Embodied-AI training data, collected in the real world
Teaching robots to do the dirty work.
In a growing pilot program, select BerryClean crews wear capture headsets while they clean real homes and offices — turning the messiest, most varied physical work on earth into the manipulation and physics data that humanoid robots need to learn, at a scale no lab can simulate.
Backed by a 600+ five-star cleaning operation across the SF Bay Area & Los Angeles.
The AI stack behind every clean
AI isn't a roadmap item at BerryClean — it's how we already operate. We're an AI-native cleaning company, using cutting-edge tools at every layer to run a safer, smarter, more responsive service for our clients and our crews.
Samsara
Samsara's AI platform gives us real-time insight into fleet safety and performance — preventing accidents, protecting drivers with drowsiness alerts, sharpening coaching with behavior data, and cutting downtime through predictive maintenance.
Front
We run every customer interaction through Front's AI. Multilingual translation lets us connect clearly with every client, and smart tagging sorts conversations by topic and sentiment so we respond faster and never miss a beat.
Quo
Quo (formerly OpenPhone) powers our phone line, and its AI receptionist Sona answers 24/7 — capturing leads, sending booking links, and handling questions in English and Spanish so no call ever goes unanswered.
Discord
We coordinate 20+ cleaning crews in real time on Discord — dispatch, scheduling, and on-the-ground communication in one place, so every job runs smoothly from first message to final walkthrough.
The robots are coming. They just can't do the dishes yet.
Every humanoid company is racing toward the same bottleneck: real-world physical interaction data. Models can write code and hold conversations, but they still can't reliably wipe a counter or fold a towel — because that knowledge lives in millions of hours of human hands doing physical work, and almost none of it has ever been recorded.
Simulation hits a wall
Synthetic data can't reproduce the friction, deformation, and chaos of a real sponge on a real countertop. Robots trained only in simulation fail the moment they meet the physical world.
Homes are the hardest environment
Unstructured, cluttered, and never the same twice — a real kitchen is a harder test than any warehouse. Master it, and almost everything else gets easier.
Cleaning is pure manipulation
Grasping, wiping, folding, scrubbing, reaching, pouring. Cleaning is hours of dexterous, contact-rich manipulation — exactly the skills humanoids are missing.
From a real clean to a robot skill
We built a capture pipeline into the day job. On pilot shifts, the work our crews already do becomes structured, consented training data for the next generation of physical AI.
Capture on pilot shifts
Crews in our pilot program wear lightweight headsets as they clean. First-person video, head and hand motion, and contact-rich interactions are recorded across real homes and offices — and we're expanding capture shift by shift.
Sync & anonymize
Footage is uploaded, faces and personal details are automatically scrubbed, and each session is tied to its task, surface, tool, and environment — with full consent from our crews and clients.
Curate & label
We segment sessions into discrete manipulation episodes, extract hand and body pose, and annotate actions, objects, and outcomes into clean, model-ready trajectories.
Train better robots
Curated datasets ship to robotics teams to train and evaluate manipulation policies — grounded in how real work actually gets done, not how a simulator imagines it.
A data moat you can't teleoperate your way to
Most robotics data is collected by paid teleoperators in a lab for a few weeks. We collect ours from a real, profitable cleaning company that has been serving the Bay Area and Los Angeles for years — with hundreds of professionals already in real homes every single day.
That means scale, diversity, and authenticity that a research lab simply can't buy: real mess, real materials, real constraints, captured by people who actually know how to do the work well.
Five-star-rated service track record
Major metros: SF Bay Area & Los Angeles
Cities and neighborhoods served
Consented & anonymized capture
Built for the teams building embodied AI
The companies chasing general-purpose robots — from Figure to Tesla to Apptronik — are all bottlenecked on the same thing: data from the physical world. BerryClean.ai is the supply line.
Humanoid robotics companies
The teams racing to ship general-purpose humanoids all face the same wall: real-world manipulation data. We supply it.
Foundation-model labs
Training a vision-language-action model? Ground it in millions of real human manipulation trajectories instead of synthetic approximations.
Robotics startups
Skip the cost and lead time of standing up your own teleoperation fleet. License curated data and move straight to training.
Academic & research labs
Access diverse, consented, real-world datasets for benchmarking and publication that lab-collected data can't match.
A flywheel, not a research project
BerryClean is two businesses that make each other stronger: a top-rated cleaning service and the real-world data engine it powers. Each clean improves the data; better robots make each clean better.
Cleaning funds capture
BerryClean.com is a real, growing cleaning business. Every booking puts another professional — and another headset — into another real home.
Capture trains robots
Those shifts become the manipulation data robotics teams use to teach machines how to handle the physical world.
Robots lift the crews
As robots take on the heavy, repetitive grunt work, our people move into higher-skilled, better-paid roles — and the loop compounds.
Our pledge to the people who make this possible
Every dataset we sell is built on the skill of real cleaning professionals. Our pledge is simple: use this technology to empower them.
We don't capture this data to put our crews out of work — we capture it because the people who clean for a living are the world's foremost experts in physical manipulation, and that expertise deserves to be valued.
As robots take on the heaviest, most repetitive tasks, our professionals move into better-paid, higher-skill roles: training, quality control, and supervising the machines they helped teach. AI is a tool to elevate people, not erase them.
Get early access to the dataset
Building a robot that has to work in the real world? Tell us what you're training and we'll show you what our crews are capturing. Partnerships, pilots, and early data access are open now.