Amazon open-sourced the climb. The hands were human.
Two releases, fourteen days apart. On July 30, Google DeepMind launched Gemini Robotics 2, whole-body intelligence spanning locomotion, manipulation, and task reasoning, with Apptronik, Boston Dynamics, Franka, and a good share of the credible Western body ecosystem renting the brain rather than building their own. Days later, Amazon FAR, the frontier robotics lab Amazon assembled around the Covariant team, published LadderMan: a humanoid that climbs real ladders with zero-shot sim-to-real transfer and works at height. Then they open-sourced the code.
Look at what the Amazon team chose to demonstrate. Not a warehouse pick. Straightening a wall painting. Retrieving a box from a high shelf. Tightening a light bulb. One of the hardest whole-body control problems in robotics, climbing under contact constraints, was introduced to the world doing household chores.
The honest detail worth more than the highlight reel
The part we respect most is a design choice the authors state plainly: the climbing policy runs autonomously, and the on-ladder manipulation is teleoperated by a human. Amazon FAR, with Covariant's lineage and Amazon's budget, shipped its flagship whole-body demo with human hands in the loop, and said so. That is how this entire field actually ships: autonomy where the model is strong, human demonstration where it is not yet, and the dial moves as the data accumulates. Driving went through the same curve in public. The teams that name the human in the loop are the teams you can believe about everything else.
It is also the register we build by. Supervision that is disclosed, permissioned by the family, and logged in the attested record is a property of a trustworthy system, not a confession. The alternative, pretending the hands are not there, is how a field loses a decade of trust in one exposé.
When giants give the motor layer away
A fortnight in which Google rents a whole-body brain to any body maker, and Amazon gives whole-body skill code away under an open license, is a fortnight that settles an argument. The motor layer is not going to be anyone's moat. It is becoming a merchant layer with multiple strong suppliers, some renting, some open-sourcing, all improving fast. For everything built on top of it, this is the best possible news: capability arrives from the supply chain, and every improvement lands in the stack the same day it ships.
And notice, again, what neither release contains. No memory. No identity. No permissions. Nothing about whose machine this is, what it has learned about this home, what it may do in which rooms, or how a family would verify what it actually did. We wrote in July that the Android shipped and the account didn't. Two of the largest companies on earth just spent a fortnight proving the first half, in public, partly for free.
The climb will be commodity soon; the code is right there, and we are glad. What sits above it is still unbuilt: the mind that remembers the home, holds the permissions, and can prove what it did. The climbing is becoming free. The trust is not. That is the part we are building.