Unmanned Devices Based on Semantic Patterns: Machine Learning-Powered Exploration and Modeling of Unknown Environments

Exploration modeling for unknown environments or inaccessible realms for humans is one of the important application scenarios for unmanned devices, and in order to complete a larger exploration range with limited endurance, there are high requirements on the efficiency of exploration mode and trajectory repetition rate. In this regard, we propose a novel exploration strategy, which characterizes the environment in multiple scales. In the global environment, a deep learning model is introduced to extract the semantic information of the surrounding environment, sparsely characterize the overall environment as a set of nodes, and obtain the shortest traversal path through the entire environment by solving the traveler's problem. In localized environments, the advantages of boundary-based and sampling methods are combined to explore the surrounding environment efficiently and carefully. The multi-scale processing of the environment enables the robot to efficiently explore its surroundings while minimizing subsequent costly backtracking exploration. The method is compared to existing state-of-the-art methods in a variety of challenging simulation and real-world environments. Experimental comparisons show that our method outperforms existing methods in spatial exploration by 40 % and shortens trajectories by more than 30 %.