What Tiangong Ultra’s 9.39-Second Sprint Shows About Humanoid Robots

Tiangong Ultra ran 100 metres in 9.39 seconds in a preliminary heat at Beijing’s World Humanoid Robot Games on 22 August 2026, faster on the clock than Usain Bolt’s 9.58-second human world record. The result is a serious milestone in high-speed bipedal locomotion, but it is neither an athletics record nor evidence that humanoid robots are ready for safe, reliable general-purpose work.

The machine was Beijing Humanoid Robot Innovation Center’s Tiangong Ultra, competing for the Tianzhuo team in a large-robot preliminary heat. The 9.39-second mark was a preliminary result, not a final medal result, and should be described as such. Beijing’s official account says the 2026 100-metre race had been upgraded to fully autonomous robots only. That matters in a sport where teleoperation has often blurred the line between robot capability and operator skill. (bbtnews.com.cn)

The Bolt comparison is mostly a headline device. Bolt’s time is a ratified World Athletics result from a regulated human competition, with defined starting, wind and timing procedures. Tiangong Ultra’s run is a robot-games result under a different technical rulebook, using a robot built and tuned for this task. The robot covered the distance faster on the clock. It did not take Bolt’s record. (worldathletics.org)

That is more than pedantry. The result shows that a large electric biped can accelerate, actively balance a body that becomes unstable while running, and sustain a highly specialised running gait at extraordinary speed. It does not show a general-purpose worker able to move safely through a warehouse, care setting or hotel while handling people, clutter, changing instructions and the badly placed cable that turns up in every real building.

The result is real. The framing needs work.

The second World Humanoid Robot Games runs from 22 to 26 August at Beijing’s National Speed Skating Oval. Organisers say 666 teams and 2,056 robots from 16 countries are taking part across competition and scenario events. The programme extends beyond running to tasks intended to resemble housekeeping, logistics, retail, emergency response and dexterous hand work. (news.cgtn.com)

That broader programme matters. A 100-metre sprint is deliberately narrow: flat surface, fixed distance, known lane and known start. There is no unfamiliar object to recognise, door to open, safe grip to choose or person about to step into the robot’s path. Narrow tests are useful; that is how engineering advances. But they produce narrow evidence.

It is worth being precise about the number, too. The official Beijing report calls 9.39 seconds a preliminary heat result; AP independently reported the same time and identified the robot as coming from Beijing-based X-Humanoid. Reports before the Games also noted an Honor robot’s claimed 9.32-second trial run. A vendor-reported test run and an event preliminary are not one definitive global record. (beijing.gov.cn)

Last year’s result provides useful perspective. The best 100-metre mark at the inaugural 2025 Games was reported as 21.50 seconds. Even allowing for changed rules, hardware revisions and a competition designed to reward progress, moving from that order of performance to 9.39 seconds in a year is substantial. Teams appear to have become markedly better at task-specific high-speed locomotion, not merely at polishing a demo reel. (beijing.gov.cn)

What a nine-second biped has to get right

Running is harder than walking quickly. In a running gait there is a flight phase: neither foot is on the ground. The robot cannot correct itself simply by shifting weight over a planted foot. It has to predict the next footfall, place the foot accurately, absorb the impact, generate the next push-off and stop its torso pitching or yawing out of control.

That demands a tightly integrated stack. Electric actuators — motor, gearbox, sensors and control electronics packaged around a joint — must repeatedly deliver large bursts of torque. Encoders report joint position and speed, while an inertial measurement unit (IMU) senses body acceleration and rotation. Control software combines those signals into state estimation, an ongoing calculation of the robot’s position, movement and orientation, then a low-level controller makes corrections many times per second. At sprint speed, a small mistake is rarely a wobble that can be fixed later. It is usually a fall.

The software matters as much as the metal. Researchers working with Oregon State University’s Cassie biped describe the challenge as finding efficient gait parameters across different speeds, then integrating them into a controller that can start, run and return to standing. Their 2022 Guinness-recognised 100-metre record was 24.73 seconds — and, notably, required Cassie to finish upright rather than merely cross a line. (arxiv.org)

Tiangong Ultra’s performance does not make those earlier efforts obsolete. It shows how quickly the field can move when motors, structural design, state estimation, simulation-trained control policies and tuning are all aimed at a visible target. A control policy is the set of rules a robot uses to choose actions; simulation-trained policies are developed using simulated environments before deployment on hardware. A robot sprint is not a magic trick. It is a difficult optimisation problem solved in hardware.

Power delivery is part of that problem. A sprint robot needs a battery capable of high discharge rates, power electronics able to handle peak current, actuators that do not overheat during the run, and a structure light enough that its joints are not constantly throwing excess mass around. The trade-off is clear: a machine optimised to cover 100 metres quickly need not be the machine best suited to lifting boxes for an eight-hour shift. Range, thermal headroom, payload, compliance and repairability pull the design in different directions.

Autonomous does not mean adaptable

The 2026 event’s fully autonomous requirement matters. A robot that needs a person beside it with a controller has not solved the same problem as one that receives a start command and completes the course itself. Organisers explicitly presented the rule change as a move towards testing perception, decision-making and whole-body control. (english.beijing.gov.cn)

But autonomy is not a single switch. On a prepared straight track, the robot can be given a known course, lighting and surface, plus a motion policy engineered for one direction of travel. Its perception task may be limited to lane keeping and state correction. That is a valid autonomous capability. It is not evidence that the same system can understand a loading bay where pallets have moved, a delivery driver is waiting, reflective packaging confuses cameras and a fire door has been left ajar.

A company considering humanoids for a Jersey warehouse should infer very little from the 9.39-second time about whether a robot can complete a safe shift. The procurement questions are less glamorous:

• Can it recover from a slip without damaging itself, stock or nearby people?
• Can it stop predictably when a human crosses its route?
• Can it identify the right parcel, grasp it without crushing it and recognise a failed pick?
• How long can it work at the required pace before charging, overheating or needing intervention?
• What happens when a sensor, joint actuator or network link degrades?
• Can staff safely maintain it, and can the supplier demonstrate repeatable uptime rather than one good run?

Stopping is part of running

One revealing detail in coverage of the Beijing heat is that the robot ran through the finish and into foam protection. That does not invalidate the 100-metre time; sprint races are timed at the line. But it exposes the difference between a track performance and a useful mobile machine. A deployable robot must decelerate within a known envelope, remain upright and avoid turning stored kinetic energy into a collision hazard. (tomshardware.com)

This is not a minor compliance detail. The existing ISO 13482 personal-care robot safety standard excludes robots travelling above 20 km/h — far below the average speed implied by a 9.39-second 100 metres — and specifically flags the lack of exhaustive, internationally recognised impact injury limits. A fast humanoid in a shared workplace therefore presents a safety case that has not been solved by showing it can run in a protected venue. (iso.org)

Legged systems have an awkward safety characteristic: removing power can itself be hazardous because the robot may fall. Graceful degradation, controlled stopping, obstacle detection, safe-speed zoning and physical separation are therefore central to practical deployment. A robot can be impressive and still be unsuitable to operate close to people at anything like its maximum capability.

Why Bolt is the wrong benchmark — and the right provocation

Comparing Tiangong Ultra with Bolt is technically untidy, but it does force readers to confront how far bipedal robotics has come. The old Guinness benchmark of 24.73 seconds, achieved by Cassie in 2022, was itself a major accomplishment. A full-sized humanoid posting a sub-10-second preliminary time is a different class of dynamic motion. (guinnessworldrecords.com)

The useful comparison is not human versus robot as a contest of species. Humans remain astonishingly energy-efficient, self-repairing, terrain-adaptive and able to combine sprinting with perception, judgement and dexterity. The more revealing robotics comparison is between a machine that can execute one aggressive, rehearsed locomotion policy on a controlled track and one that can work safely through thousands of messy, ordinary cycles.

A robot may beat a human over 100 metres yet still be unable to do a useful morning’s work without supervision. Speed does not establish hand dexterity, robust vision, long-duration battery performance, fault tolerance, economical servicing or the ability to learn a new task without an engineering team preparing the environment.

The benchmarks that would mean more

The Beijing Games are moving in the right direction by adding scenario tests and an “Integrated Five” challenge combining dexterity, stable transfer, lifting and a 100-metre run. Those are better indicators because they punish a robot tuned for one trick. (news.cgtn.com)

Still, the tests that should matter most to prospective users are brutally mundane. Publish completion rates over hundreds or thousands of cycles, not one fastest run. State how often a robot falls, how long recovery takes, how much human intervention it needs and what happens after a battery or actuator fault. Test across changed lighting, worn floors, obstacles and unfamiliar object variants. Measure safe stopping distance and collision response. Include energy per useful task, maintenance hours and the cost of a failed pick or damaged item.

Those measures do not make for a viral clip. They do make for credible robotics progress.

The sensible conclusion is neither “robots have beaten humans” nor “it is just a gimmick”. Tiangong Ultra’s 9.39 seconds is an important proof of specialised dynamic locomotion. Engineers can now make a human-shaped electric machine run very fast under controlled conditions, without a person steering it in real time. That says little, by itself, about whether the machine can be trusted to do general-purpose work around people.

That is not a disappointment. It is a useful boundary — and a necessary one for anyone trying to separate robotics engineering from robotics theatre.