Beyond Scene Description: Multi-Agent Orchestration for Non-visual Access to Virtual Worlds

Summary

The paper proposes MetaBlind, a design-stage architecture for nonvisual access by blind and visually impaired (BVI) users to live multi-user virtual environments. Its diagnosis: assistive tools solve isolated tasks (object naming, scene description, routing), while a virtual room floods the user with concurrent channels, and a single narrating assistant merely trades a visual barrier for an auditory one. MetaBlind splits responsibility across eight specialized agents — perception, navigation, social, object interaction, communication, safety/trust, memory, personalization — none of which may address the user; all publish candidate items into a shared accessibility context. An Accessibility Orchestrator scores each candidate on safety relevance, goal relevance, urgency, confidence, user relevance, and estimated listening load (Eq. 4), then selects under a load budget (Eq. 5), formulated as constrained maximization where safety-gated items are mandatory and an empty selection is legitimate. Algorithm 1 fixes the cycle: deduplication, safety gating before scoring, greedy P/L selection, context write-back. Section V defines an evaluation protocol against a model-matched single-agent baseline plus a no-budget ablation, and names three outcomes that would refute the design. There are no measurements: no prototype, no simulation, no user study; weights, budget fitting, latency, and the safety gate's policy are unspecified, and Section VI concedes environment access and deliberation delay as binding constraints. The contribution is a falsifiable formulation, not evidence.