LaMOuR introduces a novel approach for recovering reinforcement learning agents when they encounter out-of-distribution (OOD) states, which typically cause task failure.

• Leverages large language models to guide agents back to familiar territory
• Outperforms existing recovery methods in complex locomotion tasks
• Demonstrates significant improvement in task completion rates when agents face unexpected situations
• Provides a more robust solution for real-world gaming and robotics applications

For gaming applications, this research enables more resilient game AI that can recover from unexpected player actions or environmental conditions, creating more engaging and challenging gameplay experiences.

LaMOuR: Leveraging Language Models for Out-of-Distribution Recovery in Reinforcement Learning