Voxel51 announced version 3.0 of its data platform on Oct. 15, combining tools for managing multimodal datasets in a workflow aimed at robotics and other physical AI teams.
The update lets users search by meaning or text and narrow results with metadata filters. It also brings synchronized sensor views and annotation for 2D, 3D, video and mixed-modality data. Voxel51 said the release includes activity logs, records of sample edits and dataset version histories.
Voxel51 said its Agent can turn dataset discovery, preparation, labeling and evaluation into repeatable workflows. The company lists more than 45 MCP tools, more than 80 operators and 18 open-source skills. The feature set is aimed at teams handling several data types as they develop robotics and other physical AI systems.
Separately, an online survey conducted in 2026 by Dimensional Research for Voxel51 included 709 qualified participants. Ninety-seven percent said their teams faced challenges iterating on datasets, while 68% worked with at least three data modalities. The report also found 99% of teams described annotation as painful, and 36% said more than half of their annotated data never reached production.
Among participants who described their teams as exceptional, 58% said data work accounted for more than half of project time. The figure was 21% among those who described their teams as struggling.
Voxel51 Product Manager Adonai Vera said physical AI teams need an agent that “can maintain context across the model development lifecycle.” CEO Brian Moore said customers building autonomous systems and perception models moved faster after establishing data infrastructure early. He compared robotics' present needs with those of autonomous vehicles during their early development.
Voxel51 says users can explore the platform for free through its app, though the announcement did not specify production pricing. It also plans nine additional physical AI workshops with Nebius in Canada and the United States through November 2026.
