Lets robotics teams collect high-quality first-person video of everyday tasks
The Hook
“For robots to learn how to work in the real world, they need vast internet scale of first-person or egocentric video data. And those video data will be showing how people get things done.”
Product Reveal
“Today, we are announcing Memories Upload.”
Feature Teaser
“We built Luci Pin. It's a lightweight modular camera that is designed to make first-person video capture part of your everyday's work. It is truly atomic and modular. You can put it under the wrist here, or here, or around the chest here, and you can put it on a hat so that you can record egocentric video, or you can put it on the fridge to record a third-person video data. However, hardware is only a small part of the vision.”
Feature Teaser
“The Memories Upload comes with a mobile phone app first. In the app, once you claim the task, for example in our case it is making a coffee, you should just put your pin on naturally and doing the task naturally. Once you finish the task, the videos will automatically get uploaded to the cloud, and our AI will run the videos automatically. It will get analyzed and then it will return you with the feedback of how well you have performed the task and how good the video is. And then, of course, you will get your corresponding rewards.”
Feature Teaser
“There is a whole video data lake that we want to fill in, one cell at a one time. And the cell is defined by three things: one is the task, second is the environment, and the third is the hardware specs. Now, the real work starts after those videos got uploaded to our video data lake. Each clip of those videos will be automatically indexed into different structured data, including hand pose, head pose, object mesh, action segmentation, dense action captioning, and etc. And all of it is done by our own trained model, the Large Visual Memory Model. It's not a labeling tool, it's a world indexer. It indexes the physical world frame by frame.”
Feature Teaser
“Today, we have deployed over 5,000 Luci Pins all over the world, and we have collected over 1 million hours of videos in just 2 months. Every month, there are hundreds of thousands of hours of memories getting uploaded to our video data lake, and then get indexed by our Large Visual Memory Model into those structured data. And we share those indexes alongside with those videos with frontier robotics companies and frontier model labs. We help them to train their models, and we help them to train their robots.”
Call to Action
“Together, we accelerate physical intelligence. Now, let's upload your memories and train your robots.”
Synthesizable in Remotion · key components of the physical-ai-ecosystem-launch template.
Human hands and robotic hands swap frame-by-frame on a clean white table assembling various tools.
Remotion Sequence using absolute positioned high-res image plates cut every 2 to 4 frames with slight translation offsets.
Monochrome crosshairs, coordinate grids, and live metric readouts pinpointing hardware specs over 3D renders.
SVG overlay with interpolate-driven strokeDashoffset for bounding boxes and frame-based count-up for specifications.
Particle network expanding outward from a central logo into a hierarchical tree of glowing orange nodes.
Canvas-based spring-interpolated bezier curves generating child nodes progressively with opacity fades.
Connected skeletal joints on user hands and 3D wireframe bounding boxes over manipulated objects.
SVG polylines connected to animated joint coordinate arrays synced to video playback time.
Purpose. Frame first-person video as the core missing ingredient for general-purpose robotics.
Execution. Human and robot hands morphing while assembling workshop tools on a seamless white surface.
Purpose. Demonstrate that capturing egocentric data requires lightweight, unobtrusive wearable sensors.
Execution. Exploded 3D rendering of the Luci Pin combined with rapid real-world mounting demos on wrist, cap, and chest.
Purpose. Show the gamified contributor app pipeline from task assignment to AI review and payout.
Execution. Screen recordings of app UI overlaid with real-time hand-mesh tracking on barista task footage.
Purpose. Explain how the Large Visual Memory Model turns raw video clips into structured spatial data.
Execution. Massive branching taxonomy graph collapsing into a 3D point cloud and global task distribution map.
Purpose. Prove real-world momentum and invite partners to train models.
Execution. Founder walking alongside sitting humanoid robots, transitioning to product website and logo.