Intrinsic Action Disentanglement for Human-AI Coordination
A novel framework for deep hierarchical reinforcement learning, known as Intrinsic Action Disentanglement (IAD), has been introduced to enhance collaboration between humans and AI. IAD identifies unique, partner-aware low-level action sequences based on high-level latent skills, utilizing an intrinsic reward system to promote disentangled action distributions. This approach creates a clear connection between high-level choices and specific partner behaviors, allowing for adaptable responses to varying partner dynamics amid distributional shifts. The framework was tested within the Overcooked-AI environment across several layouts. Further information can be found in arXiv:2605.24343.
Key facts
- IAD is a deep hierarchical reinforcement learning (DHRL) framework.
- It learns partner-aware low-level action sequences conditioned on high-level latent skills.
- An intrinsic reward encourages disentangled action distributions across skills.
- IAD provides an interpretable mapping between high-level decisions and partner-specific responses.
- It enables adaptation to heterogeneous partner dynamics under distributional shift.
- Evaluation was conducted in the Overcooked-AI domain across multiple layouts.
- Existing methods often collapse to a single dominant behavior or learn poorly aligned skills.
- The paper is available on arXiv with identifier 2605.24343.
Entities
Institutions
- arXiv