Skild's S1 learns soccer by playing itself for 140 simulated years

In short: Skild AI's September 2026 post shows Physical Self-Play: its S1 foundation model, pretrained on human data, improved by competing against recent copies of itself in NVIDIA Isaac Sim for the equivalent of 140 years of play, then transferred the soccer policy to a real humanoid.

What S1 already claimed

In August Skild introduced S1 as an in-context learner: show one video demonstration (seen or unseen) and it executes — including long-horizon tasks up to 10 minutes — with no weight updates. On internal unseen-task benchmarks, Skild reported 66% success for ICL vs 9% for a language-prompted VLA at 100k pretraining hours, and said one in-context demo matched roughly 380 post-training episodes.

What self-play adds

• Single reward: score. Skills like dribbling, shielding and tackling emerged because they helped win, not because engineers hand-shaped them.
• Early sim months: barely walk. Later: get up after falls, then play a match after sim-to-real transfer.
• Thesis: human-demo pretraining caps at human skill; self-play is the path past that ceiling for embodied AI.

Why it matters

Most robot VLA talk still means "more teleop hours." Skild is arguing for a second loop — self-improvement in simulation — closer to how AlphaGo left human game trees behind.