Dexterity and agility are software problems before they are hardware problems, and Google is betting accordingly. Google LLC's December 24, 2024 grant US12172309B2 claims learning agile locomotion for multiped robots — using machine learning, not hand-coded gaits, to control legged machines.
Read the cost thesis in the method. The CPC tags — B25J 9/161 and B25J 9/163 robot control, B62D 57/032 legged locomotion, G06N 3/042 and G06N 3/08 neural learning, G05D 1/0891 autonomous motion — describe learned control. The inventors are a reinforcement-learning group.
The capex read is about how control scales. Hand-tuning a robot's gait for every surface and task is engineering labor that does not amortize. A learned controller that generalizes across conditions turns that recurring labor into a training cost paid mostly once. If it works, the cost curve bends down hard.
For a public-equities reader, this is Alphabet applying its core competency — large-scale machine learning — to robotics control, the same way it applies it everywhere. The bet is that software leverage beats mechanical-engineering labor as the scaling mechanism for robot capability.
The honest limit: a learned-locomotion patent does not prove the controllers generalize well enough to deploy, nor does it size the savings. Sim-to-real and robustness gaps are real. The grant signals the bet, not its payoff.
The takeaway for the money desk: the cost structure of robotics increasingly turns on learned versus hand-tuned control. Track who is patenting learned locomotion — they are wagering that training scales cheaper than engineering.