A Human-in-the-Loop Study of Eye-Movement-Based Control for Workload Reduction in Delayed Teleoperation of Ground Vehicles

Abstract

Teleoperated ground vehicles (TGVs) are widely applied in hazardous and dynamic environments, where communication delay and low transparency increase operator workload and reduce control performance. This study explores the cognitive and physiological workload associated with such conditions and evaluates the effectiveness of an eye-movement-based predicted trajectory guidance control (ePTGC) framework in alleviating operator burden. A human-in-the-loop teleoperation experiment was conducted using a 2 × 2 within-subject design, incorporating subjective ratings (NASA-TLX), objective performance metrics from a dual-task paradigm (one-back memory task), and multimodal physiological indicators (ECG and EDA). Results show that delay and low transparency significantly elevated subjective, objective, and physiological workload levels. Compared to direct control (DC), the ePTGC framework significantly reduced workload across all three dimensions, particularly under high-delay conditions, while maintaining or even improving task performance. Notably, ePTGC enabled even lower workload levels under low-delay conditions than the baseline condition. These findings demonstrate the potential of the ePTGC framework to enhance teleoperation stability and reduce operator burden in delay-prone and low-transparency scenarios.

1. Introduction

Teleoperation enables humans to remotely control objects in remote environments that are inaccessible or dangerous [1]. By ensuring operator safety, teleoperation maximizes the application of human cognitive abilities and domain-specific expertise in hazardous or otherwise unreachable areas [2]. It has broad application prospects in reconnaissance, space exploration, and telemedical robotics [3,4].

Despite advances in AI-driven autonomous systems, challenges like navigating unknown terrains, responding to unexpected events, and maintaining safety and robustness hinder full autonomy, especially in high-reliability scenarios [5,6]. Here, teleoperated ground vehicles (TGVs) offer a cost-effective and safe solution, leveraging human cognition for tasks such as reconnaissance, route clearance, surveillance, and rescue [7], where machine intelligence remains insufficient [8].

However, the development of TGVs still faces two significant challenges, i.e., communication delays and low transparency [9,10,11]. A typical schematic of TGVs is shown in Figure 1. TGVs operate in complex, hazardous environments where communication delays are common due to long-distance wireless transmission and terrain interference [12]. During operation, driver commands $\mathit{\mu }\left(t\right)$ and feedback signals $\mathit{s}\left(t\right)$ (e.g., vehicle status, video feed) experience control delay $t_{\mathrm{d}1}$ and feedback delay $t_{\mathrm{d}2}$, resulting in a round-trip delay of $t_{\mathrm{d}1}+{ t}_{\mathrm{d}2}$. This delay causes temporal misalignment between operator input and vehicle response. While short delays can be mitigated through human prediction [13], increased delay disrupts the perceived input–response relationship, raising cognitive load and impairing control adaptability [14,15]. Performance drops notably when video delay exceeds 200 ms [16,17].

Another major challenge in TGVs is insufficient transparency, i.e., the extent to which operators can perceive environmental information compared to direct interaction [10]. In TGVs, video feedback is the primary source of situational awareness. However, its limitations hinder timely and accurate operator responses to dynamic changes [18,19]. While haptic feedback can enhance transparency in robotic teleoperation, it is less effective for TGVs operating at high speeds due to delay and limited information bandwidth [20]. Moreover, the potential benefits of haptic feedback in such systems are limited by issues such as delay amplification and operator information overload. These limitations not only undermine its ability to enhance transparency but may also disrupt operational stability.

To mitigate the negative impact of delay and low transparency on TGV teleoperation, our previous work [21,22] introduced an eye-movement-based predicted trajectory guidance control (ePTGC) framework to alleviate delay issues. Moreover, a Preview-LPV controller is extended and designed to ensure the robustness of the ePTGC framework under low-transparency conditions. Experimental evaluations demonstrated improvements in system-level performance metrics such as trajectory tracking accuracy and task completion time, particularly under high-delay and low-transparency conditions.

However, a critical limitation of the previous studies is the lack of investigation into how the ePTGC framework affects operator workload—especially under high task demands and time pressure. While prior evaluations [21,22] emphasized control performance and physical steering effort (SE), they did not consider how the framework influences the human operator’s cognitive and physiological workload.

In contrast to our earlier work, the current study adopts a human-centered evaluation perspective to systematically examine the impact of the ePTGC framework on operator workload. The specific contributions of this study are twofold:

(1)

It aims to further investigate the impact of delay and transparency on operator workload during TGV teleoperation tasks using a multimodal evaluation approach, including subjective ratings, task performance, and physiological signals (e.g., ECG, GSR).

(2)

It seeks to quantitatively assess the proposed ePTGC framework’s effect on operator workload and overall teleoperation task performance.

2. Related Work

TGVs are often used in high-risk environments where operators face significant cognitive and physical demands. Unlike natural driving, teleoperation is challenged by communication delays and low transparency, leading to increased workload. Sustained high workloads can degrade performance and endanger operator health [23], making workload management essential. Standard assessment methods include subjective, task performance, and physiological measures [24].

Subjective workload assessments rely on participants’ self-reports and are typically divided into unidimensional and multidimensional scales [25]. Unidimensional tools, such as the modified Cooper–Harper scale, focus on a single factor, whereas multidimensional tools, including NASA-TLX and SWAT, are designed to capture multiple dimensions. NASA-TLX evaluates six workload components (e.g., mental, physical, and temporal demand), while SWAT assesses time demand, mental effort, and stress.

Performance-based methods use a concurrent secondary task (e.g., N-back tests) to infer cognitive workload from changes in task performance [26].

Physiological measurement methods assess cognitive workload by monitoring various physiological indicators, which are generally divided into two categories: brain-activity-based and non-brain-activity-based measures. Brain-activity-based approaches include electroencephalography (EEG), near-infrared spectroscopy (NIRS), and functional magnetic resonance imaging (fMRI), which are widely used for cognitive state classification [27]. Non-brain-activity-based measures include electrocardiography (ECG) (e.g., heart rate (HR), heart rate variability (HRV)), skin conductance (SC), respiratory metrics (e.g., breathing rate, ventilation), and ocular metrics (e.g., blink frequency, pupil diameter) [28]. These physiological signals can effectively reflect an individual’s cognitive load and psychological state. For example, Lantieri et al. [29] used EEG and a mobile eye-tracking system to assess drivers’ mental workload in real-world driving with Adaptive Cruise Control (ACC). Their approach combined subjective, EEG, eye movement, and performance data to reveal workload variations under different conditions. However, unlike passive observation studies, our framework actively incorporates real-time eye movement data as control inputs, where operators intentionally use eye gaze to express their driving intentions. This active involvement of eye behavior introduces intentional modulation, making eye-tracking data less reliable as an objective workload indicator in our context.

2.1. Workload Assessment Research in Teleoperated Robotics

With the growing use of teleoperated robots, numerous studies have explored workload assessment in these systems. Timman et al. [30] demonstrated that communication delays in a lunar sampling task increased operator workload, using NASA-TLX and heart rate variability (HRV) metrics. Zhang et al. [31] analyzed workload using eye-tracking (SRT), NASA-TLX, and task performance, revealing that head-mounted displays (HMDs) influence workload over time. Nenna et al. [32] compared button-based and gesture-based VR control interfaces, finding that interface complexity impacts workload and performance, especially under dual-task conditions. Seo et al. [33] assessed time delay effects in lunar construction tasks using NASA-TLX, eye-tracking, and electrodermal activity (EDA), highlighting the relationship between delay and cognitive load. Dybvik et al. [34] compared workload under delayed teleoperation with and without predictive display support, finding that while predictive displays reduced workload, their effectiveness was limited. Guo et al. [35] showed that eye-tracking metrics (fixation duration, saccade rate, pupil diameter) reliably reflected workload under time pressure and delay in space-based teleoperation. In a follow-up, Seo et al. [36] further examined how delay influences control strategies and workload in lunar excavation, integrating eye-gaze data and NASA-TLX.

Most existing studies on teleoperated space robots or robotic manipulators have focused on low-speed, relatively static task scenarios, such as space assembly or lunar sampling. However, these tasks differ from the operational conditions of teleoperated ground vehicles (TGVs), which typically require higher mobility and must navigate complex, dynamic environments. This fundamental difference limits the direct applicability of existing research findings to TGV teleoperation. Consequently, further research and technical adaptations are necessary to address the unique challenges of high-dynamic, complex environments encountered in TGVs applications.

2.2. Workload Assessment in Teleoperated Ground Vehicle Tasks

Recent research has increasingly addressed the unique challenges of assessing operator workload in teleoperated ground vehicle tasks, an area previously dominated by methods developed for space and manipulator teleoperation. For example, Zheng et al. [20] introduced steering effort (SE) as an indirect workload metric in trajectory tracking, demonstrating its sensitivity to communication delay and compensation strategies. Lim et al. [37] extended this by showing that increased delay leads operators to adopt more active control strategies, with SE peaking at 900 ms due to vehicle instability. However, individual driving styles can influence such indirect metrics, and their correlation with actual workload remains to be validated. To provide more robust assessments, Lu et al. [38] examined the impact of delay on operator workload in TGV tasks and proposed a model predictive delay compensation method. Through human-in-the-loop experiments, they combined subjective workload assessments (NASA-TLX), task performance measures, and physiological indicators (e.g., pupil diameter) to evaluate the effectiveness of the proposed approach under different delay conditions. Papaioannou et al. [39] used the NASA-TLX to study the effects of time delay on performance, emotional state, cognitive workload, and perceived system usability. They found that delay increased frustration and reduced usability and task success, with delay exerting a more substantial negative impact than task complexity and showing additive interaction effects.

In a related study, Musicant et al. [40] simulated a driving environment and found that delays up to 250 ms increased perceived workload. To support real-time workload estimation, Luo et al. [41] proposed a Bayesian inference model based on physiological signals such as eye gaze and pupil dilation, achieving high accuracy (F1 = 0.823) and outperforming traditional single-model approaches. Additionally, Musicant et al. [42] investigated the influence of camera viewpoint height on teleoperation performance, showing that an elevated (roof-mounted) view reduced perceived workload and frustration compared to a driver-eye view. However, neither study addressed the combined effects of delay and interface transparency on operator workload, which are two critical factors in the teleoperation of TGVs.

2.3. Summary

Existing studies on workload assessment in teleoperated ground vehicle (TGV) tasks primarily rely on subjective measures (e.g., NASA-TLX) or single physiological indicators (e.g., pupil diameter). However, physiological signals are susceptible to individual variability, and single-modality indicators often fail to capture the operator’s cognitive and emotional state fully. Moreover, current research tends to focus narrowly on the effects of communication delay and compensation strategies while neglecting the impact of low transparency—such as insufficient feedback or inconsistent task responses in dynamic environments—on operator workload.

To enhance assessment accuracy and ecological validity, research should integrate multimodal physiological data (e.g., ECG, EDA) with subjective ratings and performance metrics, enabling a multidimensional evaluation of operator workload. Furthermore, experimental designs should reflect real-world complexity by jointly considering delay and transparency, enabling a deeper and more comprehensive understanding of how multiple interacting factors contribute to workload and supporting the development of more effective mitigation strategies.

3. Eye-Movement-Based Predicted Trajectory Guidance Control Framework

The ePTGC framework is illustrated in Figure 2. The framework has two modules: trajectory prediction and trajectory tracking.

In the trajectory prediction module, the teleoperator’s commands $\mathit{\mu }\left(t\right)$, the eye movement ${\mathit{P}}_{AOI}\left(t\right)$, the contextual constraints $\mathit{c}(t-t_{d2})$ and the vehicle states $\mathit{s}(t-t_{d2})$ are employed to predict the teleoperator’s intended trajectory, denoted as $\mathit{T}\mathit{r}\mathit{a}\left(t-t_{d2}\right)$. This predicted trajectory is transmitted to the vehicle side after a delay $t_{d1}$ and denoted as $\mathit{T}\mathit{r}\mathit{a}\left(t-t_d\right)$, here ${t_d=t}_{d1}+t_{d2}$. After truncating the initial $t_d$ segment, the trajectory is aligned with $\mathit{s}\left(t\right)$ and used to guide the vehicle’s future maneuvers. Unlike traditional direct control (DC), where operator inputs directly control the vehicle, ePTGC treats control commands as features for trajectory prediction. This decouples the operator from the control loop, mitigating over- and mis-maneuvering caused by delays. Moreover, the operator’s eye movement data provides predictive and accurate information about future control intentions. By integrating eye movement into the ePTGC framework, the system enables early perception and guidance of upcoming control demands, tightly coupling the operator’s visual attention with vehicle operation. This allows operators to make decisions based solely on the current state information received from the vehicle, without needing to account for delay-induced response discrepancies, thereby reducing their cognitive workload.

In the trajectory tracking module, a Preview-LPV (Linear Parameter-Varying) controller is designed to address transparency-related challenges. This controller integrates a preview-based feedforward gain and adaptive regulation based on external conditions, ensuring that vehicle responses remain consistent with operator intentions during low-transparency teleoperation tasks. It improves vehicle response consistency and enhances control stability under uncertainty.

Previous studies have commonly employed techniques such as predictive displays or Smith predictors to compensate for communication delays in teleoperation tasks [7,43]. Predictive displays assist operators by visually extrapolating future vehicle states, enhancing situation awareness without modifying the underlying control dynamics [44]. However, they typically require real-time rendering and are more suitable for low-speed or visually simple environments. In contrast, the Smith predictor depends on accurate modeling of system dynamics and delay characteristics to reconstruct a delay-free response [45], which limits its applicability in complex or uncertain scenarios.

The proposed ePTGC framework differs from these approaches in that it offers direct trajectory guidance based on predicted operator intent, which is inferred from multimodal cues such as gaze behavior and used to generate reference trajectories for control. This intent-aware mechanism enables a model-free and human-centric solution that actively shapes the control process. These distinctions highlight the practical advantages of ePTGC in uncertain and dynamically changing environments. Overall, ePTGC addresses communication delays and low transparency in teleoperation, enabling more intuitive and stable control.

Human-in-the-loop experiments demonstrate that the framework enhances both operator efficiency and vehicle stability under challenging conditions [21,22]. However, its impact on operator workload in delayed and low-transparency scenarios remains underexplored and warrants further investigation.

4. Methodology

4.1. Apparatus

The experiment was conducted on a human-in-the-loop teleoperation simulation platform, as illustrated in Figure 3. The platform comprises four main subsystems: the driving station, the vehicle–road system, the controller, and the communication network module.

At the driving station, a Logitech G27 steering kit provides steering, throttle, and brake inputs. Visual feedback, delayed to simulate communication delay, is displayed on a monitor equipped with a Tobii eye-tracker to record the operator’s gaze behavior. Eye movement data, vehicle states, and 3D point clouds are integrated into the trajectory prediction module to generate the operator’s intended trajectory.

The vehicle–road system is built using CARLA and TruckSim. CARLA handles visual rendering and sensor simulation (e.g., RGB camera, LiDAR), while TruckSim simulates vehicle dynamics via CAN-based control input, updates vehicle states in real time, and models road conditions with varying friction levels.

Communication delay is emulated in ROS using a FIFO (First-In–First-Out) buffer. Video and vehicle state data are enqueued and dequeued according to the buffer length, allowing for adjustable delay simulation.

While the platform simulates only the vehicle–road dynamics and network latency, the human–machine interface closely mirrors real-world setups. This allows for accurate reproduction of operator responses to communication delays and sudden environmental changes.

The operator was equipped with physiological sensors to collect biosignals during the teleoperation task, as illustrated in Figure 4. Electrodermal activity (EDA) and electrocardiogram (ECG) signals were recorded using the MP150 physiological data acquisition system manufactured by Biopac Systems Inc, with a sampling rate of 2000 Hz. The corresponding sensors are also shown in Figure 4a,b. The one-back voice interaction task was implemented using PsychoPy (v2022.2.2). Audio files containing digit stimuli (0–9) were imported and configured to be randomly presented at one digit every 2 s, as shown in Figure 4c.

4.2. Participants

A total of 22 volunteers participated in the experiment, with a mean age of 24.9 years (SD = 2.9). All subjects held a valid driver’s license and had at least one year of practical driving experience. The participants were in good health, with normal or corrected vision and hearing, no history of neurological or cardiovascular diseases, and met the basic physiological stability requirements for the experiment. During the experiment, one participant could not complete the task after several attempts under high-delay conditions and voluntarily withdrew from the study. This participant did not complete the entire experimental process, so his data were excluded from the final analysis. The final valid sample size was 21.

4.3. Experimental Design

The experiment was conducted in a 2 × 2 within-subjects design to determine the effect of delay level and control method on the operator’s workload. The control method consists of two levels, i.e., {DC, ePTGC}, where DC refers to a direct control framework without any aids. To ensure the smooth implementation of the experiment and to capture a clear contrast between high- and low-delay conditions, the delay levels were selected based on pilot study results. Specifically, 200 ms (low delay) and 600 ms (high delay) were chosen as the experimental conditions, as they represent threshold values at which the impact of delay and transparency on teleoperation performance can be clearly observed. In addition, these delay values align with typical ranges encountered in communication networks [34,46,47] and are consistent with those reported in previous delay-related studies [37,38,48].

Using the results from the DC method under zero-delay and high-adhesion road conditions as a baseline, each participant completed five teleoperation driving tasks. To minimize the influence of specific scenarios or path memorization, the order of the tasks was randomized across subjects.

4.3.1. Test Road

As depicted in Figure 5, the experimental test route consists of several straight-line segments alternating with curved segments. The straight-line segments primarily evaluate the teleoperated system’s steady-state performance under constant input conditions, whereas the curved segments are designed to assess the system’s transient response under step or ramp input conditions. The total length of the test route is 588 m, including six sharp curves with radii ranging from 10 to 14 m and one gentle curve with a radius of 44 m. These variations in curvature introduce differing levels of maneuvering difficulty to increase task complexity. This design facilitates a more comprehensive elicitation of participants’ physiological responses and enables effective observation of both fluctuation patterns and recovery dynamics in physiological indicators.

The road surfaces were differentiated based on their adhesion coefficients to further investigate the operator’s workload in response to sudden environmental changes under low-transparency conditions. A higher adhesion coefficient (μ = 0.85) was applied to the straight segments to simulate typical dry road conditions, while a lower adhesion coefficient (μ = 0.25) was applied to the curved segments to simulate low-adhesion wet and slippery road conditions. The corresponding regions are highlighted with dashed ellipses in Figure 5. Participants were required to complete one lap along the centerline of the predefined trajectory, following task instructions. The driving direction could be either clockwise or counterclockwise.

4.3.2. Experimental Procedure

The entire experimental process was divided into two phases: the training phase and the testing phase. The training phase aimed to familiarize participants with the operational procedures of the human-in-the-loop teleoperation platform and to provide them with experience operating under different delay and road condition scenarios. Notably, the same test route described in Section 4.3.1 was used during both the training and experimental phases to ensure consistency in participants’ spatial awareness and control experience. During the training, participants were also asked to report any physical discomfort caused by teleoperation under delayed conditions.

Participants signed an informed consent form after completing the preparation and training phases. Next, two SC electrodes were attached to participants’ distal phalanges to obtain the SC signal (Figure 6a; [49]). Three electrodes were placed on the body in the lead II configuration (Figure 6a; [28]) to record the ECG. A five-minute baseline physiological data collection session was conducted to record their initial physiological state. Following this, a teleoperation task was randomly selected to begin the experiment. Participants completed two trials of the primary task, during which physiological data were collected in real-time. After each trial, they filled out the NASA-TLX scale to evaluate their subjective workload. A five-minute rest period followed each trial to allow participants’ physiological indicators to return to baseline levels. Upon completing the two primary trials, participants proceeded to the secondary task phase, during which physiological data were no longer recorded. Instead, they performed a one-back voice interaction task, which was also conducted in two trials. No further subjective workload assessments were conducted in this phase, as the increased cognitive load from the one-back task could potentially bias participants’ self-reports. It should be noted that the one-back verbal interaction task was not designed to induce stress but to provide an objective measure of cognitive workload during teleoperation. Its moderate difficulty level was validated through pilot testing and ensured consistency across experimental conditions. The overall experimental procedure is illustrated in Figure 7.

Upon completion of the experiment, interviews were conducted to obtain qualitative insights into task difficulty, user experience, and potential areas for improvement, thereby enriching the interpretation of the experimental findings.

4.4. Data Preprocessing and Data Description

This study assessed workload in teleoperation tasks from three dimensions: subjective workload, physiological measures, and task performance evaluation.

4.4.1. Subjective Workload-Related Dependent Variables

The subjective workload was measured using the NASA-TLX scale. The subjective workload-related dependent variables included six dimensions: Mental Demand, Physical Demand, Temporal Demand, Performance, Effort, and Frustration. In addition, an overall workload score (Score) was calculated based on an equal-weight principle, providing a comprehensive indicator of the operator’s subjective workload level.

4.4.2. Task Performance-Related Dependent Variables

Task performance was evaluated using a dual-task experimental design. The primary task consisted of a teleoperation path-tracking task, with performance assessed using three key metrics: deviation-to-centerline (D2C), heading-to-centerline (H2C), and task completion time (TCT), which reflected the control stability and maneuverability of the teleoperation system. Specifically, D2C quantifies the lateral tracking performance by calculating the enclosed area between the vehicle’s actual trajectory and the reference centerline. H2C, as a stability-related metric, represents the cumulative heading discrepancy between the vehicle orientation and the road direction over the entire path, capturing fluctuations caused by unstable control. TCT measures the time required to complete an entire loop of the driving task, reflecting longitudinal efficiency.

The secondary task was a one-back voice interaction task, where the operator workload during concurrent task execution was objectively evaluated through the Accuracy and Response Time time of the voice interaction task.

4.4.3. Physiological Measure Related Dependent Variables

In this subsection, we first preprocess the raw data (i.e., EDA and ECG) to remove noise and outliers and then perform standardization. Finally, feature extraction is performed on the preprocessed data.

The raw SC signal was down-sampled to 25 Hz and smoothed by a median smoothing filter with a 1 s width to eliminate movement artifacts [50]. To eliminate individual differences in EDA signals, they were normalized according to the following equation (Standardized EDA = (EDA value—minimum value)/(maximum value—minimum value)). The maximum and minimum values of EDA refer to the minimum value of the subject’s resting baseline and the maximum value of all task phases, respectively. The phasic component of the EDA was extracted using a 0.05 Hz high-pass filter from the standardized EDA [28]. The ECG signal was down-sampled to 200 Hz, and then a 0.5–35 Hz bandpass filter was used to remove the noise [51].

A total of 21 physiological features were extracted from the preprocessed data—11 ECG-related and 10 EDA-related. However, these features may be influenced by individual differences and environmental factors, and not all features are equally effective in capturing the overall workload state. To address this, a multi-stage feature selection strategy was developed and implemented to identify the most predictive physiological indicators of workload, enhancing the performance and interpretability of subsequent statistical modeling.

In the first stage, LASSO (Least Absolute Shrinkage and Selection Operator) regression was used to filter the initial physiological feature set. By introducing an L1-norm penalty, LASSO enables automatic variable selection, removing redundant and noisy features while retaining those most relevant to subjective workload ratings.

A five-fold cross-validation was conducted to determine the optimal regularization parameter λ. As shown in Figure 8, λ_min (0.14) corresponds to the minimum cross-validation error and retains 11 non-zero features, while λ_1se (0.53) yields a sparser model with only 3 features. Although the λ_1se model is simpler, it may exclude important variables. Therefore, the λ_min model was selected, and 12 non-zero coefficient features (see Figure 8) were used for subsequent analysis.

Although LASSO regression is effective for feature selection, it is sensitive to multicollinearity and only captures linear relationships. To address this, a second-stage analysis using a Random Forest model (n_estimators = 500) was conducted to assess the importance of the LASSO-selected features. As shown in Figure 9, the top eight features account for over 80% of the cumulative importance. Therefore, these eight features (△SCL, SCL_ER_sum, HR_max, HR_mea, ER-SCR_count, HR_med, ER-SCR_max, and SCL_mea; see

Table 1. Detailed physiological profile.

Features	Unit	Clarification
HR_mea	bpm	The average heart rate measured during a single task is usually expressed as beats per minute (bpm). It reflects the overall level of cardiac activity and is regulated by the autonomic nervous system
HR_max	bpm	Highest heart rate value measured during a task period
HR_med	bpm	Median value of heart rate measured during a mission
△SCL	µS (microsiemens)	The range of skin conductance level (SCL) variation during a task session
SCL_mea	µS	Average of SCL during a task session
SCL_ER_sum	µS	The cumulative amplitude of skin conductance level (SCL) responses corresponding to event-related skin conductance responses (ER-SCRs) during a task session
ER-SCR_count		The total count of ER-SCRs during a task session
ER-SCR_max	µS	The peak ER-SCR amplitude observed during the task session.

for details) were selected as the core physiological feature set for subsequent analysis.

In the third stage, principal component analysis (PCA) was applied to the core feature set to reduce dimensionality. As shown in Figure 10, the cumulative variance explained by the first four principal components (PC1–PC4) exceeded 90%, indicating that they preserved most of the data’s information. Beyond these components, the curve leveled off, and additional components contributed little, suggesting they could be disregarded.

As shown in the loading plot (Figure 11), P_EDA (PC1) was primarily composed of EDA-related features (△SCL, SCL_mea, SCL_ER_sum, ER-SCR_count, and ER-SCR_max), all of which had high positive loadings, while ECG-related features contributed minimally and negatively. This suggests that P_EDA predominantly captures variability in electrodermal activity and reflects the main trends in EDA-related physiological processes.

In contrast, P_ECG (PC2) was mainly influenced by ECG-related features such as HR_mea, HR_std, and HR_max, all exhibiting high positive loadings. EDA features contributed less to this component, indicating that P_ECG primarily represents patterns in cardiac activity.

In terms of variance contribution, P_EDA and P_ECG explained 45.27% and 32.80% of the total variance, respectively, reaching a cumulative variance contribution of 78.07%. This indicates that just the first two principal components effectively preserved over 70% of the original information, providing good dimensionality reduction while accurately representing the variation patterns of both EDA and ECG physiological channels. Therefore, the two principal components, P_EDA and P_ECG, were used to represent electrodermal and electrocardiac physiological variations, respectively, in subsequent analyses.

5. Results and Discussion

Table 2. Mean and SE of task performance, subjective workload, and physiological measure.

Tasks	Metrics	N	Baseline	Delay 200 ms		Delay 600 ms
				DC	ePTGC	DC	ePTGC
Primary task	D2C (m2)	205	1033 ± 342	1322 ± 521	1109 ± 312	2305 ± 732	1259 ± 297
	H2C (rad × m)	205	65.8 ± 15.8	77.8 ± 22.2	80.8 ± 16.2	116.5 ± 28.8	83.7 ± 11.5
	TCT (s)	205	59.6 ± 6.3	64.1 ± 7.7	58.7 ± 4.0	74.0 ± 9.5	60.0 ± 4.3
Secondary task	Accuracy (%)	205	88.10 ± 9.6	81.35 ± 16.3	91.06 ± 8.8	72.14 ± 15.2	91.38 ± 9.4
	Response Time (s)	205	1.52 ± 0.12	1.53 ± 0.14	1.42 ± 0.17	1.59 ± 0.10	1.46 ± 0.13
Subjective workload	Mental Demand	206	10.3 ± 3.8	13.2 ± 3.4	8.9 ± 3.2	16.0 ± 2.9	10.8 ± 3.1
	Physical Demand	206	9.4 ± 3.8	12.1 ± 3.3	8.0 ± 3.4	14.6 ± 3.8	9.5 ± 3.1
	Temporal Demand	206	10.2 ± 4.1	13.2 ± 3.4	8.6 ± 3.6	16.3 ± 3.3	10.2 ± 3.5
	Performance	206	9.4 ± 4.4	12.4 ± 3.5	8.9 ± 3.5	15.7 ± 3.4	9.9 ± 3.3
	Effort	206	10.8 ± 4.3	13.4 ± 3.5	8.2 ± 3.5	16.6 ± 3.1	10.3 ± 3.7
	Frustration	206	8.9 ± 4.4	11.9 ± 3.3	7.8 ± 3.0	15.6 ± 3.8	9.3 ± 3.5
	Score	206	9.8 ± 3.9	12.7 ± 3.1	8.4 ± 3.1	15.8 ± 3.0	10.0 ± 3.1
Physiological measure	PEDA	210	0.34 ± 0.20	0.42 ± 0.20	0.31 ± 0.18	0.60 ± 0.17	0.36 ± 0.18
	PECG	210	0.42 ± 0.15	0.47 ± 0.18	0.36 ± 0.19	0.54 ± 0.19	0.38 ± 0.20

summarizes the mean and standard error (SE) values of the primary task performance, secondary task performance, subjective measure of workload, and physiological measure of workload. Following outlier removal via the boxplot method, the resulting sample size (N) for each dependent variable is presented in Table 2.

The Kolmogorov–Smirnov test revealed significant deviations from normality for all task performance and subjective measures, except for physiological indicators (p < 0.05). Furthermore, the assumption of data independence was violated due to the 2 × 2 within-subjects design in which each participant completed all conditions. To address these concerns, a Generalized Linear Mixed Model (GLMM) was employed to analyze each dependent variable, which accounts for within-subject correlations via random effects and does not assume normality or homoscedasticity. The GLMM was specified with a Gamma distribution and a log link function. Fixed effects included Delay Level, Control Method, and their interaction (Delay Level × Control Method), while Participant ID was treated as a random effect to account for individual variability.

To further examine differences between Control Method under each Delay Level, pairwise comparisons were conducted using estimated marginal means (EMMs). Bonferroni correction was applied to control for multiple comparisons, with the adjusted significance threshold set at α = 0.05.

5.1. Task Performance

5.1.1. Primary Task

The fixed effects results from the GLMM analysis for the primary task performance metrics are presented in Appendix B—

Table A2. Results of GLMM fixed effect tests on primary task performance metrics.

Metric	Fixed Factors	DF	F	p
D2C	Delay Level	2	62.7	0.000
	Control Method	1	76.3	0.000
	Delay Level × Control Method	1	33.3	0.000
H2C	Delay Levellevel	2	57.8	0.000
	Control Method	1	24.3	0.000
	Delay Level × Control Method	1	35.2	0.000
TCT	Delay Level	2	29.9	0.000
	Control Method	1	84.4	0.000
	Delay Level × Control Method	1	16.7	0.000

. As shown, Delay Level, Control Method, and their interaction significantly affected all three primary performance metrics: D2C, H2C, and TCT (p < 0.001). Among these, Control Method showed particularly strong effects on TCT (F = 84.4, p < 0.001) and D2C (F = 76.3, p < 0.001).

The results of the pairwise comparisons for the primary task performance metrics are shown in Figure 12. As illustrated, with increasing delay, all three primary performance metrics (D2C, H2C, and TCT) significantly increased under the DC method. At the 200 ms delay, D2C, H2C, and TCT significantly increased compared to no delay (p < 0.05), with a more pronounced increase at 600 ms (p < 0.001), indicating a more substantial negative impact of higher delays on task performance.

Although at a 600 ms delay, the ePTGC method also showed increases in D2C and H2C compared to the no-delay condition, the increase was relatively minor and better than the DC method at the same delay level. This suggests that the ePTGC method effectively maintains control accuracy and stability under high-delay conditions.

Notably, at the low delay level (200 ms), there was no significant difference between the DC and ePTGC methods in H2C, while a significant difference was observed in TCT. This can be attributed to the fact that operators can mitigate the delay’s effect on control accuracy and stability at low delay levels through their adaptability, often sacrificing speed to maintain higher control quality, which explains the significant difference in TCT.

5.1.2. Secondary Task

The fixed effects results from the GLMM analysis for the primary task performance metrics are presented in Appendix B—

Table A3. Results of GLMM fixed effect tests on secondary task performance metrics.

Metric	Fixed Factors	DF	F	p
Accuracy	Delay Level	2	15.3	0.000
	Control Method	1	59.0	0.000
	Delay Level × Control Method	1	6.1	0.016
Response Time	Delay Level	2	5.3	0.007
	Control Method	1	42.9	0.000
	Delay Level ×Control Method	1	0.2	0.657

. As shown, the Accuracy metric was significantly affected by both Delay Level (F = 15.3, p < 0.001) and Control Method (F = 59.0, p < 0.001), with a significant interaction between them (F = 6.1, p = 0.016), indicating that the effect of control method on Accuracy varied across delay levels. For Response Time, there was a significant effect of both Control Method and Delay Level (p < 0.01) but no significant interaction (F = 0.2, p = 0.657). Further pairwise comparison results are shown in Figure 13.

As shown in Figure 13, the Accuracy of the secondary task under the DC method decreased progressively with increasing delay. Compared to the baseline condition, Accuracy under the DC method was significantly lower at 200 ms (p < 0.05) and declined even further at 600 ms (p < 0.001). In contrast, the ePTGC method yielded higher Accuracy than the DC method under both delay levels (p < 0.05), with performance close to the baseline. This suggests that the ePTGC method effectively reduces operators’ objective workload under high-delay and variable road adhesion conditions, enabling performance closer to the ideal state.

For the Response Time metric (Figure 13b), the DC method resulted in significantly longer response times at both 200 ms and 600 ms delays compared to baseline (p < 0.05), mirroring the trend observed in Accuracy. The ePTGC method, however, consistently outperformed the DC method, showing significantly shorter response times under the same delays. Notably, at 200 ms, Response Time under the ePTGC method was even shorter than under baseline conditions (p < 0.05) and remained near baseline at 600 ms. These findings suggest that the ePTGC method effectively alleviates operator workload, with particularly pronounced benefits under low-delay conditions.

Although the ePTGC method did not yield significant improvements in primary task metrics (D2C, H2C) under low-delay conditions, it significantly reduced the objective workload in the secondary task. In conclusion, the ePTGC method effectively improved secondary task Accuracy and reduced Response Time under varying road adhesion and delay conditions, demonstrating strong workload mitigation capabilities.

5.2. Subjective Workload

The fixed effects results from the GLMM analysis for the subjective workload metrics are presented in Appendix B—

Table A4. Fixed effect results of the GLMM for each subjective workload dimension.

Metric	Fixed Factors	DF	F	p
Mental Demand	Delay Level	2	28.1	0.000
	Control Method	1	87.0	0.000
	Delay Level × Control Method	1	0.8	0.362
Physical Demand	Delay Level	2	20.6	0.000
	Control Method	1	75.7	0.000
	Delay Level × Control Method	1	0.99	0.323
Temporal Demand	Delay Level	2	25.9	0.000
	Control Method	1	74.0	0.000
	Delay Level × Control Method	1	1.3	0.252
Performance	Delay Level	2	20.1	0.000
	Control Method	1	57.5	0.000
	Delay Level × Control Method	1	4.0	0.049
Effort	Delay Level	2	23.3	0.000
	Control Method	1	82.9	0.000
	Delay Level × Control Method	1	0.7	0.410
Frustration	Delay Level	2	26.7	0.000
	Control Method	1	75.3	0.000
	Delay Level × Control Method	1	3.5	0.067
Score	Delay Level	2	27.9	0.000
	Control Method	1	90.0	0.000
	Delay Level × Control Method	1	2.1	0.155

.

As shown in the fixed effect results, Delay Level had a significant effect (p < 0.001) on all NASA-TLX dimension scores as well as the overall Score, indicating that higher delay levels led to a significantly increased subjective workload. Similarly, Control Method also showed a significant main effect (p < 0.001) across all dimensions and the overall Score, suggesting that the ePTGC method effectively reduced workload compared to the DC method.

Notably, the interaction between Delay Level and Control Method was significant or marginally significant for Performance (p = 0.049) and Frustration (p = 0.067), implying that the influence of control strategies on these dimensions may differ depending on the delay condition. The results of the pairwise comparisons for the subjective workload metrics are shown in Figure 14.

Pairwise comparisons further confirmed that control methods significantly affected subjective workload under different delay levels. At both 200 ms and 600 ms delays, the DC method resulted in consistently higher scores than the ePTGC method, indicating that ePTGC effectively reduced perceived workload.

5.3. Physiological Measure

The fixed effects results from the GLMM analysis for the physiological measure metrics are presented in Appendix B—

Table A5. Results of GLMM fixed effects for physiological measure metrics.

Metric	Fixed Factors	DF	F	p
PEDA	Delay Level	2	50.59	0.000
	Control Method	1	76.07	0.000
	Delay Level × Control Method	1	9.76	0.002
PECG	Delay Level	2	15.90	0.000
	Control Method	1	121.20	0.000
	Delay Level × Control Method	1	4.08	0.046

.

The physiological workload metric P_EDA was significantly influenced by both delay level (F = 50.59, p < 0.001) and control method (F = 76.07, p < 0.001), with a significant interaction between them (F = 9.76, p < 0.01). Similarly, the P_ECG metric was significantly affected by delay level (F = 15.90, p < 0.001) and control method (F = 121.20, p < 0.001), also showing a significant interaction effect (F = 4.08, p < 0.05). Detailed pairwise comparison results are presented in Figure 15.

As shown in Figure 15, physiological workload levels (P_EDA and P_ECG) under the DC method increased progressively with rising delay. Compared to the baseline, both P_EDA and P_ECG significantly increased at 200 ms delay (p < 0.05), with even more pronounced increases at 600 ms. This trend was especially evident in the P_EDA metrics, where delay-induced differences in EDA-related measures were highly significant (p < 0.001).

In contrast, under the same delay and road conditions, the ePTGC method resulted in significantly lower physiological workload levels than the DC method (p < 0.05) and remained close to or even better than the baseline.

5.4. Overall Workload Analysis

To quantitatively assess the improvement of the ePTGC method over the DC method across different workload dimensions, secondary task performance and physiological measure metrics were first normalized and then separately integrated into two distinct composite metrics. The Score metric directly represented subjective workload, reflecting the operator’s perceived workload.

The Accuracy metric was reverse-normalized using (1), while the Response Time, P_EDA and P_ECG metrics were normalized using (2) and (3), respectively. Then, the two secondary task performance metrics were combined using (4) to yield the overall objective workload metric P_obj, and the two physiological workload metrics were combined using (5) to yield the overall physiological workload metric P_phy.

$${\mathit{P}}_{accu}={\displaystyle \frac{\left({r_0^a-r}_{accu}\right)}{r_0^a}}+1,$$

(1)

$${\mathit{P}}_{res}={\displaystyle \frac{\left(r_t-r_0^t\right)}{r_0^t}}+1,$$

(2)

$$P_{EDA/ECG}^n={\displaystyle \frac{\left(r_{EDA/ECG}-r_0^{EDA/ECG}\right)}{r_0^{EDA/ECG}}}+1,$$

(3)

$$P_{obj}=({\mathit{P}}_{accu}+{\mathit{P}}_{res})/2,$$

(4)

$$P_{phy}=(P_{EDA}^n+P_{ECG}^n)/2,$$

(5)

where $r_{accu}$ and $r_t$ are the Accuracy and Response Time metrics before normalization, respectively; $P_{EDA/ECG}^n$ and $r_{EDA/ECG}$ represent the P_EDA and P_ECG metrics before and after normalization, respectively; $r_0^a$, $r_0^t$ and $r_0^{EDA/ECG}$ are the mean values of the Accuracy, Response Time, P_EDA and P_ECG metrics for the baseline, respectively.

After processing, three workload metrics were obtained: Score (representing subjective workload), P_obj (representing objective workload), and P_phy (representing physiological workload). Then, the improvement of the ePTGC method over the DC method in each dimension was calculated using the following equation.

$$P_{S/obj/phy}^{c,DC}=-{\displaystyle \frac{\left|P_{S/obj/phy}^c-P_{S/obj/phy}^{DC}\right|}{\left|P_{S/obj/phy}^{DC}-P_{S/obj/phy}^{zero}\right|}}\times 100\%,$$

(6)

where $P_{S/\mathrm{obj}/\mathrm{phy}}^{\mathrm{c}}$, $P_{S/\mathrm{obj}/\mathrm{phy}}^{\mathrm{D}\mathrm{C}}$ and $P_{S/\mathrm{obj}/\mathrm{phy}}^{\mathrm{z}\mathrm{e}\mathrm{r}\mathrm{o}}$ are the Score/P_obj/P_phy metrics using the ePTGC, DC method, and the baseline condition, respectively.

The pairwise comparison results for the Score, P_obj, and P_phy metrics are shown in Figure 16, which also illustrates the improvement of the ePTGC method over the DC method.

Figure 16a,c show that under delay and variable adhesion conditions, the DC method resulted in significantly higher subjective and physiological workload levels compared to baseline, highlighting the impact of delay and transparency on teleoperation workload. In contrast, at low delay levels, the ePTGC method reduced subjective workload, objective workload, and physiological workload by 151.5%, 218.9%, and 157.9%, respectively, compared to DC, with reductions exceeding 100%, indicating a workload level better than baseline.

At high delay levels, ePTGC reduced subjective, objective, and physiological workload by 97.6%, 135.2%, and 101.5%, respectively, compared to DC, showing improvements close to baseline levels. These results suggest that ePTGC effectively reduced operator workload by replacing part of the manual steering with an eye-tracking guidance mechanism, especially achieving better-than-baseline performance at low delay levels and maintaining near-baseline workload levels at high delay levels.

Additionally, while ePTGC did not significantly improve primary task performance (D2C, H2C) at low delay levels, it significantly reduced operator workload. At high delay levels (600 ms), ePTGC not only improved primary task performance but also significantly lowered operator workload to near baseline levels.

It is worth noting that although the present study used fixed delays for experimental control, previous research [20] has shown no significant difference in operator performance between fixed and variable delay conditions. Nonetheless, future work should assess the robustness of the ePTGC framework under variable or unpredictable delays to further validate its applicability in real-world teleoperation scenarios.

6. Conclusions

This study evaluated the impact of communication delay and low transparency on operator workload in teleoperated ground vehicle tasks and demonstrated the effectiveness of the eye-movement-based predicted trajectory guidance control (ePTGC) framework. Experimental results showed that, compared with the traditional direct control method, the ePTGC framework significantly reduced operator workload across all dimensions. Under high-delay conditions (600 ms), subjective, objective, and physiological workloads were reduced by 97.6%, 135.2%, and 101.5%, respectively. Under low-delay conditions (200 ms), the reductions were even more substantial, reaching 151.5%, 218.9%, and 157.9%. These findings confirm that the ePTGC framework effectively mitigates the adverse effects of delay and low transparency, enhances control stability, and provides a promising solution for improving operator performance and reducing workload in real-world teleoperation scenarios.

While this study focused on evaluating the effectiveness of the ePTGC framework, we acknowledge that factors such as age may influence operator performance due to differences in reaction time and cognitive capacity. However, our previous findings [22] suggest that ePTGC can mitigate performance variability across different experience levels, and potentially, age groups. Future studies should include participants with a wider age range to further assess the robustness and generalizability of the framework.

In addition, future work will also focus on validating the ePTGC framework in real or semi-realistic teleoperation scenarios, to further examine its effectiveness and adaptability under real-world operational conditions. Moreover, the integration of a delay adaptation mechanism, such as real-time delay estimation and adaptive control adjustment, will be explored to enhance the framework’s robustness under dynamically changing network conditions.