Adaptive Multi-Agent Coordination among Different Team Attribute Tasks via Contextual Meta-Reinforcement Learning

In the realm of Multi-Agent Reinforcement Learning (MARL), the challenge of ensuring effective coordination in different multi -agent teams remains a significant hurdle. Existing methods often fall short in generalizing learned policies to novel team compositions, sizes, and capabilities. Addressing this gap, our study focuses on systems with variable and obscure attribute compositions, harnessing a context-based meta-reinforcement learning framework. The approach is twofold: context inference and context-based decision-making. Agents interpret historical data to identify the system's attribute composition, guiding their collective efforts. The accuracy of these inferences is crucial, prompting us to integrate contrastive learning to refine the context inference network via unsupervised training. In the process of decision-making, agents integrate the inferred context with the observation features to select the optimal strategy. Our empirical results underscore the method's efficacy in bolstering decision-making efficiency and precision amidst attribute diversity, marking a significant stride in adaptive teaming for robust multi-robot deployments in dynamic real-world scenarios.