Multi-Robot Active Mapping based on Inverse Reinforcement Learning in Unknown Environments

Multi-robot active mapping aims to efficiently explore and construct complete maps of unknown environments using multiple robots. The key to solving this problem lies in frontier-based exploration, which enables more efficient robot movements. Recent approaches have employed Reinforcement Learning (RL) to optimize robot policies for efficient exploration. However, designing a suitable reward function for $R L$ in this context is challenging because it requires balancing multiple objectives, such as rapid exploration, efficient coordination, and complete coverage. To address this challenge, we propose an Inverse Reinforcement Learning (IRL) approach that automatically updates the underlying reward function from expert demonstrations of effective multi-robot mapping. We use the performance of NeuralCoMapping on specific maps as expert demonstrations and take NeuralCoMapping as the RL policy. By learning the potential reward function optimized by these expert demonstrations, we aim to discover a more comprehensive and balanced reward function that captures the key features of efficient multi-robot mapping. The proposed IRL-based approach not only offers a promising direction for automatically learning effective reward functions and policies in complex multi-robot systems but also instills optimism for the future of robotics, reducing the need for manual reward engineering.