Multi-Agent Embodied Visual Semantic Navigation With Scene Prior Knowledge

In visual semantic navigation, the robot navigates to a target object with egocentric observations and class label of is given. It meaningful task inspiring surge relevant research. However, most existing models are only effective for single-agent single agent has low efficiency poor fault tolerance when conducting more complicated tasks. Multi-agent collaboration can improve strong application potentials. this letter, we propose multi-agent in which multiple agents collaborate others find objects. challenging that requires learn reasonable strategies perform efficient exploration under restrictions communication bandwidth. We develop hierarchical decision framework based on mapping, scene prior knowledge, mechanism solve task. The experimental results unseen scenes both seen objects illustrate higher accuracy proposed model compared model.