Robots delivered by the thousand: what's sold may not be the robot

Let me say something. Flipping through the Maker H01 launch materials, the 20+ DOF and kilogram-level payload multi-finger hand didn't really catch my eye — what stood out most was the line "a data entry point designed for the data closed loop." It nails down the product positioning.

The evidence is in the sensor layout. Head-mounted stereo RGB-D, torso and wrist for blind-spot fill, omnidirectional lidar on the chassis covering 360 degrees, dual arms plus mobile base. The real purpose of this layout is to let every failure be fully recorded. Ego view, wrist close-up, chassis global — three streams aligned, exactly the observation structure VLA training needs. These past few days I've been calibrating the wrist extrinsics on another wheeled dual-arm rig, and just unifying the camera and chassis coordinate frames took me half a day — looking back at this native multi-view arrangement, it really saves hassle. What it saves is the time of backfilling data afterward.

So that thousand-unit delivery thing — my read is it's measuring data mix ratios. What proportion each scenario type takes, how many failure samples to keep, how to blend real and sim — those matter far more than whether the robot can serve tea. Same for the world model line — what the RL behind GigaBrain gets fed depends on what the body collects.

As for the home scenario, set that aside for now.