. Latency, limited feedback, and task complexity can degrade performance (Smith and Stuart, 1990). Specifically, the lack of haptic cues impairs telepresence and accuracy (Kaber et al., 2000; Kumar et al., 2020; Louca et al., 2024). Furthermore, EEG-based studies show that spatial cognitive abilities directly influence workload and efficiency during manipulator tasks (Shao et al., 2019).

Recent fNIRS studies have shown that delayed sensory feedback alters activation within prefrontal and motor cortical regions during teleoperation, and that providing immediate haptic feedback can mitigate these neural changes while improving motor performance (Causse et al., 2017; Zhou et al., 2024).

Lunar analogs are terrestrial sites replicating lunar conditions to test equipment and human performance without the cost or risk of spaceflight (O’Keefe et al., 2025). The LUNA Analog Facility at ESA’s European Astronaut Center recreates these conditions for pre-mission testing (Casini et al., 2020).

By incorporating sensory realism—such as regolith simulant, lighting contrasts, and mobility constraints—analog facilities bridge the gap between simulation and mission. They reveal design flaws and performance bottlenecks, like physical fatigue and ergonomic interference, that virtual environments miss (Rahill et al., 2025). Thus, analog testing is essential for optimizing system design and minimizing operational failure.

Despite expanding teleoperation research, few studies examine performance, workload, and neurophysiological stress in high-fidelity lunar analogs (Sam et al., 2024; Glaros et al., 2023). Current human factors data often relies on virtual simulations, limiting external validity (Nilsson et al., 2022; Xie et al., 2022). Additionally, the sensitivity of workload-linked neurophysiological signals in rugged analog terrains remains under-studied (Pereira et al., 2025; Hopko et al., 2022). This study uses a multimodal approach in a realistic testbed to provide data-driven insights into how environmental and ergonomic constraints affect operator wellness and performance.

Held at the LUNA Analog Facility, this study assessed human performance and workload during rover teleoperation under simulated constraints. Participants navigated two scenarios—a “Flat Navigation Task” and a complex “Obstacle Course Task”—under two conditions: wearing a lightweight protective suit in a constrained environment and in a simplified control environment. Tasks were counterbalanced in a pseudorandom sequence. Metrics included completion time, movement commands, NASA-TLX scores, heart rate (via ElectroCardioGraphy, ECG), and prefrontal oxygenation (via function Near-InfraRed Spectroscopy, fNIRS), with baseline measurements taken before each trial.

We hypothesized that the lunar condition would lead to performance degradation, as well as a higher perceived cognitive load, as compared to the control condition. Furthermore, a presumed neuro-physiological response would match the perceived workload. The three hypotheses were formulated as follows: 1. The Flat Navigation Task would result in higher performance and lower subjective cognitive workload than the Obstacle Course, across both experimental conditions.

2. Performance would decrease under lunar conditions, due to the combined effect of the Lunar condition package (personal protective equipment, lighting, and reduced situational awareness).

3. Lower performance would be associated with: higher subjective workload; higher heart rate (HR) and lower heart rate variability (HRV), indicating increased stress; higher prefrontal oxygenation, indicating increased objective workload.

The experiment was conducted in the LUNA Analog Facility with operators requested to teleoperate a rover under two conditions: one that mimicked a Mission Control Center (Control condition) and the other imitating surface operations on the Moon (Lunar condition). Under specified conditions, subjects engaged in a level terrain navigation task and a more difficult task of having the rover drive through metal rings arranged in a zig-zag pattern in the deeper section of the facility, intended to mimic crater-like terrain. Since both tasks were completed in both conditions, all subjects had four task-condition combinations in total.

Experimental setup and infrastructure

To enable a real-world environment for teleoperation simulations that can be used in lunar exploration, the research was conducted using both the LUNA Analog Facility and the Argonaut Recycling Demonstrator rover (Orbit Recycling, Germany). The account that follows provides the major features of the LUNA facility as used as the working space, in conjunction with the Argonaut Demonstrator upon which participants teleoperated during the experimental work. The two comprised the main infrastructure supporting the research.

LUNA Analog Facility description

The LUNA Analog Facility in Cologne, Germany, was established by the European Space Agency (ESA) and the German Aerospace Center (DLR, by its German acronym). The facility is an advanced lunar simulation space to be utilized to test and validate technologies, procedures, and perform simulations for future missions to the Moon. It has a 700 squared-meter hall accommodating about 900 tons of a simulant, EAC-1, a substitute for lunar regolith. The hall is divided into two-halves, one level and a “Deep Floor Area” which is a region where the regolith drops to 3 m. The Sun simulator is a feature of the facility that reproduces conditions of lunar illumination. Furthermore, the Argonaut Landing Descent Element (LDE) has been placed inside the facility, and there are analog EVA suits available for testing. Everything is monitored through the centralized control system, where staff monitor and support integrated testing.

Orbit recycling rover description

The rover used for the experiments was the Argonaut Recycling Demonstrator, developed by Orbit Recycling (Germany). It was chosen due to its capability to cross lunar regolith and its teleoperability potential. It was designed to create furrows for cable lines. This task was part of a larger proposed system for recycling Argonaut aluminum, where a separate metal-casting rover would produce the cables to connect the lander to nearby systems. It features modular motor and servo control systems, which can be utilized by users to command its motion and mechanical functions via a web-interface or feed commands directly. A live video stream provides real-time imagery. The rover is controlled using ROS2 (Robot Operating System 2), which ensures the interaction of components and enables remote diagnostics and control by utilizing status indicators.

Conditions

The experiment was conducted under two different experimental conditions: the Lunar Analog Condition (hereafter referred to as Lunar) and the Control Center Condition (hereafter referred to as Control). Both conditions were implemented inside the LUNA Analog Facility Hall. While the Control Center Condition was designed to simulate operating conditions of a true mission control center, it was not run remotely, or from outside of LUNA’s control room, due to spotty connectivity during the time of the test. It was, therefore, carried out in the LUNA Hall. The two conditions reflected the contrasting physical and operation environments between field operations on the Moon and remote operations on Earth.

Lunar analog condition

In the Luna condition, as portrayed in Figure 1B, participants operated within the Argonaut area of the Lunar facility. A table and a chair were placed near the lunar lander, facing the regolith-covered test zone of the LUNA facility. Subjects wore dust-protection personal protective equipment, which imposes some of the visual and manual constraints of a lunar EVA suit. This included a full-body personal protective equipment (PPE) suit, gloves, boots, a hairnet, and a protective helmet. The helmet was equipped with a battery-powered ventilation system mounted around the waist, supplying filtered air. It was intended as a manner to imitate the internal airflow and cooling systems of a real lunar spacesuit. This protective equipment does not reproduce suit pressurization, joint torque, glove-induced dexterity loss, suit mass, or the reduced gravitational loading (1/6 g) of the lunar surface, and findings regarding physical/manual constraints should be interpreted as reflecting dust-protection PPE rather than a full EVA suit analog.

To enhance realism, the facility’s solar simulator was turned on during the Lunar condition, providing harsh directional lighting similar to what would be experienced on the Moon. Importantly, the standard hall lighting was turned off, which contributed to the high-contrast shadowing and limited visibility typical of the lunar environment.

Control center condition

The Control Center condition, as shown in Figure 1A, was conducted in the “apron” area of the LUNA facility, which is a zone adjacent to the regolith field but not covered with simulated lunar soil. The “apron” area is still inside the LUNA Hall, where the regolith is situated.

This condition was designed to mimic the working environment of personnel stationed at a mission control center on Earth. Consequently, participants in this condition were only required to wear minimal protective gear. This included the standard full-body PPE suit, a hairnet, shoe covers, and an FFP2 mask for safety, but notably excluded the gloves, helmet, and boots worn in the Lunar condition. In the Control Center Condition, only shoe covers were needed.

Additionally, the full overhead lighting of the LUNA hall was kept on throughout the Control Center condition, resulting in uniformly bright conditions across the entire regolith field. This setup allowed participants to operate without the visual restrictions imposed by the lunar light simulation.

When in this condition, the subjects could not see the lunar regolith, as they were placed behind a fake rock wall.

Task design

The experiment comprised two distinct teleoperation tasks. Both tasks were performed by each subject under both test conditions—Lunar and Control—resulting in four task runs per participant. The objective was to compare performance and subjective workload between varied environmental and ergonomic conditions.

As shown in Figure 2, teleoperation was conducted using the Orbit Recycling Rover, controlled using an Xbox gamepad. Input was limited to four basic direction keys: forward, backward, left, and right. These were reached through the D-pad (directional pad), not the joysticks. Participants sat in front of a computer console showing the rover control panel. The only relevant feedback for performing the tasks was real-time video feedback from the onboard camera of the rover.

Every task had a 15-min time constraint. If the task was not completed during this time, the trial was terminated by the experimenters. A two-minute baseline was collected prior to every leg of the experiment while the participant rested. After each task, participants filled out a paper format of the NASA Task Load Index (NASA-TLX) to report their subjective workload. A five-minute break was provided while switching between one condition and another (Lunar or Control).

Prior to performing the tasks, participants underwent a short training exercise to familiarize themselves with the control system and visual feedback. In the brief exercise, which lasted for approximately two minutes, they were asked to take the rover around one boulder.

To control for order effects, the experimental conditions were pseudorandomized across participants. Each subject completed both the Lunar and Control conditions, performing both tasks (Flat Navigation and Obstacle Course) once per condition, for a total of four task trials.

The starting condition (Lunar or Control) was counterbalanced across subjects to ensure an even distribution. Each participant completed both tasks in the first assigned condition before switching to the second. Within each condition, the order of the two tasks was pseudorandomized. This resulted in four possible task-condition sequences: Lunar first with Navigation then Obstacle, Lunar first with Obstacle then Navigation, Control first with Navigation then Obstacle, Control first with Obstacle then Navigation.

Flat navigation task

The Flat Navigation Task was conducted on the level regolith field of the LUNA hall (Figures 3A, 4). The objective was for participants to complete one and a half loops between two designated points. The task began at a large rock positioned in the corner of the hall and ended at a flag placed in front of the Argonaut lander mock-up. Participants were instructed to drive the rover from the rock to the flag, return to the rock, and then drive once more to the flag. Throughout the task, they were required to keep the rover as far to the left of the course as possible, while avoiding any craters, rocks, or other obstacles. The subjects had to maintain control of the rover without flipping, stalling, or deviating from the path, to complete the course successfully.

Deep floor area obstacle course

The second task was conducted in the Deep Floor Area of the LUNA hall (Figures 3B, 4). This 3 × 3 m area features a depth of approximately 3m and includes slopes of 15 ° and 25 ° to simulate the maneuvering required for crater or uneven terrain operations.

For this purpose, nine metal rings were inserted vertically into the regolith, forming a zigzag slalom-style obstacle course. The participant’s objective was to locate each ring using the rover’s live video feed and then navigate around them in alternating directions, as if weaving through a slalom path. Furthermore, the participants had to avoid touching the ring, or making it fall. Because the Deep Floor Area includes elevation changes and softer terrain, this task posed a higher risk of the rover flipping, stalling, or veering off-course. This task started at the flag position, and to finish, the participant had to take the rover up to the initial rock.

Subject selection and training

18 participants completed the experiment: 15 were European Astronaut Centre (EAC) interns, and 3 were experimenters. The average age of the participants was 24.7 (SD = 3.5), 7 were female. The participants enrolled voluntarily, and signed an informed consent form before their participation. The 15 EAC interns had no hierarchical relationship with any of the experimenters.

Before entering the LUNA Analog Facility, all participants went through a mandatory safety briefing delivered by ESA staff, according to operational procedures for entering the hall. The briefing covered the use of personal protective equipment, navigation restrictions, and behavior in the regolith simulation.

Although participants were recruited from the European Astronaut Centre (EAC) and may have had incidental exposure to general images of the LUNA facility through internal communications, they were not provided with maps or prior information regarding the specific terrain layout used in this study. All participants were considered unfamiliar with the experimental course and were required to navigate solely based on the visual feedback and a standardized briefing provided at the start of the session. Only the three experimenters had seen the terrain before executing the task.

Participants also went through a short training session before their first activity. This exercise was designed to accustom them to the rover’s control scheme and interface, i.e., relying on live video feedback for navigation. In this exercise, participants were instructed to teleoperate the rover in a round trip around one boulder in the flat area of the LUNA hall. This exercise was approximately two minutes long.

Data analysis

The data collected during the experiment belonged to one of these three categories: performance data, subjective data, or neurophysiological data. Complete ANOVA results for all models can be found in the Supplementary Tables, including the exact degrees of freedom, F-statistics, p-values, and 95% confidence intervals for partial eta-squared (calculated using the noncentral F method (Steiger, 2004)). Each of these data were analyzed separately and then compared across the two experimental conditions, that is, Lunar and Control; and the two tasks, Flat Navigation and Deep Floor Area Obstacle Course. Prior to the start of each trial, a two-minute resting baseline was recorded to normalize the physiological data for each participant. Data processing and analysis were carried out using MATLAB, Python, and JASP software.

Performance data

Performance was quantified using three independently analyzed metrics derived from the rover’s command log(timestamped discrete forward/backward/left/right inputs): time on task, path length, and steering reversal rate. Each metric was analyzed separately.

Prior to metric extraction, each trial’s raw command log was segmented to identify the continuous block of activity corresponding to the completed task attempt. Raw logs occasionally contained commands separated from the main sequence of activity by unusually large temporal gaps, reflecting pre-trial calibration presses, aborted attempts, or trials that were restarted after an interruption. Gaps were treated as segment boundaries when they exceeded both an absolute floor (100 s) and a multiple (8 ×) of that trial’s own median inter-command interval, to avoid mis-segmenting genuine within-task pauses (e.g., a participant pausing to assess an obstacle). Among the resulting segments, the final (most recent) segment with a duration not exceeding the protocol’s 15-min time limit plus a 60-s tolerance was retained as the completed task attempt, consistent with the interpretation that earlier segments represent calibration or superseded attempts. Trials for which no segment satisfied this criterion were flagged for manual review rather than automatically trimmed.

Time on task was defined as the duration, in seconds, of the retained segment, with shorter times reflecting better performance. Path length was defined as the count of translational commands (forward and backward) within the retained segment. Similar trajectory-based measures are widely used to quantify navigation efficiency in teleoperation, with shorter trajectories generally reflecting more efficient task execution (Odoh et al., 2024). Steering reversal rate was defined as the proportion of turning commands (left/right) within the retained segment that reversed direction relative to the immediately preceding turning command, consistent with the treatment of frequent directional correction as a marker of degraded control (oversteering/repetitive correction) rather than proficiency in prior teleoperation work. A lower rate reflected smoother, more decisive steering.

To evaluate differences in performance, separate repeated-measures ANOVA were conducted for each of the three metrics with Task (Navigation vs. Obstacle) and Condition (Lunar vs. Control) as the within-subject factors and task presentation order (four levels) as a between-subjects factor. The Bonferroni correction was applied for post-hoc comparisons.

Subjective data

Workload was assessed using the NASA-TLX, which has six subscales: Mental Demand, Physical Demand, Temporal Demand, Performance, Effort, and Frustration. Participants rated each subscale immediately after each task on a 20-point scale. Subscale scores were averaged to yield a raw workload score per trial, ranging from 1 to 20.

The overall NASA-TLX score was analyzed using repeated-measures ANOVA with Greenhouse–Geisser correction for sphericity violations. The effects of Task, Order, and Condition on overall workload were tested.

Neuro-physiological data

All participants wore an ECG sensor (a chest belt; Movesense, Finland) and a Portalite fNIRS sensor (two optodes on the forehead, one over each eyebrow; (Artinis Medical System, Netherlands).

Continuous ECG recordings at a sampling rate of 200 Hz (Kwon et al., 2018) were obtained in 18 subjects under four experimental conditions. Within each subject and condition, recording contained a 2-min rest (baseline) and task recordings of 7–15 min. Raw data were preprocessed by excluding the first and last 15 s of all tasks and baseline recordings to remove artifacts and ensure signal stability. Automatic detection of the R-peaks was performed using a peak-detection algorithm, Kubios (Tarvainen et al., 2014), with adaptive thresholds, manually confirmed, and refined with an interactive graphical interface. RR intervals were calculated as time differences between adjacent R-peaks.

Heart rate (HR) and heart rate variability (HRV), quantified as the standard deviation of RR intervals (SDNN) and the root mean square of successive differences between RR intervals (RMSSD), were computed for both the task and baseline recordings using non-overlapping 60-s windows. Because recording durations varied across participants and conditions, fixed-length windows were used to ensure that all estimates were calculated from the same amount of data. Any remaining segment shorter than 60 s was discarded. For each trial, the final value of each metric was calculated as the mean across all valid windows. Baseline recordings were processed in the same way and were matched to the corresponding task recording collected immediately beforehand.

Baseline recordings were collected while participants were wearing the same equipment used during the corresponding task: full PPE, including the helmet and ventilation system, in the Lunar condition, and minimal protective equipment in the Control condition. As a result, each baseline already included the resting physiological effects of that condition’s equipment. Expressing HR, SDNN, and RMSSD as percentage change relative to baseline therefore removes these resting differences, so the normalized values reflect the additional physiological response to the task itself rather than the combined effects of the task and the protective equipment.

To account for differences in resting physiology between participants, HR, SDNN, and RMSSD were also expressed as percentage change relative to the baseline recording collected immediately before each trial. Both raw and baseline-normalized values were analyzed. The raw values were considered the primary analysis because the repeated-measures design already accounts for stable differences between participants, whereas percentage-change measures are more sensitive to variability in the baseline estimate (Vickers, 2001). SDNN and RMSSD were both computed and reported because they are differentially sensitive over short recording windows: SDNN reflects overall variability, including respiratory sinus arrhythmia, and is comparatively sensitive to posture, movement, and breathing pattern at short durations, whereas RMSSD is the recommended index for short-term recordings and better isolates parasympathetically-mediated variability (Task Force of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology, 1996; Shaffer and Ginsberg, 2017). Reporting both allows the two established but distinct interpretations of short-term HRV to be compared directly. This is particularly relevant here because the baseline recordings were shorter than the task recordings and therefore provided fewer valid 60-s windows. The percentage-change analysis is reported as a supplementary analysis.

A repeated-measures ANOVA, using the JASP software, was used to compare the effects of Condition (Lunar vs. Control), Task (Navigation vs. Obstacle), and Order (1–4) on raw and, secondarily, percentage-change heart rate (HR), and heart rate variability (SDNN and RMSSD).

Continuous fNIRS recordings were acquired at a sampling rate of 100 Hz in 8 participants (S1–S8), under the same four experimental conditions. Only 8 of the 18 participants were recorded with fNIRS because the sensor malfunctioned partway through the data collection campaign and could not be replaced or repaired in time to complete testing with the remaining participants; the failure is presumed to be related to the dusty conditions of the LUNA Hall (see Section 4.5). Concentration changes in oxyhemoglobin (HbO) and deoxyhemoglobin (HbR) were directly exported from the Artinis OxySoft software. Consistent with the ECG analysis, the first two minutes of each recording were designated as the baseline period. All subsequent data processing was performed using Python.

First, a short-channel regression method was applied to eliminate systemic physiological artifacts. Following this correction, the first 10 s and the last 20 s of each recording were removed to avoid edge artifacts and periods of subject movement. The corrected signals were then processed with a low-pass Finite Impulse Response (FIR) filter designed with a Hamming window, a cut-off frequency of 0.4 Hz, and a 0.2 Hz transition band. For each participant and condition, signals from the three long-separation channels of each optode were averaged, and the resulting left and right hemispheric signals were further averaged into a single global prefrontal signal.

For each trial, changes in HbO and HbR concentration were calculated as the mean concentration during the task period minus the mean concentration during the baseline period. The primary analysis included all eight participants with complete datasets. Changes in HbO and HbR were analyzed using a repeated-measures ANOVA with Condition (Lunar, Control) and Task (Navigation, Obstacle) as within-participant factors.

Signal quality was evaluated descriptively for each participant and optode using the tissue saturation index (TSI) fit factor together with the cardiac signal-to-noise ratio of the long-separation channels. All optodes demonstrated adequate coupling (TSI fit factor ≥ 98.4 %), and no participant met the predefined quality criterion for exclusion.

Because increases in HbO without a corresponding decrease in HbR may reflect extracerebral or systemic physiological contamination rather than cortical activation (Tachtsidis and Scholkmann, 2016; Kirilina et al., 2012), three sensitivity analyses were performed. First, the short-channel correction was repeated using two more stringent approaches: regression using both short-separation channels simultaneously and the correlation-based signal improvement (CBSI) method (Cui et al., 2010), which removes components common to HbO and HbR under the assumption that the two signals should be negatively correlated. Second, heart rate, obtained from the simultaneously recorded ECG, was included as a measure of systemic physiology. The correlation between the change in heart rate and the change in HbO was calculated across all trials, and the repeated-measures ANOVA was repeated with the change in heart rate included as a covariate. Finally, the power spectrum of the HbO signal from each trial was decomposed into physiologically relevant frequency bands (very-low-frequency/thermoregulatory, Mayer-wave, respiratory, and cardiac). The correlation between the power of each frequency band in the long-separation channel and the corresponding short-separation channel was calculated to estimate the extent to which each band reflected systemic rather than cortical activity (Caldwell et al., 2016). To assess the robustness of the findings to violations of parametric assumptions, non-parametric analyses were performed at each of the three decorrelation levels. The main effect of Condition (paired Lunar vs. Control) was evaluated using a Wilcoxon signed-rank test, while the Condition × Task interaction was assessed using a permutation test based on 10,000 sign-flips of the participant-level interaction contrast. These analyses were conducted for both the primary mean concentration change measure and the HbO signal-variability measure to confirm that the parametric results were not driven by distributional assumptions in the small sample.

Sensitivity analysis: experimenter-participant exclusion

Three of the 18 participants were members of the experimenter team rather than European Astronaut Centre interns and were therefore familiar with both the terrain layout and the experimental protocol (Section 2.4). To determine whether their inclusion influenced the results, a sensitivity analysis was performed. For the NASA-TLX and ECG measures, which included all 18 participants, the 2 × 2 mixed-design ANOVA (Condition × Task as within-participant factors and Order as a between-participant factor) was repeated after excluding the three experimenter-participants (N = 15). To assess whether any changes were specific to these participants or simply reflected the removal of any three participants, the same analysis was repeated for all 18 3 = 816 possible combinations of three participants. This generated a distribution of p -values for each effect. Two measures were then calculated: (i) the percentile rank of the experimenter-participant exclusion within this distribution, and (ii) the fragility rate, defined as the percentage of all possible three-participant exclusions that changed the statistical conclusion (α = 0.05) relative to the full-sample analysis. The same procedure was also applied to the three performance measures. Since the experimenter-participants had already been excluded from the primary performance analysis (Section 3), only the remaining 13 3 = 286 possible three-participant combinations were considered. All analyses were performed in MATLAB using repeated-measures models fitted with fitrm and evaluated with ranova.

Performance data

A series of repeated-measures ANOVAs were conducted to assess the effects of Task (Navigation or Obstacle), Order, and Condition (Lunar or Control) on performance. For this analysis, 13 of the 18 subjects were considered (S4-S7, S9-S11, S13-S18). Five subjects were excluded from the performance analysis. S1–S3, the three experimenter-participants, were tested on the first day of data collection, before key-press logging had been correctly implemented; their command logs are therefore unusable for the performance metrics reported here. S8 and S12 were excluded because a software crash partway through data collection caused the command log to not be saved for at least one of their four task-condition combinations; specifically, S8’s Control–Navigation trial and S12’s Lunar–Navigation trial were not saved. As these two losses occurred in different Conditions (Control and Lunar) and involved different subjects and Order groups, the missing data show no discernible systematic pattern. Among the 13 subjects retained for this analysis, the between-subjects Order factor was distributed as follows: Order 1, n = 3 (S7, S9, S10); Order 2, n = 4 (S4, S11, S13, S14); Order 3, n = 2 (S5, S6); Order 4, n = 4 (S15, S16, S17, S18), and no trial reached the protocol’s 15-min time limit, meaning no trial was terminated by the experimenters on this basis.

Separate mixed-design ANOVAs (Condition × Task within-subjects, Order between-subjects, N = 13) were conducted for each of the three performance metrics. No significant main effects of Condition, Task, or Order, and no significant two-way interactions, were found for any of the three metrics. A marginal three-way Condition × Task × Order interaction was observed for Path Length (F (3,9) = 3.465, p = 0.064, η p 2 = 0.536, 95 % C I [ 0.000, 0.692 ]) and Steering Reversal Rate (F (3,9) = 2.935, p = 0.092, η p 2 = 0.494, 95 % C I [ 0.000, 0.663 ]), alongside a marginal Condition × Order interaction for Path Length (F (3,9) = 3.552, p = 0.061, η p 2 = 0.542, 95 % C I [ 0.000, 0.696 ]). Given the small per-cell sample size (n = 2–4 per Order group, per the breakdown above), these are reported as exploratory trends rather than confirmed effects (Figure 5).

Subjective data

In order to analyze the perceived workload during the different tasks and conditions, a repeated-measures ANOVA was performed for the overall NASA-TLX composite score with a 2 × 2 (Condition: Lunar vs. Control × Task: Navigation vs. Obstacle) within-subjects design. The presentation order of tasks (4 levels) was a between-subjects factor.

Significant main effects were found for Condition (F (1,14) = 30.321, p < 0.001, η p 2 = 0.684, 95 % C I [ 0.304, 0.811 ]) and Task (F (1,14) = 44.172, p < 0.001, η p 2 = 0.759, 95 % C I [ 0.430, 0.855 ]), with workload rated significantly higher in the Lunar condition than the Control Center condition (mean difference = 2.229, p bonf < 0.001) and significantly higher during the Obstacle Course than during Navigation (mean difference = − 2.940, p bonf < 0.001). A significant Condition × Task interaction was also found (F (1,14) = 6.945, p = 0.020, η p 2 = 0.332, 95 % C I [ 0.007, 0.588 ]). Post-hoc comparisons showed that the Lunar-Obstacle combination produced significantly higher workload than every other condition-task pairing (vs. Lunar-Navigation: p bonf < 0.001; vs. Control-Navigation: p bonf < 0.001; vs. Control-Obstacle: p bonf = 0.001), while Control-Navigation produced significantly lower workload than every other pairing (all p bonf ≤ 0.002). Lunar-Navigation and Control-Obstacle did not differ significantly from one another (p bonf = 0.912). The main effect of Order was not significant (F (3,14) = 0.675, p = 0.582, η p 2 = 0.126, 95 % C I [ 0.000, 0.324 ]), nor was the Condition × Order interaction (F (3,14) = 0.366, p = 0.779, η p 2 = 0.073, 95 % C I [ 0.000, 0.237 ]), the Task × Order interaction (F (3,14) = 0.943, p = 0.446, η p 2 = 0.168, 95 % C I [ 0.000, 0.376 ]), or the three-way Condition × Task × Order interaction (F (3,14) = 1.506, p = 0.256, η p 2 = 0.244, 95 % C I [ 0.000, 0.454 ]) (Figure 6).

Physiological data

Heart rate

Heart rate was analyzed using a mixed-design repeated-measures ANOVA with Condition (Control, Lunar) and Task (Navigation, Obstacle) as within-participant factors and Order as a between-participant factor (N = 18). For the raw HR values, a significant main effect of Condition was observed (F (1,14) = 5.330, p = 0.037, η p 2 = 0.276, 95 % C I [ 0.000, 0.549 ]), with heart rate higher in the Lunar condition than in the Control condition (mean difference = 2.356 bpm, p bonf = 0.037). Neither the main effect of Task (F (1,14) = 0.836, p = 0.376, η p 2 = 0.056, 95 % C I [ 0.000, 0.342 ]) nor the Condition × Task interaction (F (1,14) = 0.281, p = 0.604, η p 2 = 0.020, 95 % C I [ 0.000, 0.275 ]) reached significance.

Several interactions involving Order were significant, including the Condition × Order interaction (F (3,14) = 7.131, p = 0.004, η p 2 = 0.604, 95 % C I [ 0.118, 0.732 ]), the Task × Order interaction (F (3,14) = 4.165, p = 0.026, η p 2 = 0.472, 95 % C I [ 0.000, 0.638 ]), and the three-way Condition × Task × Order interaction (F (3,14) = 4.793, p = 0.017, η p 2 = 0.507, 95 % C I [ 0.022, 0.664 ]). The between-participant main effect of Order was not significant (F (3,14) = 0.005, p = 1.000, η p 2 = 0.001, 95 % C I [ 0.000, 0.000 ]). A CI of [0.000, 0.000] indicates an F-statistic small enough that the data are maximally consistent with a true null effect. Bonferroni-corrected pairwise comparisons were performed to follow up the significant interactions. Only two remained significant after correction: Order 2 Lunar-Obstacle versus Order 2 Control-Obstacle (p bonf = 0.018), and Order 2 Control-Navigation versus Order 2 Lunar-Obstacle (p bonf = 0.011). Given the small number of participants per three-way Order combination (approximately two to three), these results should be interpreted cautiously (Figure 7).

The same analysis was repeated using heart rate expressed as percentage change relative to each participant’s baseline. This supplementary analysis was included to facilitate comparison with previous workload studies. In contrast to the raw analysis, the main effect of Condition was no longer significant (F (1,14) = 2.431, p = 0.141, η p 2 = 0.148, 95 % C I [ 0.000, 0.444 ]). Likewise, neither the Task × Order interaction (F (3,14) = 1.018, p = 0.414, η p 2 = 0.179, 95 % C I [ 0.000, 0.388 ]) nor the three-way Condition × Task × Order interaction (F (3,14) = 0.311, p = 0.817, η p 2 = 0.062, 95 % C I [ 0.000, 0.215 ]) remained significant. However, the Condition × Order interaction remained significant (F (3,14) = 4.357, p = 0.023, η p 2 = 0.483, 95 % C I [ 0.005, 0.646 ]), consistent with the raw-value analysis. The differences between the raw and baseline-normalized analyses are discussed in Section 4.

Heart rate variability

Separate repeated-measures ANOVAs were conducted for both SDNN and RMSSD. For the raw values, neither metric showed a significant main effect of Condition, Task, or Order, and no significant two-way or three-way interactions were observed (SDNN: all p ≥ 0.175; RMSSD: all p ≥ 0.317). The percentage-change analyses showed the same pattern, with no significant main effects or interactions for either SDNN (all p ≥ 0.186) or RMSSD (all p ≥ 0.323). Full statistics for every effect, including 95% confidence intervals for η p 2, are reported in the Supplementary Material.

HbO concentration changes

The repeated-measures ANOVA (N = 8) showed no significant main effect of Condition on the change in HbO concentration (F (1,7) = 0.263, p = 0.624, η p 2 = 0.036, 95% CI [0.000, 0.399]), nor a significant main effect of Task (F (1,7) = 0.020, p = 0.892, η p 2 = 0.003, 95% CI [0.000, 0.247]). The interaction between Condition and Task was also not significant (F (1,7) = 2.461, p = 0.161, η p 2 = 0.260, 95% CI [0.000, 0.596]), although the Lunar–Obstacle condition was numerically highest (Figures 8A,C).

This null result was at least partly attributable to systemic rather than cortical physiological signals. Applying progressively more stringent short-channel correction reduced the correlation between the long- and short-separation channels (r = + 0.28 → + 0.10 → + 0.04) and progressively weakened the Condition × Task interaction on HbO signal variability, from the uncorrected signal (F (1,7) = 19.1, p = 0.003), to single short-channel regression (F (1,7) = 10.3, p = 0.015), and finally to CBSI correction (F (1,7) = 4.4, p = 0.075), indicating that the original interaction depended substantially on extracerebral signal contamination (Figure 9).

Heart rate did not appear to explain the observed HbO changes. The change in heart rate was not significantly correlated with the change in HbO before (r = − 0.27, p = 0.14) or after short-channel correction (r = − 0.15, p = 0.41), and including heart rate as a covariate did not materially alter the ANOVA results (Figure 11).

Frequency-domain analysis further suggested that the remaining effects were predominantly physiological rather than neural. The Condition × Task interaction was confined to the respiratory frequency band (F (1,7) = 5.98, p = 0.044), while the very-low-frequency band showed a significant correlation between the long- and short-separation channels (r = 0.41, p = 0.02), consistent with a shared systemic (thermoregulatory or vascular) origin of the low-frequency signal (Figure 10).

HbR concentration changes

The repeated-measures ANOVA (N = 8) showed no significant main effect of Condition on the change in HbR concentration (F (1,7) = 1.276, p = 0.296, η p 2 = 0.154, 95% CI [0.000, 0.522]), nor a significant main effect of Task (F (1,7) = 0.019, p = 0.895, η p 2 = 0.003, 95% CI [0.000, 0.244]) (Figure 8B). The interaction between Condition and Task was also not significant (F (1,7) = 2.861, p = 0.135, η p 2 = 0.290, 95% CI [0.000, 0.614]). Although this interaction did not reach statistical significance, its effect size was not negligible and followed the same pattern as that observed for HbO. The consistent direction of the HbO and HbR interactions suggests that both chromophores were influenced by a common systemic physiological signal rather than reflecting a dissociation between cerebral oxygenation and deoxygenation.

Non-parametric robustness checks

Non-parametric analyses were consistent with the parametric findings. Wilcoxon signed-rank tests on the primary mean concentration change revealed no significant main effect of Condition for either HbO (W = 9.000, p =. 250) or HbR (W = 11.000, p =. 383), consistent with the corresponding parametric results. Similarly, permutation tests based on 10,000 sign-flips of the participant-level Condition × Task contrast found no significant interaction for either chromophore (HbO: observed contrast = 0.612, p =. 165; HbR: observed contrast = − 0.458, p =. 127).

For the HbO signal-variability measure, Wilcoxon signed-rank tests likewise showed no significant main effect of Condition at any decorrelation level (base: W = 12.000, p =. 461; scr2: W = 10.000, p =. 313; cbsi: W = 10.000, p =. 313), again matching the parametric results. In contrast, the permutation tests reproduced the same progressive attenuation of the Condition × Task interaction observed in the parametric analyses. The interaction remained significant under the least stringent correction (base: observed contrast = 0.381, p = 0.007), persisted following dual short-channel regression (scr2: observed contrast = 0.231, p = 0.023), and was no longer significant after CBSI correction (cbsi: observed contrast = 0.153, p = 0.069). This closely paralleled the corresponding parametric results, which likewise showed a progressive attenuation of the interaction from the uncorrected signal (F (1,7) = 19.1, p = 0.003), to single short-channel regression (F (1,7) = 10.3, p = 0.015), and finally to CBSI correction (F (1,7) = 4.4, p = 0.075).

Sensitivity analysis: experimenter-participant exclusion

The NASA-TLX results were unaffected by the inclusion of the three experimenter-participants. Both the main effect of Condition (p = 0.001) and the main effect of Task (p < 0.001) remained significant after their exclusion, and neither effect became non-significant under any of the 816 possible three-participant exclusions (fragility rate = 0%). The Condition × Task interaction also remained significant after excluding the experimenter-participants (p = 0.030), although this effect was less robust, with 34.8% of the possible three-participant exclusions changing the statistical conclusion.

For the physiological measures, the raw HR main effect of Condition no longer reached significance after excluding the experimenter-participants (p = 0.037 to p = 0.128). However, this effect was generally sensitive to participant removal, as 65.1% of all possible three-participant exclusions also changed the statistical conclusion. The experimenter-participant exclusion therefore behaved similarly to many other possible exclusions and did not appear to have a disproportionate influence on this result. The raw HR Condition × Order interaction remained significant after excluding the experimenter-participants (p = 0.004 to p = 0.013) and showed a low fragility rate (1.1%). In contrast, the Task × Order and Condition × Task × Order interactions were less robust, with fragility rates of 56.8% and 27.0%, respectively.

A different pattern was observed for HRV. Although the raw SDNN and RMSSD Condition main effects were not significant in the full sample (SDNN: p = 0.869; RMSSD: p = 0.878), both became significant after excluding the experimenter-participants (SDNN: p = 0.022; RMSSD: p = 0.011). The corresponding Condition × Order interactions also became significant (SDNN: p = 0.004; RMSSD: p = 0.014), together with the SDNN Task × Order interaction (p = 0.006). Unlike the HR results, these changes were relatively uncommon across all possible three-participant exclusions. The experimenter-participant exclusion fell within the 7 t h − 1 7 t h percentile of the corresponding p -value distributions, and the fragility rates ranged from 16.7% to 17.2%, indicating that excluding the experimenter-participants had a greater influence on these HRV effects than excluding most other combinations of three participants.

The three performance measures, for which the experimenter-participants had already been excluded from the primary analysis, showed low fragility rates (0%–27.8%) when the same leave-three-out analysis was applied to the remaining 13 participants. This is consistent with the largely non-significant results reported for these measures.

This study examined how lunar analog constraints affect teleoperation performance, workload, and physiological stress. Task, condition, and order did not significantly affect objective performance, though a small number of marginal higher-order interactions suggest exploratory trends worth further investigation. NASA-TLX scores, by contrast, revealed clearly higher subjective workload in the Lunar condition, especially during the Obstacle task. Physiological data partially supported this pattern: heart rate was elevated in the Lunar condition in the full sample, although this effect did not prove robust to sample composition (Section 3.4). Prefrontal oxygenation showed no significant condition-related effect on the full fNIRS sample, and exploratory analyses suggest the pattern observed is more likely of systemic than neural origin. These findings suggest that while lunar constraints did not measurably degrade objective performance, they consistently elevated perceived workload, with a more mixed picture for physiological strain.

Effects of the environment on performance

This study analyzed the impact a simulated lunar setting (visual limitations, PPE, and simulated lunar lighting) has on teleoperation task performance. No significant main effects of Condition (Lunar vs. Control), Task, or Order were observed for any of the three performance measures. These results do not support H2, which predicted reduced performance under the simulated lunar condition. Similarly, because no significant main effect of Task was found, the first part of H1, which predicted better performance during the Flat Navigation Task, was also not supported. Marginal Condition × Order and Condition × Task × Order interactions were observed for Path Length, together with a marginal Condition × Task × Order interaction for Steering Reversal Rate (Section 3). However, given the small number of participants in each Order group (n = 2 − 4), these findings should be interpreted with caution. Although they may reflect learning or cognitive fatigue effects (Massimino and Sheridan, 1994; Si et al., 2021), the present study was not sufficiently powered to draw firm conclusions, and these trends require confirmation in a larger sample.

Previous studies have suggested that task sequencing may influence performance through learning and transfer effects in analog mission environments (Casler and Cook, 1999), and that task order is an important consideration in complex operational settings (De la Torre et al., 2024). While the present study did not provide evidence for such an effect, these findings support investigating task order in future studies with larger sample sizes.

Subjective workload ratings on the NASA-TLX (Hart and Staveland, 1988) demonstrated greater mental, physical, and emotional demands in the Lunar condition, especially for the Obstacle task. This trial consistently had the highest workload scores on all NASA-TLX subscales, indicating that while performance was not overtly impaired, participants were finding the Lunar Obstacle combination particularly demanding (Ryu and Myung, 2005). This subjective finding, however, does support the second part of our first hypothesis (H1), which correctly predicted that the Obstacle Course would be perceived as having a higher cognitive workload compared to the Flat Navigation task.

Physiological measures added another dimension to these findings. Raw heart rate was significantly higher in the Lunar condition than the Control condition, and several interactions involving Order were also significant (Section 3). After Bonferroni correction, the significant pairwise comparisons were limited to Order 2, specifically between Lunar-Obstacle and Control-Obstacle, and between Control-Navigation and Lunar-Obstacle. This suggests that the simulated lunar environment may increase physiological stress even when differences in task performance are not observed. Neither SDNN nor RMSSD showed a significant main effect of Condition, Task, or Order, nor any significant interaction, in the full sample. This pattern suggests that environmental demands can elicit covert stress reactions that may not be evident in scores alone, although, as detailed in Section 3.4, the raw HR Condition effect itself was not robust to the exclusion of any small subset of participants, while the null HRV effects specifically became significant once the experimenter-participants were excluded, and both results should be interpreted with these caveats in mind.

The prefrontal oxygenation measurements showed a similar pattern to the other physiological variables, with no significant effects of Condition, Task, or their interaction observed for either HbO or HbR. Additional sensitivity analyses suggested that the small condition-related differences in HbO were primarily driven by systemic physiological signals rather than cortical activity. Applying progressively more stringent short-channel correction reduced these effects, while frequency-domain analysis indicated that they were concentrated in the respiratory and very-low-frequency bands, which were shared between the long- and short-separation channels. Together, these findings suggest that the observed HbO differences were more likely to reflect systemic physiological responses than changes in cortical activation. Given the elevated temperature and altered respiratory environment inside the sealed, actively ventilated helmet used in the Lunar condition, such systemic effects are physiologically plausible and are discussed further below.

Previous studies have shown that visual restrictions, high-contrast shadows, and limited mobility can negatively affect lunar teleoperation performance (Cataldo et al., 2013). Similarly, NASA human factors studies have reported that restricted visual fields and altered haptic feedback impair fine motor control during rover driving (Massimino and Sheridan, 1994). Although we observed higher physiological stress under the simulated lunar condition, this was not accompanied by a measurable reduction in task performance. This may reflect the relatively small sample size and short task duration used in the present study rather than the absence of an environmental effect.

In comparable experiments of remote teleoperation under latency on Earth, operators were shown to change their behavior under stress—by, for instance, slowing down (Young and Peschel, 2020), braking more frequently (Jernberg et al., 2024), and reporting higher workload (Zhou et al., 2024). Our findings are in line with these observations, demonstrating that constrained environments change operator strategy and increase stress responses regardless of performance ratings.

Taken together, the results of this study contribute to a growing body of evidence that realistic environmental degraded conditions, in this case, poor lighting, suit encumbrance, and poor visibility, can increase perceived and, to a lesser extent, physiological workload, even without a measurable cost to objective task performance. Although several interactions involving task order were observed, these should be interpreted with caution given the small sample size and exploratory nature of the analysis. Previous studies have reported that latency (Kim et al., 2021; Timman et al., 2023) and task sequencing in analog mission environments (De la Torre et al., 2024) can influence operator performance and workload. While the present study did not provide clear evidence for such effects, they remain an important area for future research using larger participant samples.

A related consideration is that the Control condition did not include communication latency or the physical separation typically associated with Earth-based remote teleoperation (Section 4.4). Previous studies have shown that communication latency changes how operators control a robot, encouraging slower, more deliberate movements instead of continuous control (Timman et al., 2023; Kim et al., 2021; Jernberg et al., 2024). These changes in control strategy could influence measures such as completion time, path length, and steering reversal rate. Because participants in the Control condition operated the robot without latency and from the same location, it may not have fully represented true remote teleoperation. This may partly explain why no significant differences in objective task performance were observed between the Lunar and Control conditions, as the two conditions may have differed less in terms of control strategy than intended.

In summary, the simulated lunar environment increased participants’ perceived workload and, to a lesser extent, their physiological response, without reducing objective task performance. Future studies should include larger participant samples and longer task durations, and investigate the combined effects of environmental constraints, latency, and different forms of operator feedback to better understand their impact on teleoperation performance in operationally relevant settings.

Alignment between performance, subjective, and physiological measurements

One important contribution of the present study is its multimodal assessment of teleoperation performance, combining subjective ratings of workload, objective measures of performance, and physiological data. There was a mismatch between subjective judgment and objective performance on the task: while participants rated the Lunar condition and Obstacle Course task as requiring significantly greater mental and physical effort, subjective ratings were not significantly related to objective measures of performance. This mismatch directly contradicts the first part of our third hypothesis (H3), which predicted that lower performance would be associated with higher subjective workload.

This divergence is normal in high-stakes or novel situations. Initial research has shown that subjective ratings of workload are more likely to represent a more sensitive or predictive measure of perceived challenge, particularly under conditions of complexity or low feedback (Lin and Lukodono, 2025). Measures of performance, on the other hand, may be stable due to compensatory effort, short task duration, or task familiarity (Wickens et al., 2021; Matthews et al., 2019). Our findings follow this trend: participants may have been able to maintain performance at the cost of increased effort and cognitive load, particularly in the most difficult Lunar Obstacle condition.

Physiological measurement provides additional insights, though with the caveats on robustness noted above (Section 3.4). HR increased for the Lunar condition in the full sample, with the significant pairwise differences concentrated at Order 2 rather than broadly across task order. Prefrontal oxygenation showed a broadly similar pattern to HR and HRV. No significant effects of Condition or Task were observed in the full sample, and the exploratory sensitivity analyses described above suggest that the fNIRS signal in this study was influenced predominantly by systemic physiological processes rather than task-evoked cortical activity. Once more, these physiological responses were not tied to performance and are consistent with the interpretation that operators were expending more effort in maintaining consistent output, which matches the results found by Marlatte et al. (2025); Odoh et al. (2024) and Luque-Casado et al. (2016). The remaining components of H3 (that lower performance would be associated with elevated HR, reduced HRV, and elevated prefrontal oxygenation) are similarly not supported: since objective performance did not differ significantly across conditions, there was no performance variation for these physiological measures to track, and none of the raw physiological effects (HR, HRV, HbO, HbR) showed a relationship with performance in the present data.

Robustness of the physiological findings to sample composition

Because three of the 18 participants were members of the experimenter team rather than naive participants, a leave-three-out sensitivity analysis was performed to determine whether they disproportionately influenced the physiological results (Section 3.4).

For heart rate, the significant main effect of Condition was no longer significant after excluding the three experimenter-participants. However, this effect was generally sensitive to participant removal, as most other combinations of three participants produced the same outcome. This suggests that the loss of significance is more likely explained by the small sample size than by the inclusion of the experimenter-participants.

A different pattern was observed for SDNN and RMSSD. Although neither measure showed a significant Condition effect in the full sample, both became significant after excluding the experimenter-participants. This was relatively uncommon compared with excluding other combinations of three participants. One possible explanation is that the experimenter-participants, having prior knowledge of the terrain and experimental protocol, responded differently to the Lunar and Control conditions than the naive participants, reducing the overall difference between conditions.

These findings should be interpreted with caution. The analysis was performed post hoc on a reduced sample (N = 15), was not part of the primary analysis, and further reduced the statistical power of an already small study. Consequently, these results should be considered exploratory and confirmed in future work using a larger sample of entirely naive participants.

Limitations of the study

Several limitations should be acknowledged in this study, that might limit the generalization and interpretation of the findings.

As intended, rover operations were not possible from the actual LUNA Control Room on the first floor due to technical limitations. Due to connectivity issues, only a portion of the terrain could receive strong signals, making remote teleoperation impossible. In the “Control” condition, participants used minimal protective gear and operated from inside the LUNA hall, specifically from the apron area behind a rock wall, as a workaround. This configuration lacked important remote teleoperation features like isolation, latency, and longer-range control fidelity, even though it was intended to replicate control-center conditions. It should also be noted that this workaround introduced a difference in direct visual access to the task environment between the two conditions. In the Lunar condition, participants faced the regolith field directly, whereas in the Control condition the rock wall completely blocked this view. This difference is representative of the operational scenarios the two conditions were intended to simulate, as a surface operator would naturally have direct visual and environmental awareness of the surroundings, whereas a remote Earth-based operator would not. However, participants in both conditions were instructed to rely exclusively on the camera feed to complete the task. Consequently, any additional situational awareness available in the Lunar condition was neither controlled nor measured and cannot be separated from the simultaneous effects of the PPE and lighting manipulations.

The absence of external motion-capture systems, like OptiTrack or LIDAR, limited the ability to track performance accurately. Even though a custom script recorded timestamps and directional commands, it was impossible to reconstruct precise rover trajectories due to the lack of precise localization. Therefore, participants’ trajectories could not be evaluated. Variability was further introduced by slight changes in the initial orientation, which was manually set. Performance metrics were therefore limited to broad proxies such as task completion time, direction changes, and path length, which might not adequately represent the subtleties of teleoperation skill or error. Consequently, it is plausible that changing how performance was measured could have altered the final results. Additionally, key-press logging was not yet functional during the first day of testing, resulting in the exclusion of the three experimenter-participants from the performance analysis (Section 3).

Young Graduate Trainees, interns from the European Space Agency, and a few experimenters themselves made up the sample from which the majority of participants were selected. Although this methodology was suitable for a proof-of-concept investigation, the findings’ wider relevance is constrained by the subject pool’s homogeneity with regard to age, education, and work experience. This demographic introduces a specific selection bias relevant to interpreting the present results: as young graduate trainees, participants likely differ from trained astronauts or professional mission control operators in baseline technological literacy, familiarity with gamepad-style controllers, and stress resilience under operationally demanding conditions, any of which could influence both objective performance and physiological/subjective response to the constraints tested here. A related concern is that the three experimenter-participants also completed the NASA-TLX questionnaire. Because they were involved in designing and running the study, their workload ratings may have differed from those of the naive participants. However, the sensitivity analysis (Section 2.5.4) suggests that this did not influence the overall conclusions. The significant main effects of Condition and Task remained after excluding the experimenter-participants, and these results were unchanged across all 816 possible three-participant exclusions. Although this does not rule out the possibility of bias in the experimenters’ self-ratings, it suggests that the NASA-TLX findings were not driven by their responses. A goal of future research should be to recruit a cohort with characteristics more representative of the intended operators, especially regarding operational experience levels.

Because of the dusty conditions of LUNA Hall, electrical devices such as monitors, joysticks, and sensors were covered with protective plastic. The participants’ sight and job accuracy may have been slightly hampered by these covers’ introduction of visual artifacts such as glare or reflection, especially because of the solar simulator’s backlighting. This unforeseen ergonomic difficulty adds an uncontrollable variable to the experimental setup while simultaneously reflecting limitations seen in the actual world.

Due to connectivity constraints, both conditions were carried out in the same physical location, which diminished ecological validity even though the Control Condition sought to replicate remote operations. This might have reduced the disparities in latency and perception between situations. Another limitation is that communication latency, illuminance (lux), and ambient temperature were not directly measured during the experiment. Although the Lunar and Control conditions were designed to differ in lighting (solar simulator versus overhead hall lighting) and thermal load (full PPE versus minimal gear), these differences were not quantified. As a result, it is not possible to determine how much each factor contributed to the observed workload and physiological responses, or whether variation in these factors within each condition influenced the results. It is important to note that this study did not model secondary mission-driving constraints such as EVA life-support budgets, consumables, or long-duration shift fatigue. These factors materially affect risk and workload. Our results provide a focused view of the environmental impact (regolith and illumination) on the operator; however, future high-fidelity analogs at LUNA should integrate ‘time-on-task’ constraints to more accurately simulate the pressure of real-time lunar operations, along with direct measurements of communication latency, illuminance, and temperature, to better quantify the environmental conditions experienced during lunar teleoperation.

In addition, task-specific events, such as the rover touching or knocking over an obstacle ring, flipping, or becoming stuck, were not systematically recorded during the experiment and therefore could not be analyzed. These measures would have provided a more direct assessment of task performance than the metrics used in this study (completion time, path length, and steering reversal rate). Consequently, the performance results should be interpreted as an indirect rather than a complete measure of teleoperation performance.

As detailed in Section 2.5.3, percentage-change physiological measures reflect the physiological response to the task relative to an already-suited resting baseline, rather than the total physiological cost of each condition. This may partly explain the differences observed between the raw and percentage-change heart rate analyses.

This study was essentially exploratory in nature, being the first teleoperation experiment with human volunteers conducted within the LUNA facility. Statistical power is limited by the relatively small sample size (N = 18, with only 13 included in some analyses), particularly when it comes to identifying interaction effects. The study’s design should be regarded as foundational rather than definitive because it was not designed for testing hypotheses across numerous between-subjects factors.

Lessons learned

Based on our experience of performing the first human-subject teleoperation experiment in the LUNA Analog Facility, we present the following lessons to future researchers who plan on performing such experiments.

Condition changes, operating sensors, managing PPE transitions, and monitoring data collection take time. We recommend at least two practiced experimenters on hand to ensure smooth operation, minimize downtime, and maintain data integrity—especially when time is of the essence.

The LUNA configuration is not easily alterable for last-minute substitutions or repairs. All systems critical to the mission—laptops, ECG sensors, gamepads, ventilation fans, and power supplies—need to be backed up on site. Dust, handling, and battery failure can destroy a session effortlessly. We also suggest regular cleaning during multiday campaigns to prevent cumulative degradation.

While ambient temperature in the LUNA Hall is relatively low, complete PPE under high-power lighting may cause significant heat build-up. This will impact participants’ performance and comfort and also likely introduce artifacts in physiological sensors (e.g., ECG signal attenuation due to sweating).

The solar simulator provides realistic moonlight but creates sharp contrast and deep shadows that can impair camera vision. Video feeds, exposure settings, and display positions must be tested under realistic lighting to ensure reliable teleoperation feedback and avoid data loss.

A radio communications network links personnel in the hall to ESA control room personnel. The system is convenient for coordination, safety, and task progress—especially when the actors are behind occlusions like the apron wall or obstructions.

Implications and future research

These findings from the current proof-of-concept study provide valuable information regarding human performance and experience with lunar-analog teleoperation. While neither environmental constraints such as light nor PPE consistently degraded objective performance, both consistently elevated perceived workload; the corresponding physiological picture was more mixed, with a significant raw-HR effect in the full sample that did not survive sensitivity analysis (Section 3.4), and no significant effect of Condition on prefrontal oxygenation once the full instrumented sample was analyzed and systemic confounds accounted for. This has far-reaching implications for ensuring lunar surface operations are designed with both technical feasibility and human factors-influenced workload management in mind.

From an operational standpoint, task sequencing did not significantly affect objective performance in this study; the small number of marginal interactions observed are exploratory and would need confirmation in a larger sample before informing operational recommendations. What the data do support is that the Lunar environment reliably increases perceived workload, regardless of task order, which is itself relevant to planning EVA activity and robot control from surface or orbital bases.

Our findings also raise important questions about measurement fidelity in analog research. While physiological sensors such as fNIRS and ECG were tantalizing measures of environmental stress, their confound sensitivity (e.g., heat accumulation of PPE, motion) must be independently validated. Similarly, the discrepancy between subjective workload and objective performance suggests that independent real-time telemetry data may not fully represent operator state—highlighting the value of multi-modal monitoring systems.

While this study focused on manual teleoperation to establish a psychophysiological baseline in the LUNA facility, we acknowledge that future Mission Operations will likely employ a hybrid Concept of Operations (ConOps). This would include varying levels of autonomy and supervisory control. Establishing the operator’s workload under manual conditions is a prerequisite for evaluating how much ‘cognitive offloading’ is actually achieved by these alternative ConOps.

In the future, studies need to examine the effects of delayed communication, longer tasks, and more realistic suit simulations to better simulate lunar surface environments. Having larger heterogeneous participant groups and professional backgrounds would also maximize the external validity of the findings. Since teleoperation will play a central part in upcoming Moon missions, designing robust human-robotic systems that are resilient to physical, cognitive, and environmental stressors will be critical.

Finally, the LUNA facility also offers a unique and growing capability for calibrating systems under controlled analog conditions. We urge future research to make full technological use of the hall to advance both astronaut training and system design for future lunar surface missions.