RCT (n=47) - three short VR sessions helped autistic teens and adults respond more effectively in live police encounters, vs matched video

Background

Autistic adolescents and adults face disproportionate risk during interactions with police officers. A 20% police-contact rate by their early 20s, 5% arrested, elevated injury and death rates, and a system that is slow to retrain - all documented across the literature the authors cite. While the long-term solution is better police training, there is also value in supporting autistic individuals to practice these high-stakes encounters before having them.

Two existing approaches: video modeling (BeSAFE The Movie, used widely) and immersive VR. Video modeling teaches the rules but cannot replicate the visual, sensory, or emotional pressure of an officer in front of you. Immersive VR can - and is highly customizable across context (day/night, indoor/outdoor, calm/loud), officer characteristics, and number of officers.

This RCT is the first head-to-head test.

What the researchers did

Forty-seven verbally fluent autistic participants (ages 12-60, IQ ≥ 75 on WASI-II) were randomly assigned to one of two interventions, stratified by age (within 10 years), sex, SCQ score (within 5 points), and IQ (within 10 points):

1. Floreo Police Safety Module (PSM) - a smartphone-based immersive VR application: an iPhone worn in a lightweight headset and linked via Wi-Fi to an iPad held by an intervention monitor. The iPad displays the visual scene the participant sees, a text-based instruction window, and control buttons for adjusting the virtual officer’s behavior in real time. In this study, intervention monitors were study staff with a bachelor’s degree in psychology or neuroscience. Three 45-minute sessions, ~12 minutes of active VR per session (six 2-minute police-interaction trials with breaks between trials), with real-time implementer feedback driven by a cognitive-behavioral framework: collaborative goal-setting, post-trial data-driven feedback, end-of-session debrief. Dialogue ranged from easy (perceived as adult: “How’s your day going? What’s your first name? Do you have any ID on you?”) to difficult (perceived as minor: “I need to see your hands. What’s your birthday? Where were you last night?”). Sessions averaged 9.25 days apart (SD 6.2; range 1-37).

2. BeSAFE The Movie - an established video modeling intervention with educational videos plus role-play with a familiar adult (parent, teacher, clinician). The current study used the first four BeSAFE lessons: Laws Help us BE SAFE; Law Enforcement Officers Help Us BE SAFE; Uniforms and Safety Tools; Stay Calm When You Meet the Police. Three 45-minute sessions matched VR in dose.

Pre-test (Visit 1): A 3-5 minute live simulated police interaction with a study-staff confederate wearing a badge. Audio/video recorded for masked coding. Post-test (Visit 3): A 3-5 minute live interaction with an actual uniformed police officer or security guard. The authors explicitly flag the pre/post differences as adding noise but argue this could not account for the between-condition effects they observed.

Videos were coded in 15-second intervals using ELAN, by study staff masked to condition. Reliability procedure: two coders viewed and discussed 5 recorded sessions to reach consensus on code definitions, then separately coded 20 sessions for inter-rater reliability. Cohen’s kappa: Overall Behavior 0.67, Appropriate Response 0.60, Orienting 0.53, Fidgeting 0.68. All remaining sessions were coded by one reliable coder. The four primary video-coded behaviors were:

• Fidgeting - frequency and size of repetitive actions
• Appropriate response - listening, verbally responding, following directions
• Orienting - eyes and body oriented to the officer
• Overall behavior - gestalt of the above during each 15-second interval

Linear mixed-effects models tested condition × time interactions, controlling for age (centered), sex, and full-scale IQ (centered), with random effect of participant ID. Power analysis indicated N = 47 detects effect sizes of Cohen’s d = 0.54 or larger. Community Involvement: family members of autistic people were involved in conceptualization, design, and implementation of the study; autistic people and family members of autistic people were involved in the development of the VR intervention itself; and family members of autistic people contributed to interpretation of the findings.

What they found

Primary outcome - fidgeting: Significant condition × time interaction (estimate 0.47, p = 0.005). The VR group’s fidgeting decreased significantly (estimate -0.44, p = 0.002); BeSAFE group did not change (p = 0.99).

Appropriate response: Interaction term not significant (p = 0.07), but planned comparisons showed significant pre-to-post improvement only in the VR group (p = 0.02).

Overall behavior: Same pattern - interaction non-significant overall (p = 0.23), but VR group improved significantly (p = 0.02) and BeSAFE did not.

Orienting: No effect either group. Authors attribute this partly to limitations of single-camera video coding, which made it hard to score where participants were looking.

Self-reported knowledge and comfort: Improved significantly in both groups (no group difference). Knowing about police interactions helped how participants felt regardless of format.

Police officer ratings: VR group rated as showing better eye contact post-intervention (p = .05). No other significant officer-rated differences.

The authors interpret reduced fidgeting as a likely marker of reduced anxiety/arousal during the live encounter - a conclusion they hedge appropriately, since they did not directly measure physiology.

Why this matters

For the Evidence Hub, this is an important addition because it tests something most VR-in-communication-differences research does not:

• A direct comparison with an active, evidence-based control rather than a waitlist or no-treatment condition. The VR-specific effects survived this comparison.
• An ecologically valid post-test. Most VR studies measure within-VR or with the same examiners pre/post. Here the post-test was with real police officers, in person. This is one of the strongest direct-transfer demonstrations in the immersive-VR-CSD literature.
• Pre-registration and masked coding - the design closes off the most common methodological criticisms.
• A clinician-monitored, cognitive-behavioral protocol. The VR was not used as a stand-alone tool; the implementer was actively shaping the practice. This matches the social-model framing already used in Therapy withVR clinical work.

For Therapy withVR specifically: this study reinforces a thesis we have argued elsewhere on the basis of more limited evidence - that immersive VR practice with a clinician alongside can carry over to real-world encounters that involve emotional pressure. It does not test Therapy withVR (the product is Floreo’s PSM), but it adds to the broader evidence base for the approach.

Limitations

The authors are explicit about what this trial does and does not establish:

• Sample size is small (N = 47). Several outcomes had non-significant interaction terms despite within-group improvements; replication with larger samples will help.
• The sample was predominantly White (n = 34 of 47). The authors acknowledge this is exactly the demographic where police interactions carry less differential risk than for urban-living autistic people and autistic people of color, who they identify as the most urgent population for future research. Race/ethnicity did not differ by condition.
• Co-occurring intellectual disability was excluded via the IQ ≥ 75 inclusion criterion. The authors note this is “a common co-occurring condition for autistic individuals” and limits generalizability accordingly.
• Pre/post officers differed - pre-test was study staff with badges, post-test was actual officers/security personnel. This adds noise but, the authors note, cannot account for the between-condition differences that emerged.
• Effect sizes are modest. Even the primary outcome is a relatively small absolute change in fidgeting; the authors argue that even small reductions in suspicion-evoking behavior during high-stakes encounters can have significant real-world impact.
• Coding was largely single-coder. Only 20 sessions were double-coded for the reliability assessment; the rest were coded by a single reliable coder. Cohen’s kappa for orienting (0.53) was the lowest of the four outcomes, which the authors attribute to single-camera video setup making it hard to score where the participant was looking. Future work should use multi-camera setups or head-mounted eye trackers.
• Significant conflict of interest. Three authors are Floreo Inc. employees (Solorzano, Turnacioglu, Ravindran) and a fourth (McCleery) has served as a paid consultant for Floreo. The Author Contributions show that these same authors were involved in Conceptualization, Methodology, Funding Acquisition, and Supervision - not just passive co-authorship. The trial was pre-registered (NCT03605368, 30 July 2018), coders were masked, and the comparator was an established video modeling intervention, all of which mitigate but do not eliminate the COI risk. It should be flagged when the study is cited.
• No direct anxiety measurement (skin conductance, heart rate). The interpretation that reduced fidgeting reflects reduced arousal is plausible but not directly tested. The authors explicitly raise an alternative interpretation: “VR-based practice somehow enabled autistic participants to temporarily mask their anxiety in order to get through police officer interactions with less fidgeting despite high levels of physiological arousal.” They note that because the cognitive-behavioral protocol let each participant choose their own skill targets, any masking would be “a conscious decision of their own volition” - but argue that the real-world benefit (reduced suspicion-evoking behavior during police interactions) accrues either way.

How this fits with the wider Evidence Hub

This RCT is the most methodologically robust direct demonstration to date of immersive-VR practice transferring to real-world social interactions in autism. It complements:

• Nudelman et al. 2026 scoping review - confirms that ecologically valid responses in VR are a consistent finding across populations; this RCT extends that to a comparative trial design.
• Brundage et al. 2007/2015/2016 - established the foundational claim that speech behavior in VR audiences correlates strongly with speech in real audiences. This RCT shows the corresponding real-world transfer for non-speech behavior (fidgeting, response) in autistic encounters with police.
• Moïse-Richard 2021, Brassel 2023, Hansa 2025 - other VR studies in adjacent populations.
• Al-Nafjan et al. 2021 - autism + VR feasibility work.

Where the field’s evidence base previously rested mostly on within-subject designs and case studies (per Nudelman 2026), this RCT contributes one of the genuinely comparative, pre-registered, masked-coded data points.