Constraint‐based multi‐agent reinforcement learning for collaborative tasks

In order to be successfully executed, collaborative tasks performed by two agents often require a cooperative strategy learned. this work, we propose constraint-based multi-agent reinforcement learning approach called constrained soft actor critic (C-MSAC) train control policies for simulated performing multi-phase tasks. Given task with n $$ phases, the first − 1 n-1 phases are treated as constraints final phase objective, which is addressed centralized training and decentralized execution approach. We highlight our framework on tray balancing including phases: lifting target following. evaluate proposed compare it against its unconstrained variant (MSAC). The comparisons show that C-MSAC leads higher success rates, more robust policies, better generalization performance.