Voxel51’s wager is that consolidating the toolchain, not adding another point product, is what physical AI teams will pay for. Teams training on robot, vehicle and factory sensor feeds shuttle data among applications that share no format.
Release 3.0 pulls search, curation, annotation and evaluation into one platform. Visual data, its original niche, is now a subset.
Several sensor streams can be viewed together on a shared timeline, thanks to time-series multimodal support, and segment-level embeddings let an analyst jump to an exact moment inside a long recording. Labels can be applied across 2D, 3D, video and mixed data, with automated QA, annotator metrics and a human review step. Vision-language models also generate labels directly from natural-language instructions and examples under the agentic labeling feature.
Among the supported formats are MCAP, images, video, 3D point clouds, meshes, DICOM, NIfTI, geospatial and audio. Queries run against a persistent hybrid index that constructs embeddings on its own, so semantic retrieval can be combined with lexical search and structured filters over timestamps, sensor fields and metadata.
LanceDB underpins vector retrieval, keeping storage and compute separate so teams decide which samples get embedded and when. New samples or schema changes are handled incrementally, without a full rebuild.
The centerpiece is Voxel51 Agent, which arrives with more than 45 MCP tools, 80 operators and 18 open-source skills. It can chain search, curation, labeling and evaluation into repeatable runs and write its own plugins in code. Activity is logged by a provenance layer at the deployment, dataset and sample level.
Deployment is either the hosted app or self-hosted across cloud, hybrid, on-premises and air-gapped setups. The free tier bundles the agent with 30,000 inference requests and 30 compute hours per month, limited to three datasets of 10,000 samples and 20 GB of storage.
