NestRL: Nested RL Training for Adaptive Human-AI Teaming Accepted at NeurIPS

rao2z · x · 2026-09-25

Researchers from ASU and Colorado State published NestRL, a nested RL training regime for mutual adaptation in Human-AI teaming, accepted at NeurIPS.
Problem: human teammates adapt to an AI agent's behavior, but existing approaches use static training partners; standard joint multi-agent training converges to opaque, partner-specific coordination strategies that generalize poorly.
Method: formulate human-AI teaming as an Interactive POMDP (I-POMDP) and train agents at each nested level against adaptive agents from the level below, exposing them to adaptive behavior while avoiding opaque coordination.
Results: in Overcooked, NestRL beats state-of-the-art baselines with unseen adaptive agents and real human teammates, showing significantly greater adaptability, backed by theoretical analysis.

More from Embodied