Atlas Robot Masters the 'Ghost Rabona Kick'... Hyundai Motor Unveils Behind-the-Scenes Training

Motion Capture, Retargeting, and AI Reinforcement Learning
From Simulation Verification to Real-World Application
Learning Natural Movement Through Soccer

The training process behind how Boston Dynamics’ humanoid robot Atlas was able to demonstrate advanced soccer skills has been revealed.

A Boston Dynamics researcher teaching soccer skills to Atlas. Hyundai Motor Company

On June 4, Hyundai Motor Company released a making-of film on its official YouTube channel, showcasing the development of its FIFA World Cup 2026™ campaign, “School of Football.” In addition, Boston Dynamics introduced content on its official blog explaining how Atlas trained to acquire soccer skills.

“School of Football” is a video that is part of Hyundai Motor Company’s FIFA World Cup™ campaign “Next Starts Now,” documenting Atlas’s journey as it learns soccer movements and expands the frontiers of robotics technology.

In particular, the robot precisely executes not only fundamental soccer movements such as footwork, passing, and shooting, but also advanced techniques like the “Ghost Rabona Kick”—a variation of the cross-legged Rabona kick. Since its release, the project has attracted global attention from soccer fans and the robotics industry alike.

The Boston Dynamics research team explained that for humanoid robots to move more naturally, they must learn balance, timing, coordination, and adaptability all at once. To achieve this, the researchers focused on soccer as a way for humanoid robots to learn human movements.

Evaluating learning outcomes through simulation before performing the kick motion. Hyundai Motor Company

Soccer is a representative sport that demands a complex combination of balance, timing, coordination, and precise movements. The research team leveraged these characteristics as the optimal environment for humanoids to learn natural movement.

Hyundai Motor Company and Boston Dynamics designed the Atlas training program, drawing inspiration from the biomechanics and movement patterns of world-class soccer players. The movements of actual players were referenced and converted into trainable motion data and action protocols, which were then applied to Atlas’s learning process.

The researchers first used a motion capture system to record soccer player movements, then performed a retargeting process to adapt these movements to the physical structure of Atlas.

This process is a crucial step in overcoming the differences between human and robot anatomy. Although humanoid robots may appear similar to humans, their joint structures and ranges of motion differ, requiring precise conversion work.

Reinforcement learning was then used to have the robot repeatedly practice these movements. In this stage, Atlas not only imitates human actions but also learns its own physical mechanics and motor control, independently optimizing balance and force transmission.

Atlas also trains by running thousands of simulations simultaneously in a cloud GPU environment. This allows the robot to experience the equivalent of approximately one year’s worth of human trial and error within just 24 hours. Through such large-scale parallel learning, Atlas rapidly acquires complex movements.

The learned movements are then applied to the actual Atlas robot, and in most cases are stably executed from the very first attempt. Any errors detected during testing are fed back into the learning process, leading to continuous performance improvements.

School of Football Campaign Key Visual. Hyundai Motor Company

The video also reveals the development process of the advanced “Ghost Rabona Kick” technique performed by Atlas.

The Ghost Rabona Kick is a high-level soccer move that combines the traditional Rabona kick with a defender-deceiving feint. The research team documented a soccer player performing this move, converted it to fit Atlas’s body structure, and then implemented it on the robot using artificial intelligence learning.

This technique is a complex move that requires rapid directional changes for the feint, dynamic balance during takeoff and landing, and powerful force delivery at the moment of the kick. It is regarded as a technology that tests the physical control limits of humanoid robots.

The research team explained that the movements Atlas learns through soccer are not limited to sports skills, but also play a crucial role in advancing robotic technology.

For example, learning the kicking motion helps the robot develop timing, power generation, and coordination, while more complex movements improve rotational motion, weight shifting, and full-body control abilities.

Environments like soccer, where movement and manipulation are required simultaneously, can be directly translated to work capabilities in logistics and manufacturing, where robots need to handle and transport objects. Through this, robots can help reduce human burden by taking on difficult, dangerous, or repetitive tasks in industrial settings.

Previously, Atlas demonstrated its exceptional full-body control by lifting a refrigerator weighing approximately 23 kg and accurately placing it on a table, showcasing the robot’s ability to stably handle heavy objects. By mastering both object manipulation and highly dynamic movements, Atlas is recognized as possessing world-class humanoid robot control technology.

Hyundai Motor Company and Boston Dynamics plan to continue enhancing Atlas’s movement capabilities through a variety of challenging tasks, such as soccer, in the future.

This content was produced with the assistance of AI translation services.