Boston Dynamics deployed two of its robots at the 2026 FIFA World Cup in the United States in distinct roles. The Atlas humanoid was trained to perform the "Ghost Rabona" — a complex soccer trick — using motion capture and reinforcement learning in simulation. The Spot quadruped is patrolling four stadiums, scanning for hazards and suspicious packages.
Key takeaways
- Atlas mastered the "Ghost Rabona" soccer trick in hours after motion-capturing a professional player's movements
- Boston Dynamics abandoned classical predictive balance control in favor of reinforcement learning (RL) about a year ago
- Spot is operating at 4 stadiums: 2 in Dallas and 2 at Citi Field in New York, in both autonomous and teleoperated modes
- Spot Cam loadout: 360° panoramic camera, 4K imaging, thermal imaging, LiDAR and an enhanced autonomy package
- Alberto Rodriguez (Boston Dynamics): the technology for safe interaction with other people near the robot does not yet exist
The "School of Football" campaign and why the World Cup made sense
Boston Dynamics has been part of Hyundai Motor Group since 2021. Hyundai has partnered with FIFA for 27 years. When Hyundai proposed a soccer campaign, Alberto Rodriguez, director of robot behavior, was direct: the World Cup was too good an opportunity to pass up.
The target move was the "Ghost Rabona" — a trick in which the player crosses one foot behind the supporting leg to surprise the goalkeeper. Interacting with the ball and the ground proved more technically demanding than the acrobatic jumps Atlas had mastered earlier. The rapid lower-body footwork required particularly precise control policies.
A skilled soccer player wore a mocap suit and performed the trick in a studio. The data went into a simulation engine. As Rodriguez noted, training with this technique was "way faster" than before — the resulting behavior policy was ready within hours, at most a day.
A methodological shift — RL replaced predictive control
About a year ago, the company decided to abandon its own predictive balance control implementation — at the time, arguably the best in the world — and start from scratch with reinforcement learning.
Atlas moving a fridge — footage the company released — would have been difficult to achieve with classical techniques. RL enables the robot to go beyond what can be directly demonstrated. Rodriguez notes the same methods that enabled the Ghost Rabona translate directly to more reliable stepping in varied industrial terrain.
The robot's mechanics make this easier. Atlas uses only rotary actuators, no linear ones. This design simplifies simulation and shortens the training cycle — from capturing data to a working policy on the physical robot.
Grass was a challenge. A soccer pitch is more compliant than concrete, plywood, or other typical test surfaces. The robot's foot can sink, slip, or catch on turf. The algorithm had to be robust enough to handle this variability even though it was absent from the simulation environment.
Spot at the stadiums — a working dog, not a photo op
Spot is operating at four venues: two in Dallas and two at Citi Field in New York. It walks the perimeter when most of the audience has left, scanning for hazards and suspicious packages.
"Spot is a working dog — it's not a photo op."
— Merry Frayne, senior director of product for Spot, Boston Dynamics
The hardware loadout is significant. Spot Cam provides a 360° panoramic view, 4K imaging, thermal imaging, and LiDAR. An enhanced autonomy package is included — the robot can operate fully autonomously or under teleoperator control.
The robot has no facial recognition capabilities. Security teams at the stadiums operate the robots themselves — not Boston Dynamics employees. This mirrors how Spot is deployed at nuclear facilities, mines, and first responder organizations worldwide.
Why it matters
The FIFA World Cup is a visibility test in public environments — significantly harder than the closed industrial facilities where Spot typically operates. Hundreds of thousands of people, varied infrastructure, and changing terrain conditions create challenges different from pipeline inspection or nuclear site monitoring.
For Atlas, the more significant takeaway is technological. The mocap → simulation → behavior policy pipeline completed within a single day. This shortens the iteration cycle by weeks compared to the model-based approach. In industrial settings — warehouses, factory floors — it means faster adaptation to new tasks without rebuilding the physical model.
Rodriguez acknowledges the limits directly: the technology for interacting with other players or running safely alongside people does not yet exist. That honest framing matters — and it identifies clearly where the next engineering challenges lie.
What's next
- Boston Dynamics has indicated it will announce additional deployment markets for Spot in the coming months
- RL techniques developed for the FIFA project will continue to be refined for industrial navigation in varied terrain
- The 2026 FIFA World Cup final is scheduled for July 19, 2026 — both robots will continue operating at stadiums through the end of the tournament
Sources
- The Robot Report — Boston Dynamics brings its legged robots to the FIFA World Cup
- Hyundai Newsroom — Hyundai Motor Announces Largest-Ever Mobility and Robotics Deployment for FIFA World Cup
