Voxel51 released the findings of the inaugural State of Visual and Physical AI Survey. The study, which surveyed more than 700 professionals working across visual, spatial, and physical AI, looks at how organizations are adopting visual and physical AI, and the challenges they face in bringing these systems into real-world deployment.
A majority of respondents (78%) reported seeing value from their visual and physical AI investments, and expect its importance to grow over the next three years. Yet, 74% state that the industry is underinvested, with the biggest reason being GenAI and LLMs absorbing budget and attention. In fact, only a third say their deployments have made it to production.
Model failure is a major challenge as visual and physical AI projects move to the real-world. Of those surveyed, 100% reported experiencing underperforming models, with 85% saying model failures are common. The survey identified the primary reasons for these failures, and all focused on one area—data.
When asked about challenges, many of those surveyed highlighted insufficient training data, data quality issues, and costly annotation errors as having the biggest impact on projects. Eighty-nine percent reported data as the primary driver of visual and physical AI success, yet 97% continue to struggle to iterate on these more complex datasets.
“Physical AI will be one of the defining technology advances of the next decade. But for safety-critical systems, autonomous vehicles, factory robots, and surgical AI, there is zero margin for error,” said Brian Moore, co-founder and CEO of Voxel51. “The bottleneck isn’t algorithms anymore. It’s data. The quality, coverage, and observability of training data are what separate organizations that deploy reliable physical AI from those that don’t. The companies that treat data understanding as a core competency, not an afterthought, will win.”
Key findings from the survey include:
- World models and spatial intelligence will lead the next wave of AI: 92% highlighted that world models and spatial intelligence represent the next frontier in visual AI and are ready to break into mainstream awareness.
- Annotation workflows are under strain: Annotation remains one of the most significant operational pain points. Just over a third (36%) of respondents say less than half of their annotated data ultimately makes it into production training pipelines. Meanwhile, 44% expect annotation costs to increase in the coming year.
- Data, not models, are the biggest blocker to success: 89% identify data as the key driver of success, yet most struggle to iterate effectively. Human involvement remains a critical component of model development, reinforcing the need for better tools and workflows.
- Multimodal complexity is the new standard: With the majority of teams working across multiple data types, organizations are navigating increasingly complex AI pipelines. This shift is driving demand for better infrastructure to manage, process, and integrate a mix of data sources.
- A defining moment for physical AI: Confidence in visual AI is high, but success will depend on improving data quality, scaling annotation processes, and building infrastructure that supports multimodal development. Organizations that can bridge the gap between experimentation and deployment will be best positioned to lead the next wave of AI innovation.
To learn more or download the survey, please visit: 2026 State of Visual and Physical AI Survey.
