Meet Voxel51 3.0: A Unified Multimodal Data Platform for Physical AI
MarkTechPost Asif Razzaq
Voxel51 3.0 puts search, labeling and evaluation into one multimodal platform. It’s aimed at physical AI teams, and the free tier now includes Voxel51 Agent.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Voxel51 has rolled out Voxel51 3.0, a broader version of its data platform that folds search, curation, annotation and evaluation into one system. The company is pitching it at teams building models from sensor data in robots, vehicles and factories, where the hard part is often not the model but the mess of data around it.
The biggest shift is that Voxel51 is no longer treating visual data as the whole story. Voxel51 3.0 adds time-series multimodal support, so teams can line up multiple sensor streams on a single timeline. It also adds segment-level embeddings, which means users can hunt for specific moments inside long recordings instead of only searching whole samples.
Annotation gets a wider reach too. The platform now covers 2D, 3D, video and multimodal data, and it adds automated QA, annotator metrics and human-in-the-loop review. On top of that, Voxel51 is pushing Agentic Labeling, where vision-language models produce labels from natural-language instructions and examples. Supported data types include MCAP, images, video, 3D point clouds, meshes, DICOM, NIfTI, geospatial data and audio.
Search is built around a persistent hybrid index. Voxel51 says datasets are indexed and embeddings are computed automatically, so search starts working from day one. The system mixes semantic retrieval with lexical and structured queries, while vector retrieval runs on LanceDB. Teams can choose what to embed, which model to use and when to update, and incremental updates are meant to avoid full rebuilds when data changes.
The new Voxel51 Agent is the automation piece. It chains search, curation, annotation and evaluation into repeatable workflows, and it ships with 45+ MCP tools, 80+ operators and 18 open-source skills. Voxel51 also says the platform is model-agnostic, so teams can start with a foundation model and later swap in a domain model. Free self-service access is part of the release, though the free tier is capped at 3 datasets, 10,000 samples each and 20 GB of storage.
My take — AI-written commentary, not fact-checked reporting
This is the rare AI platform update that sounds built for actual grunt work, not demo theatre. The real tell is provenance: activity logs, edit history and versioning are what matter when a model goes wrong and everyone suddenly claims the data was “fine.”
Read more about this at: MarkTechPost
