AgiBot wins 18 golds in humanoid robot debut

The robots had just finished working in a factory when they were “given time off” and sent to the arena. AgiBot made its debut at the World Humanoid Robot Games, ultimately winning 18 gold, 16 silver and 12 bronze medals. More importantly, all of the participating models were mass-produced machines, not customized robots built specifically for the competition. As reported by ifanr, these “factory workers” brought experience from real-world scenarios to the event.

That gave the demonstrations a work-related context. The dexterous hand of the Wuhan University-AgiBot joint team weighed 20 grams of salt in 27 seconds, while also picking up beans and building with blocks. The Expedition A3 performed movements including a flying kick, an outward lotus kick and a horse stance in the tai chi event. The Lingxi X2 navigated around obstacles, cleared barriers and passed through restricted spaces in the 100-meter obstacle race. The tai chi movements were learned from motion data and then repeatedly refined through Sim2Real—training first in a simulated environment before transferring the results to a real machine—to correct deviations caused by friction, joint gaps and sensor errors.

The details most closely tied to production appeared in the firefighting event. The fire extinguisher weighed 4.5 to 5 kilograms, exceeding the payload limit of the end-effectors on the vast majority of humanoid robots available on the market. Rather than developing an expensive new gripper, AgiBot added structural components at the wrist, using wrist strength to hook the extinguisher and place it in a load-bearing container on the robot’s base. Combined with VR-based beyond-line-of-sight teleoperation, inverse-kinematics mapping and force-feedback whole-body control, this allowed the robot to absorb impact when handling door handles, valves and firefighting equipment, reducing the risk of an unexpected contact triggering a shutdown.

Book sorting tested the stability required for continuous work. AgiBot formed a joint team with Tsinghua University and Shanghai Jiao Tong University to transport books out of storage, place them on shelves, and identify and correct misplaced books. The vision model not only read the text on book covers, but also calculated grasping points based on each book’s thickness and tilt angle, allowing both arms to work together to pull out books and insert them into the correct gaps. AgiBot combines task intelligence, motor intelligence and interaction intelligence into a full-stack model architecture, using the GO-2 action chain of thought to break down long-horizon tasks and the GE-2 world model to predict changes in the environment after an action.

Specifically, factory deployment says more than the medals do. In June this year, several AgiBot Genie G2 units entered Longcheer Technology’s mass-production factory in Nanchang, Jiangxi, completing 17,625 operations with a 99.99% task success rate. This means the robots are no longer merely performing individual actions in a lab, but are taking on repetitive tasks in an existing production environment. Human intervention, time between failures, automatic recovery, customer expansion and unit economics will nevertheless determine whether they can move from a handful of factories to large-scale deployment.