On your marks: humanoid robots have left the lab
A few years ago, a humanoid robot meant a video of a machine falling over and getting back up, slowly. Now there are starting guns.
We have been watching this race closely, and we think it is one of the most exciting things happening in technology. Machines with two legs, two arms and a head, built to work in spaces made for people, are walking out of the research lab.
Why a human shape at all? Because the world is built for us: stairs, door handles, shelves, tools. A robot shaped like a person can, in theory, use all of it without anyone rebuilding the building.
The first lap: a robot half-marathon in Beijing
In April 2025, Beijing held the world's first humanoid robot half-marathon. About 21 km, with robots and human runners in separate lanes. Robots could have their batteries swapped mid-race, at the cost of a time penalty.
The winning robot reportedly took 2 hours 40 minutes. Human runners finish far quicker. That is slow, and it hardly matters: machines ran a real course, with battery swaps and a crowd, and finished.
The medal events: the World Humanoid Robot Games
Four months later, in August 2025, Beijing held the first World Humanoid Robot Games. Reports say 280 teams from 16 countries took part, with events including athletics, football and combat sports.
Sport is a clever way to test a robot. A track or a sparring partner demands balance and fast reactions, in public, with a clear winner. Engineers get blunt feedback fast, and the crowd gets to fall in love. Robot football is wobbly and wonderful.
Robots that learn by watching
What makes us sit up is how the robots are being taught. The old way was rules. A programmer wrote out, step by step, how to pick up a cup. A new cup or a new table meant more rules.
Generative robotics flips that. In general terms, robots learn tasks from video of people doing them and from huge numbers of practice runs in simulation, where a virtual robot can fall a million times without breaking anything. The skills then move across to the real machine.
It is an active, widely reported research direction, and it is early. But it explains why progress feels like it is speeding up. It is the same shift that has changed software: less hand-written instruction, more learning from examples.
A sprint on the sidelines: Tesla's next Roadster
A quick detour to the car park. After repeated delays, reports said Tesla planned to unveil its next-generation Roadster on 1 October 2026. We are not covering the event itself here.
What we can share is what Tesla itself says it is aiming for: 0 to 60 mph in under 1.9 seconds, a top speed above 250 mph, a range of about 620 miles, and an optional SpaceX cold-gas thruster package. Those are Tesla's claims, not tested results. Production was reported as not expected before mid-to-late 2027.
Why a car in a robot report? Because of the mood. Rocket-assisted road cars and robots running half-marathons sound like fan fiction from ten years ago, and now they sit in company plans. Whether the Roadster's claims hold up is for the road to settle.
The pit lane: why China's hardware capital is good at robots
Shenzhen is widely known as China's hardware and electronics manufacturing hub, with an exceptionally dense local supply chain. UBTech, one of the big humanoid makers, is based there.
A robot is hundreds of parts: motors, sensors, batteries, cameras, chips, gears. When the people who make each of them are close by, a team can order a part, test it, change it and order it again in a fraction of the usual time. Years of building phones and drones in one region grew that web of suppliers and skilled workers.
Who is in front: the leaderboard, with a health warning
Sources vary and sometimes conflict, so read every number here as a reported figure, not an official scoreboard.
Reports say Chinese makers account for the large majority of humanoid robots shipped worldwide. One report put their share of H1 2026 units above 90%, with Unitree, UBTech and AgiBot the biggest. Rest of World reported that Unitree sold about 5,500 units in 2025 and AgiBot about 5,168, while US makers Figure AI, Agility Robotics and Tesla were reported to have sold only about 150 each.
The US side has money and big plans. Figure AI reportedly raised $1 billion in September 2025 at a $39 billion valuation, and has robots working with BMW, though reports differ on how far that goes. Tesla's Optimus reportedly missed a 2025 target of 5,000 units, and a July 2026 shareholder update reportedly said first-generation Optimus lines were being installed at Fremont.
In China, some makers are aiming at ordinary buyers. Unitree's G1 was listed from about $13,500, and Unitree has been reported to be pursuing a stock market listing. UBTech's Walker S2 is reportedly in car-factory pilots, and reports say a consumer model began home deliveries in China in September 2026.
What it means for jobs
The honest bit first: work will change. Some jobs will look very different, and pretending otherwise helps nobody.
The hopeful bit, which we believe: robots go first where the work is dull, dangerous or heavy. Lifting the same boxes all shift, or crawling into a cramped, hot space. Handing that to a machine can be a good trade, if the person who did it is given the next role.
New roles appear around every new machine: people who train, supervise, maintain, programme, design and sell them, and jobs nobody has named yet. Earlier waves of machines changed work more than they ended it, and the people who pick up the new tools are in the strongest position.
Software Bosh now has a new hardware and robotics division, and this is the field we are watching.
The next lap
We expect the next race to be quicker. The field is crowded, and the gaps are still being worked out.
Nobody has crossed the line yet. The leaders keep changing, and so does the line.