Humanoid robots leave the lab: 2025 is the year deployments go real

Why now

Several things changed roughly simultaneously. Large language models improved to the point where a robot could receive natural language instructions and convert them into actionable steps without requiring every task to be hand-programmed. Vision-language models let robots interpret visual scenes semantically rather than just detecting objects with fixed classifiers. And the physical hardware, particularly the actuators and power systems in humanoid form factors, reached a maturity level where machines could reliably handle objects without breaking them or themselves.

The result is that training a robot to do a new task in a warehouse or factory now takes days to weeks rather than months. The bottleneck shifted from engineering to business development and safety certification.

The key players in mid-2025

Amazon and the warehouse bet

Amazon's interest in humanoid robotics is not surprising. The company already operates one of the world's most automated logistics networks, with hundreds of thousands of traditional industrial robots handling shelves, sorting, and conveyor systems. The problem with those systems is that they are inflexible. A traditional industrial robot is optimised for one task and requires significant downtime and engineering work to be repurposed.

Humanoid robots with general manipulation capabilities promise a different model. A robot that can pick arbitrary items, pack boxes in different configurations, and handle exceptions without falling back to human intervention would dramatically reduce the engineering overhead of automation. Amazon has been running Agility's Digit robots in fulfilment centres and evaluating several other platforms.

The economic case is clear when a humanoid robot that costs $150,000 and requires no benefits, breaks, or scheduling works 24 hours a day on tasks that currently require a human earning $40,000 annually. The payback period, even at current robot reliability levels, is measured in years not decades.

Stargate UAE: the geopolitics of robotics AI

In June 2025, the Stargate initiative announced plans for a major AI and robotics research centre in Abu Dhabi. The UAE had been positioning itself aggressively as a global AI hub, with significant investments in compute infrastructure and partnerships with US AI companies. The announcement signalled that robotics AI was now part of the same geopolitical investment race as language models and chips.

This matters because humanoid robots require enormous amounts of training data and simulation compute. A country that controls that training infrastructure for the next generation of physical AI has a significant industrial advantage. The race for dominance in AI is no longer just about software.

What still does not work

Current humanoid robots are slower than humans at most manual tasks. A trained warehouse worker picks items faster, handles unexpected packaging better, and costs less upfront than most of the current commercial platforms. The robots succeed in environments with high repetition and predictable objects, and they fail badly when conditions are not what was expected during training.

Battery life is a consistent constraint. Most platforms run for two to four hours on a charge before needing to dock, which limits their usefulness in continuous operations without investment in charging infrastructure. Dexterity for small objects and irregular surfaces remains a challenge that LLM-powered planning cannot fully compensate for.