Voxel51 launches version 3.0 platform for multimodal physical AI data
The new release unifies search, annotation, and evaluation for sensor data from robots and vehicles, adding an agentic workflow layer.
Key points
- Voxel51 3.0 merges search, curation, annotation, and evaluation into a single multimodal data platform.
- The update introduces Voxel51 Agent with 45+ MCP tools and a free self-service tier for teams.
- New features include time-series multimodal support and agentic labeling using vision-language models.
Voxel51 has released version 3.0 of its data platform, marking a significant expansion of its capabilities for physical AI teams. According to MarkTechPost, this update consolidates search, curation, annotation, and evaluation functions into one unified system designed for organizations training models on complex sensor data.
The platform now supports a broader range of data types beyond visual inputs, including time-series multimodal data. Teams can view multiple sensor streams on a single timeline and use segment-level embeddings to locate specific moments within long recordings. This allows for more granular analysis of data from robots, vehicles, and factory environments.
Platform Capabilities and Features
Voxel51 3.0 introduces Voxel51 Agent, which chains various data processing steps into repeatable workflows. This agent ships with more than 45 Model Context Protocol tools, over 80 operators, and 18 open-source skills. It enables robotics teams to search behaviors across thousands of episodes and curate balanced training sets efficiently.
Unified annotation features now cover 2D, 3D, video, and multimodal data. The system includes automated quality assurance, annotator metrics, and human-in-the-loop review processes. Additionally, agentic labeling allows vision-language models to generate labels based on natural-language instructions and examples, streamlining the labeling workflow for diverse data types.
Search and Storage Architecture
The platform’s search functionality relies on a persistent hybrid index that computes embeddings automatically. This ensures search capabilities are available immediately upon dataset ingestion. Semantic retrieval can find visually or semantically similar moments even without shared labels, while lexical and structured queries handle exact terms and metadata.
Vector retrieval runs on LanceDB, supporting vector, scalar, and full-text indexes. The architecture separates storage from compute, allowing teams to choose which models and compute resources to use. Incremental updates prevent the need for full rebuilds when new samples arrive or schemas change, maintaining performance as datasets grow.
Accessibility and Limitations
Voxel51 3.0 is available as a hosted web application or can be self-hosted in cloud, hybrid, on-premises, and air-gapped environments. The company has introduced a free self-service tier that includes access to Voxel51 Agent, 30,000 inference requests, and 30 compute hours per month.
However, the free tier has strict limitations. It caps usage at three datasets of 10,000 samples each and provides only 20 GB of storage. These constraints may restrict its utility for larger projects, though it offers a viable entry point for smaller teams or initial evaluation purposes.
What should you do?
Security and technology professionals should evaluate whether their current data workflows align with the new multimodal capabilities of Voxel51 3.0. Teams handling diverse sensor data should test the platform’s agentic labeling features to assess efficiency gains. Organizations with air-gapped requirements should verify that the self-hosted option meets their specific security and compliance standards before deployment.
What to do and how to stay safe: Voxel51
- Review current data annotation workflows to identify bottlenecks that agentic labeling might resolve.
- Assess storage needs against the free tier limits to determine if a paid plan is necessary for production use.
- Test the self-hosted installation in a sandbox environment to ensure compatibility with existing air-gapped infrastructure.
- Monitor vendor updates for any changes to the 45+ MCP tools or operator suite as the platform evolves.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What new features does Voxel51 3.0 introduce?
Voxel51 3.0 adds Voxel51 Agent, time-series multimodal support, unified annotation for various data types, and agentic labeling using vision-language models.
Is there a free version available?
Yes, a free self-service tier includes Voxel51 Agent, 30,000 inference requests, and 30 compute hours per month, with caps on datasets and storage.
Can Voxel51 3.0 be deployed on-premises?
Yes, the platform can be self-hosted in cloud, hybrid, on-premises, and air-gapped setups, in addition to the hosted web app option.
