Training Pipeline

Six stages from raw hardware to a trusted housekeeper

One closed loop, run continuously. Every deployed robot feeds new data back into stage one, so the fleet gets better without waiting for a hardware refresh.

STAGE 01

Capture

Tele-operated demonstrations collected inside 24 real furnished testbeds and partner homes — real light, real clutter, real interruptions.

≈ 4,800 new episodes / week
STAGE 02

Annotate

Semantic labelling of objects, surfaces, fragility and household preference. Human reviewers resolve ambiguity before it reaches the model.

3-stage QC · 99.2% label agreement
STAGE 03

Simulate

Each episode is cloned into thousands of physics variants — lighting, friction, object pose, room layout — and trained in parallel.

1 episode → 3,000+ variants
STAGE 04

Verify

Candidates are replayed on physical robots under a standardised home benchmark. Nothing ships on simulation scores alone.

HomeBench-200 · 200 tasks, 6 room types
STAGE 05

Deploy

Verified skills are packaged and pushed over the air, versioned, rollback-safe, and scoped to the specific robot platform.

OTA · per-skill versioning · instant rollback
STAGE 06

Learn

Failure cases, tele-op interventions and user corrections are auto-triaged back into Capture — the flywheel never stops.

Median 11 days, failure → fixed skill
∞

The data flywheel: more homes → more edge cases → better skills → fewer interventions → more homes. Competitors can buy the same arm we train on. They cannot buy the loop.

Training Infrastructure

Where the training actually happens

Three physical assets that software-only competitors cannot replicate: real homes, a human tele-op fleet, and a benchmark that refuses to lie.

Household robot mopping the floor in a real living-room testbed

Real-Home Testbeds

24 furnished apartments in Suzhou and Shenzhen — studio to three-bedroom, with pets, children and deliberately imperfect layouts.

Household robot loading a washing machine during a teleoperation demonstration

Teleoperation Fleet

Trained operators collect demonstrations and take over on failure. Every takeover is logged, triaged and converted into training data.

Household robot wiping a kitchen counter during HomeBench validation

Simulation & HomeBench

A GPU cluster replays each episode across thousands of physics variants; HomeBench-200 scores the result on real hardware before release.

24/7

Tele-op safety net with a median takeover time under 8 seconds.

3,000×

Simulation multiplier per captured real-world episode.

11 days

Median time from field failure to certified skill fix.