A Benchmark of Planning-based Exploration Methods in Photo-Realistic 3D Simulator

Autonomous exploration is an essential ability of mobile robots and has been widely investigated. A reliable evaluation is required to compare exploration methods in different aspects. However, existing benchmarks are mostly implemented on gird-based simulators which are lack of real-world environmental factors. Besides, most benchmarks only focus on single-agent exploration or multi-agent exploration, lacking comprehensive analysis on both of them. In this paper, we select representative planning-based exploration techniques in both single-agent and multi-agent settings and evaluate them on large and highly cluttered maps. The experiment is conducted on a photo-realistic 3D simulator, Habitat, and the results indicate the performance of these methods in aspects of map size, cooperation, efficiency and agent's team size.