Skild AI's S1 robot learns soccer via 140 years of self-play

Skild's demonstration rests on a post-training extension of S1, the robotics foundation model it introduced in August. S1 is pre-trained on human videos, glove data, simulation and teleoperation, and is designed to learn tasks from in-context demonstrations. Skild's argument is that because such models are pre-trained on human data, they are capped at human capability, and that physical self-play offers a way past that ceiling.

The self-play method itself is borrowed from software AI. DeepMind used it extensively in AlphaGo, and it later contributed to systems that play StarCraft II and Dota 2. Skild is trying to extend the same principle into physical AI, where a robot must learn the movements as well as the strategy. Soccer serves as the test because it combines locomotion, manipulation, recovery from physical disturbances and interaction with opponents.

The company frames the simulation-to-humanoid transfer, alongside the complexity of the task, as a bright green flag for self-play on the road to physical AGI. It is also experimenting with larger groups of agents, and says early four-agent experiments have begun to produce passing and coordination. The broader implication Skild draws is an alternative way of scaling robot learning: use human data to establish foundations, then let robots improve autonomously in simulation — which, if it transfers to practical tasks, could reduce the human-generated data that physical AI currently requires.