Fully Decentralized ADMM for Coordination and Collision Avoidance
Felix Rey, Zhoudan Pan, Adrian Hauswirth, John Lygeros · 2018
We utilize the alternating direction method of multipliers (ADMM) to devise a communication and control protocol for fully decentralized coordination of moving agents. In particular, we consider a model predictive control (MPC) framework for a group of agents. Each agent has linear dynamics with convex state and input constraints. Nonconvex collision avoidance constraints constitute inter-agent coupling. We develop an algorithm that, if applied by all agents, mediates individual objectives while satisfying constraints. The resulting procedure exhibits several attractive features, including (i) fully decentralized, parallel, and aggregator-free operation, where each agent is only aware of its closest neighbors; (ii) adaptive linearization for handling the nonconvex collision avoidance constraints; and (iii) the treatment of uncooperative agents.