Event
Skild AI Uses Self-Play Training to Teach Robots to Play Soccer
1 reports · 1 sources
Skild AI demonstrated a self-play training method: its robot foundation model S1 learned to play soccer through more than 140 years of simulated soccer matches in Nvidia Isaac Sim simulation, without relying on human demonstrations, and transferred the trained policy to a physical humanoid robot to complete a soccer-playing demonstration. This is post-training extension work for the S1 model released in August this year. In simulation, the policy successively learned to recover after falling, dribble past opponents, shield the ball, and tackle. The company says the method has been extended to everyday robotic tasks, and early experiments on four-agent passing and receiving coordination have been conducted.
Report timeline
- Skild AI trains S1 model to play soccer via over 140 years of simulated self-play, then transfers it to humanoid robots
Skild AI demonstrated a self-play training method in which its S1 robot foundation model learned through more than 140 years of simulated soccer matches in Nvidia Isaac Sim, without relying on human demonstrations. In simulation, the policy successively learned to recover from falls, dribble past opponents, shield the ball, and tackle, and was then transferred to physical humanoid robots for soccer-kicking demonstrations. The company says the method has been extended to everyday robotic tasks, and early experiments with four-agent passing plays have been conducted.
Robotics & Automation News ·