Bigger Models Won't Fix Physical AI. 700 Practitioners Say Data Will.

A 2026 survey of more than 700 AI practitioners finds that teams shipping visual and physical AI in production spend nearly three times as much time on data work as teams whose projects stall. The 3x ratio is the cleanest operating signal in the report.

"Visual and physical AI" describes systems that perceive and act in physical space: warehouse robots, autonomous vehicles, drones, video-understanding tools, and the sensor pipelines that feed them. The category has absorbed most of the post-LLM frontier in applied AI, and it runs on a different kind of data than text models. A single training run for a self-driving system can consume millions of hours of synchronized camera, LiDAR, and radar streams, all of which must be time-aligned, labeled, and checked before they teach a model anything useful.

The 3x figure, drawn from the sponsored white paper, compares successful teams against stalled ones on data-work hours. Teams that ship invest roughly three times more hours in selecting frames, annotating point clouds, validating labels, and iterating on edge cases than teams whose projects never reach production. The unsuccessful teams treat data as a one-time acquisition, hand the corpus to a model team, and discover too late that the failure modes live in the labels, the sensor calibration, and the missing edge cases rather than in the architecture.

The sponsor is Voxel51, a data-curation tooling vendor, and the survey was distributed through IEEE Spectrum and Wiley. The framing aligns with the sponsor's commercial position, and the Wiley page is a download-gated teaser: the full methodology, sample composition, and exact wording of the 3x question are not visible in the hydrated source. The headline number should be read as a practitioner signal inside a vendor survey, not as a settled industry measurement.

The other survey numbers fit alongside that caveat. Seventy-eight percent of respondents say their teams already see measurable value from visual and physical AI. Seventy-four percent consider the field underinvested relative to its opportunity. Ninety-two percent believe the field is heading in a particular direction, though the direction sentence is truncated in the public teaser and the survey's authors do not complete it in the visible copy. The combination reads less like a vote of confidence than like an industry that knows the work matters and is not yet sure how to budget for it.

The under-the-hood failure mode is annotation. Teams that treat labeling as a throughput problem, label everything and sort the wheat from the chaff later, burn the budget twice. Teams that ship build curation into the pipeline from day one: which frames are worth labeling, which sensor windows are noise, which labels are wrong, which edge cases are missing. The report's central thesis is that data work, not data collection, separates the teams that ship from the teams that stall.

A 3x data-work ratio does not translate to a larger labeling team in the abstract. It means a small curation team with feedback loops into the model team, plus the tooling to throw away mislabeled data before it trains on it. A robotics group that ships is not labeling more frames than a robotics group that stalls. It is choosing different frames, rejecting different labels, and rebuilding the training set each time the model surfaces a new failure mode.

The 3x ratio is also the figure most likely to be tested outside the survey's sample. The full white paper, when consulted, will reveal how "data work" and "model work" were defined for respondents and how the figure was computed. If the ratio survives independent benchmarks and non-sponsor surveys, it becomes a usable diagnostic for engineering leads. If it does not, the figure remains a practitioner signal from a vendor-sponsored sample, and the underlying point that data curation is the deployment bottleneck is still the part of the report worth keeping.