A Beijing Robot Broke Usain Bolt's 100m Record, Then Crashed

On the opening day of the second World Humanoid Robot Games, held August 22 at Beijing's National Speed Skating Oval, a two-legged robot built by the Beijing-based company X-Humanoid ran 100 meters in 9.39 seconds. Usain Bolt's world record, set in Berlin in 2009, is 9.58 seconds. A robot from a separate team, Honor, had already run 9.32 seconds at a preparatory trial before the Games began, meaning two machines beat the human record inside the same week.

The finish line was not the hard part. Organizers had placed a thick padded barrier several meters past it, and the sprinting robots slammed into that barrier to stop, some falling apart on impact and needing to be carried off on stretchers. Bringing a running biped to a controlled stop turns out to be a harder control problem than making it accelerate: it requires dissipating a large amount of kinetic energy while continuously rebalancing a top-heavy frame, and no event at the Games scored a robot on how well it braked.

The same afternoon, a second X-Humanoid robot cleared 2.88 meters in the standing high jump, well past Cuban high jumper Javier Sotomayor's 2.45-meter human record from 1993 and a sharp jump from the 0.95 meters that had been the best result in the same event a year earlier, at the Games' first edition.

The 400 meters supplied the clearest explanation of how these gains are being produced. X-Humanoid's Tiangong Omni won the race running with its arms held up near its face rather than swinging at its sides, a posture that went viral online with viewers speculating the company had copied a child's running style. Han Gang, the company's motion-control lead, said the pose was not designed by any engineer: the robot arrived at it on its own, through reinforcement learning, by iterating movements in simulation under a reward that favored speed. Keeping the arms near the chest turned out to reduce the load on the shoulder joints over the distance, letting the stronger hip and leg actuators carry more of the work.

That is also why deceleration lagged behind acceleration at these Games. A reinforcement-learning system optimizes for whatever gets measured and rewarded; sprint time and jump height are easy to score, so training pushed hard on both. Stopping safely was not a scored event, so nothing pushed on it, and the robots that ran fastest arrived at the same unsolved problem: a body tuned entirely for going forward has no equivalent instinct for stopping.

Games organizers and competing companies were candid that the sprint and jump numbers, however shareable, are not the point of the exercise. "Only when it can work in those end scenarios does it have real value," one participating executive said of practical tasks the robots also performed, such as connecting cables, packing warehouse boxes, and manipulating tools — jobs that demand vision, fine force control, and error recovery rather than raw speed. "Simply running and jumping does not improve efficiency."

More than 2000 humanoid robots from teams in over a dozen countries took part across the five-day event, up sharply from the roughly 500 robots that competed at the first Games in 2025, alongside combat and skill disciplines including kickboxing, table tennis, tai chi and power-tool assembly. The scale suggests the event is becoming a fixture rather than a one-off spectacle — one where the marquee records are also, so far, the events best at exposing what these machines still cannot do.