Real-world training data for Physical AI
Curatrix delivers precisely structured robotics datasets captured from skilled German tradespeople—in real working environments, with explicit usage rights.
Teams training robots on real manipulation work
Curatrix is the demonstration-data layer for groups that need recordings of skilled physical work, not synthetic approximations of it.
- Physical-AI startups
- Robotics foundation-model companies
- Robot manufacturers
- Industrial robotics integrators
- Embodied-AI research teams
- University robotics labs
- Head of Robotics
- Robotics Research Lead
- Embodied AI Lead
- ML Lead
- Data Lead
- Director of Automation
- Research Scientist
Three ways to get the data
License what already exists, commission what does not, and manage both from one portal.
Dataset Library
Browse and license existing datasets. Every release ships with episode manifests, annotations, provenance records and licence terms.
Custom Data Collection
Commission a dataset for a specific robotics task. We write the capture protocol, recruit qualified tradespeople and authorized sites, and deliver to your specification.
Data Platform
A secure portal to preview episodes, monitor running collection projects, download licensed versions, and manage team access and API keys.
Capture, process, structure, deliver
Every dataset moves through the same four stages, under the same documented protocol.
- 01
Capture
Authorized first-person recording at approved sites. Head, helmet and chest-mounted cameras, fixed protocol, documented conditions.
- 02
Process
Personal and confidential material is detected and removed. Every clip passes automated privacy detection and a human privacy review.
- 03
Structure
Footage is segmented into task episodes and annotated with tools, objects, actions, phases and outcome states.
- 04
Deliver
A versioned release with manifests, annotations, provenance, licence terms and documentation, in the formats your stack expects.
Featured datasets
Representative samples from the library. The figures shown illustrate the structure and volume of a release.
Electrical Assembly
Fine-grained assembly sequences from professional electrical work.
- Episodes
- 1,420
- Accepted hours
- 56 hr
- Contributors
- 12
- Size
- 428 GB
Cable Routing
Egocentric cable preparation and routing across active worksites.
- Episodes
- 880
- Accepted hours
- 34 hr
- Contributors
- 8
- Size
- 276 GB
Industrial Fastening
Tool-rich fastening procedures with inspection and rework states.
- Episodes
- 1,100
- Accepted hours
- 41 hr
- Contributors
- 10
- Size
- 351 GB
Built on paid, authorized skilled work
Contributors are paid for their time and agree to every recording. Cumulative figures to date.
Specify a task. We collect, process, and deliver the dataset.
A custom collection runs as a defined engagement, from data specification through to a measured result on your benchmark.
- 01
Define the task and the data specification
- 02
Design the controlled capture protocol
- 03
Recruit tradespeople and authorize sites
- 04
Record the demonstrations
- 05
Privacy-clear every clip
- 06
Segment into episodes and annotate
- 07
Build the versioned dataset
- 08
Deliver in the required format
- 09
Measure model and benchmark improvement
Rights, privacy, and provenance are built into the data pipeline.
Designed for GDPR-conscious physical-data collection.
Privacy-by-design architecture
Data minimisation and access limits are properties of the pipeline, not policies bolted on afterwards.
Controlled, authorized capture
Recording happens only under an agreed protocol, at agreed times, for an agreed task.
Clear worker participation
Contributors know what is recorded, what it is used for, and are paid for their time.
Site authorization
Every capture location is authorized in writing by the party responsible for it before recording begins.
Pseudonymous contributor identities
Delivered data references contributors by pseudonymous identifiers only.
Automated detection plus human review
Automated privacy detection is followed by human privacy review before any clip enters a release.
Dataset provenance
Each episode carries its capture context, protocol version and processing history.
Purpose-based access control
Access is granted against a stated purpose and licence scope, not by blanket account permission.
Versioned dataset releases
Releases are immutable and versioned, so a training run can always be tied to an exact dataset state.
Full disclosure logging
Every disclosure of data is logged, with recipient, scope and time.
EU-region data infrastructure
Storage and processing run on infrastructure located in the EU region.
No worker-performance analytics
Recordings are never used to rate, rank or monitor the people who contribute them.
Designed for GDPR-conscious physical-data collection. Rights, privacy, and provenance are built into the data pipeline.
This describes how the pipeline is designed and operated. It is not a claim of certification, and no dataset is presented as fully anonymous.
Video + structure + annotations + provenance + licence + documentation
What you receive is a documented dataset release, not a folder of videos.
| Format | Contents | Status |
|---|---|---|
| MP4 | Video | Available |
| Parquet | Episodes and annotations | Available |
| JSON | Manifests, calibration, provenance and licensing | Available |
| MCAP | Synchronized sensor streams, where applicable | Available |
| Python loaders | Customer-specific, written against your training pipeline | Available |
| LeRobot export | LeRobot-compatible export | Planned |
| ROS export | ROS-compatible export | Planned |
Dataset releases and methodology notes
Occasional updates on new releases, capture protocols and annotation schema changes. No product marketing.
Start with a scoped pilot
Bring a task. On the call we scope the capture protocol, the episode volume, the annotation schema and the delivery format.
Pilots typically start from approximately €25,000. This is an indicative figure only—final scope and pricing are set after a technical qualification call.