Blind and Low Vision Users Took the Lead in Virtual Reality

A group of blind and low-vision adults entered two virtual parks, learned what was around them, and then took the lead.

In a CHI 2026 study, 16 participants explored two researcher-built social-VR environments and tried guiding small groups through them. Thirteen chose to use a conversational AI guide; three preferred to rely on residual vision. The authors report that every participant shared accurate information about the parks during their tours. Ten guided their group to at least two landmarks—including three who reached more than two—while six reached one.

That result points toward a richer idea of accessibility. The goal is not merely to help someone follow a route. It is to give them enough command of a place to become the person others follow.

Research references: [1]

From assistance to agency

The research team built a prototype guide that participants could speak with inside virtual reality. It could describe the surrounding scene, answer questions about nearby objects and suggest navigation instructions. Participants could choose a dog, robot or human-like guide persona.

The system combined speech recognition, a large language model and screenshots from the virtual environment. It ran on Meta Quest 2 headsets in two custom parks, plus a tutorial environment. After exploring, each participant led a tour with researcher confederates acting as group members.

This structure matters. Many accessibility tools are evaluated on whether a person can complete a task. Here, the study also asked whether participants could build a mental model of a place, decide what mattered and communicate that knowledge socially.

The authors report that all 16 participants shared accurate information about the parks. Some also improvised extra details or supplemented what the system had told them, so the finding should not be read as a claim that every statement in every tour came directly from the guide or was independently verified.

Research references: [1]

What the prototype got right—and missed

Across 476 participant questions, the prototype returned 301 answers the researchers classified as correct: 63.2%. That is useful enough to show the direction of travel, but far from the reliability a mature mobility or safety tool would need.

Response time also depended on the task. The paper reports averages of 6.1 seconds for scene descriptions, 11.2 seconds for navigation answers and 6.8 seconds for other queries. In conversation, that delay can shape whether an assistant feels like a companion or an interruption.

Among the 13 participants who used the guide, nine rated it useful, nine said they enjoyed using it and nine said it helped them understand the scene. Navigation was harder: five rated it easy, while eight were neutral or disagreed. Nine said they felt comfortable using the guide in front of others, while four expressed some discomfort.

The denominator is important. These ratings do not include the three participants who chose to navigate with residual vision instead of the AI guide. Their choice is not a failure of accessibility—it is part of accessibility. A useful system should offer support without forcing people to disclose a disability or adopt a single way of moving through virtual space.

Research references: [1]

A prototype, not a finished product

This was a small, one-session study: 16 adults recruited with help from LightHouse for the Blind and Visually Impaired San Francisco, each participating for about 90 minutes. The environments were two researcher-built parks, the social companions were research confederates, and the study had no randomized control condition or long-term deployment.

The technology stack also began with 2023-era components, so the exact accuracy and latency figures should be treated as measurements of this prototype—not permanent limits on what newer systems can do.

The work was supported in part by a U.S. National Science Foundation grant. The paper states that its findings and conclusions are those of the authors and do not necessarily reflect the agency's views.

Research references: [1]

Design for the person in charge

The most promising lesson is not that an AI can narrate a virtual world. It is that accessibility can change a person's role inside that world.

When an interface supplies useful context, respects a person's preferred way of navigating and leaves disclosure under their control, assistance can become agency. A visitor can become a host. A participant can become a guide. And a space designed around sight can become somewhere more people can lead.

That is still a research result, not a product guarantee. But it offers a concrete design target for social VR: build systems that help people choose how they participate—and make it possible for them to take the lead.

: Collins J, Lin S, Liu T, Won AS, Azenkot S. “AI Guide: Supporting Blind and Low Vision People in Social Virtual Reality.” *Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems*. DOI: 10.1145/3772318.3791143. : The authors’ open manuscript and revision history are available on arXiv:2603.09964.

Research references: [1] [2]