AgiBot's Olympic Debut: Factory Workers Turned Athletes Sweep 18 Gold Medals

Cyber workers took time off to attend the "Robot Olympics"!

At this year's World Humanoid Robot Games, the most viral contender was undoubtedly the AgiBot robot, taking a leave of absence from the factory to compete.

It ultimately topped the medal table with 18 golds, 16 silvers, and 12 bronzes. Among these, its dexterous hands secured 7 out of 8 golds in specialized competitions, the operational track won 6 out of 12 golds, and the athletic events captured 5 golds.

The Wuhan University-AgiBot joint team's dexterous hand picked up a small spoon, weighed out 20 grams of salt in 27 seconds, and even used tweezers to pick up beans and stack building blocks; the AgiBot Elf G2 entered a fire rescue scenario and also completed book organizing in a library; the AgiBot Expedition A3 effortlessly executed highly difficult moves like a 540-degree outward lotus kick...

Who would have thought that just before competing, these robots were diligently working their shifts at the factory.

This was AgiBot's first time participating in the games. All competing models were mass-produced units, with no custom-built machines specifically tailored for the competition.

Transitioning from "workers" to "athletes," AgiBot stayed true to its working genes, fully leveraging its application accumulation in real-world deployment scenarios. This demonstrated the robots' deployment capability, evolving from being "able to demonstrate" to "able to complete real work.

Without further ado, APPSO is here to break down how this gold medal magnet of the humanoid robot games dominated the arena.

Usually "Workers," These Days Sports Masters

At the opening ceremony, AgiBot's Expedition A3 engaged in a human-robot battle with Olympic champion Ding Ning.

Table tennis is perhaps one of the hardest sports for a robot to "master." The ball moves fast with heavy spin, and the hitting window is measured in milliseconds; worse still, the opponent is an Olympic champion capable of improvising and changing the ball's placement at any moment.

Yet, in this match, Expedition A3 played with a sense of on-the-fly adaptability.

One second it was a table tennis expert, and the next, the vibe shifted completely—

Enter the Tai Chi Master.

In the martial arts Tai Chi competition, Expedition A3 smoothly completed a full set of highly difficult moves, including a flying kick + 450-degree outward lotus kick, a 360-degree outward lotus kick + horse stance, and a 540-degree outward lotus kick + horse stance.

Even an elderly man doing his morning exercises would remark: Professional.

This earned AgiBot its third gold medal at this World Humanoid Robot Games.

The Tai Chi competition isn't just about whether the robot can execute the moves; the most important factors boil down to three words: realistic, stable, and difficult.

The arms, waist, and legs cannot move independently of each other. The transitions, pacing, and rhythm of the movements must be seamlessly connected.

Take the horse stance, for example: not only must the body lower itself, but the knees, hips, and ankles must also bear the load together. If even one joint fails to keep up, the scene can quickly turn from "handling with ease" to "stumbling and crawling."

Faced with new routines like Tai Chi, the A3 primarily learned universal motion patterns from the motion data accumulated on the Lingchuang platform. The system then adapted the motion trajectories and rhythms based on its existing motor capabilities.

During training, AgiBot also employed the Sim2Real method. The robot first practiced jumping, spinning, and landing in a simulated environment. The engineering team could adjust motion parameters and repeatedly test different states.

Subsequently, the trained policies were transferred to the physical robot. Real-world friction, joint clearances, sensor errors, and landing impacts differ from the simulated environment. The control system required further adjustments to ensure the movements could be executed stably on the actual machine.

The "relaxed control" emphasized in Tai Chi translates, for the robot, into a public test of its physical structure and motion control capabilities.

Additionally, the games awarded the gold medal for the 100-meter obstacle course, claimed by AgiBot Lingxi X2. While moving at high speeds, the Lingxi X2 simultaneously completed a series of actions including dodging, crossing obstacles, and passing through confined spaces.

Powered by its self-developed AGILE generative universal cerebellum engine framework, the AgiBot Lingxi X2 demonstrated significant speed advantages and outstanding stability when navigating irregular obstacles, capable of making autonomous or semi-autonomous decisions across various terrains.

When it came to the scenario competitions, it was essentially AgiBot's home turf. As a native "worker," the AgiBot Elf G2 had already been deployed on a large scale at factories like Longqi and SAIC before attending the games, boasting a wealth of "work experience."

The fire emergency scenario competition replicated a real fire environment, where participating robots had to sequentially complete three core tasks: hazardous material identification, abnormal valve shutoff, and fire extinguisher grasping and firefighting.

The fire extinguisher weighed 4.5 to 5 kilograms, far exceeding the payload limit of the end-effectors of most humanoid robots on the market.

Instead of customizing expensive grippers for this, AgiBot added structural components to the wrist. By leveraging the structural strength of the wrist itself, the robot hooked the extinguisher and inserted it into a dedicated carrying bucket on its chassis with a 20-degree tilt, cleverly bypassing the payload bottleneck with a smart mechanical design.

Working in tandem with these actions was AgiBot's self-developed beyond-visual-range teleoperation system. Based on a VR solution, the system uses inverse kinematics mapping to translate the operator's body movements into robot commands. It employs whole-body control to map the operator's physical motions to the entire robot, supplemented by force control strategies.

This gives the robot compliance and fault tolerance when interacting with the environment, meaning the operator doesn't have to deliberately avoid every collision, allowing for more natural and continuous movements.

In traditional robot control, any physical contact exceeding the preset trajectory could trigger overload protection in the joint motors, leading to a system shutdown.

By introducing a whole-body control algorithm based on force feedback, AgiBot enabled the robot to produce slight elastic deformations, much like human joints, when contacting door handles, valves, or firefighting equipment, thereby absorbing the impact of collisions.

The library scenario competition tested end-to-end stability. AgiBot formed a joint team with Tsinghua University and Shanghai Jiao Tong University, tasked with completing three missions: outbound transportation, book shelving and placement, and the identification and correction of misplaced books.

During the actual book organizing process, the reflective covers of books and the disorderly arrangement posed extremely high requirements for visual recognition. The visual model adopted by the joint team could not only extract text information from the book covers but also...