Humanoid robots at the second World Humanoid Robot Games in Beijing beat Usain Bolt's human 100-meter record and the standing high jump world record, then in the same heats crashed into barriers, snapped at the waist in showers of sparks, and in at least one case caught fire. The event, held August 22–26, 2026 at the National Speed Skating Oval and backed by Chinese state media and the Beijing city government, drew viral clips of an Honor humanoid breaking apart mid-sprint next to two Tian Gong Ultras. Investors poured more than $6 billion into humanoid robot companies in 2025, and Beijing wants the world to see what that money bought.
The record-breaking sprints represent real engineering progress from the inaugural 2025 event, when no robot came close to human times. This year robots also competed in kung fu, dance, kickboxing, soccer, and table tennis, plus a new slate of practical challenges that ranged from laundry to firefighting. The practical events are where the story gets more honest about the state of the field.
Dipam Patel, a PhD candidate in computer science at Purdue University and a research assistant at the US Army DevCom Army Research Lab, told Ars Technica that beating a sprint record genuinely requires coordinated whole-body control across every joint, and that the engineering leap between the 2025 and 2026 games is real.
Key facts
- 01Humanoid robots surpassed Usain Bolt's 100-meter human record and the standing high jump world record at the Beijing games held August 22–26, 2026.
- 02Only 3 of 12 competing robot teams finished a 30-minute outdoor firefighting and rescue event that required identifying hazards, shutting valves, and extinguishing a fire.
- 03Investors poured more than $6 billion into humanoid robot companies in 2025, with Chinese firms leading deployment.
- 04X-Humanoid's small-group 400-meter winner ran with an unusual arms-up posture learned through reinforcement learning in simulation.
- 05Many competing robots relied on cloud computing over 5G modules rather than onboard hardware, and some events still permitted scored-penalized human teleoperation.
The catch is that the sprinting robots were engineered for exactly one task. They ran in a straight line and could not stop or turn at the finish, which is why so many of them ended the race in pieces. Different companies fielded different robots for different events rather than one machine that could do everything, a design choice that inflates the medal count without saying much about general capability.
Beijing-based X-Humanoid drew separate viral attention by winning the small-group 400-meter race with a robot that held its arms up near its face and swung them via hip rotations. The developers told the Global Times that the posture was learned through reinforcement learning trials in simulation, replacing a conventional human-style arm swing. Patel suggested a more meaningful benchmark would be running the same race 10 times in a row, forcing robots to demonstrate reliability and safe deceleration across trials rather than a single hero run.
The chore events told a harder story. Robots washed and hung laundry, folded clothes, cleaned living rooms, and had to break from those tasks to receive a delivery package — a test of task-switching that humans do without thinking. The Beijing Continental Grand Hotel ran a hospitality track with timed challenges to move wheeled luggage to assigned rooms, clean beds, and restock supplies. A separate library event tested visual recognition and reasoning by asking robots to shelve returned books correctly.
The firefighting and rescue event was the clearest stress test. Robots had 30 minutes to identify hazardous substances, locate and shut off three open valves of different types, spot a fire, and find a fire extinguisher to put it out. Only 3 of the 12 competing teams completed the whole challenge, according to the Global Times. Six hand-dexterity events — picking up beans, driving screws, hammering nails into a corkboard without angling them — separated the machines that can grip from the ones that cannot.
Patel noted that many of the competing robots were still being teleoperated by humans directly controlling their motions. The games allowed teleoperation but docked the final score for it, an acknowledgement that fully autonomous performance is still the exception rather than the rule at the frontier. That is the gap between a highlight reel and a product.
The autonomy problem is not just a data problem, it is a generalization problem. Training robots through simulation or human demonstration data has proven expensive and time-consuming, even as companies pay ordinary people to strap cameras to their heads to collect footage. Patel told Ars that algorithmic development would have to advance considerably before a robot could generalize to eight-hour work shifts in an unfamiliar environment.
Even the more autonomous entrants lean on cloud compute rather than running AI models onboard, according to the South China Morning Post. Robots at the games were equipped with 5G modules that transmit data to an off-board embodied-intelligence system and receive instructions on what to do next. That works fine inside a stadium in Beijing. It works less well in a warehouse with spotty coverage, and it complicates any export pitch — particularly given that the US government has banned the import of foreign-made robots, including China's most popular humanoid models.
The competitive frame outside the arena is that Chinese companies are testing and deploying humanoid robots aggressively while US firms like Boston Dynamics and Agility Robotics push their own commercialization plans in factories and warehouses. The $6 billion invested in 2025 is chasing a market that does not yet reliably exist, and the honest read of the Beijing games is that the sprint records are marketing while the 3-of-12 firefighting completion rate is the ground truth.
The commercial case for humanoids rests on one bet: that a general-purpose bipedal robot will eventually be cheaper and more flexible than a task-specific machine on wheels or rails. The Beijing games gave that bet a mixed grade. Whole-body control is measurably better than a year ago, which matters for anyone selling to logistics buyers who care about mobility across stairs and unstructured spaces. But the persistence of teleoperation, the cloud dependency, and the collapse of purpose-built sprinters at the finish line all argue that the near-term deployments will be narrower and more supervised than the highlight clips suggest. For investors trying to underwrite the next $6 billion, the useful benchmarks are the boring ones.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
