Skild AI announced a breakthrough in physical AI using self-play. A strong base model like S1 can now learn complex tasks such as soccer by competing against itself in simulation. This approach allows robots to surpass human capability. The model started unable to walk but learned to dribble, shield, and tackle over time. After 140 years of simulated play, the policy transferred successfully to a humanoid robot. Soccer was chosen as a testbed for its physical and strategic demands. The method is general and will extend to other robotics tasks. Skild plans to scale this to larger teams and new domains. The work revives self-play as a path to physical AGI. The results build on Skild’s S1 robotics foundation model.
What changed
S1-class models were capped at human capability due to pre-training on human data.
What it unlocks
Robots can now exceed human capability in dexterous and dynamic physical tasks through self-play.
- 140 years of simulated play
- 100 to 0 AlphaGo Zero vs AlphaGo
What you need to act on it
- a strong base model like S1
- simulation environment (e.g., NVIDIA Isaac Sim)
Sources