Skild AI demonstrates physical self-play in robotic soccer
Skild AI, Inc. →Skild AI announced physical self-play research on September 23, 2026, extending its S1 work into reinforcement-learning post-training. The company trained a robotic soccer policy in simulation by having it compete against recent versions of itself with the objective of scoring. It reported that skills such as dribbling, shielding the ball and tackling emerged without separate handcrafted rewards for each behavior. Following 140 years of simulated play in NVIDIA Isaac Sim, the team transferred the policy to a physical humanoid robot. Skild described the work as an early research result and said it was extending the approach to everyday robotics tasks. Further work would explore larger teams and collaborative behaviors.