Robot sprints to a world record, then crashes into the wall: the other story from Beijing's humanoid games

BEIJING — A Chinese humanoid robot crossed the 100-meter line in 8.64 seconds on Wednesday, breaking the record for the second straight day at Beijing's World Humanoid Robot Games. Then it slammed into a padded barrier, and sparks flew from its waist as staff rushed over with fire extinguishers.

For context, Jamaican sprinter Usain Bolt's men's world record is 9.58 seconds, set in 2009. So the machine — the Tiangong Ultra, built by the Beijing Innovation Center of Humanoid Robotics, also known as X-Humanoid — beat the fastest human ever clocked. The same model set an 8.86-second mark on Tuesday and a 9.39-second one four days earlier, all inside the five-day event.

The twist is that beating a human over a flat, straight track is the easy part. The falls, the smoke and the stretchers that followed the finish line tell a more honest story about where humanoid robots actually stand — and why the industry is racing to move them out of the arena and onto factory floors.

A showcase, with more than 2,000 robots

The World Humanoid Robot Games wrapped its second edition on Wednesday at the National Speed Skating Oval in Beijing. Organizers said more than 2,000 humanoid robots competed across 1,300 contests in 51 events — running, long jump, table tennis, soccer, boxing and even ballroom dancing — in what has become a high-profile display of China's push in robotics.

Alongside the sports, teams ran scenario-based competitions that mimicked real jobs: housework, hotel services and emergency response. That split matters. It mirrors a broader argument inside the industry over whether these machines are ready for anything beyond a carefully prepared demonstration.

"The market debut of Unitree, China's leading humanoid robot company, has stoked fears that Beijing's ambitions for robots exceed demand," Chatham House noted in a September analysis, pointing to a stock that soared past a $66 billion valuation on listing day in Shanghai before sliding around 45 percent below its post-IPO peak. Unitree raised 6.1 billion yuan in an offering more than 8,000 times oversubscribed.

Why sprinting is the easy part

The track conditions are predictable, and experts are quick to say that is exactly what makes the speed records less impressive than they look. The hard problems show up once a robot has to pick its way around a cluttered living room, a warehouse aisle or a factory floor.

"True progress will be measured in warehouses and factories where humanoids do work autonomously for hours and generate real economic value," said Jonathan Hurst, a robotics professor at Oregon State University, in comments carried by the Associated Press.

Dennis Hong, a professor of mechanical and aerospace engineering at the University of California, Los Angeles, spelled out the flip side of the wins. As speed rises, he said, balance, control, impacts, actuator performance and mechanical robustness all get tougher. "Whether robots can move and stop safely, recognise unexpected obstacles, recover from mistakes and continue working" — that, he argued, is what ultimately makes them useful.

The finish-line pileup illustrated the gap. Several robot sprinters, including the winner, crashed into the padded barrier after crossing, and almost all were eventually carried off on stretchers. Staff sprayed fire extinguishers over machines lying on the ground.

China's official People's Daily struck a measured tone in an editorial on Monday. "Housekeeper and caregiver robots are still in the pilot and validation phase," it said, adding that "challenges are precisely where the focus of development lies."

The pivot from spectacle to shop floor

Beijing is betting heavily that these showpiece abilities translate into paid work. The government classified robotics as a priority strategic emerging industry in its 15th Five-Year Plan for 2026-2030 and is standing up a venture capital fund focused on robotics, AI and innovation expected to raise around $138 billion.

The numbers back up the scale of the push. China's makers accounted for more than 90 percent of global humanoid shipments in the first half of 2026, according to Chatham House. Separate estimates differ on the totals — one industry committee said China shipped more than 40,000 humanoid robots in the period, about 97 percent of the world, while Counterpoint Research put global shipments above 22,000 units, or roughly 300 percent higher than a year earlier — but both agree Chinese firms dominate the supply side, with AgiBot, Unitree, Galbot, UBTech and Leju Robotics leading.

The direction of travel is also shifting inside the events themselves. At the World Robot Conference in Beijing in mid-August, robots were no longer selling backflips and piano recitals. They were screwing bolts, sorting parcels and inspecting electrical control cabinets on simulated production lines, and crowds were asking how long a machine could run stably and how high its error rate was.

Regulators are moving in step. On September 1, China's national standard GB/T 47245-2026 on the general architecture of robot intelligent control systems took effect, the first national standard for general-purpose intelligent control across industrial, service and humanoid robots. The same day, the National Development and Reform Commission said it would back small and mid-sized robotics firms crossing the line "from prototype to mass production."

The road ahead

Unitree's founder and CEO, Wang Xingxing, gave a characteristically hedged time line for when robots reach what the industry calls a "ChatGPT moment" — a point where a humanoid could be placed in an unfamiliar home and complete about 80 percent of tasks from text or voice commands alone. He said that could arrive "in two to three years if things move fast, or five to ten years if they move slowly."

Wang conceded that today's humanoids are still less efficient than human workers and generally must be retrained from scratch for each task, and that they struggle with the precision required in the "last few centimetres or millimetres" of a job. On the market's froth, he argued short-term valuation swings mattered less than the longer trajectory.

The economics behind the hype are worth untangling. Unitree's prospectus shows its humanoid revenue still leans heavily on universities, academic bodies and research laboratories, which often buy the machines with state subsidies and sometimes sell the data they generate back to the robot makers. Industry analysts have flagged the circular nature of that arrangement as a risk to its durability. Only a small slice of Unitree's revenue comes from direct sales to factories, a sign that commercial demand from real production lines remains thin.

That revenue mix is why the debate over valuations and the debate over the field's readiness to work overlap so closely. Beijing wants humanoids to behave like its EV industry did — scale up at home on the back of heavy support, then export. But overseas markets are already tightening the door. Washington banned imports of new Chinese humanoid and quadruped robots in late July, citing national security and the risk to critical AI infrastructure, and it has pressured allies to consider similar limits.

Hype versus hardware reality

Robots may be getting dramatically quicker, but several core shortcomings still hold them back. Battery life is often too short for a full working day. Many models struggle with navigation, obstacle avoidance and knowing when to stop — the exact failure that turned sprint wins into a barrier-crashing spectacle. Manual dexterity remains hit or miss, and a robot's perception systems are a bottleneck in areas such as understanding what hides behind a partial view or in low light.

The list of improvements coming down the line is just as long. Chinese maker AgiBot claims a battery life of up to 10 hours on a single charge. Linkerbot says its dexterous hand can place an object with a margin of error of 0.2 millimeters and lift up to 30 kilograms. And Hikvision and others are working on 3D machine vision that captures an object's depth, shape and spatial coordinates. Each of those gains, if it holds up outside the lab, chips away at the gap between a robot that can sprint and one that can work.

Some companies are already putting machines to the test in live settings. Lumos Robotics, a two-year-old startup, has deployed humanoid robots at a Mitsubishi factory in Suzhou to handle mechanical parts, and the founder said the bar for judging these machines has shifted. The old benchmarks asked how fast they could run or how well they could dance. Now, he said, people are asking about practical applications, success rates and return on investment.

Not everyone is convinced the trajectory is smooth. Ni Tao, founder of Yangtzeer.com, a site that tracks technology innovation in China's Yangtze River Delta, said the industry has yet to produce convincing milestones that show robots moving past "flashy showcases of their dexterity and athletic potential into real-world deployments that create genuine commercial and societal value."

For the machines that just shattered records in Beijing, the finish line is the easy moment. The harder test — working an eight-hour shift without bursting into flames — has only just begun.

Humanoid ground is still shifting fast, and China is pushing harder than anyone. For more coverage of the machines and makers shaping this space, follow our robotics and drones section. The AP's full report on the record run is here, and Chatham House's take on the valuation question is here.