The present study explores how training and affective context (via aversive electrodermal stimulation) influence orientation-averaging performance in an immersive virtual reality (VR) environment. Specifically, we investigate:

1. Whether perceptual training improves ensemble processing via early-stage sampling or late-stage decision efficiency.

Our hypotheses are as follows:

• H1: If training improves early-stage sampling, then increasing set-size will result in shallower performance declines over time (i.e., improved slope). If training improves late-stage decision processes, accuracy will increase overall, but the slope of set-size degradation will remain constant.

We employed a 10-day longitudinal design, in which 17 participants completed a VR-based shooting task involving orientation averaging of Gabor stimuli. Gabor arrays were positioned in fixed circular configurations to ensure consistent visual sampling across trials. This design allowed for controlled manipulation of set-size while minimising confounds due to spatial unpredictability, which could affect attention allocation.

Participants were randomly assigned to one of three groups: no shock, performance-contingent shock, or random shock, and we measured accuracy, response time, and anxiety levels using validated instruments (e.g., STAI). While the 10-day period aligns with prior perceptual learning studies indicating reliable gains over such durations (Moerel et al., 2016), future work may examine shorter or longer timelines.

This study seeks to clarify the mechanisms by which perceptual learning unfolds and the influence of stress and motivation on visual decision-making in immersive environments—critical insights for adaptive training and human performance optimisation in VR.

Twenty-two participants were recruited from the Western Sydney University community via convenience sampling between 1st April and 31st July 2022. Two participants withdrew from the study—one due to illness (COVID-19) and one voluntarily—leaving 20 who completed all testing sessions. Three additional participants were excluded: one for non-compliance with task instructions, and two for receiving an unusually high number of shocks due to technical error. The final sample comprised 17 healthy adults (8 female; 9 male; mean age = 26.2 years, SD = 12.4), all with normal or corrected-to-normal vision.

All participants provided informed consent prior to participation. The study was approved by the Western Sydney University Human Research Ethics Committee (H13736) and conducted in accordance with the Declaration of Helsinki. Participants were naïve to the study’s aims and received $250 AUD for their participation.

The experiment used a 3 (Shock Group: No Shock, Random Shock, Performance-Contingent Shock) × 10 (Training Day) × 5 (Set-Size: 1, 2, 3, 4, or 8 Gabor elements) mixed factorial design. Shock group was a between-subjects factor; training day and set-size were within-subjects factors. Each participant completed 10 training sessions on consecutive weekdays.

Participants were randomly allocated to one of three shock conditions. In the performance-contingent shock group, participants received an electrodermal stimulus following every 10th incorrect response. In the random shock group, shocks were administered with a 3% chance per trial, independent of performance. This probability was determined via pilot testing to approximate the average frequency of shocks received by the performance-contingent group. Participants in the no-shock group received no stimulation.

Average daily shocks were 10.9 (SD = 2.3) in the performance-contingent group and 13.1 (SD = 3.0) in the random shock group.

The key outcome measures were orientation discrimination accuracy, correct response time (RT), and state anxiety scores (via the STAI).

Participants performed a custom-built virtual reality orientation-averaging task embedded in a first-person shooter (FPS) game. On each trial, they viewed an array of Gabor elements (stimulus set-size varied across trials) and judged whether the average orientation was tilted clockwise or counter clockwise from vertical (Figure 1). Participants responded by shooting one of two visually identical virtual agents, each standing to the left or right of a central post, depending on the inferred average orientation of the Gabors.

Correct responses prevented hostile fire, while incorrect responses triggered either a visual consequence (loss of health bar) or, in shock conditions, electrodermal stimulation (depending on shock group). The task structure was designed to promote integration of ensemble visual information and translate perceptual decisions into motor actions in an immersive environment.

The virtual environment included a central pole flanked by two stationary soldier agents (98 × 256 pixels), one of whom was designated “hostile” each trial. Hostility was randomly assigned on each trial and signaled solely via the mean orientation of the Gabor array: −5° indicated the left agent was hostile; +5° indicated the right.

Stimuli appeared in a semi-transparent 1,024 × 1,024 pixel square, fixed to the participant’s visual field. The health bar in the lower-right corner reduced by 10% after incorrect responses for all participants, acting as performance feedback and shock countdown (in the performance-contingent group). Following any trial in which the health bar reached 0%, it was reset to 100% at the start of the subsequent trial.

All participants completed 10 testing sessions on consecutive weekdays. On Day 1, participants were briefed, consented, and randomly assigned to a shock group (No Shock: n = 5; Random Shock: n = 5; Performance-Contingent: n = 7). After measuring IPD and fitting the headset, participants completed a practice session (∼20 trials) to ensure task understanding.

Shock group participants underwent a daily calibration to identify their preferred stimulation level. Each testing session consisted of two blocks of 210 trials (42 per set-size), separated by a 5–10 min break. Participants completed the STAI before and after each session.

The experiment was approved by the WSU Human Research Ethics Committee (H13736). All procedures complied with the Declaration of Helsinki. This study was not preregistered. De-identified data and analysis scripts are available upon request.

The authors declare no competing interests.

Whether participants performed differently across training sessions was assessed by fitting linear mixed-effect models with shock condition and training session as fixed effects and set-size and participants as random effects. The Kenward-Roger degrees of freedom approximation (Kuznetsova et al., 2017) was used to calculate p values for the fixed-effect factors, and the ANOVA function from package car (Fox and Weisberg, 2018) was used to calculate F. Pairwise comparisons were conducted using the lsmeans package (Lenth, 2016) in R when necessary. Linear mixed-effect ANOVAs on our performance measures were undertaken using Type-II Wald F tests with Kenward-Roger degrees of freedom.

Mixed-effects models revealed main effects of training session on both accuracy (F (9, 727.01) = 53.634, p < 0.0001) and RTs (F (9, 727.01) = 86.607, p < 0.0001). Set-size also produced significant main effects for both accuracy (F (4, 727.00) = 471.307, p < 0.0001) and RTs (F (4, 727.00) = 111.968, p < 0.0001).

No significant training × set-size interactions were found for accuracy (F (36, 727.00) = 0.660, p = 0.9441) or RTs (F (36, 727.00) = 0.125, p = 1.0), indicating that training improved performance uniformly across set-sizes.

For the next series of analyses, we investigated whether performance feedback visible to participants via the ‘health bar’ (present throughout each training session) affected participant performance (see Methods for information on health bar performance feedback. Figure 6 shows the relationship between health-bar level (low (≤20%), medium (21%–80%) and high values (>80%)) on average accuracy and response times in each of the three group training conditions.

Health-bar level had a significant main effect on accuracy (F (2, 922) = 497.380, p < 0.0001), with higher visual feedback levels corresponding to better performance. All pairwise comparisons were significant (all p < 0.0001).

No effect of health-bar level was observed on RTs (F (2, 922) = 0.310, p > 0.05). Shock condition did not significantly influence accuracy (F (2, 922) = 0.001, p > 0.05) or RTs (F (2, 922) = 0.475, p = 0.632), and no interactions were found between health-bar and shock condition (accuracy: F (2, 922) = 0.326, RTs: F (2, 922) = 0.299, both p > 0.05).

The health-bar provides performance feedback to the participant, dropping 10% following each incorrect trial, and refreshing after each 10th incorrect trial. It is worth noting that the health bar status is predictive and consequential only for participants in the performance-contingent shock group who received a shock following nine previous errors. The health bar level is completely inconsequential (i.e., unpredictive of the of a shock stimulus) for participants in both the no-shock and random shock groups. To determine whether there exist additional effects of anticipating and/or receiving a physical shock, we conducted additional analyses evaluating the effects on performance (accuracy and response times) on the five trials immediately preceding and the five trials immediately succeeding a physical shock (Figure 7). Given that shock was only presented to participants in the random and performance-contingent shock conditions, participants in the no-shock condition were omitted from these analyses.

Analyses of the five trials before and after shock events in random and performance-contingent groups revealed no main effects between these groups on accuracy or RTs (all p > 0.05). A significant interaction between trial number and shock condition was observed for RTs (F (4, 99) = 3.313, p = 0.016), but no pairwise comparisons survived Bonferroni correction (all p > 0.05).

Effects of pre-vs. post-session training and shock condition on subjective stress are shown in Figure 8. No main effects were found for shock condition (F (2, 12.99) = 0.186, p = 0.832) or training session (F (9, 245.01) = 0.326, p = 0.966) on state anxiety. A significant effect of test time (pre vs. post) was detected (F (1, 345.00) = 22.032, p = 0.0001), with post-training anxiety higher than pre-training.

A significant interaction between time (pre vs. post) and shock condition was found (F (2, 245.00) = 3.359, p = 0.036). Post hoc analyses revealed no change in the no-shock group (p = 0.996), but significant post-training increases in anxiety for the performance-contingent (p < 0.001) and random shock conditions (p = 0.02).

This study addressed two key questions in perceptual learning. First, we investigated whether training enhances the human visual system’s capacity to integrate orientation information across the visual field. Second, we explored whether aversive electrodermal stimulation (used as acute stress or punishment) modulates perceptual learning in an orientation-averaging task implemented in virtual reality.

Our findings demonstrate (1) consistent performance degradation with increasing set-size, (2) significant training-related improvements in accuracy and response times, and (3) no evidence that training altered the slope of the set-size effect. These results suggest that training does not enhance early-stage sensory integration but likely operates at a post-integration or decisional stage. We also observed (4) no robust effects of shock condition on learning outcomes, and (5) an unexpected, systematic influence of visual performance feedback (“health bar”) on accuracy.

Whilst broadly consistent with prior work, we show for the first time that orientation-averaging performance degraded with larger set-sizes. This suggests that the visual system’s capacity for averaging orientation information across 2D space is fundamentally constrained. Whilst training led to significant gains in accuracy and response speed, there was with no interaction between training and set-size, implying that the quantity of integrated information did not increase. These gains, therefore, are likely to arise at decision-related stages—possibly by increasing the signal-to-noise ratio during integration or decision execution (Dakin et al., 2009; Moerel et al., 2016).

Our task design departed from prior orientation-averaging studies in several important respects, particularly in how local Gabor orientations were sampled on each trial. While previous studies typically draw orientations randomly from Gaussian distributions with a defined mean and standard deviation (e.g. (Moerel et al., 2016; Dakin et al., 2009)), we used a deterministic method to fix the global mean orientation across trials. Specifically, we first sampled all but one Gabor from a Gaussian distribution and then adjusted the final Gabor’s orientation to ensure that the mean orientation of the array exactly matched a predefined target (e.g., ±5° from vertical). This approach eliminated trial-by-trial variability in global mean orientation that can arise in probabilistic sampling schemes.

However, one potential drawback of our method is that it may have encouraged participants to adopt a “max-rule” strategy—relying on the most tilted element in the array rather than computing a true average. Such a strategy would introduce substantial variability in responses and lead to a higher rate of incorrect choices. If participants relied on this approach, at least on a subset of trials, task performance may have been supported in part by relational visual search mechanisms (e.g. (Becker et al., 2025)), rather than global ensemble processing. Future work could address this limitation by drawing stimulus sets from a large library of pre-generated Gaussian samples with fixed means and variances, which would preserve statistical consistency across trials without encouraging max-rule heuristics.

While visual crowding can impair averaging by reducing access to local features, we argue that crowding is unlikely to account for our set-size effects. First, the inter-Gabor spacing in our design exceeded Bouma’s limit for crowding interference (Kurzawski et al., 2023; Van der Burg et al., 2024; Bouma, 1970). Second, participants were free to move their eyes, which reduces crowding by maintaining central fixation and reducing retinal eccentricity. Nonetheless, because eye movement patterns were not monitored, comparisons with fixed-eccentricity paradigms should be interpreted with caution as the distinction between early and late selection frameworks typically hinges on whether orientation averaging depends on retinotopically local filters—such as V1 neurons—which are spatially specific and sensitive to gaze position. In our task, Gabor elements were visible for extended durations, allowing participants to freely move their eyes. This limits our ability to isolate the contribution of early, retinotopically specific encoding mechanisms, as participants may have used saccades to serially sample the array. That said, it is nonetheless noteworthy that performance consistently declined with increasing set-size, suggesting that averaging was at least partially capacity-limited despite the opportunity for foveal sampling. Future studies should address this issue by using brief stimulus presentations (<200 m) to ensure that saccades are not utilised during encoding. In addition, gaze-contingent presentation could be employed to fix the eccentricity of the Gabor elements relative to gaze, enabling more direct assessment of peripheral ensemble processing. Finally, integrating eye-tracking measures would allow quantification of oculomotor strategies and their impact on orientation averaging.

Although perceptual learning can improve orientation discrimination through early-stage enhancements (Matthews et al., 1999; Zhang et al., 2010), our task introduced random variation in both Gabor location and orientation across trials. Such variation likely prevented low-level learning. Prior work shows that perceptual learning generalises across location and orientation when exogenous attentional cues are present (Donovan et al., 2020). In line with this, we propose that learning in our task reflects enhancements in attentional efficiency at a post-integration or decisional stage—rather than early-stage sensory encoding.

Contrary to predictions, we observed no main effect of shock condition on task accuracy or learning rate. However, a shock condition × session interaction was observed for response times, with participants in the performance-contingent shock condition improving faster than those in no-shock or random-shock groups. This may reflect increased urgency or motivation to avoid aversive outcomes.

Temporal patterns differed across groups: performance-contingent shock yielded improvements by Day 3, while other conditions lagged by 1–2 days. These differences suggest subtle effects of punishment contingency on response preparation or motivational state, although not robust enough to affect accuracy.

Self-reported anxiety, as measured by the STAI, was elevated post-training in shock conditions, particularly in the performance-contingent group. Yet, no correlation was found between anxiety scores and behavioural performance. This suggests that STAI may lack sensitivity to trial-level changes or that anxiety’s influence is indirect. Future studies should incorporate objective physiological stress markers (e.g., salivary cortisol (Hellhammer et al., 2009), pupillometry (Ginton et al., 2022)) to clarify these relationships.

A key and unexpected finding was that orientation-averaging accuracy varied systematically with health-bar status—even in conditions where the health bar had no consequences (no-shock, random-shock). Higher displayed “health” predicted better accuracy, regardless of shock contingency. This effect was not accompanied by changes in response times, ruling out a speed-accuracy trade-off.

We propose that the health bar acted as a motivational or attentional cue. Visual feedback has been shown to enhance perceptual learning by directing attention and reinforcing effort (Seitz and Watanabe, 2003; Posner, 1980; Ahissar and Hochstein, 1993). In our task, the health bar may have operated as an implicit performance incentive, encouraging attentional persistence and error avoidance. This aligns with prior findings showing that feedback—even when decoupled from consequence—can shape perceptual strategies (Blank et al., 2013; Eisma et al., 2021; Choi and Watanabe, 2012). While some studies dispute the necessity of feedback for learning (Asher and Hibbard, 2020), our findings suggest that continuous visual feedback can enhance perceptual accuracy independently of external reinforcement.

Time-course analyses further support this view: performance was not different immediately before or after shock events, ruling out the possibility that shock anticipation alone explains the health bar effect.

While the present study offers novel insights into perceptual learning and the effects of aversive stimulation within a VR environment, several limitations should be acknowledged.

First, although our findings suggest that training-related gains occur at a post-integration stage, the possibility that eye movements may have contributed to orientation averaging performance limits our ability to draw definitive conclusions about early-stage mechanisms. Because Gabor arrays were visible until response and no gaze-contingent control was implemented, participants were free to foveate individual elements. Future studies should adopt brief (<200 m) stimulus presentation times and/or gaze-contingent displays to ensure that ensemble averaging relies on peripheral processing and to isolate the contributions of early, retinotopically specific filters.

Second, the lack of objective physiological markers of arousal limits interpretation of the observed elevation in STAI-measured anxiety following shock exposure. While the STAI is widely used, it may not capture transient fluctuations in arousal that occur on a trial-by-trial basis. Incorporating direct physiological metrics such as salivary cortisol, galvanic skin response, or pupillometry would enable more precise characterization of stress responses and their relationship to perceptual performance.

Third, while our findings suggest that the health bar acts as a visual feedback signal that improves performance, we did not include a condition in which the health bar was absent. As such, although the statistical association between health bar magnitude and accuracy was robust—even in conditions where it had no consequences—causal inferences are constrained. Future research should include a no-health-bar control condition and manipulate the timing, visibility, or relevance of feedback to better understand its role in attentional persistence and performance modulation.

Finally, although our immersive VR platform enabled a more engaging and ecologically valid testing environment, it also introduced some variability in response timing and user interaction. Comparing outcomes from VR-based tasks with those obtained from conventional 2D psychophysics paradigms would help assess the generalisability of results across modalities and help establish VR as a reliable platform for vision science.

Taken together, these limitations suggest several promising directions for future work. Larger sample sizes, objective arousal measures, constrained visual exposure, and systematic feedback manipulation will be key to refining our understanding of the interplay between motivation, stress, and perceptual learning.