Inside the Data Bottleneck Slowing Visual and Physical AI
Wiley Science and Engineering Content Hub Voxel51
AI teams say the real bottleneck isn’t models, it’s data. A survey of 700+ pros says data work, not bigger systems, decides who ships.
Based on reporting by Wiley Science and Engineering Content Hub, Voxel51 — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
The AI frontier has moved off the page and into the physical world. Video, LiDAR point clouds, sensor streams, and other messy real-world data are now what people are using to build systems that can see, reason, and act in space.
A new IEEE Spectrum and Wiley white paper, sponsored by Voxel51, is based on a 2026 survey of more than 700 professionals working in visual and physical AI. Its core message is blunt: data problems are behind most model failures, and the teams that get products out the door spend much more time on data than the teams that don’t.
The report says 78% of practitioners already see measurable value from visual and physical AI, yet 74% still think the field is underinvested relative to its opportunity. That gap matters. It suggests the technology is delivering enough to convince teams to keep going, while still being starved of the work needed to make it reliable at scale.
And that work is not glamorous. The report says annotation is expensive and often wasteful because teams label everything, only to throw away a lot of it before production. The sharper lesson is that curation beats brute force: the winners don’t just collect more data, they spend far more time making it usable.
The white paper also says 92% of practitioners believe this is where the field is headed next. That sounds less like a forecast than a verdict. Text may have powered the last decade of AI, but in physical AI the boring stuff — sorting, cleaning, choosing — looks like the whole game.
My take — AI-written commentary, not fact-checked reporting
This is the part of AI everyone wants to skip, because data work doesn’t demo well and never makes a glossy launch video. But the industry keeps relearning the same lesson: if the inputs are a mess, the model becomes an expensive way to make the mess prettier. The hype machine loves bigger architectures; the field keeps paying for better spreadsheets.
Read more about this at: Wiley Science and Engineering Content Hub