Part of the Robotic Movement series. Project page section: Manipulation.
The longer-term goal
Finding the optimal movement is the first step. The next abstraction is knowing which movement to use for which task, and the far end of that road is a robot folding laundry. Doing the laundry and other everyday tasks with a bottom-up framework, built from the robot’s own body rather than from human demonstrations, is what would set this apart from how other humanoid companies approach it.
Making it solid
Early manipulation scenes did not hold up to scrutiny. Surfaces and objects were not solid, and objects were not genuinely being picked up by the hands. So the rule became: take one example, a sponge or brush wipe, and make it rock solid, with no glitches and no unrealistic physics, before going wider.
The result so far is one real finger grasp: the G1’s Dex3 hand wrapping a brush handle and then wiping with controlled force. It holds in 6 of 10 stress tests. Every other hold in the project is an attachment, and it is labelled as one.
Why chores waited
I set chores aside deliberately. Until the method reliably derives the optimal movement from a body’s own components, tasks like catching or wiping are the wrong shape of test: they mostly measure grasping and perception, not movement. The order is movement first, then the choice of movement, then tasks. Household tasks remain the longer-term goal.
Previous: Carrying with the body. Next: No demonstrations.