Voxel51 announced Voxel51 3.0., a major expansion of its platform for building and managing physical AI datasets.
The next wave of AI is moving from the internet into the physical world. In particular, robotics is entering a phase that resembles the early days of autonomous vehicles: rapid investment, increasingly sophisticated foundation models, and an urgent need for infrastructure built around complex real-world data.
Voxel51 3.0 brings the core data workflows for physical AI into a single, extensible platform built for teams moving from prototype to production. New capabilities include:
- Voxel51 Agent: An AI agent that works across the physical AI data loop to search, curate, label, evaluate, and chain actions into repeatable workflows.
- Voxel51 Search: A persistent, hybrid search index that combines semantic search, lexical queries, and metadata filters as datasets grow.
- Voxel51 Annotation: Annotation workflows across 2D, 3D, video, and multimodal data, including automated labeling with vision-language models through Agentic Labeling.
- Multimodal time-series support: Synchronized visualization of sensor streams and segment-level embeddings for finding specific moments within long recordings.
- Data provenance: Activity logs, sample edit history, and dataset versioning that track changes and help teams identify exactly what data went into a model.
- Bring your own models: A model-agnostic architecture that supports foundation models and domain-specific models across embeddings, inference, auto-labeling, and evaluation.
- Free self-service access: The ability to get started with Voxel51 for free at app.voxel51.com
Voxel51 is built for the physical AI data flywheel
Models are getting easier to build. Data work is not. And teams that consistently get models into production invest heavily in the work of systematically finding, fixing, and improving their training data. According to the State of Physical AI 2026 report, exceptional teams are three times more likely than struggling teams to spend more than half their project time on data work.
Voxel51 3.0 brings the core workflows for physical AI into an end-to-end data platform, helping teams move from raw multimodal recordings to searchable, curated training data and support active learning as models improve.
Voxel51 Agent is the most capable agent for the physical AI data flywheel. The agent draws on more than 45 MCP tools, 80+ built-in operations, and dozens of skills to chain actions across search, curation, labeling, evaluation, and other data workflows.
Nebius and Voxel51 partner to connect the physical AI data and compute stack
Nebius and Voxel51 address two of the hardest infrastructure problems in physical AI: compute and data.
Nebius provides predictable performance and reliable cloud infrastructure as an integrated system engineered for AI, while Voxel51 helps teams turn large volumes of raw recordings into model-ready datasets.
Together, the companies are hosting a series of hands-on workshops where developers can run Voxel51 workflows on Nebius AI Cloud. The goal is to help teams build a tighter physical AI data flywheel while spending less time stitching together the underlying data and compute stack.
For more information, visit voxel51.com.
The post Voxel51 3.0 expands platform for physical AI datasets appeared first on Engineering.com.