TL;DR
China's Tiangong Ultra ran 100m in 8.64 seconds at the World Humanoid Robot Games, beating Usain Bolt's 9.58-second human world record and signaling how fast embodied AI is closing the gap with human physical performance.
What happened
- Tiangong Ultra, developed by the Beijing Humanoid Robot Innovation Centre, clocked 8.64 seconds in the 100m final, nearly a full second faster than Bolt's 2009 Berlin record of 9.58 seconds.
- The robot improved across three runs: 9.39s and 8.86s in qualifying, then 8.64s in the final, averaging roughly 42 kph.
- The feat happened at the second World Humanoid Robot Games in Beijing, a five-day event with 51 disciplines, more than 2,000 robots, and 666 competing teams, most of them Chinese.
- Developers credited redesigned torso and waist components, upgraded motors, expanded range of movement, and improved control software for the speed gains.
- The robot takes more than 50 steps to cover 100m versus Bolt's 41, using a shorter, faster cadence tuned to what its electric motors do best rather than mimicking human gait.
Why it matters
- Electric motors do not fatigue: unlike human sprinters who slow in the final 20m due to neuromuscular fatigue, Tiangong maintained top speed through the finish, a structural advantage that compounds as hardware improves.
- The games function as a public stress test for embodied AI, combining athletic events with factory, restaurant, office, and emergency-response simulations, giving developers real-world performance data at scale.
- China is running a concentrated national effort: the overwhelming majority of the 666 teams were Chinese, and the Beijing Humanoid Robot Innovation Centre is a state-backed entity, framing this as industrial policy as much as sport.
- The record exposes where robots still lag: Tiangong started from an upright posture, waited nearly a second after the start signal, cannot use starting blocks, and struggles to stop or change direction, pointing to gaps in perception and decision-making.
- Experts note that transferring skills between robot bodies remains hard: different joints, sensors, and balance constraints mean a speed breakthrough for one platform does not automatically lift the broader fleet, though projects like Open X-Embodiment are pooling data across 22 robot types to address this.
What to watch next
- Whether Tiangong Ultra or rivals close the start-reaction and stopping-control gaps, which sport scientists say could push times significantly lower and signal genuine locomotion intelligence.
- How task-based benchmarks from the games, such as factory simulations and emergency-response challenges, translate into commercial deployment timelines for Chinese humanoid makers.
- Progress in cross-embodiment learning: if pooled training data lets one robot's skills transfer reliably to another platform, the pace of capability gains across the entire industry accelerates sharply.
Originally published on Present of AI, a daily source-linked AI news timeline. Read the full timeline or browse the open dataset.