3 Of 12 Teams Finish Beijing Robot Games Firefighting Test As Sprints Go Viral

Humanoid robots at Beijing's World Humanoid Robot Games turned speed records and visible failures into the same product demonstration, Ars Technica detailed, with sprinting machines beating human benchmark times while others crashed, broke apart or caught fire after crossing the line.

The second edition of the event ran from August 22 to 26 and mixed spectacle with worksite-style tasks.

Robots raced, danced, performed kung fu, fought in kickboxing matches and played soccer or table tennis, but the programme also added household chores, hotel service work, library shelving, firefighting and hand-dexterity contests.

That split made the games less a simple robotics showcase than a benchmark of where humanoid systems still need engineering discipline.

The fastest runners showed progress in whole-body control for a narrow task, yet several could not slow down safely after a straight sprint.

Purdue University computer science PhD candidate Dipam Patel, who is also a research assistant at the US Army DevCom Army Research Lab, told Ars that the record jumps from last year's results showed a large engineering push for specific motions.

The same expert framed the limitation just as sharply.

A robot that can complete one straight-line sprint may still be engineered only for that task, and some teams used different robots for different events.

The practical weakness is reliability across repeated trials, not a single viral pass; a ten-race format would test whether software and hardware can repeat the performance and stop without damaging the machine or nearby people.

The less dramatic contests gave a clearer view of commercial readiness.

Household events asked robots to wash laundry, fold clothes and clean living rooms, while one unscheduled interruption required a robot to switch from chores to receiving a delivery package.

At the Beijing Continental Grand Hotel, robots moved luggage to assigned rooms, cleaned beds and restocked supplies.

Library and emergency-response tasks added more perception and sequencing pressure.

Chinese state media publication Global Times described a shelving challenge in which robots had to collect returned books and place them correctly, and an outdoor firefighting event gave teams 30 minutes to identify hazardous substances, shut off three valves, find a fire and use an extinguisher.

Only three of 12 teams completed the full firefighting challenge.

Fine motor control remained another separate bottleneck.

Robots picked up beans, drove screws and tried to hammer nails into corkboard without sending them in at an angle.

Those tasks are closer to warehouse, factory or service work than a sprint, and they demand manipulation, visual recognition and recovery from small mistakes.

Autonomy was also uneven.

Many competitors were still teleoperated by humans, and the rules allowed remote control in some events while applying scoring penalties.

Patel's assessment was that robots need far more training across edge cases than humans, who can adapt quickly to unfamiliar tasks.

Simulating 100,000 nail-hammering scenarios still covers only one activity among many.

The computing architecture behind some entrants also pointed away from a self-contained humanoid worker.

South China Morning Post described more autonomous robots using 5G links to send data to an embodied-intelligence system in the cloud and then receive instructions, rather than running all decision-making on onboard hardware.

The investment backdrop is large enough to keep the demonstrations commercially relevant.

More than $6 billion has gone into humanoid robot companies in 2025, with Chinese firms testing aggressively and US companies including Boston Dynamics and Agility Robotics pushing factory and warehouse deployments.

The Beijing games left the strongest evidence in the chores and rescue courses: useful humanoid robots need repeatable autonomy, dexterous manipulation and reliable task switching more than one record-breaking sprint.