Survey of 700 practitioners finds data work drives success in visual and physical AI

The white paper synthesizes responses from over 700 professionals working on visual and physical AI, revealing a clear divide in outcomes based on data practices. Teams that successfully deploy systems dedicate significantly more time to data curation, annotation, and iteration than those that stall. This suggests that the bottleneck in physical AI is not computational power or model complexity but the ability to prepare and refine datasets for real-world conditions. The report implies that chasing larger architectures without addressing data quality may yield diminishing returns.

Data work emerges as the critical factor separating successful deployments from failures. The survey highlights that annotation remains a costly and inefficient process, with teams often labeling large volumes of data only to discard much of it before production. This inefficiency points to a need for better tools and methodologies to identify high-value data subsets. The finding that 92% of practitioners see data curation as the next frontier underscores the urgency of improving these workflows, particularly for high-dimensional data like video, LiDAR, and sensor streams.

The shift from text-based AI to physical AI introduces new challenges in data observability and quality. Unlike text, physical-world data is multimodal, noisy, and often lacks clear ground truth. The report suggests that teams must prioritize dataset iteration and edge-case identification to avoid model failures in deployment. This requires a cultural shift in AI development, where data work is no longer an afterthought but a first-class engineering concern. The survey’s findings serve as a call to action for better tooling and processes to support this transition.