Answers are grounded in this paper and cite their sources.
In this paper we propose a novel approach to distributed predictive formation control for a network of autonomous vehicles (agents) with collision avoidance. The interaction topology through which neighboring agents exchange information is usually represented by a connected and undirected graph, where the nodes are associated with agents and the edges represent communication links. Since the optimal control problem (OCP) that each agent needs to solve in each model predictive control (MPC) step requires information from the whole network, a standard way of handling this is to use iterative MPC schemes. These schemes require multiple solutions of the OCP in each MPC step, which might render them impractical for real‐time implementation. We propose to consider instead the auxiliary problem of controlling a multi‐agent system (MAS) where the agents are associated not with the nodes but with the edges of the interaction graph. This allows to decouple the individual OCPs and thus avoids the need of multiple solutions in one MPC step. Iteration can however not be avoided completely: agents need to run an iterative steepest descent algorithm with their neighbors in order to recover the physical control inputs from the solutions obtained on the edge system. The iterative steepest descent algorithm is computationally much cheaper than multiple solutions of the OCPs; a comparison with a standard Alternating Direction Method of Multipliers (ADMM) distributed MPC approach in a simulation example suggests that our proposed method speeds up the computation by almost two orders of magnitude. While in this approach it is not possible to include absolute constraints on agents states or inputs in the OCPs, it is possible to consider collision avoidance constraints, which are relative constraints between agents.
No takes yet. Share an insight, caveat, or question.
Aditya et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: