Physical Self-Play (www.skild.ai)

A groundbreaking advancement in physical AI has been announced with the introduction of self-play for the S1 model, which allows robots to master complex tasks such as soccer through competitive simulation against themselves. Last month, S1 was unveiled as a robotics foundation model that learns tasks similarly to language models through in-context demonstrations. This new method signifies a shift in how robots can learn beyond human capabilities, enabling them to develop novel strategies and skills autonomously without relying solely on human-derived data. By engaging in self-play, S1 demonstrated an impressive trajectory of improvement, starting from basic movement to sophisticated game strategies such as dribbling and tackling over 140 years of simulated play. This approach not only challenges the traditional reliance on reinforcement learning by removing predefined rewards but also revives interest in self-play as a potential pathway toward artificial general intelligence (AGI). Beyond soccer, the implications of this technique extend to various everyday robotic applications, suggesting a promising expansion into domains like construction and urban navigation as the methodology scales.