Skild AI’s S1 learned soccer in 140 simulated years of self-play, with goals as its only reward

Skild AI said on September 23, 2026 that its S1 robotics foundation model learned to play soccer through more than 140 years of simulated self-play in NVIDIA Isaac Sim, then transferred the policy to a real humanoid and played matches. In the self-play stage, scoring goals was the only objective; Skild says dribbling, shielding the ball and tackling emerged on their own.

Why it matters: self-play made game AI superhuman because an improving opponent is a curriculum that never runs out. Skild is claiming the same loop works for whole-body physical skills, and it says it is applying it beyond sport.

How did Skild AI teach a robot to play soccer?

In two stages, according to Humanoids Daily’s account of the release. First, the model practised drills such as dribbling and kicking, with individual rewards and human reference motion. Second, it played matches against progressively stronger versions of itself, with goal-scoring as the sole objective. Skild’s own post describes the second stage plainly: “The policy then learned to score by playing against recent versions of itself. As it improved, its opponents improved alongside it.”

S1 did not start from zero. Skild says the base model is pre-trained on human videos, data gloves, simulation and teleoperation, then post-trained with reinforcement learning. The soccer work is that post-training step. Skild’s post says that in its first few simulated months the model could barely walk. Humanoids Daily reports the policy outputs joint angles 50 times a second and training used NVIDIA Isaac Sim with Omniverse libraries.

What behaviours emerged from self-play?

Skild lists dribbling past a defender, shielding the ball and tackling, and says these appeared “simply because they helped the model score.” In early four-agent games, passing and coordination began to appear, according to Humanoids Daily and GamesBeat. After the 140 simulated years, the policy was transferred to a real humanoid that played against people and other robots. Humanoids Daily identifies the hardware as the Unitree G1.

Robot soccer through reinforcement learning is not new. Google DeepMind trained miniature humanoids for one-on-one soccer in 2024, including fall recovery and tactical responses. What is different here is the combination of commercial humanoid hardware, a general foundation model and open-ended self-play, rather than a policy trained for one game. Our sim-to-real explainer covers why policies trained in simulation can now survive the jump to hardware.

What has Skild not disclosed?

Most of what a researcher would need to judge the result. There is no paper, no peer review, no code and no released checkpoint, as Mixed News noted. Skild gives no win rate, goal count or baseline comparison, no wall-clock training time and no compute budget; 140 years is simulated time, not elapsed time. Humanoids Daily notes the release does not show gains on commercial tasks and has no quantitative evaluation of teamwork. Skild says a fuller account of the training recipe and sim-to-hardware transfer will follow, and that its next release will cover behaviours that emerge in larger teams.

The commercial context explains the timing. Bloomberg reported on September 10 that Skild had reached a $100 million annual revenue run rate, which Skild says came ten months after its first commercial deployment. That figure is company-reported. Soccer is a demonstration; the claim that matters is Skild’s statement that the approach is general and is being extended to everyday robot tasks. How S1 compares with other general-purpose models is covered in our robot foundation model comparison, and the competing bet on learned simulators in our world models analysis.