Skild AI unveils Messinator, a humanoid robot trained by 140 years of simulated World Cup self-play

Skild AI, a US embodied-AI company, has published a demonstration of Messinator, a humanoid robot that dribbles past defenders and finishes with a shot while wearing an Argentina kit. The skills come from roughly 140 years of simulated World Cup self-play, as reported by QbitAI on September 25.

Messinator's brain is S1, Skild AI's flagship robot foundation model. S1's core capability is in-context learning for robots: given a short demonstration video, it treats the footage as a prompt, works out the intended task and translates the intent and motions into an execution strategy suited to its own body — no retraining needed when a new task appears.

On top of that, Skild AI added reinforcement learning and self-play. The robot was placed in a virtual soccer pitch built with NVIDIA Isaac Sim with a single objective: score and get a high reward. No separate reward was designed for dribbling, shielding, tackling, shooting or getting back on its feet, and its only opponent was a previously saved version of itself.

Each winning policy becomes a tougher opponent in the next round, so every improvement automatically manufactures the next challenge. According to Skild AI's technical blog, the robot could barely walk during its first months in the virtual stadium; by the time it "reached college age" it had taught itself to stand up after falling.

As opponents grew stronger, more complex behaviours appeared — dribbling around defenders, shielding the ball with its body, actively tackling — none of them specified by engineers. In the final demo Messinator walks up to the ball and shoots, and has also learned to carry and protect the ball, tackle, and adjust its positioning in live match-ups.

The approach builds on a longer track record. Founded in 2023 out of Carnegie Mellon University's robotics research ecosystem, Skild AI launched Skild Brain, a cross-hardware robot foundation model, in July 2025 under the slogan "Any robot, any task, one brain."

Less than two months later it published omni-bodied training results: about 100,000 robot bodies with different structures simulated, roughly 1,000 simulated years of cumulative training, with the model then required to take over robots it had never seen. Even when a robot suddenly loses a leg, locks a joint, jams a wheel or carries extra load, the model adjusts its centre of gravity and locomotion online, sometimes regaining balance within seconds.

That work matters because the more bodies a model has seen, the harder it is to pass by memorisation, forcing it to learn general physics around balance, joints, centre of gravity and actuation. Skild AI was co-founded by CEO Deepak Pathak, known for curiosity-driven exploration research, and president Abhinav Gupta, who has worked on large-scale visual learning, self-supervised learning and robot manipulation.

The demonstration is a simulation result, and how reliably it transfers to real hardware and messy real-world opponents remains the open question. If it does hold up, robot skill acquisition shifts from manual teleoperation and hand-coded rewards toward compute and simulation — a much faster and cheaper loop for embodied AI.