Partner-Aware Hierarchical Skill Discovery for Robust Human-AI Collaboration
Researchers have developed a new framework called Partner-Aware Skill Discovery (PASD) to improve human-AI collaboration. This method addresses limitations in existing hierarchical reinforcement learning by conditioning skills on partner behavior, rather than just agent-centric rewards. PASD uses a contrastive intrinsic reward to identify patterns in partner interactions, promoting adaptive coordination and mitigating shortcut learning. Evaluations on the Overcooked-AI benchmark demonstrated that PASD significantly outperforms other methods in transferring skill learning across diverse partner behaviors, including human proxy models. AI
IMPACT Enhances AI's ability to adapt and coordinate effectively with novel human partners, crucial for robust human-AI teaming.