"Chinese robots can fly." Over the past two days, this phrase has appeared frequently across social media. From August 22 to 26, the second World Humanoid Robot Olympics was held at the National Speed Skating Oval, known as the "Ice Ribbon." A total of 666 teams and 2,056 robots from 16 countries across six continents competed in 51 events over five days. Compared to the inaugural edition, the number of participating teams grew by 138%, while the robot count quadrupled.

At last year's first Games, the best time for the humanoid robot 100-meter sprint was 21.50 seconds, and the standing high jump champion reached 95.641 centimeters. Just one year later, the fastest robot clocked an astonishing 9.32 seconds in the 100 meters—not only over 12 seconds faster than last year but also surpassing Usain Bolt's human world record. The high jump soared to 2.8843 meters, nearly three times last year's height. If judged purely by results, humanoid robots appear to have entered a "cheat mode" of performance.

Yet what truly deserves attention at the second Games is not just that robots are running faster or jumping higher. This year, organizers introduced more delicate dexterous manipulation tests for humanoid robots, along with competitions set in real-world scenarios such as industrial, landscaping, catering, and emergency response. The Olympics are transforming into a comprehensive examination of whether robots can genuinely integrate into the real world.

Why Can Chinese Robots Fly?

The answer lies first in the increasingly powerful "bodies" of robots. The repeatedly broken competition records may look like mere numbers, but behind them lies a collective upgrade of robot joints, power supplies, control algorithms, and training methods. Guo Yijie, head of the TienGong Ultra robot development team at the Beijing Humanoid Robot Innovation Center, stated that compared to last year, this year's TienGong Ultra underwent upgrades in robot body structure, joint capabilities, electrical component reliability, and "cerebellum" motion control. Joint torque and speed were significantly improved, requiring higher-capacity power sources to support high-dynamic movement.

From a broader perspective, he explained that running and high jump correspond respectively to a humanoid robot's generalized mobility and whole-body action imitation capabilities. One addresses "how a robot can reach any place a human can go," while the other tackles "how to perform any complex action a human can do." If a humanoid robot can establish both core capabilities, it becomes equivalent to a car that humans can drive at will. In the future, as its "brain" develops, it will evolve into a self-driving vehicle.

Details on this year's track also confirmed this shift. The obstacle events featured 100-meter and 400-meter courses. The 100-meter obstacles included S-curve bridges, symmetrical ramps, spiral staircases, and continuous hurdles; the 400-meter added complex challenges like climbing boxes, crawling under ground nets, tire arrays, and plum blossom posts. Notably, this year even eliminated the previous single-lane format, allowing two robots to race side-by-side on the same track. Robots could no longer simply pursue speed; they had to maintain balance under high-velocity motion, handle disturbances, and achieve whole-body coordination.

Weightlifting posed similar demands. Lifting heavy objects appears to test motor torque, but it actually involves end-effector gripping, joint output, structural strength, torso coordination, and dynamic balance under load. These abilities are extending beyond the track to more events, including table tennis, tennis, tai chi, and freestyle gymnastics. In the tai chi competition, the Shanghai Zhiyuan team ultimately claimed the gold medal.

"The overall performance wasn't perfect, but it met expectations," a team member told The Paper. The team spent two months refining their routine for this event. What impressed him most was witnessing a robot complete a single-leg support movement. According to the team, tai chi tests not only action imitation but also balance, the gap between simulation and reality (Sim2Real), and the dynamic performance of robot joints. Meanwhile, the Xinghaitu team's full-size humanoid robot executed multiple difficult maneuvers in freestyle gymnastics, winning the championship through movement fluidity and landing stability. Behind this lies not any single joint's performance, but whole-body coordinated control, posture switching, and dynamic balance capabilities.

The test for robots has shifted from "possessing athletic ability" to "having a truly usable body." As robot bodies grow stronger, another question emerges: beyond running and jumping, can they actually work?

Robots Begin "Working" in Park Management, Book Arrangement, and Charging Plug Insertion

These tasks don't sound like competitions; they resemble job postings for robots. This year's Games featured numerous scenario-based events. In industrial scenarios, robots had to handle a pharmaceutical warehouse "packing and warehousing" role: stacking medicine boxes into cartons, inserting medicine blister packs and instructions into boxes and sealing them, and folding flat cardboard into three-dimensional packaging. In office settings, robots needed to arrange name cards, bottled water, and hard pads, load paper into printers, retrieve printed materials, and feed documents slated for destruction into shredders. In new-energy charging scenarios, robots had to complete charging or plug-unplug operations for three vehicles in sequence within 30 minutes. In catering, they had to understand voice commands, use food tongs to serve meals, place food in microwave ovens, and then collect beverages for delivery. In emergency response, they had to clear obstacles from passages, open windows, and set up safety barriers.

These tasks collectively point to a shift: robots are moving from "completing a single action" to "completing a full job." But the difficulty of real work lies in the "trivial" details that humans take for granted. According to Gong Xiao, committee member and deputy director of the China Software Testing Center, loading paper into a printer requires precision: the paper feed slot is only a few millimeters wide, and paper is a flexible material. Robots must accurately identify the position, control force, and align the paper into the narrow slot. Similarly, inserting a charging gun into a vehicle requires the robot to locate the charging port, plan a path within confined space, control the gun head's angle, and handle soft, easily tangled cables—testing multi-task switching and fine manipulation stability in unstructured environments.

What takes humans mere seconds requires robots to string together visual recognition, path planning, end-effector control, and action execution. As the arena increasingly resembles the real world, robots are increasingly resembling true "workers."

From Remote Control to Autonomous Decision-Making

Another significant change this year: multiple competitive events, including the 400-meter run, abolished manual remote control entirely, relying on full autonomous operation throughout. In the 400-meter race, operators could only issue the start command; the rest depended entirely on the robots' own capabilities. In scenario events, fully autonomous operation carried a score weight of 1.0, while teleoperation was only 0.5.

This is no simple rule change. It signals a shift from "watching whether it can do something" to "watching whether it can decide how to do it on its own." The champion of the 400-meter large-class event finished in 38.15 seconds, surpassing the human world record, with operators only able to issue the start command.

The dexterous hand special competition was especially telling. Robots had to complete tasks including powder weighing, block stacking, picking beans with tweezers, and cable connection within five minutes. The error tolerance for weighing 20 grams of powder was ±0.5 grams; the bean task required using tweezers to pick up soybeans one by one. Humans perform these actions almost without thought. But robots must first perceive, then judge, then plan, then execute—and the real world doesn't always match training data.

Gong Xiao noted that in preliminary rounds, fully autonomous robots performed beyond many expectations. However, subtle changes in object material, softness, or placement angle could still cause deviations in autonomous decision-making. This is one of the most critical technical challenges in embodied intelligence today: models in the digital world can be trained infinitely, but the physical world offers no standard answers.

Gong Xiao emphasized that behind all events lies a single question: Can robots step out of the comfort zone of "standard answers" and complete small tasks thoroughly from start to finish in real, ever-changing production and life scenarios? "The arena is both a touchstone for testing capabilities and a compass for technological progress. We hope people pay attention not just to who wins first place, but to the algorithmic breakthroughs and engineering iterations behind every grasp, every plug-in, and every recognition."

Olympics 2.0: No Longer About Who Runs Fastest

Another notable phenomenon at this year's Games: robots are increasingly competing not on a single athletic metric, but on an entire integrated capability set. Unitree Robotics provides a telling example. In the 100-meter preliminaries, the Unitree robot finished toward the back of its group. The company issued a statement that day explaining that due to time, energy, limited numbers of new-generation robots, and testing constraints, it had scaled back some registered events, with primary focus remaining on mass production.

Unitree founder and chairman Wang Xingxing attended the Games as a spectator. At times his expression was serious as he photographed the proceedings; at others, he smiled watching robot performances. For robotics companies already in mass production, the Olympics may not be a competition that must be won, but rather a mirror showing the industry where humanoid robots currently stand.

Over the past year, the most visible industry change was robots running faster and jumping higher. But what will truly determine whether robots enter thousands of industries may be less headline-grabbing capabilities: stability, reliability, autonomy, and generalization in real environments. Guo Yijie also believes that robot athletic performance has reached a certain height, and the next critical phase is converting high performance into high stability and high reliability. Future development of robot body capabilities may gradually slow, with the industry exploring and training robots' abilities in real-world scenarios.

This is what Robot Olympics 2.0 truly showcased. On the surface, it produced 100-meter records, high jump records, weightlifting results, and football scores. Looking deeper, it revealed a directional shift in the humanoid robot industry: from "what can be done" to "whether tasks can be completed reliably," from athletic performance demonstration to real-world scenario validation, from manual remote control to autonomous decision-making, and from standardized tracks to the complex physical world.

So "Chinese robots can fly" may just be the most sensational phrase of this year's Games. What's truly worth anticipating is a day when robots don't fly as high or run as fast, yet enter our lives and quietly complete ordinary tasks well. Only then will the Robot Olympics have truly produced its results.