Skild AI Teaches S1 Robot to Play Soccer Using Physical Self-Play

Skild AI has announced a new post-training result for physical artificial intelligence by teaching its S1 robot to play soccer through self-play. The breakthrough shows that a strong base model like S1 can learn to complete dexterous and dynamic tasks by competing against itself in a simulation. Last month, Skild AI introduced S1 as a flagship robotics foundation model that learns tasks from in-context demonstrations similarly to how language models learn from prompts. Because these models are pre-trained on human data, their capabilities are typically capped by human limits. However, the company believes that physical self-play will enable robots to far exceed human capability.

Self-play previously drove major artificial intelligence milestones, such as AlphaGo beating world champions and AlphaStar and OpenAI Five mastering complex multi-player strategy games like StarCraft II and Dota 2. By discovering novel strategies absent from their initial training data, self-play allowed systems to operate at super-human levels. Skild AI hopes to revive interest in this approach and transform it for physical artificial general intelligence.

For this project, Skild AI gave the model a single objective to score. The model then competed against previous versions of itself without any human demonstrations during the self-play phase. Initially, inside NVIDIA Isaac Sim, the model could barely walk. Over time, it learned to stand back up after falling, dribble past defenders, shield the ball, tackle opponents, and perform fancy footwork. Early four-agent matches also showed emerging teamwork, passing, and coordination.

The simulation ran for the equivalent of 140 years of simulated play before Skild AI successfully transferred the policy into a physical humanoid robot for a match. Soccer serves as a testbed because the sport requires both physical and strategic ability, though Skild AI is now extending the approach to day-to-day robotics tasks beyond sports, including virtual construction sites, factories, and homes.