Startup by former Meta researchers launches platform for training robots
Memories.ai launches global platform Memories Upload: participants will record paid tasks on wearable cameras to train robots.
The platform will collect videos from people cooking meals, restocking shelves, or repairing equipment. For filming, the company developed the LUCI Pin camera, weighing less than 45 grams: it can be attached to the head, chest, or wrist. Two cameras allow for stereo filming and assessing the depth of the scene.
According to Memories.ai, in the first month of beta testing, the network recorded over 1 million hours of first-person video. The company expects to increase the volume to more than 10 million hours by March 2027. The startup was founded by former Meta researchers Sean Shen and Ben Zhou; in total, it has raised $16 million in venture funding.
Such recordings can show a robot the details of actions—for example, how a person picks up a cup, places it under a coffee grinder, and moves it without spilling the drink. Video materials from different kitchens, with various tools and angles, help developers expand the set of examples, though translating human movements into the actions of a specific robot remains a separate task.
Memories.ai uses Nvidia tools for processing and searching videos, as well as Cosmos models for describing and labeling actions. This allows finding specific moments in recordings by action without reviewing the entire material. According to the company, three of the top five frontier AI labs are among its clients, but their names are not disclosed due to non-disclosure agreements.
Filming may involve the private information of colleagues, passersby, and clients. Before transferring materials outside the country or to clients, the company requires consent from operators and regularly filmed colleagues, obscures faces and data such as badges, screens, and license plates, and then manually checks a portion of the footage.
