Voxel51 3.0 brings sensor-data search and annotation into one platform
The release connects data selection, labeling and model evaluation for teams building AI for robots and vehicles.
Voxel51 announced version 3.0 on October 15. It brings together work on multimodal data: recordings containing different kinds of sensor information. The company describes instruction-based AI labeling with human review, alongside search and evaluation. Hosted self-service access is available free; teams unable to upload their data can contact Voxel51 about self-hosting.
The company's search explanation, published in July, shows the underlying workflow. After indexing their recordings to make them searchable, teams can combine a written query, visual similarity, defined sensor events and metadata filters. A failed robot grasp can serve as an example for finding related moments. Search can locate portions of a recording rather than only whole episodes. Opening a result brings cameras, three-dimensional data and sensor signals onto a shared timeline for inspection, before selecting material for training or evaluation.
The new release also makes data history visible. Voxel51's provenance description distinguishes activity across a deployment from changes within one dataset. Individual sample histories record edits to labels, metadata and other properties. Dataset versions preserve unchangeable snapshots for comparing earlier states and identifying what data a model used. These records connect changes to people or automated work, helping teams investigate how a dataset reached its current form. The documented capability concerns the data's history and state.
MarkTechPost also covered the release.