Zhejiang University Publishes First Survey of Egocentric AI for Smart Glasses

A team from Zhejiang University has published what it says is the first systematic survey of "first-person intelligence" for smart glasses. The title of the work sets out its trajectory: from "seeing" to "acting." The Chinese outlet Jiqizhixin (机器之心) reported on the survey. The authors' starting point will be familiar to anyone who has tried on Ray-Ban Meta or similar devices: the average user still thinks of smart glasses as a camera with speakers and a voice assistant. The survey argues that this picture is out of date.

What the survey covers

According to the announcement, the authors connect four technology areas: egocentric vision (video from a head-mounted camera), multimodal large models, human-computer interaction and embodied intelligence. At their intersection, glasses stop being an accessory for capturing footage and displaying information. They become an intermediate link between a person's perception, long-term memory and actions in the real world. The end of that sentence is cut off in the available excerpt, so the full structure of the taxonomy cannot yet be reconstructed.

The announcement does not name the number of papers reviewed, a list of benchmarks or specific models. That matters for assessment: a survey is valuable for its coverage and honest classification, and neither can be verified without the full text.

Why first-person view is a separate problem

Models trained on third-person video transfer poorly to head-mounted footage. The camera jerks along with the wearer's neck, hands block half the scene, and an object can appear and vanish for minutes at a time. As a result, the field depends on dedicated datasets. Ego4D, for example, contains about 3670 hours of video from 931 participants in nine countries. It has become the de facto standard for tasks such as "where did I leave my keys" or "what was I doing before the call."

Memory is the central theme here. Glasses that simply describe a scene are useful for a few seconds. Glasses that accumulate a continuous stream and answer questions about the past change how the device is used. They also raise the sharpest questions about privacy and battery drain.

What is known about the survey:

• Authors: a research group at Zhejiang University.
• Format: a survey paper, presented as the first on this topic.
• Framework: "from seeing to acting" — perception, memory, interaction, execution.
• Technology base: egocentric vision, multimodal LLMs, HCI, embodied AI.
• Not disclosed: the number of papers, benchmarks, model comparisons, and estimates of latency and power consumption.

What developers and buyers should check

For teams building applications for smart glasses, the survey is useful as a map of tasks. Start by deciding which level your product operates at: recognition, memory or action. These put three different loads on the model, the battery and the network link. For anyone choosing a device, the practical criterion is simple: don't ask about camera resolution, ask where the stream is stored, how long it is kept and what is processed on the device itself.

Researchers should wait for the full text and check three things: how well the survey separates academic results from industry demos, whether it offers a unified scheme for comparing benchmarks, and how it describes privacy risks. If those sections are thorough, the work will serve as an entry point for newcomers to the field. If not, it is more a manifesto for the direction than a reference.

For the industry, the signal is clear: Chinese universities are systematically claiming the "AI on the body" niche while the market is still taking shape.