Manipulation & household tasks
Capture varied human demonstrations and robot interactions across natural environments.
Physical AI data collection
Turn real-world capture into data your models can learn from.
Design and run multimodal collection programs, then carry every session through synchronized ingest, labeling, QC, and versioned dataset release—ready for model evaluation.
One operating record from capture plan to published dataset.
Capture program
Ready to run- RGB video
- Depth
- IMU
- Robot state
- 01
Capture in context
Synchronized - 02
Qualify the hard cases
QC passed - 03
Release and improve
v12 ready
Evaluation findings
Hard cases define the next targeted capture cycle.
Egocentric capture
Robot teleoperation
Custom sensor rigs
Field operations
One operating loop
Collection is the first mile—not the final deliverable.
Keep sensor setup, raw sessions, annotations, QC decisions, and dataset releases in one traceable workflow—then use evaluation findings to plan the next collection cycle.
- 01
Plan for the model
Define target tasks, environments, sensor schemas, and acceptance criteria before collection begins.
- 02
Capture in context
Coordinate egocentric, teleoperation, fleet, or custom-hardware sessions without separating the sensor streams.
- 03
Qualify the hard cases
Ingest, inspect, label, and review the data that matters before it reaches training.
- 04
Release and improve
Publish a versioned dataset, export it to your stack, and use downstream evaluation findings to scope the next collection cycle.
Dataset operations
Give model teams an operating record—not a folder of files.
Package source sessions, quality state, and export history into a versioned release that model teams can evaluate and use to plan follow-up.
- Synchronized sensor and interaction sessions
- Structured annotations and review decisions
- Versioned dataset releases and export history
- Evaluation findings ready for targeted follow-up
Dataset releaseExport ready
- Dataset
- manipulation-kitchen
- Release
- v12
- Modalities
- RGB · Depth · IMU · Robot state
- Quality gate
- QC passed
Programs
Collect against the behavior your model must learn.
Run focused programs across human demonstration, robot interaction, mobile autonomy, and industrial work.
Warehouse automation
Build coverage for pick-and-place, sorting, palletizing, and inventory workflows.
Mobile autonomy
Collect edge cases across navigation, weather, terrain, and human interaction.
Industrial robotics
Record tool use, assembly, inspection, and contact-rich tasks with the context training teams need.
Start with one model objective
Design the collection program your next model run needs.
Bring your sensors, target tasks, environments, and current data bottleneck. We’ll scope the capture plan, quality gates, and path into training.