Robust Barrier-Certified Safe Learning-based Adaptive Control for Multi-Agent Systems in Presence of Uncertain Environments

This paper develops a decentralized safe learningbased adaptive control for multi-agent systems operating in uncertain environments. Due to the fact that the safe set of the local agent depends on the other agents’ states, the uncertainty of these external systems leads to an uncertain safe set. As a result, the safe control design of the local agent system in a multi-agent setting becomes intractable. To address this challenge, a neural network (NN) based adaptive observer is developed to estimate the state of the unknown external agents. Based on the state estimation of external agents, an adaptive interplay control barrier function (AI-CBF) is formulated. The AI-CBF is designed by considering both the local agent’s state and the NN-based estimated states of external agents. Notably, the limitation of forward invariance for the approximated safe set without guaranteeing the same for the actual safe set is acknowledged in AI-CBF design. The AI-CBF incorporates the bounds on state estimation errors of external agents to guarantee the strict safety requirements of the local agent while learning external agent dynamics. Then, a control framework is formulated using a quadratic programming (QP) method that integrates the safety and stability of the system.