Physical AI has a data problem. Voxel has released Voxel51 3.0, an end-to-end multimodal data platform built for teams training models on sensor data from robots, vehicles, and factories. It combines search, curation, annotation, and evaluation in one system—because sending data through a parade of disconnected tools remains a popular way to lose time.
Voxel51 3.0 supports time-series multimodal data, segment-level embeddings, unified annotation, and agentic labeling. Teams can also bring their own models to the platform, giving them a way to work with existing systems instead of treating every new product launch as a demand for total replacement.
The platform ships with Voxel51 Agent, a data provenance layer, and free self-service access. Voxel51 Agent includes 45+ MCP tools, 80+ operators, and 18 open-source skills, then chains search, curation, annotation, and evaluation into workflows.
A search layer built for messy sensor data
Voxel51 Search serves as a persistent hybrid index for datasets. It indexes datasets and computes embeddings as part of that process, supporting semantic retrieval alongside lexical and structured queries.
The search system runs on LanceDB and is designed to connect different ways of finding data inside the same workflow. Semantic retrieval can locate samples by meaning, lexical queries can match terms, and structured queries can target defined data fields. That combination matters when teams work with sensor collections that do not behave like neat document libraries.
Voxel51 Agent uses the search layer as part of a broader chain rather than treating retrieval as an isolated feature. Search leads into curation, annotation, and evaluation, allowing teams to move through the data cycle inside one platform.
More access, more provenance, and a funding plan
Voxel51 3.0 runs through a hosted web app at app.voxel51.com or as a self-hosted deployment in cloud, hybrid, on-prem, and air-gapped setups. The range of deployment options covers teams that need a hosted service as well as those that keep their environments under tighter control.
The best free tier includes Voxel51 Agent, 30,000 inference requests per month, and 30 compute hours per month. It caps usage at 3 datasets of 10,000 samples each and 20 GB of storage, which offers a meaningful starting point without pretending that industrial-scale workloads fit inside a generous demo account.
Voxel51 3.0 adds four provenance layers: a deployment-level activity log, a dataset-level activity log, sample-level edit history, and dataset versioning. Together, these layers track activity across the platform, from deployment events down to changes made to individual samples.
The focus on data work reflects a problem Voxel says its own research has measured. According to Voxel’s 2026 State of Physical AI survey, 58% of “exceptional” teams spend more than half their project time on data work.
That figure gives the product a clear target: reduce the time teams spend preparing and checking data before models can do useful work. Voxel is not presenting another model as the answer; it is building the machinery around models, where much of the labor actually sits.
The company is also preparing for its next phase. Ventral AI was acquired by Voxel, and Voxel plans a Series C raise of at least $50 million next year as it expands into new markets.
“Voxel is eyeing a Series C raise of at least $50 million next year,” said Vernon O’Donnell, CEO of Voxel. The funding plan and Ventral AI acquisition point to a company pushing beyond a single data workflow—while betting that physical AI teams will keep needing better ways to find, label, evaluate, and trace the data behind their systems.
Based on
