This week, a Sony robot beat professional table tennis players, humanoid machines took centre stage at Germany's biggest industrial fair, and a Chinese robotics company opened its U.S. headquarters in Dallas with 15,000 units already deployed across the Americas. Taken individually, each story has the feel of a corporate announcement. Taken together, they suggest that physical AI is crossing a threshold: the era of the impressive demo is giving way, slowly and unevenly, to something that looks more like an industry.
Sony's robot, called Ace, was detailed in a paper published this week in the journal Nature. It was pitted against professional and highly skilled athletes at an Olympic-sized table tennis court Sony built at its Tokyo headquarters, and it won often enough that the company's AI division called it "the first time a robot has achieved human, expert-level play in a commonly played competitive sport in the physical world." The technical achievement is genuine: table tennis requires the kind of fast, adaptive response to an ever-changing environment that robots have historically struggled with. Ace uses nine cameras and reinforcement learning to track ball spin and respond in real time. What makes it interesting beyond the headline, says Sony AI president Michael Spranger, is the underlying lesson: you can train robots to be fast and adaptive in environments that constantly change, not just repeat the same trajectory.
At Hannover Messe 2026, Germany's flagship industrial fair, humanoid robots shared the floor centre for the first time. Among the exhibitors was Accenture, which presented results from a pilot it ran with Vodafone Procure & Connect and SAP at a warehouse in Duisburg. A humanoid robot received inspection tasks through SAP's warehouse management system and autonomously carried them out: finding misplaced products, assessing pallet stacking, flagging aisle obstructions, and feeding findings back into the system in real time. It is a narrow set of tasks, but it is real deployment in a real warehouse with real enterprise software integration, which is a step beyond most humanoid robot pilots that conclude with a press release and no published outcomes.
Pudu Robotics, which makes service robots for hospitality, cleaning, and industrial logistics, opened a U.S. headquarters in Dallas this week and reported that nearly 15,000 of its robots are already deployed across the Americas, with regional revenue up 285% year over year. These are not humanoid robots in the science-fiction sense: they're purpose-built service machines that carry food trays and scrub floors. But Pudu's scale points to where the sector is mature and where it isn't. Repetitive, spatially simple tasks in constrained environments: solved. Flexible general-purpose labour: still ahead.
That gap is worth sitting with. A piece in Stuff South Africa this week offered a useful corrective to the week's more exuberant coverage. Honor's Lightning humanoid robot recently ran a half-marathon faster than any human, but it did so on a prepared track with extensive engineering support. The choreographed demonstrations that dominate robotics marketing are a real indicator of what's possible, but they're not indicative of what robots can do unsupervised in messy real-world environments. "You can't hand an Amazon-bot a broom and tell it to clean up a mess," the piece notes, "not without significant engineering involvement." Functional speed is improving, but the ability to switch tasks without retraining remains a hard problem.
The market is pricing in faster progress regardless. A report this week from Precedence Research estimated the humanoid robot market at $2.16 billion in 2026 and $8.78 billion by 2035, driven by more than 610 investment deals in 2025 alone. Tesla has signalled it will spend $25 billion on AI this year, with its Optimus humanoid program a significant part of that. The money is real and so is the pace of improvement. The honest answer to whether humanoid robots are arriving is: yes, in specific, narrow applications, with considerable caveats still attached.
What's shifting is less the capability ceiling and more the commercial architecture around it. Pudu's Dallas HQ, Ati Robotics' rebranding around "material orchestration," Accenture's SAP integration work: these are the unglamorous signals that serious capital is learning how to sell and service physical AI rather than just demonstrate it. The demos get the headlines. The integrations get the contracts.