THG News · Research

Skild AI Uses Simulated Self-Play to Teach a Humanoid Robot Soccer Skills

· Updated

Skild AI said on September 23, 2026 that it had used a post-training method it calls physical self-play to teach a humanoid robot to play soccer, without human input during that stage of training. In a post on its website, the company said its S1 foundation model learned by competing against progressively stronger versions of itself in NVIDIA's Isaac Sim simulator, and that the resulting policy transferred to a physical robot.

From standing start to tackling

According to Skild, the simulated robot began unable to walk. Over what the company describes as 140 years of simulated play, behaviors such as dribbling, shielding the ball and tackling emerged. Skild said it did not hand-engineer rewards for those skills. In the company's words, "Rather than hand-crafting rewards for these behaviors, they emerged simply because they helped the model score."

Skild noted that S1 was pre-trained with human video demonstrations, so people still shape the base model. The self-play stage itself, the company said, requires no human input.

Why self-play is being tried on robots

Self-play is the technique behind several well-known game-playing AI systems. Skild pointed to Go and StarCraft as cases where agents trained against themselves found strategies that were not present in their training data. The company argues that the approach had not previously been applied effectively to physical robots, and says a strong pre-trained base model is what makes it workable for dexterous and dynamic tasks such as soccer.

The method contrasts with imitating human demonstrations gathered by teleoperation and with reinforcement learning driven by rewards that engineers design for each behavior. Skild's stated aim is to let robots improve beyond the level of the human data they start from. Skild has presented the work in a company post rather than an independent evaluation, and the results described here are the company's own.

The S1 model underneath

The self-play work builds on S1, which Skild introduced on August 18, 2026 as its flagship foundation model. Skild describes S1 as an in-context learner that can perform a new task after watching a single video demonstration, with no fine-tuning. The company reported 96% success on in-distribution tasks and 66% on unseen tasks, compared with 9% for language-prompted policies, at 100,000 hours of pre-training. Skild also said one in-context demonstration is worth roughly 380 post-training examples, equal to about 50 to 100 hours of teleoperation.

According to Skild, S1 is deployed with the company's commercial partners and is open to early-access sign-ups. Skild raised a $1.4 billion Series C in January 2026, led by SoftBank, at a valuation of more than $14 billion, the company said at the time.

Sources