An examination of active inference in autistic adults using immersive virtual reality

The integration of prior expectations, sensory information, and environmental volatility is proposed to be atypical in Autism Spectrum Disorder, yet few studies have tested these predictive processes in active movement tasks. To address this gap in the research, we used an immersive virtual-reality racquetball paradigm to explore how visual sampling behaviours and movement kinematics are adjusted in relation to unexpected, uncertain, and volatile changes in environmental statistics. We found that prior expectations concerning ball ‘bounciness’ affected sensorimotor control in both autistic and neurotypical participants, with all individuals using prediction-driven gaze strategies to track the virtual ball. However, autistic participants showed substantial differences in visuomotor behaviour when environmental conditions were more volatile. Specifically, uncertainty-related performance difficulties in these conditions were accompanied by atypical movement kinematics and visual sampling responses. Results support proposals that autistic people overestimate the volatility of sensory environments, and suggest that context-sensitive differences in active inference could explain a range of movement-related difficulties in autism.


Post-Hoc Analyses of Gaze Tracking Behaviours
In our main analysis, we found that autistic people use interceptive gaze behaviours that are typically associated with high-uncertainty conditions. While neurotypical participants tended to pursue expected ball trajectories more closely than unexpected ones, autistic individuals appeared to sample both cues with similar levels of accuracy (or 'error'; Fig. 5). These differences were not a result of any significant gaze tracking deficits, nor were they accompanied by any alterations in the timing or amplitude of key saccadic eye movements ( Fig. 3 & 4). Instead, they likely reflect aberrant behavioural surprise computations (see discussion). However, it is possible that atypical gaze responses in ASD stem from underlying attentional and/or oculomotor impairments that determine one's ability to engage, disengage, and shift attention in coordination with fast-moving sensory cues. This supplementary analysis evaluated such a possibility, through a series of exploratory gaze data comparisons.
First, we examined whether there were any broad, autism-related differences in the frequency of gaze shifts during the study. The total number of saccades and fixations were assessed for each trial and averaged for both conditions, before being entered into separate mixed-model ANOVAs. Here, atypically-low frequencies might indicate impaired disengagement or shifting of visual attention. Conversely, any inaccuracies in continuous smooth pursuit or goal-directed saccades would likely demand a relatively high frequency of corrective gaze shifts 41 . Neither of these data patterns emerged, with ANOVAs showing null significant group ( Next, we explored whether autism-related gaze differences in our task simply reflect impaired motion tracking abilities. If this was true, then we would expect particular difficulties to emerge on trials with the greatest ball velocities. As such, we extracted the average postbounce distance (i.e., tracking error) between gaze and ball pitch vectors during all 'bouncy' ball trials. Here, any fundamental motion tracking impairments would result in generally high gaze-ball differences, regardless of whether the high-elasticity ball speeds are expected or uncertain. Therefore, we averaged bouncy-ball trial values from both conditions and compared them between groups using an independent t-test. Group averages were not statistically significant (t(70) = .41, p = .68, BF10 = .27), indicating that autistic and neurotypical participants had similar post-bounce tracking abilities with regards to the fastmoving ball cues.
Overall, this exploratory analysis finds little support for the notion that autism-related gaze differences in our study result from broad impairments in attentional and/or oculomotor control. Instead, they reinforce proposals that sensorimotor difficulties are likely related to context-sensitive mechanisms (e.g., trial-by-trial computations about likely ball bounciness probabilities and dynamic volatility estimations). Further empirical scrutiny is required, however, before any definitive conclusions can be made.

Exploratory Analyses of Gaze Fixation Variability
Autistic participants predictively positioned their gaze at a higher location than neurotypical individuals when the virtual balls were bouncing (Fig. 3). Such data patterns may be consistent with proposals that autistic people overestimate environmental volatility 31 : agents who perceive that the world is more changeable will increasingly update their long-term predictive models according to recent (high-elasticity) sensory information. Computationally, this would reflect an increase in learning rate 15 , though such conclusions require further scrutiny (see main discussion).To initiate this enquiry, we conducted an exploratory analysis into the variability of participants' gaze fixation behaviours. Specifically, we looked at the standard deviation of bounce fixation locations (pitch angles) shown in each condition. If participant's visual sampling behaviours were being heavily driven by long-term prior expectations, then this trial-by-trial variability should be relatively low. On the other hand, larger standard deviations would indicate that gaze fixations are being strongly influenced by recent sensory data (i.e., changeable ball elasticity profiles from preceding trials).
ANOVA highlighted a significant effect of condition (F(1,70) = 5.63, p = .02, np 2 = .07, BF10 = 3.03) for these standard deviation values. Participants showed increased variability between stable and volatile conditions (Supplemetary Fig. S6), as indicative of an increased updating of prior models (i.e., a higher learning rate). While no significant interaction effects emerged (F(1,70) < .001, p = .99, np 2 < .001, BF10 = .27), the ASD group displayed generally higher trial-to-trial variability than their neurotypical counterparts (F(1,70) = 38.47, p < .001, np 2 = .36, BF10 = 3.11*10 5 ). This tendency to increasingly update bounce fixation locations on a trial-by-trial basis is in line with proposals that autistic people are over-reactive to environmental change and reinforces the potential role of aberrant volatility processing in ASD 31 . Research may wish to explore this topic further, by using sophisticated computational models of gaze fixation behaviours to estimate volatility-based learning rate parameters. Figure S6: Trial-by-trial standard deviation values corresponding to the spatial location (pitch angle, °) of bounce gaze fixations in each condition. NT: neurotypical; ASD: autism spectrum disorder. Two extreme values were identified and are represented as light grey circles (note: removal of these cases do not affected the overall pattern of results).