Learning to Navigate Sequential Environments: A Continual Learning Benchmark for Multi-modal Navigation

Existing works for robot navigation typically focus on performance in specific tasks, overlooking the adaptability of navigation strategies to sequential environments. In this paper, we present CLBMN, a continual learning (CL) benchmark for multi-modal robot navigation, which aims to enhance robots' navigation capabilities in handling sequential tasks. The benchmark is composed of a multi-modal dataset, a CL-based algorithm, and three comprehensive evaluation metrics for CL-based multi-modal navigation. Specifically, we construct a multi-modal dataset with 10,000 samples by collecting images, point clouds, and odometry data from five scenarios, including daytime, nighttime, rainy conditions, highway, and field. Then, we propose an end-to-end navigation framework with momentum update to mitigate catastrophic forgetting across various environments, which provides a baseline approach for CL-based robot navigation. Furthermore, we define three evaluation metrics specifically designed to evaluate the robot's continual navigation performance across different scenarios. Experimental results demonstrate that the proposed benchmark is effective for studying multi-modal continual robot navigation in sequential environments. Our project page is available at https://vsislab.github.io/clbmn/.