A humanoid runs faster than Bolt: Tiangong Ultra sets the 100m record in 8.86 s — and RL invents a gait that no one programmed

A humanoid runs faster than Bolt: Tiangong Ultra sets the 100 m record in 8.86 s — and RL invents a gait no one programmed

🔎 A record that hides the real revolution

A Chinese robot just ran the 100 meters in 8.86 seconds. That's faster than Usain Bolt. But if you take away one thing from this competition, it's not the time.

It's that another robot, in the 400 meters, invented on its own a running technique that no one had taught it. Arms raised in front of its face, an absurd posture at first glance, perfectly logical once analyzed: this position reduced the thermal load on its actuators.

Second edition of the World Humanoid Robot Games in Beijing, August 2026. Two thousand robots, 51 events, 1,300 competitions. The sprint makes the headlines. The reinforcement learning that emerges from simulations into the real world makes history.

The key points

• Tiangong Ultra ran the 100 m in 8.86 s on August 25, 2026, beating Usain Bolt's human record (9.58 s in 2009).
• Three days earlier, the same robot had already broken the 9.58 s barrier with a 9.39 s time in the preliminary heats.
• Tien Kung Omni won the 400 m in 45.66 s with an emergent gait via reinforcement learning — arms raised in front of its face, never programmed by the engineers.
• Only 3 out of 12 teams completed the fire scenario. Several events remained teleoperated. The gap between sprinting and general autonomy remains immense.

Recommended tools

The 100 m: 8.86 s, sparks, and a deceiving image

Tiangong Ultra lowered its own record of 9.39 s to 8.86 s in three days, according to data compiled by Reuters and Al Jazeera. The human record of Usain Bolt, set at the 2009 World Championships in Berlin, is 9.58 seconds.

The difference is 0.72 seconds. Over 100 meters, that is an eternity.

But take a closer look at the footage reported by Ars Technica: at the finish line, the robot fell and produced sparks upon hitting the ground. This is not an athlete in control of their race. It is a machine pushed to its thermomechanical limits.

The robotic 100 meters is a test bench of raw power, not of motor sophistication. The actuators run at their maximum regime, the gait is optimized for one thing only: linear speed on a flat track. No wind adaptation, no race management, no block start with reaction time.

The leap in one year

The previous year, at the first World Humanoid Robot Games in 2025, the 100 m winner had crossed the line in 21.50 seconds according to Reuters. In twelve months, the time was more than halved.

That is significant. But it must be kept in mind that the starting baseline was very low. Going from 21.50 s to 8.86 s is mostly a sign that the first edition was a competition prototype, not that the second one solved humanoid locomotion.

The real news: a robot that invents its own running technique

Tien Kung Omni, another robot from the Tiangong family, ran the 400 m in 45.66 s according to ECNS. The time is interesting. The posture is infinitely more so.

During the race, the robot adopted an unusual position: arms raised in front of its face, like a boxer in a high guard. No engineer had asked it to do this. This gait emerged spontaneously during reinforcement learning (RL) simulation training, and was then transferred to the real world.

TechTimes reports that the thermal constraint was the key factor: the shoulder actuators were overheating with a classic arm swing. The RL simulation discovered that raising the arms reduced the torque exerted on the scapular joints, limiting the heating.

What this actually means

Reinforcement learning works like this: an objective is defined (running as fast as possible) along with physical constraints (the robot's thermomechanical model). The agent explores millions of trajectories in simulation. The best ones are kept, the others eliminated.

Here, the agent found a solution that the human designers had not envisioned. This is the principle of emergence: the system finds a local optimum that exceeds the imagination of its creators. It is not intelligence in the human sense. It is optimization in a solution space so vast that exhaustive human exploration is impossible.

It is also proof that sim-to-real transfer works in complex motor contexts. The gait discovered in simulation held up in the real world, with its friction, its vibrations, and its inaccuracies.

The World Humanoid Robot Games: 2,000 robots, 51 events, a lot of teleoperation

The figures from this second edition are impressive on the surface: 2,000 robots competing, 51 different events, 1,300 competitions in total according to reports from Channel News Asia. Beijing has clearly made this event a national technological showcase.

But behind the sprint records, the reality is more nuanced.

The fire scenario: 3 out of 12

The fire intervention event is perhaps the best indicator of the actual level of autonomy. Result: only 3 teams out of 12 managed to complete the entire scenario. The others failed at various stages — navigation, opening doors, handling a fire extinguisher.

This is revealing. The 100-meter dash requires power in one direction. A fire scenario requires perception, planning, the manipulation of non-standard objects, and real-time adaptation. The gap in complexity is astronomical.

Masked teleoperation

Ars Technica notes that several events, presented as autonomous, were actually partially or entirely teleoperated. A human controlled the robot remotely, sometimes using exoskeleton-type interfaces.

This is common in the industry. The line between "autonomous robot" and "teleoperated robot" is intentionally blurred in manufacturers' communications. A robot that "completes a task" could very well have had a human in the loop at every critical decision.

Why the sprint is a gimmick (and why that's normal)

Let's honestly compare the 100m robot sprint with the human 100m sprint.

A human sprinter has to manage: the block start with a reaction time under 0.15 seconds, phase-by-phase acceleration, the transition to the speed maintenance phase, resistance to crosswinds, neuromuscular fatigue in the final meters, and the coordinated stop after the finish line.

Tiangong Ultra has to manage: going as fast as possible in a straight line on a perfect track, with no wind, no opponents, and no starting rules.

It's not the same discipline. It's a test of maximum actuator speed with pre-calculated kinematic optimization. The robot doesn't "run" in the sense that a human runs. It executes an optimized motor pattern.

The parallel with autonomous cars

It's exactly like comparing the top speed of an F1 car with an autonomous car's ability to navigate a city. The F1 goes 370 km/h on a closed circuit. The autonomous car drives at 50 km/h in a chaotic environment. One isn't "better" than the other. They solve different problems.

The robot sprint is a demonstration of mechanical power and kinematic optimization. The real measure of autonomy is the fire scenario. And there, we are at a 25% success rate.

Reinforcement learning as a driver of non-human innovation

What happened with Tien Kung Omni on the 400m is a fundamental case study for robotics. RL found a solution that classic human engineering had not envisioned.

Historically, the design of robot gaits follows a top-down process: an engineer models biomechanics, writes control equations, adjusts parameters, tests, iterates. It's long, expensive, and limited by human imagination.

RL reverses this process. You give an objective, a physical model, and let the algorithm explore. The results can be counterintuitive — like arms raised in front of the face while running — but optimized for the real constraints of the system.

Sim-to-real: the bottleneck

The transfer from simulation to reality remains the weak point. The simulation is always an approximation. Friction, elasticity, communication delays between sensors and actuators — all of this differs between the virtual world and the physical world.

That the gait of Tien Kung Omni worked in reality is a positive signal. But Ars Technica points out that other robots were not so lucky: falls, sparks, and erratic behavior were common outside of the optimal trajectories.

RL is powerful. It is also fragile as soon as you step outside the training domain.

The Chinese humanoid ecosystem: industrial sprint, nuanced reality

China has clearly chosen the humanoid as a strategic axis. The World Humanoid Robot Games are part of this policy: create a media event, bring companies together, generate viral images.

The Tiangong robots are developed by Robot Era, a Beijing-based startup. The family is expanding: Tiangong Ultra for the sprint, Tien Kung Omni for the endurance events, other variants for manipulation.

If we compare with what Boston Dynamics is doing with Atlas, the approach is different. Atlas bets on whole-body dynamics, the ability to chain diverse tasks, robustness in the face of perturbations. Tiangong bets on specialization by event and advanced RL optimization.

Both approaches are valid. They measure different things.

The broader industrial context

Humanoid robotics is evolving in a context of massive transformations. AI is not only replacing office jobs — as shown by the Cloudflare restructuring that eliminated 1,100 positions despite $639M in revenue, the era of AI-driven restructuring now affects all sectors. Humanoids are positioned as the logical next step for physical jobs.

But the distance between a robot running in a straight line and a robot replacing a worker on an assembly line is still measured in years, not months.

What current AI models say about this evolution

The most performant language models in 2025-2026 — Claude Opus 4.7, Gemini 3.1 Pro, GPT-5.5 — are used to support robotic RL. Not to directly control motors, but for high-level task planning, scene interpretation, and generating intermediate rewards for RL agents.

In particular, agentic benchmarks show that GPT-5.5 reaches 98.2 and Claude Opus 4.7 (Adaptive) 94.3 in complex autonomous tasks. These reasoning capabilities are progressively being integrated into robotic control pipelines, but the path between "reasoning about a scene" and "executing a coordinated movement in that scene" remains long.

The RL that made Tien Kung Omni run with its arms raised probably did not use an LLM. It was direct optimization in a physical action space. The complementarity between low-level RL and high-level LLMs is the true current research frontier.

❌ Common mistakes

Mistake 1: Directly comparing the robot's time with Bolt's

A human runs under strictly regulated conditions: reactive start, measured wind, anti-doping. The robot runs on an optimized track, without opponents, without starting constraints. The number is real, the comparison is misleading. The only honest comparison is intra-robotics: 21.50 s in 2025 versus 8.86 s in 2026.

Mistake 2: Confusing emergence and intelligence

Tien Kung Omni's arms-raised gait is emergent, not intelligent. The algorithm explored a solution space and found an optimum. It does not "understand" why this position works. It could not generalize this logic to another thermal problem. Emergence is a phenomenon of optimization, not of understanding.

Mistake 3: Extrapolating sprinting to general autonomy

A robot that runs fast in a straight line does not know how to navigate a burning building. The 3 out of 12 in the fire scenario proves it. Fast locomotion is a necessary but largely insufficient condition for generalized autonomy.

Mistake 4: Ignoring the role of teleoperation

Many viral videos of "autonomous" robots hide a human operator. At the World Humanoid Robot Games, several events were teleoperated. Always check before concluding that it is autonomous.

❓ Frequently asked questions

Can a robot really beat Usain Bolt over 100 m?

In pure track time, yes: 8.86 s compared to 9.58 s. But the conditions are not comparable. No starting blocks, no wind, no athletic regulations. It is a linear speed record for a robotic biped, not an athletic 100-meter record.

What is reinforcement learning in robotics?

A learning algorithm where an agent explores actions in an environment, receives rewards or penalties, and adjusts its behavior to maximize its score. In simulation, this allows for millions of rapid trials before transfer to the physical robot.

Why do raised arms reduce heat?

Classic arm swinging creates significant torque on the shoulder joints, heavily stressing the actuators. Raising the arms in front of the face reduces this torque and therefore the heat dissipation in the scapular motors. The aerodynamic trade-off is offset by the thermal gain.

How many robots participated in the World Humanoid Robot Games 2026?

Around 2,000 robots spread across 51 events and 1,300 competitions, according to figures reported by Channel News Asia. This is a significant leap compared to the first edition in 2025.

Is the emergent gait reproducible on other robots?

Not directly. The gait is optimized for the specific geometry, actuators, and thermal constraints of Tien Kung Omni. Transferring this solution to another robot would require RL re-training adapted to its own physics.

✅ Conclusion

The 8.86 s of Tiangong Ultra is a perfect marketing figure. The emerging gait of Tien Kung Omni is an important scientific result. Both occurred on the same weekend in Beijing, but they do not hold the same value.

One shows that actuators are progressing fast. The other shows that RL is starting to find solutions that engineers do not see. The second is the signal that matters — robots are starting to solve their own physical problems, and that is a paradigm shift even if it remains localized and fragile.

The day a robot invents a gait to cross a burning room without anyone having taught it, we can talk about a revolution. For now, it has invented a posture to run 400 meters in a straight line. That is already considerable.