Skild AI announced on September 22 that its S1 model has developed soccer-playing capabilities through a self-play training method carried out entirely within NVIDIA's Isaac Sim platform. The training centered on a single objective—score goals—with no tailored rewards, no task-specific demonstrations, and no direct guidance from human trainers. The resulting policy transferred directly to the physical world, where the robot can now play football against both humans and other robots.

How 140 years fits into a few weeks

Modern GPU clusters can run thousands of parallel simulated environments simultaneously, compressing what would amount to more than a century of real-time experience into a small fraction of that in wall-clock time. Self-play, the technique Skild employed, has an established track record: DeepMind famously used it to build AlphaGo and AlphaZero, systems that mastered board games by competing against copies of themselves. The method also sidesteps one of reinforcement learning's most labor-intensive steps: hand-crafting the reward signals that define success for an agent.

The difference here is that Skild applied the method not to a board game with discrete moves, but to the complex physics of a bipedal robot on a football pitch—an environment demanding continuous motor control, balance, object tracking, and real-time decision making. The direct transfer matters because getting simulation-trained policies to hold up on physical hardware—the sim-to-real gap—has long been one of robotics' hardest problems.

The robot's training objective was stripped to its essence: score the goal. Everything else—the footwork, positioning, and ball control—emerged as learned behaviors rather than programmed instructions.

The S1 model and Skild's broader ambitions

Football serves as the showcase demo, but the S1 model's commercial value lies in its versatility. Launched in late August 2026, the S1 can perform complex manipulation tasks lasting up to 10 minutes after learning from a single video demonstration, with no fine-tuning or parameter updates required. The company has shown the model handling tasks that range from pancake flipping to kit assembly. Conventional industrial robots, by contrast, typically require explicit programming for each task they perform.

Skild AI's development philosophy draws explicit parallels to large language models: the company trains on extensive data drawn from human videos and physics simulations to build general-purpose motor intelligence. NVIDIA's simulation infrastructure plays a central role in that strategy. Isaac Sim supplies the physics engine and rendering pipeline that make high-fidelity robotic training feasible at scale, helping narrow the sim-to-real gap.

A $14 billion bet on general-purpose robots

Skild reached a $100 million annual recurring revenue run rate shortly after its first commercial deployment earlier in 2026, and its robots now operate across more than 60 client companies. In January 2026, the company raised $1.4 billion in a Series C funding round, achieving a valuation exceeding $14 billion, with SoftBank and NVIDIA participating in the round.

ABB Robotics and Teradyne—which owns Universal Robots and MiR—are both working with Skild. The company has also deployed robots to assemble NVIDIA Blackwell GPU systems at Foxconn. NVIDIA, notably, appears on both sides of that relationship: it supplies the Isaac Sim training infrastructure while also having participated in the Series C. Together, the ABB, Teradyne, andconn engagements give Skild's simulation-trained approach real-world proving grounds spanning industrial automation, collaborative and mobile robotics, and electronics manufacturing—settings that will put its general-purpose approach to the test at commercial scale.