Thoughts on geo data collection, field data, GIS and AI workflow automation — from the team that runs the programmes.
A recent operational disclosure from data-infrastructure vendor Unidata shows what egocentric capture looks like once it stops being a research project and becomes a running facility.
A warehouse robotics company just said the quiet part out loud: the mess is the product. Here's why that logic applies far beyond robotics.
A RoboBusiness talk on humanoid motion just named the real bottleneck: hands, hips, and torsos need completely different data. Field research has the same problem.
ESOMAR just formalized rules for AI-generated survey respondents. The accuracy numbers are real, and so are the gaps underneath them.
Hugging Face crossed one million public datasets this week. For teams building physical AI, that milestone measures noise almost as much as it measures progress.
SoftBank just backed a construction robotics company whose machines survey the site while they dig it. That's not a small detail.
Three separate reports published this August put real dollar figures on robot training data for the first time. The numbers explain why data collection just moved from the lab to the field.
A multi-point turning feature grabbed the headline. The real story is buried in the data-throughput number nobody asked about.
A new arXiv dataset turns 204,000 clips of human hands into 17 million affordance labels, without a single human annotator marking a frame.
AI can summarize ten thousand survey responses before lunch. It still can't tell you whether the shelf tag was actually there.
A startup put cameras on a thousand real heads. A new retail dataset just proved that where you point the camera matters more than how much footage you collect.
A dataset announcement from Figure AI, paired with fresh moves from LG, Nvidia, and SEER Robotics, shows the real bottleneck in physical AI isn't models — it's field-grade data infrastructure.
A new synthesis pipeline claims to be the largest ego-to-robot dataset to date. Here's what it actually measured — and what it didn't.
A new arXiv paper on converting human video into robot training data quietly explains why raw field data almost never arrives ready to use.
A headset startup just raised real money to record human hands. The harder problem it's actually solving isn't the camera.
Two funding announcements and a dataset paper from the last month show the same thing: robotics companies are now building physical operations, not just models.
Public egocentric datasets crossed a million hours in 2026, but teams report discarding 90% of bulk footage before training. Why coverage-designed collection is the complement bulk data can't replace.
The EU pushed high-risk AI obligations to December 2027. What the delay does and doesn't change for teams collecting training data now.
What it actually takes to capture first-person training video, and why robotics labs can't get enough of it.