Meta's 'super-sensing' glasses may fuel robot AI training

Meta is testing a prototype smart glasses model, internally described as "super-sensing", that could capture photos every few seconds and continuously sample audio throughout a wearer's day, according to the Financial Times. The newspaper reported that while Meta says raw footage will not be stored, the metadata extracted from it could still be used to feed AI training systems. Industry analysts argue that this is, in fact, the entire point of the venture.

Google and Meta return to smart glasses

Both Google and Meta previously stumbled in earlier attempts at smart glasses, but both companies are now re-entering the space. At its I/O 2026 developer conference, Google announced Android XR-based glasses built in partnership with Samsung and Gentle Monster.

The smart glasses category was initially marketed as a smartphone accessory for tasks such as navigation and translation. However, the current thinking among analysts is that the real aim of the business is to gather information needed to train robots.

According to industry insiders, egocentric video, which involves recording activities performed by humans from a first-person perspective, holds immense value for training robots in manual dexterity. Tesla, despite possessing around 16 billion kilometres' worth of driving data, does not have this type of information, as its data is vehicle-centric rather than human-centric.

Research conducted by the Shanghai Artificial Intelligence Lab found that robots trained only on simulated data were able to perform sorting tasks successfully 33.3% of the time. When first-person human video was included, this success rate rose to 40%, and when both simulated and human data were combined, it climbed further to 53.3%, even though the first-person videos made up just 8% of the total training data.

A separate study by Peking University, the National University of Singapore, and MIT found that training using first-person video was 24% more efficient at reducing inference errors in robot actions than data collected from remotely controlled robots operated by humans.

DS Investment & Securities researcher Choi Tae-yong argues this explains why Big Tech continues to build cameras into glasses despite thin near-term profits and mounting privacy concerns. He said the companies are not really in the eyewear business, but are instead collecting the data needed to dominate robotics in the future.