Memories.ai turns gig workers with cameras into robot trainers

Training a humanoid robot to make a cup of coffee or stock a warehouse shelf sounds straightforward until you realize how little data actually exists. Teleoperating robotic arms with VR rigs is wildly expensive and painfully slow.

Memories.ai thinks the solution is already all around us: everyday humans doing their jobs. The San Francisco startup founded by ex-Meta researchers Shawn Shen and Ben Zhou just launched Memories Upload, a platform that turns daily work footage into robot training data.

Instead of building robotic arms, Memories.ai built a global gig-work engine. Workers get paid to film real tasks, and robotics engineers get structured data ready to feed into their models.

A 45-gram magnet pinned to your wrist

To capture clean data without getting in the way, the startup engineered a wearable camera called the LUCI Pin. It weighs under 45 grams and snaps magnetically into mounts for your head, chest, or wrist.

That wrist view turns out to be crucial. Head cameras often lose sight of tiny contact moments when a worker grips a tool or slides a cup under a grinder. Putting a camera directly on the wrist captures the exact micro-adjustments robots struggle with.

Two pins can also pair up for stereo recording to calculate 3D depth. During its first month in beta, the network pulled in more than 1 million hours of POV video. The team plans to hit 10 million hours by March 2027, already spanning over 1,000 tasks across 20 industries on five continents.

Making human chores searchable for neural nets

A massive pile of video is useless if engineers have to watch it all. Memories.ai teamed up with Nvidia to turn the flood of footage into a queryable action library.

Using Nvidia Metropolis search tools and models fine-tuned from Nvidia Cosmos, the platform labels each hand frame by frame. It tracks 3D joint poses, writes dense natural-language captions, and maps action boundaries.

If a lab needs 500 examples of a human stabilizing a slippery mug while flipping a toggle, they do not scrub through hours of footage. They query the moment via an API or MCP server and pull the exact clips into their training pipeline. Memories.ai says three of the top five frontier robotics labs are already customers.

Why it matters

Physical AI has run into the exact data wall that text models hit a few years ago: running out of high-quality internet data. Robots cannot learn common sense by reading web pages; they need to see how gravity, friction, and human hands interact with real materials.

Memories Upload is a clever end-run around the bottleneck. Instead of manufacturing thousands of robots to collect their own clumsy data, the company turns millions of gig workers into distributed data collectors using cheap, lightweight hardware. For roboticists, it replaces slow teleoperation with an on-demand API.

The real test will be embodiment translation. A human wrist moves with soft tissue and twenty-seven bones; a robotic gripper is rigid metal and motors. Seeing an action is not the same as executing it on very different hardware, and developers still have to bridge that kinematic gap.

Still, with players like Figure and Physical Intelligence racing to turn human demonstrations into robot motor skills, the bottleneck is shifting entirely to high-volume, high-diversity real-world video. Memories.ai just built the first scaled pipeline to sell it.