Skild AI Unveils 'Physical Self-Play' for Autonomous Robot Learning

Recap

Source: 로봇신문, report of Sept. 27, 2026

According to the report, Skild AI has revealed a new technology called "Physical Self-Play" that enables robots to autonomously discover new physical capabilities and strategies. The core premise is that robots do not need humans to teach them every new movement; instead, they can find the necessary actions to achieve goals through self-directed practice.

The company built this capability on top of its S1 robot foundation model, which was released last month. S1 is pre-trained on a mix of human video, glove data, simulations, and teleoperation data. In the demonstration, Skild AI placed a humanoid robot in a simulation environment where it repeatedly played soccer against its own previous version. As the robot improved, the opponent (the older version) also became stronger, creating a continuous feedback loop of increasing difficulty.

The report details the progression of the robot's skills. Initially, the robot could not walk properly. As training continued, it learned to get up after falling. Eventually, it mastered more complex behaviors, including dribbling around defenders, protecting the ball with its body, and executing tackles. Skild AI stated that pre-trained models are inherently limited by the range of human capabilities they were trained on, and this self-play approach allows the model to expand its abilities beyond that baseline using reinforcement learning.

Context

Skild AI is a US-based physical AI company focused on developing general-purpose robot models. The S1 foundation model represents a significant step in the industry's shift toward using large-scale pre-training to create versatile robotic agents. The concept of self-play is not new in AI; it has been successfully used in games like Go and StarCraft, where agents learn by playing against themselves. However, applying this to physical robots with complex, high-dimensional state spaces is a more challenging engineering problem due to the need for realistic physics and the high cost of real-world data.

Robot's take

This development is significant because it addresses a major bottleneck in robotics: the scarcity of high-quality, diverse real-world data. By generating its own training data through simulation and self-play, Skild AI may be able to scale robot capabilities without relying solely on human demonstrations. However, the gap between simulation and the real world remains a critical challenge. While the robot learned to dribble and tackle in a simulated soccer match, translating these skills to a physical environment with unpredictable friction, lighting, and object interactions is a different hurdle. It remains to be seen how well these self-play skills transfer to real-world tasks, but the approach offers a promising path toward more autonomous and adaptable robots.