Memories.ai builds egocentric video pipeline for robot training

Memories.ai is positioning itself at the data layer of physical AI. The company obtains egocentric video through data collectors and partners, providing standardized capture hardware—described as a modular camera setup—and a uniform process to collect footage consistently. It then processes and annotates that footage with its own models and makes it available through its data infrastructure.

That infrastructure is described as built for physical AI, robotics, world models, vision-language-action (VLA) models, and humanoids. The company has published multiple elements of its research on GitHub and at ICLR, NeurIPS, and EMNLP events.

A key piece of the pipeline is its use of NVIDIA Cosmos. Memories.ai said last month that it has used Cosmos to interpret video entering its data lake, then curates that data to post-train Cosmos for customers' robotics and industrial applications. This loop turns real-world video into models tailored to the tasks customers need to solve.

CEO Shawn Shen said the company is focused on processing its own data rather than collecting video from a wide range of devices and sources. He described the company's vision as accelerating physical intelligence and being part of the robotics game. The longer-term aim is to be a data-infrastructure partner for robotics and physical AI, not just a data vendor.