Skild AI Robot Learns Football Through Self-Play

Skild AI says its S1 robot learned to play football through roughly 140 years of simulated self-play in NVIDIA Isaac Sim. The model trained against copies of itself with one objective: score goals. It received no task-specific demonstrations, custom rewards or direct human coaching. The resulting skills transferred to real-world matches against people and other robots. The achievement highlights advances in robot learning, simulation training and sim-to-real transfer. Thousands of virtual environments can run in parallel on GPU clusters, compressing more than a century of experience into a few weeks of computing time. The system learned balance, positioning, ball control and motor coordination as part of the broader goal. Football is a public demonstration of the S1 model’s wider capabilities. Skild AI says the model can complete manipulation tasks lasting up to 10 minutes from a single video demonstration, without fine-tuning. Applications have included pancake flipping, kit assembly and manufacturing work. Skild AI reportedly reached a $100 million annual recurring revenue run rate in 2026, with robots deployed at more than 60 companies. The company raised $1.4 billion in a Series C round in January at a valuation above $14 billion, with SoftBank and NVIDIA among the participants. Its industrial partners include ABB Robotics and Teradyne, and its robots have reportedly assembled NVIDIA Blackwell GPU systems at Foxconn.