By Ethan Lee
A humanoid robot named Tiangong Ultra delivered a striking performance at the second World Humanoid Robot Games in Beijing, completing a 100-meter sprint in 9.39 seconds on Aug. 22.
The time was 0.19 seconds faster than Jamaican sprint legend Usain Bolt’s 9.58-second world record, set in 2009. Another robot, Blitz, reportedly finished in 9.47 seconds.
On Aug. 26, Tiangong Ultra broke its own record, finishing a 100-meter dash in just 8.86 seconds.
The most revealing moment, however, came after the race. Tiangong Ultra could not slow down after crossing the finish line and crashed into a protective barrier. Blitz also struck the padding and fell. Staff eventually carried both machines from the arena on stretchers.
For the best of our weekly content！
You are now signed up for our newsletter
Check your email to complete sign up
The spectacle captured the contradiction at the center of humanoid robotics: Machines can now outperform humans in tightly defined physical tasks, yet still struggle with basic control, adaptability and judgment.
Remarkable gains–and obvious limitations
Tiangong Ultra had required 21.50 seconds to finish the same distance a year earlier. Its performance more than doubled within a year. It also reportedly completed 400 meters in 39.70 seconds and demonstrated a standing high jump of 2.8843 meters.
Yet speed and strength do not necessarily translate into practical ability. During an Aug. 24 weightlifting event, another robot raised a 16-kilogram barbell to its chest before losing balance and moving toward the judges’ area. It fell and was also removed on a stretcher.
READ MORE:
A USB-plugging contest offered an even clearer illustration. Of the 24 teams that entered the preliminary round, only 14 advanced to the final.
Inserting a USB cable requires a robot to identify the connector’s orientation, locate the port, position its hand, adjust the angle and apply the correct amount of force. A deviation of only a few millimeters can cause the attempt to fail.
For humans, this seemingly effortless action combines vision, spatial awareness, hand-eye coordination, force control and the ability to adjust after failure. These are precisely the capabilities robots still find difficult to integrate.
From sporting events to household work
This year’s games attracted 666 teams and 2,056 robots from 16 countries across six continents. The program expanded from 26 events last year to 51.
Thirty events resembled conventional sports, including running, soccer, table tennis, weightlifting and tug-of-war. The other 21 focused on possible workplace applications, such as industrial assembly, warehouse logistics, hotel service, household cleaning, emergency response and precision tasks.
More than 40 percent of the assignments required robots to operate autonomously. Rather than testing whether a machine could perform a movement under remote control, organizers wanted to know whether it could interpret its surroundings, recover from mistakes and respond to unfamiliar conditions.
In a hotel-service event, robots had 30 minutes to move luggage, deliver supplies, remove used linens and make a bed. Handling fabric proved particularly difficult because soft materials constantly change shape. Too much force can pull a sheet off the bed, while too little leaves it wrinkled.
The contrast is revealing: Running faster than Usain Bolt may be easier for a robot than handling a bedsheet.
Training for the physical world
At training centers in Beijing, workers repeatedly teach robots to perform ordinary actions, such as picking up an onion, placing it on a scale and returning it. A single movement may be repeated hundreds or thousands of times under slightly different conditions.
Unlike language-based artificial intelligence, embodied AI must interact with an unpredictable physical environment. Cups can slip, water can spill and furniture can block a robot’s path. A task easily completed by a young child may require tens of thousands of examples before a robot can perform it reliably.
China has made embodied AI a national industrial priority, with plans to deploy humanoid robots in manufacturing, logistics, retail, health care and emergency services.
Some machines have already entered factories. In April, robots developed by Shanghai-based AgiBot reportedly completed an eight-hour shift at an electronics plant in Jiangxi, performing 2,283 pick-and-place operations without human intervention.
BMW has also tested Figure 02 robots at its Spartanburg, South Carolina, plant. According to BMW, the machines worked about 1,250 hours, moved more than 90,000 components and contributed to the production of more than 30,000 BMW X3 vehicles.
Commercial promise meets industrial reality
Despite these demonstrations, widespread industrial adoption remains distant. Wang Xingxing, founder of Unitree Robotics, said his company had worked with an automobile factory in 2024 but could not expand the project because the robots remained slower than human workers and lacked adaptability.
Robots trained in fixed environments can achieve high success rates, but small changes in lighting, object placement or materials may sharply reduce performance. Many industrial jobs also require submillimeter accuracy, while current robots generally operate at centimeter- or millimeter-level precision.
The industry also lacks common standards. Differences in sensors, control systems and data structures make it difficult to reuse costly training data across machines and tasks.
Concerns about an investment bubble are growing as China’s humanoid-robot sector expands. The country reportedly has nearly 370 robotics startups and more than 90 completed or planned training centers. Yet the economic value created by practical applications remains far below many company valuations.
Technical progress and financial excess can exist simultaneously. The real test will not be whether a robot can break a sprint record, but whether it can work safely, consistently and economically in changing environments.
Original article
Comments