Plan and learn to Map for Joint SLAM and Navigation
Published in , 2023
The problem of active simultaneous localization and mapping (SLAM) is an important and challenging problem in the field of autonomous operations. The task involves actively guiding the agent to explore an unknown environment and build a map of the environment while localize the agent within that environment. Although several recent works have showcased the capacities of reinforcement learning(RL) method and the success of active slam based on RGB sensor, there is a lack of model-based method that has planning ability. Monte carlo tree search(MCTS)-based reinforcement learning has been shown to be highly effective on tasks where planning is required. We proposed a Neural Tree Search(NTS) method for joint SLAM and navigation, where a novel search process and a new loss function are used. Comparison experiments on Gibson exploration tasks are conducted through Habitat platform. The proposed method realized competitive performance against the state-of-the-art method with significant less training times, demonstrating the efficiency of the model-based method.