Voxel51 has released version 3.0 of its platform, merging search, curation, annotation and evaluation of multimodal data into a single system aimed at teams building physical AI — systems such as robots and vehicles that must understand the real world.
The release ships with a built-in Voxel51 Agent, more than 45 MCP tools, over 80 operators and 18 skills, plus activity logs, edit history and versioning so training data stays traceable — features the company positions against rivals such as Encord and Foxglove.
Data plumbing rarely makes headlines, but it is where applied AI projects quietly succeed or fail: models are only as reliable as the labelled, searchable datasets behind them, and physical AI multiplies that problem across cameras, LiDAR and time-synced sensor streams.
The platform offers a free tier covering three datasets of up to 10,000 samples each with 20 GB of storage, with paid team and enterprise options including on-premise and air-gapped deployment.
