Skild AI trains robot to play football using 140 years of self-play

Summary

Skild AI has successfully trained a robot, named the #Messinator, to play football by utilizing 140 years of self-play in Nvidia Isaac Sim, without relying on human demonstrations. This method reflects a growing trend in simulation training, where robots improve their interactive skills, such as soccer, through competitive self-improvement in virtual environments. Such approaches are part of a broader development in physical AI, enabling robotic models to learn complex tasks through simulations and minimal demonstrations.

Analysis

Nvidia: Nvidia supplies simulation and AI infrastructure platforms including Isaac Sim for robotics development and training. Its tools enabled the large-scale self-play experiments described in the news for advancing robot behaviors in virtual environments. Recent updates to Nvidia's robotics stack continue to support physical AI model training and deployment. Skild AI: Skild AI develops general-purpose robotic foundation models for physical AI applications across industries. Its S1 model was trained via self-play in simulation to acquire complex skills such as playing football without human demonstrations in post-training. The company partners with Nvidia on simulation and deployment infrastructure for these capabilities. Simulation Training: Self-play in virtual environments allows robots to develop interactive skills like soccer through competitive improvement without human demonstrations. Physical AI Development: Robotic foundation models can acquire new multistep tasks using in-context learning from simulation or minimal demonstrations.

Categories

techmachine_learningaiai_agents

Related sources

View Original Tweet