Voxel51 presents Unified Data Platform 3.0 aimed at physical AI teams: Unite.AI

Voxel51 on October 15, 2026 introduced Voxel51 3.0, an end-to-end multimodal data platform for physical AI that brings together search, curation, annotation and evaluation in a single extensible system. Co-founder and CEO Brian Moore, writing in the company’s announcement, called it the “most significant release yet,” and teams can now try the platform on their own data for free at app.voxel51.com.

The release also renames the company’s enterprise product: Voxel51 is the new name for FiftyOne Enterprise, while FiftyOne continues to refer to the open source project. Moore wrote that the company has spent years working with teams implementing AI in autonomous systems, robotics, manufacturing and healthcare, and has rebuilt the platform in recent months around the data work that begins when a model goes beyond the prototype: investigating faults, curating training sets and evaluating models in real-world scenarios.

Multimodal search and time series support

Voxel51 Search is a persistent hybrid search index that combines semantic retrieval with lexical queries and filters on unstructured content and structured metadata in multimodal datasets. According to the announcement, Search automatically indexes datasets and calculates embeddings so that queries are available from day one, and keeps embeddings and queryable structured fields in their native form, combining them upon retrieval. For vector retrieval, the platform uses LanceDB, whose columnar architecture supports vector search along with scalar and full-text indexes, while separating storage from compute.

Teams control the embedding pipeline itself, including what data is embedded, what model and computation to use, and when embeddings are generated, with incremental updates that don’t require rebuilding the entire dataset. The release adds first-class support for multimodal time series data: multiple sensor streams can be viewed on the same timeline, and segment-level embeddings allow teams to retrieve specific moments within long recordings rather than representing an entire episode with a single embedding.

Voxel51 agent and agentic labeling

Voxel51 Agent chains search, curation, annotation, and evaluation into repeatable workflows powered by 45+ MCP tools, 80+ operators, and 18 open source skills. The announcement describes use cases: a robotics team looking for specific behaviors across thousands of episodes to curate targeted training and assessment sets; a perception team that moves from model failure to similar longtail scenarios, re-annotations, and re-evaluation; and annotation preparation that routes uncertain cases to human reviewers. The agent can also write code to add new functionality, and all workflows inherit the permissions, security, and governance controls set by an organization.

In a post published on October 8, 2026, ahead of launch, Voxel51 said the agent is built on the Model Context Protocol, an open standard for connecting AI applications with external tools, and that it connects to more than 100 model vendors, including Anthropic, OpenAI, and Google, with availability across cloud, hybrid, on-premise, air-gapped, and managed deployment options. The company reported that comparing the same agent tasks on the Claude Sonnet 5, Claude Opus 5, GPT-5.1 and Gemini 3.1 Pro revealed cost differences of up to 15 times.

Version 3.0 also unifies annotation workflows for 2D, 3D, video, and multimodal data, with automated QA, annotator metrics, and human-in-the-loop review management tools. Its Agentic Labeling capability uses visual language models to generate labels from natural language instructions and examples, with review workflows to validate the results. Citing its 2026 State of Physical AI survey, the company said that exceptional teams spend more than half of their project time on data work at a rate nearly 3 times higher than struggling teams; the survey graph places these shares at 58% of exceptional teams and 21% of teams in difficulty.

Data provenance on four levels

Voxel51 3.0 adds four-tier data provenance, detailed in a companion post by Chief Architect Ritchie Martori on October 9, 2026. A deployment-level activity log shows who performed an action, when it occurred, and what it affected during a deployment; a dataset-level activity log tracks changes within individual datasets; sample-level change history tracks changes to labels, metadata, and other properties; while dataset versioning captures immutable snapshots so teams can compare historical states, replay workflows, and identify the exact data used for training or evaluation.

The underlying event architecture uses fifty-one-sqm for job scheduling and Activity Core for event monitoring, providing a shared foundation for audit logs, annotation metrics, and job monitoring. Workflow permissions and visibility link each data change to the person or automated process that made it.

Model-agnostic design workshop and Nebius

The platform is designed to be model-agnostic, so teams can use their own baseline and domain-specific models for embeddings, automatic labeling, inference, and evaluation without being locked into a predefined stack. The announcement states that general-purpose baseline models have learned primarily from Internet text and visual data, with limited exposure to embedded multimodal sensor data generated by physical systems, and that outside of a handful of well-funded domains such as autonomous driving, physical AI teams often need to combine baseline models with domain-specific models and their own training data.

Historically, experiencing Voxel51’s business capabilities meant talking to its team first; with version 3.0, self-service access is free at app.voxel51.com, and self-hosted deployments remain available by contacting the company. Voxel51 also said it is working with AI cloud company Nebius, with developers running workflows such as embeddings and automatic labeling on Nebius AI Cloud during hands-on events. The two companies will host nine additional AI workshops in Canada and the United States through November 2026, with open registration.