Memories.ai is solving robotics’ biggest bottleneck by paying people to film their chores

Large language models had the entire internet to read, but if you want to teach a humanoid robot how to fold laundry, restock a shelf, or fix a leaky pipe, scraping YouTube isn’t going to cut it. Most internet video is shot from a third-person perspective, completely missing the nuanced hand-eye coordination, tool usage, and real-time decision-making required for physical labor.

It’s a data drought that is actively stalling the deployment of physical AI. According to Stanford’s 2026 AI Index, robots can successfully complete 89.4% of household tasks in simulations, but that success rate plummets to a dismal 12% in actual homes.

Enter Memories.ai, a startup founded by former Meta Reality Labs researchers that is building what it hopes will be the foundational data infrastructure for frontier robotics. 

This week, the company announced a three-pronged approach to solving the egocentric data gap: an app called Memories Upload, a lightweight wearable camera dubbed the LUCI Pin, and a Large Visual Memory Model (LVMM) to process it all.

The premise is essentially a gig economy for robot training data. Users strap on the LUCI Pin, a tiny, sub-45-gram camera that can be mounted on a wrist, chest, or cap, and record themselves doing everyday physical tasks. Through the Memories Upload app (available now on iOS and Android), contributors are paid for their accepted footage, with higher payouts allocated for high-demand, hands-on tasks.

Memories.ai claims to have already quietly amassed over 1 million hours of first-person video, with an aggressive target of hitting 10 million hours by March 2027.

“Anyone can accumulate hours of video,” Memories.ai CEO Shawn Shen said in a statement. “What moves a robot forward is video that has been checked, indexed, and described well enough to learn from, and that is the layer we build.”

Shen and his co-founder, CTO Ben Zhou, know a thing or two about first-person data. Both came from the AI team behind Meta’s Ray-Ban smart glasses, giving them a front-row seat to exactly how much first-person footage a machine needs to understand human work. Shen also brings academic pedigree to the table, holding a PhD from Cambridge and running a machine intelligence lab at the University of Bristol.

For robotics labs, raw video is mostly useless without heavy annotation. Memories.ai addresses this with its Video Data Lake and LVMM. Once a user uploads a clip, the model automatically scrubs it for quality and indexes the footage, generating step boundaries down to milliseconds, action labels, object masks, and hand-contact states. To handle the massive compute pipeline, the startup is leaning on NVIDIA’s Metropolis and Cosmos models, alongside on-device processing powered by Qualcomm’s Snapdragon platforms.

Memories.ai says privacy is baked into the pipeline: contributors are strictly barred from filming private screens or documents, must secure consent from co-workers or housemates, and can delete any clip before payout. Furthermore, the company claims all faces are blurred before the footage ever touches a training dataset, and audio is never sold.

The company is currently offering its LUCI Pins to enterprise partners and research labs via an academic program, alongside its consumer app. 

The company is backed by a roster of investors including Susa Ventures, Samsung Next, Crane Venture Partners, Fusion Fund, Seedcamp, and Creator Ventures.

Memories.ai is betting that the path to generalized robotics isn’t going to be built in a simulation, it’s going to be crowd-sourced, one dishwashing session at a time.