Skild AI trains robots through physical self-play

Physical self-play could reduce robotics’ dependence on expensive human demonstrations, but soccer remains a controlled showcase rather than proof of general-purpose autonomy.

• –A single scoring objective produced diverse behaviors without hand-crafted rewards
• –Self-play creates progressively harder opponents, enabling continual skill improvement
• –Simulation-to-real transfer suggests the learned policy can survive some hardware and environment differences
• –The approach could extend to manipulation, warehouse work, and navigation
• –Scaling compute and improving simulator fidelity will determine whether this generalizes beyond sports