Ahead of the 2026 World Cup, Boston Dynamics released a video showing its Atlas humanoid robot studying football movements from past tournament matches displayed on a large screen.

The robot closely observes the players before moving to a practice area, where it recreates the actions it has just watched. In one sequence, Atlas shifts its weight and swings its leg forward to guide a ball smoothly across the floor.

The footage also shows the humanoid performing a series of basic football drills designed to improve balance, coordination, and movement control.

Recently, Boston Dynamics showed how Atlas learned heavy lifting through reinforcement learning and simulation, carrying shifting loads like mini-fridges just weeks after its public debut.

Atlas learns football

In the new demonstration, the Atlas robot stands in front of a giant display screen, carefully observing player movements, body positioning, and reactions during different moments of the game. After each clip finishes, Atlas heads to a practice area where it immediately recreates the actions it has just watched.

The footage highlights Atlas performing a range of football-inspired movements. In one sequence, the robot shifts its weight, swings a leg forward, and smoothly guides a ball across the floor with controlled contact. Atlas then moves through a series of rapid drills designed to improve balance, coordination, and movement timing.

The robot’s motions appear increasingly fluid as the session progresses, suggesting that the exercises are intended to test not only strength but also agility, reflexes, and real-time adaptation, reports Techeblog.

Some of the most striking moments come from Atlas mimicking player emotions and reactions. After completing a drill, the humanoid raises its arms in celebration like a footballer after scoring a goal. Moments later, it drops to one knee and pauses, recreating the reaction to an on-field injury. According to the video, these gestures were copied directly from match footage Atlas had viewed only moments earlier.

Hyundai, of which Boston Dynamics is a subsidiary, described the demonstration as Atlas’s first experience with soccer and introduced the project as part of a new series called “School of Football.” The company has also hinted at plans to bring Atlas and Spot, Boston Dynamics’ quadruped robot, to the World Cup, although their exact role has not been revealed.

Simulation trains Atlas

The soccer showcase follows another recent Atlas demonstration in which the humanoid lifted and carried a roughly 99-pound (45 kilograms) mini-fridge while maintaining perfect balance, highlighting the robot’s growing combination of strength, control, and mobility.

The demonstration marked a shift from choreographed humanoid motions toward adaptable industrial behavior aimed at real-world use in factories, warehouses, and construction environments. Boston Dynamics described Atlas as a “general-purpose tool for physical work,” reflecting its broader development direction.

Unlike vision-only systems, Atlas relies heavily on proprioception—internal sensing of body position, load, balance, grip, and resistance—while manipulating objects. This allows the robot to adjust in real time when handling unstable or shifting loads. According to the company, Atlas was trained using reinforcement learning, repeatedly practicing lifting and carrying tasks in simulation under varying conditions.

These simulations altered key variables such as object weight, floor friction, grip strength, and object placement, forcing the robot to develop adaptive control strategies. Boston Dynamics said Atlas accumulated millions of simulated practice hours running in parallel on graphics processing units (GPUs).

Training begins with a reference trajectory, which may come from animation or teleoperated demonstrations. The robot is rewarded for completing tasks such as maintaining grip and balance while external disturbances are introduced. Once performance stabilizes in simulation, the behavior is transferred to the physical robot for real-world testing and refinement.

The company highlighted that this pipeline reduces the long-standing “sim-to-real gap,” where simulated skills fail in reality due to friction, latency, or sensor noise. Atlas benefits further from a simplified hardware design, featuring symmetrical limbs and just two actuator types, making simulation and transfer more accurate and efficient.