Where humanoid robots stand in 2026

Where it stands. Makers including Figure, Tesla (Optimus), Unitree, Boston Dynamics and Agility moved from prototypes towards scaled manufacturing and paid pilot deployments — mostly narrow, supervised, industrial work. KraneShares (May 2026) described the shift as the race moving from 'can we build it?' to 'can we scale it?'. Most deployments involve moving goods in warehouses or controlled factory tasks. General-purpose robots remain years away.

The accelerant. AI: large neural networks replacing hand-coded control — giving robots far better ability to handle unpredictable real-world situations — plus world models (like Decart's Oasis 3, launched June 2026) that generate simulated training environments, cutting the cost and time of preparing robots for new tasks. Nvidia sits underneath much of it as a compute-and-simulation supplier rather than a robot maker.

What to watch. The signal to track is 'second-deployment time': how fast a maker signs a second paying customer in a different task. When that shrinks to weeks, the AI-learning curve has bent. Until then, be sceptical of 'general purpose' framing and verify unit-shipment claims independently. The companies that win the next phase will have the most boring, repeatable shipping cadence — not the most impressive backflip. Software generalisation speed, not hardware polish, is the deciding variable.