Within a total of 13 sessions in pairs of two, 26 able-bodied participants [age: 23.8 (mean) ± 2.2 (SD) years; age range: 20 to 28 years; 13 male, 13 female] underwent a car driving task employing varying vehicle operation conditions. Of all participants, each reported normal or corrected-to-normal vision, and 15 confirmed prior participation in an EEG study. We assessed 23 persons as right-handed, and three as left-handed, according to the Edinburgh Handedness Inventory (Oldfield, 1971). Each participant gave their written informed consent and received compensatory payment for their time. All measurements were conducted at the Graz University of Technology and approved by the local ethics committee.

During each session of recordings, the EEG and electrooculogram (EOG) of two participants – each of which operated as an Observer and as an Executor once - were recorded simultaneously on 64 active electrodes per person (actiCAP, Brain Products GmbH, Gilching, Germany). The total setup corresponded to a 60-channel EEG according to a modified 10–10 system layout (Supplementary Figure S1), as well as a 4-channel EOG, both recorded at a sampling rate of 200 Hz. As in previous work (Martínez-Cagigal et al., 2020; Pulferer and Müller-Putz, 2022) six electrodes from frontopolar and temporal sites (Fp1, Fp2, FT9, FT10, TP9, TP10) were removed, and reallocated to two parietooccipital electrode positions for increased channel density above the cuneus and precuneus area (PPOh1, PPOh2; according to the 10–5 system), as well as four EOG locations (above and below the left eye and the outer canthi of both eyes). Ground and reference electrodes were placed at the electrode position Fpz and on the participants’ right mastoid, respectively.

All data were recorded and synchronized via lab streaming layer (https://github.com/sccn/labstreaminglayer). The presented main paradigm was created and displayed utilizing MATLAB 2017b (MathWorks Inc. United States); the eye artifact correction package we employed (Kobler et al., 2020) additionally utilized Psychtoolbox (Brainard, 1997; Pelli, 1997; Kleiner et al., 2007).

Before each session, we assigned each of the two participants an initial role at random, corresponding to either Executor or Observer. The setup involving two participants at the same time was chosen as previous literature reported considerable differences between executed and observed erroneous action in the processing of discrete error stimuli (van Schie et al., 2004). The participant acting as Executor then took a seat in front of an Xbox Wireless Racing Wheel (Microsoft, 2006/2007; sans foot pedals) inside our measurement box, while the Observer was positioned outside (Figure 1A). This spatial separation prevented direct visual contact between the two participants; however, due to the noise caused by the steering wheel’s mechanics, we decided against closing the door to the measurement box to establish identical auditory paradigm conditions for Executor and Observer. Both participants faced their computer screens at a comfortable distance (~1 m).

The whole measurement session then consisted of two identical iterations through our paradigms. In detail, for each of the iterations, we first presented the participants with a run unrelated to the main hypothesis (6 min), designed to capture the rotational movement of the corneo-retinal dipole. This run specifically recorded vertical and horizontal eye movement, blinks, and resting state data (open-eyed), cued by a moving, pulsating, or static target on-screen, to create an eye artifact removal model for offline artifact correction (Kobler et al., 2020). No interaction with the steering wheel occurred during this first run. Afterward, the car-driving paradigms commenced, consisting of four different vehicle operation conditions labeled Passive Steering, Proactive Steering, Reactive Steering (Distinct), and Reactive Steering (Indistinct). Each condition was presented in three separate runs (3.5 min each), adding up to around 15 min of recordings per condition including instructions, and totaling to approximately 1 h of recordings for one paradigm iteration (Figure 1D). To capture the subjective perception of the Executor’s steering performance per condition and run, we additionally asked both participants to rate the perceived performance on a visual analog scale ranging continuously from 0 (bad performance) to 1 (excellent performance).

After the first paradigm iteration, the participants took a short break before the roles were reversed and the second paradigm iteration started up. Specifically, the person previously acting as Executor left the box and switched positions with the former Observer to henceforth observe from the outside as the new Observer. Likewise, the former Observer took their place at the steering wheel, acting as the new Executor for the rest of the measurement. This switch allowed us to capture both observation and execution data of each participant. The measurement commenced again with a second eye movement recording run, followed by the four car-driving paradigm conditions.

As previous research reported contributions of various frequency bands to error processing in the brain (Yordanova et al., 2004; Koelewijn et al., 2008; Carp and Compton, 2009), we chose to bandpass filter both EEG and EOG signals at 0.2-30 Hz (Butterworth, 10th order). Subsequently, we identified and spherically interpolated noisy channels via visual inspection, abnormal kurtosis, and abnormal joint probability (exceeding five standard deviations of the mean for both measures). We used EEG and EOG data of the eye movement recordings to train an eye artifact correction model to correct the hypothesis-related EEG data of the main paradigm conditions for rotational movements of the corneo-retinal dipole (Kobler et al., 2020), followed by ICA [FastICA algorithm (Hyvärinen, 1999), https://research.ics.aalto.fi/ica/fastica/index.shtml] in the case of a persisting influence of ocular movement on the EEG. To this end, the independent components were first automatically labeled using the ICLabel classifier available in EEGLab (Pion-Tonachini et al., 2019). Components attributed to eye, heart, or muscle artifacts with a classification accuracy of at least 90% were then automatically removed. The scalp distribution of the ICA weights, as well as the time series of the components, were then visually inspected to exclude remaining eye, heart, or muscle artifacts. The EOG and anterior frontal row electrodes were then excluded from further analysis, leading to a remaining total of 55 channels to consider which were then re-referenced to their common average. The whole processing approach is depicted in Figure 1E.

In terms of non-biological signals, the car’s x-position signal was analyzed as recorded without further processing. To alleviate jitters in the recorded steering wheel signal, we smoothed the data with a Savitzky–Golay filter utilizing second-order polynomials to fit windows of 351 samples (1.75 s) (Savitzky and Golay, 1964).

To obtain an understanding of the differences in cognitive processing between the four car-driving paradigm conditions, as well as differences between execution and observation of the task, we analyzed the recorded data both in sensor space and in source space. Due to the lack of normality in the distribution of single evaluation metrics, non-parametric tests were selected to assess significant differences between populations. To this end, significant differences between pairs of populations (e.g., Executor vs. Observer) were evaluated using a Wilcoxon signed rank test; for comparisons between more than two populations (e.g., the four car-driving conditions) were assess with a Friedman-Nemenyi test – a many-sample extension of the Wilcoxon test (Zimmerman and Zumbo, 1993). Both tests were followed by subsequent False Discovery Rate (FDR) correction for multiple comparisons.

To gain a general understanding of how both the car’s position (for both roles) and the steering process (Executors only) are represented in the brain, and to specifically test which of both non-biological signals can more accurately be inferred from the EEG, we chose to slightly modify the EEGNet architecture for regression on our dataset (Lawhern et al., 2018). The architecture, a three-layer convolutional neural network, was chosen as EEGNet reportedly provided high performance within a range of tasks in brain signal decoding (Lawhern et al., 2018). As the goal of this approach was predominantly to provide first results for regressing car positions and steering information from the brain signals, other undoubtedly equally suitable deep learning architectures, such as deep believe networks (An et al., 2014), recurrent neural networks (Tortora et al., 2020), or even transformer models (Song et al., 2023), have not been investigated in the current work.

For each of the four conditions, a 1 s-window (corresponding to T = 200 samples) was slid through the continuous data of each run with a stride of 15 samples (75 ms) to obtain a total of 7,920 frames per condition. To prevent possible contributions of muscular activity at EEG channels near the neck to the regression performance, all EEG channels comprising the outermost circle were excluded, yielding windowed data in C = 43 EEG channels of the form X ∈ R 7920 × 43 × 200 . To retain a causal relationship between training and testing data, we then split the frames into six folds, the last of which we retained for testing. Both training and testing data were standardized with respect to the training dataset.

The network architecture processed batches of inputs of the size C × T in three consecutive layers. In the first layer, inputs were zero-padded with 32 samples before and after the input in the time dimension, followed by temporal convolution with a kernel size of (1,64) which corresponded to a high pass cut-off at approximately 3 Hz. Within the second layer, we applied depthwise convolution across all available channels [kernel size (43,1)], followed by average pooling with a kernel size of (1,4). A third (final) convolutional layer with kernel size (1,32) further extracted temporal features, followed by a second instance of average pooling with a kernel size of (1,8). Finally, the resulting features were flattened and passed through a fully connected layer returning one real-valued output corresponding to the final prediction of the regression model. Each convolutional layer was followed by a batch normalization layer to accelerate convergence (Ioffe and Szegedy, 2015), as well as a dropout layer (0.25 dropout rate each) to impede overfitting. Exponential Linear Units were used for activation functions (Clevert et al., 2015).

We generated separate models utilizing the same hyperparameters to regress the car’s x position and the steering wheel signal for each condition in the Executors’ data. In detail, the models were trained for 100 epochs at a learning rate of 10–4 with a batch size of 256. We used the Adam optimization algorithm implemented in PyTorch to optimize the model parameters with respect to the mean squared error loss between the prediction and dependent variable (car position or steering signal).

To analyze the importance of each input channel to the model’s performance, we further calculated saliency maps on the testing data. To this end, the highly non-linear mapping between a two-dimensional input frame I (i.e., the C × T frames of windowed EEG data) and the scalar regression output S I (predictions) through the neural network is linearly approximated in a first-order Taylor expansion around a given input frame I 0 as

In practice, for each (flattened) given input frame of the test set, the model weights w ∈ R 43 ∗ 200 × 1 are found via backpropagation, yielding a final saliency map M ∈ R 43 × 200 . Each entry of M then corresponds to the absolute of the matching entry of w (Simonyan et al., 2013) and indicates the contribution of a specific channel and time-point to the network prediction. A topographical feature importance map is then obtained by summing all entries along the time dimension.

The network was implemented using PyTorch (https://github.com/pytorch/pytorch); feature importance analyses were conducted with the Captum library (https://github.com/pytorch/captum).

To analyze differences in steering behavior, we time-locked the EEG signals to zero-crossings of the steering wheel’s signal for all four conditions. As only the Executor performed the steering, we omitted the Observer’s data from further analysis utilizing this specific time-lock. Results for the Proactive Steering and the Reactive Steering (Indist.) conditions are provided in the Supplementary Figures S2, S4, upper panels. A first analysis of the scalp topographies at the time-lock (i.e., steer zero) and the subsequent time points maximum steering wheel deviation (steer max) revealed comparable frontocentral positivities across all conditions, disregarding the corresponding type of control (Passive, Proactive, or Reactive Steering). The results for the Passive Steering and Reactive Steering (Dist.) conditions are shown in the lefthand side topographical maps of Figures 3A,B, respectively. Significant differences only arose in comparisons with the Proactive Steering condition, while a slight frontal shift in activity occurred for all other conditions when compared to the Passive Steering condition. The full time courses of the grand averaged EEG signals in all conditions within the steer zero time-lock for Execution and Observation are provided in the Supplementary Figures S8, S9.

Considering the focal frontocentral activity across conditions, we selected the EEG channel at electrode position FCz for sorting the trials according to their steering latencies. The resulting grand-averaged sorted and interpolated trials of the Passive Steering and Reactive Steering (Dist.) conditions are shown in the image plots of Figures 3A,B, respectively. We observed a clear modulation of brain activity recorded at position FCz with the steering activity, with positivities in the brain signals phase-locked to instances of maximum steering wheel deflection (Figures 3A,B, red line in the image plots). For the Passive Steering condition, which followed a pre-defined periodicity, the patterns visibly repeated with the steering motion throughout the epoch length, whereas the slower (self-paced) steering within the Reactive Steering (Dist.) condition occurred at a larger time frame. Interestingly, we observed very distinct ERD/ERS in the Passive Steering condition (Figure 3A, topographical maps in grey boxes). Delta and theta bands exhibited weak central scalp topographies, whereas alpha and specifically beta bands showed focal activity for EEG channels at electrode positions C3 and C4. Similar, though more frontal delta ERD and ERS topographies arose for the Reactive Steering conditions (Figure 3B, topographical maps in grey boxes), whereas no clear patterns were observed for the theta frequency band. We also noted a clear shift in alpha activity to occipital regions for both ERD and ERS maps, as well as additional occipital activity in the beta band.

An analysis of the cortical regions of interest comprising the Desikan-Killiany atlas is depicted in Figure 4; cortical maps between the single condition results indicate regions of significant pairwise difference at a significance level of 0.05 (dark red patches, FDR corrected for multiple comparisons). The cortical activity within the Proactive Steering condition (second panel from the left) showed strong attenuation and proved significantly different from all other conditions for most regions of interest; in contrast, no overall significant differences were seen for pairwise comparisons in other conditions. Interestingly, strongest activations arose in the bilateral caudal anterior (cACC) and posterior cingulate cortices (PCCs), parahippocampal and precentral gyri, and precuneus area across all conditions, with additional strong unilateral engagement of prefrontal cortex areas.

To analyze directional connectivity within all EEG channels, we utilized the standardized participant averages in the steer zero time-lock, yielding a multivariate time series matrix X ∈ R M × N × S , with M = 55 channels, N = 1601 time points [(−2,6)s epoch], and S = 26 participants. The optimum model orders per condition and role were exclusively found to be p opti ∈ 21 22 . In line with the time-frequency analysis, we then averaged the outflow across the four frequency bands of interest (delta to beta). Topographical maps displaying the total information outflow per band in the different paradigm conditions are shown in Figure 5. In detail, we observed large-scale information outflow at central and parietal regions during the Passive Steering condition across all frequency bands. In contrast, these regions condense increasingly toward frontocentral sites as the frequency is lowered; specifically in the delta band, little significant information outflow remains apart from regions at approximately electrode position FCz. Notably, an increase in information outflow from these frontocentral sites within the beta band arises in the Reactive Steering conditions.

To assess differences in visual and feedback processing between conditions and roles, we additionally time-locked the EEG signals to zero-crossings of the car’s x position for all four conditions. In the case of the Passive Steering trials, this merely implied time-locking to every instance where the car passed through the screen’s center, while for Proactive and Reactive Steering conditions, it additionally corresponded to the car’s maximum deviation from the straight road. In contrast to the steering time-lock employed before, we here considered both Observer and the Executor data, as the visual input related to the car position time-lock comparably affected both roles.

To analyze cortical responses in Observers and Executors to a moving stimulus sans error connotation, we first analyzed the scalp topographies during zero crossings (car zero) and maximum deviation from the screen’s center (car max) in the Passive Steering condition. The results are displayed in Figure 6A (Executor) and B (Observer). As can be seen, the topographical maps for both time points largely coincide for both roles, with slightly increased central activity for the Executors and overall attenuated signal amplitudes in the Observers (see scale). Statistical analyses revealed significant differences between this Passive Steering condition and both Reactive Steering conditions (small topographical maps; channels displaying significant differences are mapped as black circles) at a significance level of 0.05 (FDR corrected). The full time courses of the grand averaged EEG signals in all conditions within the car zero time-lock for Execution and Observation are provided in the Supplementary Figures S6, S7.

To analyze the relation between brain activity and car position, we sorted and interpolated all EEG trials at electrode position Pz (highest activity as observed in the topographical maps); the results are shown in the image plots of Figures 6A,B. Interestingly, we observed clear modulations of the parietal brain activity with the phase of the car signal, i.e., positivities in the EEG coinciding with the car’s center position, negativities coinciding with the car’s maximum left/right deviation from the center. In contrast to the steering time-lock, neither ERD nor ERS maps show similar C3/C4 activity in the alpha and beta bands. However, similar ERD/ERS patterns arose between Observers and Executors, with high parietal theta and alpha activity.

Similar analyses for the Reactive Steering (Dist.) condition are shown in Figures 7A,B; results for the remaining conditions are provided in the Supplementary Figures S2, S4, lower panels; Supplementary Figures S3, S5. Compared to the Passive Steering condition void of any error connotation, we observed clear frontocentral (Executors) and centroparietal (Observers) scalp distributions in the grand-average EEG signals. Overall, there appeared to be a frontal shift during task execution compared to simple task observation. Apart from the mentioned differences in comparison to the Passive Steering condition, additional significant differences arose in comparison with the Proactive Steering condition for both roles; however, no differences between the two Reactive Steering conditions were found. Considering the topographical maps, we chose to investigate the single-channel EEG activity at electrode position Cz in this case; the sorted and interpolated grand-average trials are shown in the image plots of Figures 7A,B. As can be seen, we observed strong modulations of Cz EEG activity with the car’s specific x position on the screen in the Executors’ data; a similar though much more attenuated behavior also translated to the Observers’ data.

As in the Passive Steering condition, we observed similar ERD/ERS patterns in the alpha band in the Reactive Steering. In addition, fronto-central delta band activity arose in the Executors, which could not be observed in the Observer data.

In line with analyses regarding the steer zero time-lock (see previous section), we analyzed the cortical activity within each condition in the car zero time-lock for both Observer and Executor data; pairwise significant differences between conditions at 0.05 (FDR corrected) are once again shown as red areas on the cortical maps (Figures 8, 9, resp.). We observed very clear modulations with the car’s x position in the Passive Steering condition that largely coincided for both roles. In detail, grand-average signals in the Desikan-Killiany atlas revealed shared activity in the bilateral pericalcarine cortices, lingual gyri, the isthmus of the cingulate cortices, precunei, and the superior parietal gyri.

We interestingly observed stronger overall activity within the Passive Steering condition compared to other conditions in the Observers than in the Executors; however, we report increased activity in the posterior cingulate cortex for the Executors that was not found for Observers. As in the steer zero time-lock, little activity was found within the Proactive Steering condition; however, both roles showed increased engagement of the bilateral posterior and caudal anterior cingulate cortices, as well as the parahippocampal gyri, and temporal regions within both Reactive Steering conditions.

To estimate directional connectivity between distinct brain regions, we utilized the standardized participant averages of the form X ∈ R M × N × S in the car zero time-lock for both roles, with M = 55 channels, N = 1601 time points [(−2,6)s epoch], and S = 26 participants. The optimum model orders per condition and role were again exclusively found to be p opti ∈ 21 22 . Topographical maps displaying the total information outflow per band (target-averaged DTF results) in the different paradigm conditions are shown in Figures 10A,B for Observers and Executors, respectively. As in the steer zero time-lock, we observed large-scale information outflow at central and parietal regions during the Passive Steering condition in all frequency bands, whereas the total information outflow within the other conditions increasingly condensed toward fronto-central sites with decreasing frequency.

To analyze differences between observed and executed car driving within error-free and erroneous steering conditions at moments of zero and maximum car deviation from the center of the screen (car zero and car max), a Wilcoxon test along with FDR correction at a significance level of 0.05 was consulted. Considering these two time points of interest in four conditions and 55 EEG channels, we corrected for a total of N = 440 tests. Interestingly, no significant differences arose for a comparison between Observer and Executor data in the Passive Steering condition for either time point of interest (see Figure 11, leftmost topographical maps). However, within the Proactive and Reactive Steering conditions, we observed significant differences between the roles at fronto-central electrode sites. Specifically in the Reactive Steering (Dist.) condition, an area of significance around the electrode position FCz emerged, which similarly arose for the Reactive Steering (Indist.) condition as well.

We trained separate models for all participants in each condition to regress the car’s x position on the screen and the steering wheel signal from the Executors’ EEG data, respectively. The training curves of both the root mean square error and Pearson’s correlation coefficient for the regression of both signals are provided in the Supplementary Figures S10, S11. For both signals, the maximum testing correlation between signal and prediction is shown in Figure 12 (A: car position, B: steering wheel signal). A Wilcoxon signed rank test revealed significantly higher regression performance for the Passive Steering condition (FDR-corrected at 0.01), and significantly decreased performance for the Proactive Steering condition (FDR-corrected at 0.001) when regression the car’s x position compared to the steering wheel signal. No significant differences arose between both performances for the Reactive Steering conditions, although we observed generally higher maximum correlation values per participant when regressing the steering wheel signal (see Tables 1, 2). Notably, the performance considerably dropped for all conditions apart from the Passive Steering condition, where median values of 0.84 and 0.81 for Pearson’s correlation arose.

An analysis of the importance of each channel unveiled largely identical grand average (normalized) saliency maps for both the regression of the car’s x position (Figure 12C) and the steering wheel signal (Figure 12D) in each condition (Passive Steering, Proactive Steering, Reactive Steering (Dist.), Reactive Steering (Indist.), f.l.t.r.). Overall, considerable feature importance lies on parietal and occipital channels for all conditions, with increased engagement of central channels during the Proactive and Reactive Steering conditions. Within the Reactive Steering conditions, we further observed increased channel importance at electrode positions C3 and C4.

To investigate brain signals relating to the steering process, we time-locked the EEG data to zero-crossings of the steering wheel signal. Within this time-lock, we observed largely coinciding scalp topographies across all steering conditions (Passive, Proactive, and Reactive Steering, Figure 3 and Supplementary file). Significant differences only arose in pairwise comparisons with the largely attenuated signals of the Proactive Steering condition, which most likely traces back to the minimal steering necessary during this condition. For all other conditions exhibiting comparable levels of steering movement, no significant differences arose, indicating matching executive activity in the brain that presumably relates to the steering motion itself and disregards error-free or error-prone driving performance. Nonetheless, we observed a qualitative frontal-shift within Reactive Steering compared to Passive Steering (large topographical maps, Figure 3B), substantiated by directional connectivity analyses unveiling increasingly focal fronto-central areas of the scalp as sites of highest information outflow during Proactive and Reactive Steering conditions. Source analysis traced these observations back to increased activity in the caudal anterior cingulate cortex (Figure 4, third and fourth panel), a region reportedly involved in the cognitive control of motor behavior (Holroyd et al., 2004) and performance monitoring (Ursu et al., 2009). In addition, increased activity in the bilateral parahippocampal gyrus, as well as the right caudal middle frontal gyrus and frontal pole arose; regions involed in visuospatial processing (Aminoff et al., 2013) and reorientation of attention (Japee et al., 2015) in humans, as well as evaluation of self-generated decisions in monkeys (Tsujimoto et al., 2010). However, shared regions of cortical activity could be found for all steering conditions (Figure 4). In this context, we specifically report engagement of the posterior cingulate cortex, which has been linked to both the integration of reward processing, attention, and motor control systems and the dynamic modification of strategy (Pearson et al., 2009). Furthermore, notable engagement of temporal regions arose, which we however hypothesize relates to the increased noise generated by the wheel during the conditions involving expansive steering motion (Passive and Reactive Steering).

Interestingly, we additionally observed differences in event-related desynchronization/synchronization (ERD/ERS) between the steering conditions. In detail, we observed prominent alpha and especially beta activity at electrode positions C3 and C4 during the perfectly periodic pre-defined car movement within the Passive Steering condition, both sites situated bilaterally above the hand regions of the sensorimotor cortex. In contrast, activity during the Reactive Steering conditions moved noticeably towards occipital regions especially in the alpha frequency band. This indicates increased emphasis on processing visual information within Reactive Steering conditions compared to the Passive Steering condition and aligns with previous findings of Garcia and colleagues, who likewise reported a transition from delta and beta activity during proactive brain states to prominent alpha activity during reactive brain states (Garcia et al., 2017).

Sorting all trials at electrode position FCz with respect to their steering latency, we further found clear modulations in the brain signals with the steering wheel signal for all conditions, with positivities in the EEG visibly coinciding with instances of maximum steering wheel deflection (image plots, Figure 3), coinciding with previous reports in the context of error processing by Burle et al. (2008) who showed similar fronto-central positivities in the EEG at instances of corrective button-presses.

Employing an additional time-lock to instances of zero crossings of the car’s x position on screen, we further analyzed brain responses phase-locked to the visual input delivered within our car-driving paradigms. As both Observers and Executors were exposed to the same visual stimulants, we investigated both the Observers’ and the Executors’ data within this time-lock.

Notably, the brain patterns matched closely between Observers and Executors within the Passive Steering condition (see Figure 6), substantiated by a lack of observed significant differences between the roles (Figure 11, leftmost topographical maps). Both roles showed strong modulations with the visual feedback (moving car) especially in parietal regions and displayed large-scale significant differences in fronto-central and occipital channels when compared to both Reactive Steering conditions (Figure 6), indicating a clear distinction between the processing of error-free and error-prone driving performance. Additionally, the ERD/ERS topographies (Figure 6, grey boxes) showed clear occipital engagement for both roles in the theta and alpha frequency bands, contrasting with the central (motor) activity observed in the steering time-lock (Figure 3, grey boxes).

As the error-free driving performance within the Passive Steering condition was entirely fake, the Executors’ steering movement within this condition did not correlate directly with the car’s position on-screen, yielding largely distinct patterns for both time-locks (compare Figures 3A, 6A) and thus both separate processes – motor activity and visual processing – in the brain. In contrast, the high-level correlations between steering and car position during the other conditions lead to matching scalp topographies in the Executors for both time-locks (compare Figures 3B, 7A), disallowing a separation of brain regions engaging for both individual processes. A comparison to the Observers’ data within Proactive and Reactive Steering revealed significant differences especially in the fronto-central areas of the scalp, presumably due to the slightly more attenuated and parietally shifted potentials during observation despite overall similar patterns (see Figure 11, second, third and fourth columns). The lack of a similarly significant difference during Passive Steering (first column) indicates that the observed disparities are not solely related to the additional motor activity within the Executors. This assumption is strengthened by findings of increased bilateral activity in the caudal anterior cingulate cortex (Holroyd et al., 2004; Ursu et al., 2009) within Reactive Steering conditions in the Executors in both the steering and the car position time-lock (see Figures 4, 8, 9), as well as more focal frontocentral sites of information outflow within out connectivity analyses (Figures 10A,B).

For both Observers and Executors, the Passive Steering condition yielded the highest correlations in our regression, most likely due to the perfectly periodic modulations evoked in the brain by the strict periodicity of the car movement within this condition. However, the synchrony between visual input and motor output within this steering condition may additionally contribute to high regression performances. In detail, the participants observed a periodically moving car programmed to always remain perfectly on the winding road. As such, target (the white centerline) and feedback (the moving car) concurred throughout the condition, allowing the focus of attention to condense strictly to the moving car. Interestingly, we nonetheless observed significantly higher performance for regressing the car position compared to the steering wheel signal in the Passive Steering condition within Executors, tentatively indicating a more prominent representation of the visual aspect of the task compared to the mostly automated steering (see Figure 12).

In contrast, the correlations within the Reactive Steering conditions merely reached half of the values for the Passive Steering, displaying no significance in the differences in the regression of both signals. This might suggest that both processes – visual feedback processing and motor control – commanded equal levels of related brain activity, yet the direct correlations between steering and car position within these conditions disallows a clear assessment. As the Reactive Steering conditions did not follow the strict periodicity of the Passive Steering condition, the observed drop in performance was to be expected to a certain degree. However, within Reactive Steering, the car continuously deviated from the road, constantly splitting the participants’ attention between the intended and the observed target (car) position, which might have additionally led to lower correlations. Furthermore, while a regression of the car position corresponded simultaneously to a regression of the intended position on-screen during the Passive Steering condition, it rather equaled a regression of the continuous feedback error (the deviation from the road) in the case of the Reactive Steering condition. As such, the results are indeed interesting despite the moderate correlations and indicate the feasibility of regressing continuous error signals as observed within our previous work (Pulferer et al., 2023) even during real-life situations.

Lastly, we observed the lowest correlations when regressing within the Proactive Steering condition, presumably due to both the minimal motor output and the barely visible deviations from the road. Nonetheless, we obtained significantly higher correlations when regressing the steering wheel signal, implying that the minute deviations from the road provided less relevant information than even the minimal steering movement.

Despite the exploratory nature of the current work to analyze the recently reported continuous correlates to error processing in more detail, several limitations of the present study need to be mentioned.

First and foremost, as has been the modus operandi for most studies reporting on electroencephalographic correlates in the past decades (Larson and Carbine, 2017), prior sample size calculations were neglected. As a result, the exact statistical power of the observed effects cannot be stated; a shortcoming which needs to be alleviated in follow-up studies.

Secondly, the lengthy measurement sessions and restricting nature of the paradigm during revoked steering agency gave rise to various superimposing factors; specifically, fatigue, frustration, and boredom may have impacted the recorded cortical information – a recurring issue in neuroscientific studies which nonetheless remains problematic. In addition, the limited spatial resolution within source projection methods further demands conservative interpretation.

Thirdly, the utilized car driving paradigm was displayed in a most minimalistic fashion, only incorporating a bird view of the lane instead of immersive driving simulations. While this reduction of distractors proved to be necessary for interpreting our first results presented in the current work, future work will need to close the gap to more realistic driving experiences to observe similar effects in a real-world scenario.

And lastly, the approach used for investigating the feasibility of regressing both the steering and the car position signals via a neural network could serve only as a proof-of-concept. Different architectures would need to be consulted in future work to obtain a reliable estimate for the optimum performance obtainable within this research question.