Rayform vs Voxel51
A data platform for search, curation, annotation and evaluation of multimodal data, versus Rayform's processing of robot recordings into synced, checked training episodes.
What Voxel51 does.
Voxel51 makes FiftyOne, an open-source toolkit for exploring and curating visual and multimodal datasets, and a commercial platform around it. Voxel51 3.0, released in October 2026, brings search, curation, annotation and evaluation for physical AI into one platform, with dataset versioning, edit history and an AI agent. FiftyOne can open MCAP robot logs directly as multimodal samples.
They can work together: Rayform's exported episodes can be explored and labelled in FiftyOne.
Where they are strong
- A widely used open-source core with a large community.
- Strong tools for finding, visualising and fixing problems in datasets, including embeddings and model evaluation.
- Dataset versioning and provenance that record which data went into a model.
How Rayform differs
- Voxel51 works on datasets once they exist. Rayform produces them: it aligns every stream on one clock, checks each episode soon after upload and cuts recordings at interactions.
- Rayform is built around the collection session, with usable rates per station and operator and a list of episodes to re-record.
- Rayform is a managed pipeline from upload to training loader; FiftyOne is a toolkit your team runs and extends.
Side by side.
| Capability | Rayform | Voxel51 |
|---|---|---|
| Works with your own recordingsUpload from your own robots, rigs or phones. | Yes | Yes |
| Sensors and robot state alignedRGB, depth, LiDAR and joints on one clock and one coordinate frame. | Yes | Partly |
| Interactions on the timelineGrasps, placements and moved objects detected and marked. | Yes | Partly |
| Search across recordingsFind every grasp, scene or condition across all your data. | Yes | Yes |
| Versioned training slicesFrozen selections that always return the same data. | Yes | Yes |
| Export to training formatsLeRobot, RLDS or WebDataset for training; USD for simulators. | Yes | Partly |
| 3D scenes for simulationSplats and collision meshes of the place, built from your recordings. | Yes | No |
| Self-serve accessStart without a sales call. | Partly | Yes |
Choose Voxel51 when
- Your data is already in dataset form and the main job is exploring, labelling and evaluating it.
- You work across many kinds of visual data, not only robot recordings.
Choose Rayform when
- You need raw robot recordings turned into synced, episode-level training data without writing the pipeline.
- You want collection feedback while operators can still re-record.
Sources
- Voxel51 3.0 announcement, October 2026 (PR Newswire)
- Auditing MCAP robot logs with FiftyOne (Voxel51)
Based on public information as of October 19, 2026. If anything here is wrong, write to contact@rayform.world and we will fix it. Names and trademarks belong to their owners.