Data for machines that operate in the real world.
Multi-sensor capture, annotation and evaluation for any system that has to perceive, move or decide in the physical world.
- Manipulation
- Humanoids
- Autonomous vehicles
- Aerial autonomy
- Maritime
- Counter-UAS
- Industrial inspection
- Infrastructure
- Surgical and medical devices
- Mapping and change detection
- Manipulation
- Humanoids
- Autonomous vehicles
- Aerial autonomy
- Maritime
- Counter-UAS
- Industrial inspection
- Infrastructure
- Surgical and medical devices
- Mapping and change detection
What the data looks like
Every modality, on one timeline.
Sensors are synchronized and calibrated at capture, so labels line up across RGB, depth, thermal and LiDAR.
Solutions
Built for every place models meet hardware.
Six domains, one standard for how data is collected, labeled and checked.
Robotics
Demonstrations, multi-view capture and annotation for manipulation, navigation and humanoids.
Explore robotics → SolutionAutonomy and mobility
LiDAR, camera and radar data for ground, aerial and maritime vehicles.
Explore autonomy → SolutionDefense and security
Perception data for degraded visual environments, low light, clutter and GNSS-denied movement.
Explore defense → SolutionIndustrial and infrastructure
Inspection data for plants, pipelines, grids, and construction and mining sites.
Explore industrial → SolutionHealthcare and life sciences
Device and procedure capture, where consent and compliance allow.
Explore healthcare → SolutionGeospatial and environment
Aerial and satellite imagery, mapping and change detection.
Explore geospatial →Market
Training data is a multi-billion dollar business. Physical data is the gap.
Vendors of human and expert data for text and code have grown into some of the fastest-growing software companies. Data from the physical world has not had its equivalent yet.
Why the gap exists
Incumbents were built for screens, not sensors.
- Built for textLeading vendors scaled on text, code and expert judgment. Physical data needs hardware, sites and safety.
- Cannot be bought or scrapedRobot and sensor data has to be created for each task and environment. There is no corpus to license.
- Capture competes on volumeMany capture vendors sell hours. Volume alone is a weak moat.
- Controlled domains are underservedDefense, healthcare and infrastructure cannot use open crowd pipelines.
- Synchronized by designHardware triggering or a shared clock, with sensor and host timestamps logged.
- Failures are keptFailed attempts are labeled and retained, so policies learn recovery.
- Every label has an ownerVersioned guidelines, independent review and documented adjudication.
- Handling comes firstControlled requirements are scoped before collection begins.
Left: why the gap exists. Right: how Cato Labs is different. Market figures are third-party estimates. See the full investor overview.
How it works
From a spec to a dataset you can train on.
- Step 1
Scope
We define the task, sensors, environments, label schema and acceptance criteria with your team before anything is recorded.
- Step 2
Collect
Field capture against a written plan. Every session logs device, calibration, operator and conditions.
- Step 3
Annotate
Labeling against the agreed schema, with guidelines that are versioned and shared with you.
- Step 4
Verify
Multi-pass review and expert adjudication of disagreements. Quality results ship with the data.
- Step 5
Deliver
Versioned releases in the formats your pipeline already reads, with a changelog for every update.
Research
Notes from the field.
How we think about synchronization, demonstrations and labeling in conditions that are not ideal.
Timing beats resolution
Why a few milliseconds of sensor offset can cost you more than a resolution upgrade.
What makes a demonstration usable
The details of teleoperation data that decide whether imitation learning works.
Labeling when you cannot see well
Guidelines for smoke, low light and occlusion, where forcing a label adds noise.
Standards
Every frame has a provenance. Every label has an owner.
Datasets for physical systems are only as good as the record of how they were made. We keep that record with the data, and we scope handling requirements before collection begins.
- ProvenanceCapture metadata, calibration, device and operator travel with every release.
- ReviewIndependent review, expert adjudication, and documented outcomes.
- HandlingScoped, logged access. Controlled delivery options agreed up front.
FAQ
Questions teams ask first.
What formats do you deliver?
The formats your pipeline already reads. For example, MCAP or ROS bags for sensor logs, Parquet or HDF5 for episodes, and COCO-, KITTI- or nuScenes-style annotations. Every release includes calibration files, split definitions and a changelog.
Can you collect on our hardware?
Yes. We can record on your platforms, on ours, or a mix. We agree on sensors, mounting and calibration before capture starts.
How do you handle controlled or restricted work?
Requirements are scoped before collection begins. Some programs need data to stay on your infrastructure or to follow agreements specific to the work. We will tell you plainly what we can and cannot take on.
How long does a pilot take?
It depends on the task, environment and volume. You get a written plan with milestones and acceptance criteria before you commit to anything.
Do you sell off-the-shelf datasets?
The work is mostly custom collection against your spec. If you need a particular scene or condition, ask and we will tell you the fastest way to get it.