Field report N-55 / K2 · sheet /expansion
Expansion & MarketsFiled from the territory
Ex-Meta Researchers Pay Gig Workers to Train Robots With Body Cameras
Memories.ai, founded by ex-Meta researchers, pays gig workers to record first-person task video and claims three top-five AI labs as customers. Nvidia helps make the footage searchable.
- Grid
- N-55 / K2
- Surveyed
- November 17, 2026
- Reading
- 4 min read
- Correspondent
- Sophie Lindqvist

Field notes
- Memories.ai recorded over 1 million hours of POV video in its first beta month and targets 10 million hours by March 2027.
- The startup raised an $8 million seed round in July 2025 led by Susa Ventures, with Samsung Next participating.
- Its LUCI Pin camera weighs under 45 grams and can be paired in twos for stereo depth capture.
- Customers reportedly include three of the top five frontier AI labs, unnamed due to NDAs.
- The company uses Nvidia's Cosmos models and Metropolis video search tools to label and index footage.
Memories.ai recorded more than 1 million hours of first-person video in its first month of beta operation, and it wants 10 times that — more than 10 million hours by March 2027. The San Francisco startup, founded by former Meta researchers Shawn Shen and Ben Zhou, is betting that gig workers wearing tiny cameras will supply the raw material for the next generation of robots.
The mechanism is straightforward. The company is launching Memories Upload, a global gig-work platform where contributors take paid video-recording tasks. Its LUCI Pin camera weighs under 45 grams, attaches magnetically to mounts on the head, chest or wrist, and captures point-of-view video of people preparing food, stocking shelves and repairing equipment. Two cameras can pair for stereo capture, the dual-view setup that helps models estimate depth. What the company does not say is how much contributors are paid, or how it will sustain interest long enough to reach the 10-million-hour figure.
Who is buying this footage?
Memories.ai says its customers include three of the top five frontier AI labs, alongside robotics developers — but it declined to name any of them, citing non-disclosure agreements. That claim rests entirely on the company's word. The startup does have verifiable backing: an $8 million seed round announced in July 2025, led by Susa Ventures, with Samsung Next among the investors.
The demand-side logic is real, even if Memories.ai's part in it is not independently confirmed. Figure, the humanoid robotics company, pays people to record everyday tasks through its Index platform for its Helix AI models. Physical Intelligence, which builds AI software that controls different types of robots, has tested combining human video with teleoperation data (footage recorded as people control robots remotely). Its co-founder Sergey Levine wrote in a blog post that human video could provide "a source of knowledge about what can happen in the real world," complementing data robots collect themselves.
For a robot learning to make coffee, recordings show how a person grips a cup, positions it under a grinder and moves it without spilling. Footage from different kitchens, tools and working styles exposes models to more variety. One gap remains unresolved: developers still work out how to translate observed human movements into actions suited to a robot's own body.
What does this change for robot training?
The short answer: searchability. A million hours of video is useless if developers must watch it all. Memories.ai uses Nvidia's Metropolis Video Search and Summarization tools for video streaming and processing. Its captioning model is fine-tuned from Cosmos, Nvidia's family of models for understanding the physical world, and Cosmos-based models label actions in the footage. The company also post-trains Cosmos (the process of further training a model on curated video for a customer's specific tasks).
The result, the company says, is descriptions tied to specific moments, letting developers search for an action and retrieve examples without hours of viewing. Rick Zhuang, director of product and engineering for Cosmos at Nvidia, which works with Memories.ai, told Fortune: "Recognizing an activity is different from understanding exactly how a hand grips a tool, moves an object or completes a step, especially when the camera is moving and objects are partly hidden."
That difficulty is the honest part of the story. Zhuang said trustworthy labels are particularly hard when hands make subtle movements, objects are partly hidden, or the boundary between two actions is unclear — cases that still require human review. Adam Scraba, Nvidia's head of physical AI ecosystem for vision agents, called specialized industrial tasks one of the field's hardest challenges. "The training dataset for Vision AI and video analysis becomes critically important for basically all Physical AI solutions," he told Fortune. "Developers need to find the right corner-case examples, trust their labels and trace each annotation to the moment it describes."
CEO Shawn Shen said customers increasingly want footage covering a wider range of work environments, tasks and camera angles — two people making coffee in different kitchens supply useful variation even for the same task. Developers also request wrist-level footage to catch interactions a head-mounted camera misses, such as exactly when a hand touches an object.
The privacy caveat
Paying contractors to record workplaces carries an obvious risk: footage can capture coworkers, passersby and sensitive information. Memories.ai requires camera wearers and coworkers who regularly appear to sign consent forms. Before footage leaves its country of capture or reaches customers, automated tools obscure other people and mask identifying details such as badges, screens and license plates. Teams also manually review a sample of frames per batch. Whether that process catches every leak of sensitive detail at 10-million-hour scale is a question the company cannot yet answer.
via fortune.com/feed/ (Source)
More from Sophie Lindqvist
Show full bio
Correspondent covering money, funding and market shifts around corporate strategy and chief-executive leadership at CEO Frontier.
483 articles
Adjacent sheets
- Behind-the-Scenes Videos Become Advertising's Anti-AI Weapon
- VCs question neolab math as funding targets reach $50 billion
- AI Capex Nears $700 Billion, and the Cash Burn Worries Investors
- Cursor Acquisition Puts US Startup M&A on Record Pace
- Bain backs physical AI amid a $27.6 billion robotics funding surge