Skild AI trained its S1 robotics foundation model to play soccer through self-play in simulation, then transferred the resulting policy to a physical humanoid, first reported by Roboticsandautomationnews. Skild says the approach could eventually help robots acquire increasingly complex skills without people demonstrating every action, while the physical demonstration involved soccer play.
Skild created the environment in Nvidia Isaac Sim and gave the policy one objective: score a goal. Instead of specifying dribbling, tackling or recovery after a fall, it let the policy compete against progressively more capable versions of itself.
During the first few simulated months, the robot could barely walk; it later learned to recover from falls, dribble around defenders, shield the ball and tackle opponents. Skild says these behaviors emerged from the scoring objective rather than individual engineered rewards.
After more than 140 years of simulated play, Skild transferred the resulting policy to a physical humanoid and demonstrated it playing soccer. Soccer combines locomotion, manipulation, recovery from physical disturbances and strategic interaction with opponents.
The work is a post-training extension of S1, which Skild introduced in August. S1 is pretrained using human videos, glove data, simulation and teleoperation, and is designed to learn tasks from in-context demonstrations.
Skild argues that models trained primarily on human-generated data are constrained by the capabilities represented in that data, while physical self-play could let robots exceed human capability. In this case, S1’s mixed-data pretraining and the policy’s later self-play training formed distinct stages.
Skild says the approach is already being extended to everyday robotics tasks. The company says early four-agent soccer experiments have begun to produce passing and coordination; future work is expected to explore collaborative manipulation and navigation as the number of interacting robots increases.