Voxel51 3.0 unifies sensor data workflows for physical AI teams
The platform release combines search, annotation and evaluation tools with a new agent layer, offering a free tier for robotics and autonomous vehicle developers.
Key points
- Voxel51 3.0 merges multimodal data search, curation, annotation and evaluation into a single system.
- The update introduces Voxel51 Agent, which includes more than 45 MCP tools and supports time-series data.
- A free self-service tier provides limited compute hours and storage for teams to test the platform.
Voxel51 has released version 3.0 of its data platform, consolidating search, curation, annotation and evaluation functions into one system for teams working with physical AI. According to MarkTechPost, the update targets organisations training models on sensor data from robots, vehicles and factories. The release adds a new agent layer and a free self-service access option for users.
The platform now supports time-series multimodal data, allowing teams to view multiple sensor streams on a single timeline. It also introduces unified annotation capabilities covering 2D, 3D, video and multimodal data, alongside agentic labeling using vision-language models. These features aim to streamline workflows for perception and robotics teams handling complex sensor inputs.
What we know
Voxel51 Agent chains search, curation, annotation and evaluation into repeatable workflows, shipping with more than 45 MCP tools, 80 operators and 18 open-source skills. The platform runs as a hosted web application or can be self-hosted in cloud, hybrid, on-premises and air-gapped environments. No local GPU requirement was disclosed in the release details.
The free tier includes access to Voxel51 Agent, 30,000 inference requests and 30 compute hours per month. However, it caps users at three datasets of 10,000 samples each and 20 GB of storage. Voxel51 cites its own survey data regarding performance rather than publishing independent product benchmarks for this release.
Voxel51 Search uses a persistent hybrid index that computes embeddings automatically, enabling semantic retrieval from day one. It supports vector, scalar and full-text indexes via LanceDB, separating storage from compute. Teams can choose which models to embed and when, with incremental updates avoiding full rebuilds when new samples arrive.
Supported data modalities include MCAP, images, video, 3D point clouds, meshes, DICOM, NIfTI, geospatial data and audio. The platform allows lexical and structured queries for exact terms, timestamps and sensor fields. Embeddings and structured fields remain in native form and are combined at retrieval time for more accurate results.
Defender's checklist
- Verify that self-hosted deployments comply with internal air-gapped network security policies.
- Review data retention limits on the free tier to avoid unexpected storage caps.
- Ensure that automated embedding processes do not expose sensitive sensor metadata externally.
- Confirm that vision-language model outputs for labeling meet internal quality assurance standards.
The background in plain language
Physical AI systems rely on diverse sensor data, such as video, lidar and audio, to understand their environment. Managing this data requires aligning different formats into a coherent timeline for training models. Annotation involves labeling these data points so algorithms can learn from them. Agentic workflows use automated tools to chain these tasks together, reducing manual effort and improving consistency across large datasets.
What to do and how to stay safe: Voxel51
- Assess whether your current data curation tools support the specific sensor formats you use.
- Check if your team needs automated embedding and search capabilities for large datasets.
- Evaluate the storage and compute limits of free tiers before committing to a trial.
- Monitor vendor updates for changes in supported modalities and integration options.
Step-by-step guide: AI Governance for Small Teams: Practical Controls and Cost
General security guidance from the TechWire Beat newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What new features does Voxel51 3.0 include?
It adds Voxel51 Agent with 45+ MCP tools, time-series multimodal support, unified annotation and a free self-service tier.
Can Voxel51 3.0 be self-hosted?
Yes, it can be hosted in cloud, hybrid, on-premises and air-gapped setups, or used via the hosted web app.
What are the limits of the free tier?
The free tier caps at three datasets of 10,000 samples each, 20 GB of storage, 30,000 inference requests and 30 compute hours per month.
