AgiBot Explains Lingxi X2’s Gold Medal Win in the 100-Meter Hurdles, Completed Autonomously

AgiBot (Shanghai) Technology Co., Ltd. published an account on its website on September 9, 2026, of how its mass-produced Lingxi X2 won gold in the 100-meter hurdles at the second World Humanoid Robot Games. The company said the Lingxi X2 cleared obstacles autonomously without modifications to the robot and completed the course in a personal-best 1 minute, 28.35 seconds. The article describes the 100-meter hurdles as a test of high-speed maneuverability, dynamic balance, real-time perception and navigation, obstacle clearance over complex terrain, and whole-body posture adaptation.

AgiBot says the Lingxi X2 cleared obstacles autonomously despite being a mass-produced model. Its use of the production robot rather than a specially equipped competition model may be relevant to companies focused on moving research into operational use. The design, which combines a depth camera with AGILE’s three-layer model to integrate obstacle recognition and posture control, also raises questions about how well safety and repeatability can be ensured in real-world settings.

The key point is not the gold medal itself, but that the mass-produced Lingxi X2 competed in the 100-meter hurdles without specialized modifications and won autonomously in 1 minute, 28.35 seconds. What appears to be new is not a one-off demonstration, but the use of a production robot to demonstrate an integrated system for perception, including a depth camera, obstacle recognition, whole-body control and obstacle-clearing learning. By contrast, this builds on AgiBot’s existing emphasis on integrated perception and control and the real-world deployment of mass-produced robots; the competition provided a visible test environment for that approach. In [AgiBot unveils GE-Act 2.0, tests capability gap between 300 and 30,000 hours of training for zero-shot performance on a real robot](/en/news/agibot-unveils-ge-act-2-0-testing-real-robot-zero-shot-performance-fro-6a0ed2), the company had trained on embodied AI data ranging from 300 to 30,000 hours and tested zero-shot performance on a real robot. The Lingxi X2 article can be read as connecting that data-expansion effort to a demanding real-robot environment: competition. The focus, then, is not only on increasing training data, but also on how well the control system, retained in the mass-produced robot, can integrate obstacle recognition and motion generation. The industry implications are not limited to the robot’s performance alone. Depth cameras, perception-and-motion models, real-robot version updates and mass production are linked, making the conversion of environmental inputs into control—and the frequency of updates—potential areas of competition. The competition demonstrated not only high-speed maneuvering and clearing a raised obstacle, but also a control design that adapts to differences in body shape, such as autonomously learning upper-body movements to avoid the risk of touching a pole. The next questions are which actions the Lingxi X2 can reproduce reliably outside competition, how its 16 version iterations will be reflected in mass-production operations, and how broadly each layer of AGILE can be applied in real-world use.

According to the company, the Lingxi X2 is equipped with a depth camera and determines its movements after recognizing physical information, including obstacle heights and uneven road surfaces. The article describes this setup as the AGILE (AgiBot Generative Intelligent Locomotion Engine) integrated perception-and-control model, with three layers—BFM-2, a general-purpose behavior foundation model; GCFM, a generative motion control model; and REACT, a perception-and-motion model—connecting visual perception to whole-body control. For obstacle clearance, the company said it shifted the upper body from an initial motion-capture approach to autonomous learning for pole slaloming, addressing a problem in which the hands and arms touched the poles. For clearing a raised obstacle, it said it trained the robot on “King Kong Sideways Pass,” based on the movements of a trainee. The company said the robot trained at a practice site for 50 days before the competition, was updated to a new version approximately every two days and underwent 16 version iterations by the opening of the event.

For Japanese companies and research institutions, the operational approach of autonomously clearing obstacles in a real-robot competition using an unmodified mass-produced model may offer a point of comparison with evaluations centered on prototypes. Key questions include how the integration of depth-camera perception and whole-body control, as well as a system for frequent version updates on a mass-produced robot, connects with component supply and the development of testing facilities.