Beyond logging and playback: query, index, and curate robotics data

Voxel51 is the multimodal data platform for physical AI. It turns robot recordings in MCAP or LeRobot format into training data that helps VLA and perception models generalize. Rerun is an open-source SDK and viewer for logging, querying, and replaying multimodal sensor data, with a commercial Hub for teams.

Voxel51 vs Rerun: building training datasets compared with inspecting robot logs for robotics workflows.

Rerun Alternative

Why teams choose Voxel51 over Rerun

Rerun gets multimodal, multi-rate data out of a running robot and onto a scrubbable timeline. Voxel51 goes beyond visualization with an end-to-end machine learning platform that connects data to models.

Move beyond inspection to model iteration

While Rerun helps you understand what your robot did, Voxel51 transforms those recorded episodes into datasets that improve models.

Query across your entire dataset

Voxel51 creates a persistent indexed representation of your dataset for fast, interactive queries across episodes.

Unified data platform for VLA development

Bring visualization, data understanding, curation, and multimodal annotation together into a single continuous physical AI data flywheel.

Voxel51 Features

Voxel51 turns sensor logs into your core model advantage

Stop drowning in unindexed episodes and writing custom scripts just to understand your data. Voxel51 helps teams inspect model behavior at scale, turning raw recordings into better models, faster.
Query massive datasets by metadata, temporal events, and annotation labels.
Compute and visualize embeddings across sensor data to find visually similar samples beyond keyword label matches.
Tag arbitrary time intervals to curate data for investigation.
Camera streams, LiDAR point clouds, and numeric sensors play back in sync on a single, shared timeline.
Labels exist as tracks over time. Scrub the timeline to see exactly when an object appears or an event occurs across the full scene.
Query massive datasets by metadata, temporal events, and annotation labels.
Compute and visualize embeddings across sensor data to find visually similar samples beyond keyword label matches.
Tag arbitrary time intervals to curate data for investigation.
Camera streams, LiDAR point clouds, and numeric sensors play back in sync on a single, shared timeline.
Labels exist as tracks over time. Scrub the timeline to see exactly when an object appears or an event occurs across the full scene.

Features

Voxel51 vs Rerun at a glance


Rerun is built for logging multimodal sensor data and replaying it on a timeline, live or from recordings. Voxel51 expands that visualization layer into an end-to-end data platform for VLA and perception models.

Data modalities
Voxel51 logorerun logo
Primary focus
Turning data into better models: dataset curation, model evaluation, recorded log analysis
Logging, real-time visualization, and streaming multi-rate physical data toward training
Core architecture
Built on a database framework optimized for querying, indexing, and subsetting massive datasets.
Built on an Entity Component System using Rust, optimized for real-time visualization.
Live vs. recorded
Reads recorded data to curate datasets & evaluate models
Both live streaming from running robots and recorded data
Visualization
✅ Supported
✅ Supported
Fleet-wide querying
✅ Supported. MCAP indexing is in beta
Yes. SQL and dataframe queries over a local catalog in the open-source SDK, and over a managed catalog across object storage with Rerun Hub (priced through sales)
VLA development support
✅ Supported. LeRobot v3 import and export, plus episode-level curation and temporal tags
LeRobot export and an experimental PyTorch dataloader in the open-source SDK. Rerun Hub streams dataset mixes to GPUs. No VLA-specific tooling listed
Embeddings and similarity search
✅ Supported. Segment embeddings for MCAP recordings are in beta
No built-in vector search listed. A Rerun example runs semantic frame search through external stores such as LanceDB
Scenario mining
✅ Supported via queryable Segment Embeddings and Events projections (beta)
SQL queries into recording values with Rerun Hub. No dedicated scenario mining listed
Annotation & auto-labeling
✅ Supported
Annotation in the shared viewer with Rerun Hub, and annotations as columns with the SDK. No auto-labeling listed
Model evaluation
✅ Supported
Not listed on rerun.io
Versioning & dataset lineage
✅ Supported
Schema changes without breaking history in Rerun Hub. No dataset versioning listed
Data quality and label issue detection
✅ Supported
Not listed on rerun.io
Extensibility framework
Full-Stack JS/Python plugins. Customize anything from front-end panels and Python operators for backend compute.
Log custom 2D and 3D data, coordinates, and tensors into the ECS data model, and configure views with blueprints

“FiftyOne has helped us speed up investigations by 3x. For example, if we see a wrong suction cup grasping an item, we can quickly visualize the issue across all data sources and identify what went wrong.”

Dimitry Pechyoni
Senior Principal Machine Learning Engineer at Berkshire Grey

"With FiftyOne, we were able to cluster 45,000 images from production for an analysis, and instantly found anomalies: products stacked on top of each other, unexpected artifacts, etc. This would've taken days to detect manually. The value that FiftyOne provides is priceless!" — Principal AI Engineer/Data Scientist

Principal AI Engineer, Fortune 500 Health Tech

“FiftyOne has become an important element of our dataset and model development pipeline. Since incorporating the tool into our workflows, we gained a better overview over what data is the best to be used to train and evaluate our models. As a result, we can deliver continuous dataset optimizations into production faster, as our model training pipelines read data directly from FiftyOne.”

Dmytro Prylipko
ML Engineer, EvoLogics
EvoLogics logo

“What really stands out about FiftyOne is the flexibility. The plugin framework lets us customize our workflows based on our unique needs, and the mature SDK lets us consolidate more of our pipeline into one tool, avoiding the cost of stitching together multiple systems. FiftyOne integrates directly into our production pipeline to drive 80% reductions in workplace incidents.”

Patrick Rowsome
Head of Computer Vision Operations, Protex AI

"Some of our dips in model performance were because of edge cases. With FiftyOne, we were able to catch them the same day we analyzed model performance, something that would have taken a week otherwise."

Terrance Whitehurst
ML Researcher, FloVision
Flovision sees 7x model faster model analysis with FiftyOne

Break down data silos with a flywheel you won't outgrow

Point solutions might get your first prototype off the ground, but ad-hoc stacks turn into fragmented data silos as your data scales.
Visualization is only the first step. Voxel51 turns raw robot recordings into queryable, curated datasets for training and evaluation. It reads MCAP and Hugging Face LeRobot datasets natively, and curated episodes export back to LeRobot format for training.

Turn robot data into better models

Join leading robotics and physical AI teams using Voxel51 to query, curate, annotate, and evaluate multimodal sensor data at scale. Turn recorded episodes into the datasets that train better VLA and perception models.