Skild AI unveils 'Messinator': a soccer robot trained via 140 years of self-play

The core of this news is not "robots can play soccer," but a method validation: moving AlphaGo-style self-play into robot learning in the physical world. Skild AI puts the robot into a virtual soccer field built in NVIDIA Isaac Sim and gives it only one final objective—score a goal and get a high score. Throughout training, no reward is set separately for dribbling, and no one scores shielding the ball, tackling, shooting, and getting up after falling one by one.

The opponent the robot faces is only its own previously saved model versions. Through repeated matches between old and new versions, the winning new strategy becomes a tougher opponent in the next round of training, and every improvement automatically creates the next problem. According to Skild AI's technical blog, in the first few months after entering the virtual soccer field, the robot could not even walk well, but by the time it "grew to college-student age," it had taught itself to stand back up after falling; as opponents grew stronger, more complex behaviors such as dribbling around defenders, using its body to shield the ball, and actively tackling appeared one after another. At no point did engineers prescribe that it must do this.

What needs to be distinguished is this: these behavior descriptions come from Skild AI's technical blog, relayed by QbitAI, and are statements from the source party; the QbitAI article itself is a secondhand relay, and the technical blog and X platform links it cites are first-party materials, but this article can only judge based on the given sources. That self-play played 140 years of World Cups in the virtual world and transferred soccer strategies to a real humanoid robot is the source party's claim, not a conclusion verified by a third party.

The first half of the name Messinator comes from Messi, and the second half comes from the machine Terminator in "The Terminator"; together it means "Messi Terminator." But its real brain is S1, the flagship robot foundation model Skild AI previously released. S1's most core ability is to let robots do "in-context learning" like a large model: show it a video demonstration of a task, and it treats the video as a "prompt," gradually understands the task intent, and converts the intent and actions in the video into an execution method suitable for its own body; when it encounters a new task, it does not need to be retrained from scratch and can directly get started.

Skild AI believes that foundation models like S1 mainly learn from human data, so their capability ceiling is inevitably limited by human experience. To continue getting stronger, robots must create their own experience and try to discover actions humans have not taught. So the team added reinforcement learning and self-play on top of S1.

From this, Messinator's positioning is a continuation rather than a fresh start: Skild Brain is responsible for giving it the basic skills to control its body, maintain balance, and handle collisions, while self-play is responsible for making it think "what should I do next to do better." This division of labor is the source party's formulation and is an explanation of the technical route, not independently verifiable mechanistic detail.