Multi-agent cooperative manipulation with uncertainty: a neural net-based game theoretic approach

This paper proposes a novel planning method for multi-agent dynamic manipulation on a plane. The objective of planning is to find optimal forces exerted on the object by agents with which the object can follow a given trajectory. The main contributions of the proposed approach is: First, through integrating of noncooperative game and neural-net approximation, the planner can deal with unknown pressure distribution effectively. Second, by introducing cooperative game between agents, the forces exerted by agents distributed optimally. Based on the dynamic model of the pushed object, the planing problem is solved in two levels hierarchically. In the lower control level, generalized force inputs are designed by using minimax technique to achieve the tracking performance. In the coordination level, cooperative game is formulated between agents to distribute the generalized force, and the objective of the game is to minimize the worst case interaction force between agents and object. Simulations are carried out for the three-agent cooperative manipulation, results demonstrate the effectiveness of the proposed planning method.