Autism social robots not ready as main therapy, a review of pilot evidence finds

The authors say risks include sensor discomfort in children with sensory sensitivities, cognitive overload and fatigue during prolonged VR exposure, continuous collection of sensitive physiological data, and unpredictable AI dialogue. They also warn that training data overrepresent English-language and neurotypical interaction, which could widen inequity across racial, linguistic and socioeconomic lines.

The authors propose a five-module validation roadmap: pediatric sensor calibration standards, standardized outcome measures, multisite randomized trials, follow-up of at least 6 months, and human-supported safety design. The roadmap is a theoretical construct built from fragmented pilot data and still needs prospective testing.

The review builds on a bibliometric co-occurrence map of 6,907 PubMed records on pediatric autism and social robotics, searched on 4 December 2025. The authors used the network structure to pick three blind spots: areas visible in the literature but not yet validated or standardized for clinical use. They describe the three technologies as experimental adjuncts, not independently validated interventions.

For wearable-integrated robots, feasibility has been shown, and home deployments sustained engagement over weeks. Combining several sensors increases computational load and signal noise, which reduces robustness outside controlled settings. The authors add that expensive wearables could favor well-resourced families, and suggest open-source platforms with caregiver dashboards as a route to wider access.

For VR-robot hybrids, children often prefer VR in short sessions, but long-term engagement appears to peak under hybrid conditions. Controlled pediatric studies have not yet compared outcomes such as narrative engagement between VR-robot storytelling and robot-only approaches. No consensus exists on age-stratified VR exposure limits, and anxiety reduction from reinforcement-learning optimization remains unproven.

For generative AI robots, a scoping review cited by the authors found small samples, data biases, limited validation and model hallucinations, with most applications at proof-of-concept stage. The authors call for pediatric-specific training data, multi-layer semantic filtering, real-time clinician override and use within applied behavior analysis or speech-language pathology. They also say a 6-month minimum follow-up is needed to separate lasting skill gains from novelty effects.