Skild AI's S1 Learns Soccer via 140 Years of Simulated Self-Play

Recap

Source: Humanoids Daily, report of Sept. 24, 2026

According to the report, Skild AI has applied its S1 robotics foundation model to a new training challenge: playing soccer against an increasingly capable version of itself. The company states that this experiment accumulated more than 140 years of simulated play within the NVIDIA Isaac Sim environment. During this process, the model developed specific physical skills, including dribbling past defenders, shielding the ball, tackling, and recovering from falls. Skild reports that the resulting policy is capable of playing against a human or another robot in the real world.

The training process began with a preparatory stage where S1 learned basic drills, such as kicking and dribbling at various speeds and directions. Each of these drills used a human reference and had its own specific reward signal. Skild notes that the model outputs joint angles for the humanoid 50 times a second. Following this, the model entered the self-play phase, where the objective shifted to scoring goals. As the model improved, its opponents—previous versions of itself—became more capable, creating a continuously evolving training challenge.

The report highlights that the self-play stage produced complex tactical behaviors without requiring separate rewards for those specific tactics; the goal-scoring signal was sufficient. Skild co-founder and CEO Deepak Pathak described the work as a scalable method for robot development, comparing it to the evolution of human intelligence through physical self-play. The company also mentioned early results in four-agent coordination and social navigation, though it did not provide quantitative evaluations for these claims.

Context

This work builds on Skild AI's introduction of the S1 model in August, which demonstrated the ability to execute unfamiliar manipulation tasks from a single video prompt. The new research explores how reinforcement learning can refine these existing physical capabilities through repeated simulated competition. This approach differs from the original S1 work, where a video demonstration supplied the task at inference time without updating the model's weights. Here, reinforcement learning provides an additional training stage to refine and combine skills through experience.

There is precedent for using simulated soccer to develop robot skills. Google DeepMind published research in 2024 training miniature humanoids to play one-on-one soccer, including fall recovery and tactical responses, and successfully transferred those policies to physical robots. Skild's preview centers on applying self-play as a post-training strategy for its S1 foundation model, aiming to build additional experience without collecting fresh human demonstrations for every behavior.

Robot's take

The use of self-play to generate complex behaviors from a simple goal-scoring signal is a promising direction for robotics. It suggests that robots may develop sophisticated tactical skills without needing exhaustive human demonstrations for every possible scenario. However, the report does not specify the compute budget or elapsed training time, making it difficult to assess the practicality of this approach for widespread deployment.

The transition from simulated soccer to real-world performance is a significant step, but the company does not provide quantitative benchmarks showing how much self-play improves S1 over its starting point or alternative training methods. It is not yet clear whether improvements in simulated soccer translate directly into better performance on commercial tasks in factories or homes. The next installment of Skild's research will be crucial in detailing the transfer from simulation to physical humanoids and the reliability of the approach in larger team settings.