Observant spectators will notice that in running, obstacle courses, and high jump, one or two operators often stand behind the robot holding a remote control. At last year's robot marathon, nearly every robot needed three engineers to "run along," prompting jokes that it was more like a remote-controlled car race. By the second edition, however, the event separated autonomous-navigation and remote-controlled teams, scored them with weighting coefficients of 1.0 and 1.2, respectively, and ranked them together.
That shift reflects technological progress. This year, 40% of robots navigated and ran autonomously, and the top three finishers all completed the course without human control. The Honor Monkey King team's autonomous navigation robot, Lightning, won with a net time of 50 minutes 26 seconds—nearly 1 hour 50 minutes faster than the previous champion. Among the 30 competitive events at the Second World Humanoid Robot Games, 23 explicitly required "full autonomy."
Over the past year, humanoid robots' biggest advances have centered on perception and localization. Last year, robots relied on cameras to identify the track and often drifted into the wrong lane. This year, many teams added lidar to the head, built a 3D map of the venue before the race, and matched radar data with the map to achieve centimeter-level global positioning—the foundation for completing the course autonomously.
But remote controls have not disappeared. Even robots competing autonomously still need an operator to issue the start command. In the 400-meter obstacle course, where real-time judgment of unknown space is required, most teams still chose remote control, though their results were weighted by a factor of 1.2. More importantly, remote control does not mean puppetry: engineers only issue high-level commands such as move forward or switch modes, while the robot autonomously decides how to step, adjust its center of gravity, and clear obstacles.
Viewed over a longer horizon, remote control is actually "infrastructure" supporting the evolution of autonomous capability—every human operation generates training data from the real world. Only when this data flywheel starts turning can robot autonomy continue to climb. That, in turn, requires stable and scalable training compute: lightweight decisions are made on-device, while training and model iteration happen in the cloud. StarWar Technology's work on GPU compute platforms and device-cloud collaboration is precisely aimed at creating a sensible division of labor between distributed device-side tasks and centralized training resources.
The bottlenecks are also clear. Wang Xingxing's "double 80%" standard holds that a robot must be placed in 80% unfamiliar scenarios and complete about 80% of tasks via voice or text commands before it reaches the threshold of generalization. He Xiaopeng classifies humanoid robots from L1 to L5 in the same way as autonomous driving; most products on the market remain in early L2, and reaching L3 is necessary for commercialization.
This year's games accumulated 2,500 hours of real-world operational data, covering 12 application scenarios, 44 work tasks, and more than 100 skills. These real-machine data will become "fuel" for the next round of embodied large-model training. Ditching the remote control is not merely a bonus—it is a necessary path for embodied AI to move toward industrialization. Until the generalization problem is truly solved, it remains a necessary transitional stage.