Skild AI's S1 humanoid learns soccer via 140-year self-play

Skild AI says its S1 humanoid learned to play soccer by playing against itself for roughly 140 simulated years inside NVIDIA's Isaac Sim, then transferred the policy to a physical robot. In a blog post published September 23, the company describes handing the model a single objective, "score," and letting dribbling, ball-shielding, and tackling emerge on their own.

The framing leans explicitly on AlphaGo Zero, which the post notes beat the original AlphaGo "100 to 0" after self-play. A board-game policy is one thing; controlling a humanoid body is another. The post says the model started barely walking, learned to stand back up after falling, and eventually developed passing and coordination in four-agent games.

"We believe physical self-play will enable robots to far exceed human capability," Skild writes, positioning the soccer demo as preliminary and asking what emerges when the same recipe runs "for orders of magnitude longer, on a virtual construction site, factory, or home."

The post is a company blog, not a paper. It publishes no per-task hardware success rates, no head-to-head numbers against a baseline policy, and no compute figure for the 140 simulated years. GamesBeat reports the run took "over a century" of simulator time; CryptoBriefing notes Skild raised a $1.4 billion Series C in January 2026 at a valuation above $14 billion, with SoftBank and NVIDIA in the round.

Skild's own summary of the recipe: "score. The policy then learned to score by playing against recent versions of itself."