Dexterous manipulation
Grasping bottles, folding cloth, squeezing sponges, and changing tools mid-task.
Embodied-AI data from real cleaning work
Train robots on the physical world.
In a growing pilot, select BerryClean crews capture contact-rich cleaning work in real homes and offices—creating practical manipulation trajectories that laboratory demonstrations and simulation cannot fully reproduce.
Select pilot crewsConsented sessionsReal tools and surfaces
Real tools, varied surfaces, and complete tasks in context.
Robot training at BerryClean.ai
Teaching robots to do the dirty work.
Humanoid robots can write code and hold conversations, but they still struggle to wipe a counter or fold a towel. The missing ingredient is real-world physical interaction data—the kind produced every day by skilled cleaning professionals.
In a growing pilot program, select BerryClean crews wear capture headsets during real cleaning shifts. Homes and offices provide the friction, clutter, deformable objects, changing tools, and unpredictable layouts that simulation cannot fully reproduce.
The pilot grows from BerryClean’s people-first cleaning mission and the real work of its professional crews.
Contact-rich motion
Wiping, scrubbing, reaching, pouring, and applying the right force to varied surfaces.
Long-horizon work
Complete, multi-step cleaning jobs captured in context—not isolated lab demonstrations.
From a real clean to a robot skill
The pilot turns everyday professional work into structured, consented training data for embodied AI.
- 01
Capture
On select pilot shifts, participating crews wear lightweight headsets while completing real cleaning work.
- 02
Protect
Sessions are consented and processed to remove faces and personal details before use.
- 03
Structure
Tasks are segmented and labelled by action, object, surface, tool, and outcome.
- 04
Train
Model-ready trajectories help robotics teams train and evaluate physical manipulation policies.
Building embodied AI?
We are open to conversations about pilots, data requirements, and early access with robotics and physical-AI teams.
Privacy and scope
A selective, consented pilot.
Not every cleaner, shift, customer, or property participates. BerryClean’s current process states that pilot sessions are consented and processed to remove faces and personal details before use.
Sensor modalities, available volume, schemas, delivery formats, licensing, and any additional privacy requirements are confirmed directly for each prospective pilot.
Partnership fit
Start with the task your robot needs to learn.
A useful first conversation covers:
- Tasks and objects your models need to learn
- Required labels, trajectories, and outcomes
- Privacy and consent requirements
- Pilot scope and evaluation goals
- Delivery, licensing, and access expectations
Bring us the tasks your robot cannot learn in a lab.
Tell us what you are training, which physical skills matter, and what a useful pilot would need to prove. We’ll compare those requirements with the work BerryClean crews already perform.