Why We’re Fixing Reality Capture Data
From hacking together data management software for autonomous robots at Amazon to starting LumiDB, this is the story of how we set out to fix reality capture data. Learn how we’re tackling the challenges of exploding data volumes, outdated tools, and scattered workflows to build a future where reality capture data is easily accessible.
Do not index
Do not index
Almost four years ago, we—Sampo and Jasin—found ourselves in uncharted waters. Umbra, the company we’d been building for the past 10+ years, had just been acquired by Amazon. This brought us into the world of autonomous robots, specifically Amazon Scout, a last-mile delivery bot. Our mission was to create high-definition, millimeter-scale 3D maps of residential areas from reality capture data, to support robot autonomy. We had plenty of experience building scalable real-time 3D systems for AAA games like Call of Duty and DOOM, but the sheer scale of the operation at Amazon was something we’d never seen before. So of course we were stoked to get to work on such an interesting problem.
The map data was collected using robots equipped with a high-resolution LiDAR and RGB cameras. Our job was to take this data and refine it into a high-fidelity 3D reconstruction. Each mission generated terabytes upon terabytes of reality capture files, but managing these large datasets was painful, to put it mildly. And the tools available… didn’t help. We ended up building complex data pipelines, stitching together a mix of proprietary and open-source tools. While some components worked smoothly, the overall system was finicky, clumsy, and prone to errors. There had to be a better way, but we were too busy to do much about it at the time.
Fast forward to early 2024, coincidentally we both found ourselves free of obligations, planning our next moves. Meanwhile, the idea of fixing the poor state of reality capture data management wouldn’t leave us alone. It was baffling to see reality capture becoming more common across industries, such as construction, mining and autonomous driving, and the sensors getting better and better, while people were still shipping hard drives around and using FTP, and relying on local processing of individual files. All the while other industries had adopted modern, cloud-native tools for dealing with big data. So, we did what any reasonable person would do and decided to start a company to fix it. We simply wanted to make reality capture data easier to work with, and do our small part in bringing the future a tiny bit closer to us. Put more selfishly, we set out to build what we ourselves would have wanted but didn’t have.
From an economic perspective, early 2024 was probably the worst possible time to start a startup. But it was painfully clear that reality capture data was about to explode, while the tools were collapsing under this weight. So we had to do something about it. Now, we’re heads down, building a tool that will make working with reality capture data easy and fun! In the summer of 2024, we added Alex, our awesome founding engineer, to the team, and he has been instrumental in turning our crazy ideas into an actual product. Our grand vision is to become the universal data layer for all these datasets, optimizing the entire lifecycle of scan data, from capture to applications.
We’re building in close collaboration with industry leaders, but we’re still early, so get in touch if you’d like to join us!
PS. We’re hiring!
Written by
Written by
