Voxel51 3.0 adds multimodal data platform with AI agents

What changed

Voxel51 released version 3.0 of its multimodal data platform on October 15, 2026. The update consolidates search, curation, annotation, and evaluation of sensor data into a unified system designed for teams building physical AI applications including robots, autonomous vehicles, and factory automation.

The platform introduces three core capabilities. First, time-series multimodal support enables visualization of multiple sensor streams on a single timeline with segment-level embeddings that locate specific moments within extended recordings. Second, unified annotation covers 2D, 3D, video, and multimodal data with automated quality assurance and human-in-the-loop review tools. Third, agentic labeling uses vision-language models to generate labels from natural-language instructions and examples.

Voxel51 Agent represents the centerpiece of this release, chaining together search, curation, annotation, and evaluation into repeatable workflows. The agent ships with 45+ Model Context Protocol (MCP) tools, 80+ operators, and 18 open-source skills. A new search feature called Voxel51 Search uses a hybrid index that automatically indexes datasets and computes embeddings on day one, leveraging LanceDB for vector retrieval.

Data provenance now tracks information across four layers: deployment-level activity logs showing user actions and timestamps, dataset-level activity logs capturing how datasets changed over time, sample-level edit history recording label and metadata modifications, and immutable dataset versioning enabling reproduction of previous training and evaluation runs.

Supported data modalities span MCAP, images, video, 3D point clouds, meshes, DICOM, NIfTI, geospatial data, and audio. The platform operates as a hosted web application at app.voxel51.com or deploys on-premises, in hybrid environments, or in air-gapped networks.

Why it matters

Physical AI teams typically manage fragmented workflows across multiple point tools for searching, labeling, and evaluating sensor data. This tool sprawl creates friction at each handoff and loses context between stages. Version 3.0 consolidates these functions into one system, reducing integration complexity.

The inclusion of AI agents with 45+ MCP tools enables teams to automate end-to-end data workflows without writing custom code. The company describes example use cases: a robotics team searching behaviors across thousands of episodes to create balanced training sets, or a perception team investigating model failures by identifying and pulling similar edge cases for annotation. This automation becomes increasingly valuable as datasets scale.

The free tier removes barriers to testing. It includes Voxel51 Agent, 30,000 inference requests, and 30 compute hours per month at no cost. However, the tier caps at 3 datasets of 10,000 samples each with 20 GB of storage, meaning production deployments require paid plans.

Data provenance addresses a practical production requirement: teams need to trace how training data changed, identify which labels were modified by whom, and reproduce earlier results. This traceability matters when production models fail and teams must determine whether the issue stems from data quality, model behavior, or both.

The model-agnostic design allows teams to begin with foundation models and later substitute domain-specific models as their systems mature. This flexibility prevents vendor lock-in while letting teams benefit from pre-trained embeddings and inference capabilities.

What to test

Before committing to Voxel51 3.0, validate these key areas:

Search performance: The company has not published benchmarks on query latency or retrieval accuracy. Test search speed against your typical dataset sizes and query patterns. While LanceDB’s columnar design targets large-scale workloads, real performance depends on your hardware configuration and query complexity.

Agent functionality: Confirm the 45+ MCP tools and 80+ operators actually cover your workflows. Request specific examples from Voxel51 that match your use cases. Test whether agent-generated labels meet your quality standards before integrating them into production pipelines.

Provenance tracking: Validate that activity logs capture all events necessary for your team’s data lineage requirements. Verify that dataset versioning enables reproducible training with your existing models and infrastructure.

Deployment specifics: For on-premises or air-gapped deployments, confirm technical requirements, support for your infrastructure, and whether all features remain functional in offline mode.

Tool integration: Test how cleanly the platform integrates with your current labeling tools, model training pipeline, and evaluation frameworks. Validate the claim of model-agnosticism by testing model swaps with your own domain-specific models.

Cost projection: The free tier caps at 3 datasets of 10,000 samples each. Calculate your actual storage and compute needs and request specific pricing for your anticipated scale.

The conclusion

Voxel51 3.0 addresses a genuine pain point: physical AI teams waste cycles transferring data between different systems. By consolidating search, annotation, and evaluation with agentic automation, the platform reduces operational friction and makes data changes traceable throughout the pipeline. The free tier with agent support lowers entry costs for evaluation.

The release’s primary strength is consolidation. Teams already managing multimodal sensor data benefit most from replacing multiple specialized tools with one integrated system. The agent layer, supported by 45+ MCP tools, delivers genuine automation for repetitive tasks like identifying edge cases or balancing training datasets.

Key limitations center on validation. Without published performance benchmarks for search speed, agent label quality, or provenance accuracy, teams must test thoroughly on their own data before making infrastructure commitments. The free tier’s storage and dataset restrictions mean meaningful evaluation requires moving to a paid plan.

For teams currently using Voxel51, the multimodal time-series support and persistent search layer represent direct platform enhancements. For teams building physical AI systems from scratch, testing the free tier offers a low-risk way to evaluate whether the platform fits your requirements before scaling.

Related reading

• Perplexity SPACE: Sandbox platform for AI agents
• Nemotron 3.5 Super VL 120B: Open Multimodal Agent Model
• Grok Voice 3.0 ranks first on benchmarks, adds tool-calling