Skild AI's Robot Scores Goals Against Humans After 140 Years of Simulation Training

Startup Skild AI, focused on developing "general robotic intelligence," has unveiled new capabilities of its S1 model. Last month, they released a model whose key feature was the ability to fine-tune without traditional methods, learning simply from a few demonstrations.

Now, the team decided to revive Google's experience, which once trained AlphaGo using self-play — the model playing against itself. Football was chosen as the testing ground.

Initially, S1 was taught basic kicks and dribbling, then "set free": the humanoid robot played against previous versions of itself for the equivalent of 140 simulation years. Interestingly, there were no separate rewards for learning specific tricks; the only reward the robot received was for scoring goals.

Despite this, the robot learned to dribble past defenders, shield the ball with its body, intercept passes, and develop many new kicking techniques. In preliminary matches with 4 agents, passing and coordination have already emerged. Given that the action space in football is huge, with the model needing to output angles for each joint 50 times per second, this is impressive.

Essentially, self-play created an automated curriculum that increases in difficulty on its own. The authors believe this technique can be scaled and transferred to any other domain where there is a clear goal and a simulator.

🔗 Skild AI Blog on Physical Self-Play