Skild AI trains robot to play football using 140 years of simulated self-play

Skild AI revealed on September 22 that its S1 model developed soccer-playing capabilities through a self-play training method conducted entirely within NVIDIA Isaac Sim. The training was focused on a single objective: score goals. No tailored rewards, no task-specific demonstrations, no hand-holding from human trainers. The resulting policy transferred directly to real-world scenarios, where the robot can now play football against both humans and other robots. How 140 years fits inside a few weeks Modern GPU clusters can run thousands of parallel simulated environments simultaneously, compressing what would be over a century of real-time experience into a fraction of that in wall-clock time. Self-play, the technique Skild used, has a proven pedigree. DeepMind famously used it to create AlphaGo and AlphaZero, systems that mastered board games by competing against copies of themselves. The twist here is that Skild applied this approach not to a board game with discrete moves, but to the messy physics of a bipedal robot navigating a football pitch, involving continuous motor control, balance, object tracking, and real-time decision making. The robot’s training objective was stripped d...