Adaptive Human-AI Coordination via Hierarchical Action Disentanglement
Researchers have developed a new framework called Intrinsic Action Disentanglement (IAD) to improve human-AI collaboration. This deep hierarchical reinforcement learning approach learns distinct action sequences that adapt to different partner behaviors and skill levels. IAD uses an intrinsic reward to encourage disentangled action distributions, creating an interpretable link between high-level decisions and partner-specific responses. Evaluations in the Overcooked-AI domain demonstrated that IAD outperforms existing methods in achieving reliable and adaptive coordination with various simulated and human partners. AI
IMPACT Enhances human-AI collaboration by enabling more adaptive and interpretable coordination, potentially improving performance in complex joint tasks.