Voxel51 3.0 adds a data agent and a free tier for physical AI teams

What Voxel51 released

Voxel51, the company behind the open-source FiftyOne toolkit for computer-vision datasets, released Voxel51 3.0 on 15 October 2026. Co-founder and chief executive Brian Moore called it "our most significant release yet" in a launch post. The update pitches the platform as a single place for robotics and other physical AI teams to search, curate, label and evaluate the data from cameras, lidar and other sensors.

The release adds an AI agent that can run multi-step data jobs, a persistent hybrid search index, annotation tools with automatic labelling by vision-language models, support for synchronised time-series sensor data, and audit logs with dataset versioning. For the first time, anyone can sign up and try the hosted platform for free without talking to sales. As Unite.AI noted, Voxel51 is also now the name of the paid product formerly called FiftyOne Enterprise, while FiftyOne remains the name of the open-source project.

The company's press release names Berkshire Grey, Ford, Porsche and Precision Planting among its customers. Moore argued that robotics is entering a phase like the early days of self-driving cars: "The teams that got their data infrastructure right from the start were able to move faster."

Search across sensor recordings

Physical AI data is awkward to search. A robot demonstration or a test drive produces hours of video alongside joint angles, lidar sweeps and other readings, and the useful part is often a few seconds where something went wrong. Voxel51 3.0 adds first-class support for this kind of time-series data: several sensor streams can be shown on one timeline, and the platform can compute embeddings for short segments rather than one embedding for a whole recording, so a search can land on the exact moment that matters.

Voxel51 Search, the new index, automatically indexes datasets and computes embeddings so queries work from the start. It mixes three kinds of query: semantic similarity from embeddings, lexical text search, and filters on structured metadata such as timestamps or sensor fields. Voxel51 says it keeps embeddings and structured fields in their native form and combines them only at query time, which preserves exact range filters. Vector search runs on LanceDB, an open-source columnar database that separates storage from compute.

Teams control which data gets embedded, with which model, on which compute and when. Voxel51 says the index updates incrementally as new samples arrive or embeddings are recomputed with a newer model, without rebuilding the whole dataset.

Voxel51 Agent and Agentic Labeling

The headline addition is Voxel51 Agent, an assistant that can chain search, curation, labelling and evaluation into repeatable workflows. Voxel51 says it ships with more than 45 tools exposed over the Model Context Protocol (MCP), more than 80 built-in operators and 18 open-source skills. Its examples include a robotics team searching thousands of episodes for a behaviour and checking whether it is over- or under-represented, and a perception team starting from one model failure, finding similar rare scenes, sending them for labelling and evaluating the retrained model.

The agent can also write code to build custom plugins when a workflow needs something new, and Voxel51 says every workflow inherits the organisation's existing permissions and governance controls. The skills live in an Apache-2.0 GitHub repository with plugin packaging for Claude, Codex, Cursor and Gemini tools, so outside assistants can drive FiftyOne as well. "Physical AI teams don't need another chat interface sitting on top of their data," product manager Adonai Vera said in the release.

Annotation now covers 2D images, 3D point clouds, video and multimodal data in one workflow, with automated quality checks and annotator metrics. Agentic Labeling uses vision-language models to generate labels from plain-language instructions and examples, then routes results through review, so uncertain cases can go to human annotators instead of straight into a training set.

Tracking which data trained a model

A companion post on data provenance by chief architect Ritchie Martori describes the audit features. They work at four levels: a deployment-wide activity log of who did what and when, a log for each dataset, an edit history for individual samples, and dataset versioning that captures immutable snapshots.

The point is reproducibility. When a model underperforms, a team needs to know exactly which samples and labels went into training and what has changed since. With dozens of annotators, reviewers and automated pipelines editing the same data, that history is easy to lose. Voxel51 says snapshots let teams compare historical states and rerun training or evaluation on the exact data used before. Under the hood, a job scheduler called fiftyone-mq and an event tracker called Activity Core feed the audit logs, annotation metrics and job monitoring.

The platform is also model-agnostic. Teams can register foundation models and their own domain models in one catalogue and swap them into embedding, auto-labelling, inference and evaluation steps, so a model trained on a team's own data can improve the next round of data selection.

Free tier, pricing and the open-source FiftyOne

Until now, trying Voxel51's enterprise features meant a sales call. The free tier at app.voxel51.com, per the pricing page, allows three datasets of up to 10,000 samples each, 20GB of hosted storage, 30,000 model inference requests and 30 hours of compute a month, and includes the agent. No credit card is required. Paid Team plans start at eight user seats and 2,800 compute hours a month with unlimited data and inference; prices are not listed, so buyers must contact sales. Teams that cannot upload data to a hosted service can ask for a self-hosted deployment.

The open-source FiftyOne repository remains Apache-2.0 licensed and had about 11,200 GitHub stars when we checked on 18 October. Voxel51 also announced a partnership with Nebius, whose AI cloud is being used at hands-on workshops; the two companies plan nine more Physical AI Workshops in Canada and the United States through November 2026.

Why data tooling matters for robotics

Robotics models are hungry for real-world data, and most of the effort goes into finding the right pieces of it. Voxel51 cites a 2026 State of Physical AI survey in which 58% of the teams it rates as exceptional spent more than half their project time on data work, against 21% of struggling teams. Model builders such as Nature Will, which pretrained its Natura-1 robot-hand model on a million hours of first-person video, show how large these datasets are becoming.

Voxel51 is not alone in chasing that work. Labelbox recently raised more than $100 million while shifting from labelling services to selling AI agents, and cloud providers are packaging their own pipelines, such as the open-source Physical AI Toolchain from AWS. Voxel51's pitch is that one platform covering the whole loop, from raw recordings to evaluation, beats a chain of point tools.

Limits and open questions

Almost everything in the launch is Voxel51's own description. The company has not published benchmarks for the agent's accuracy, the quality of Agentic Labeling against human labels, or search speed on large datasets, and the survey figures it cites come without a named methodology in the announcement. AI Lookout did not test the platform.

The free tier is sized for evaluation rather than production; 10,000 samples is a small slice of a robotics dataset. Paid pricing is unpublished, and the more advanced agent and annotation features sit in the commercial product rather than open-source FiftyOne. Agents that can write and run code on company data also raise questions about review and permissions that Voxel51 answers only in general terms. We will update this article if Voxel51 publishes independent results, pricing, or wider open-source releases.