A continual imitation learning benchmark for mobile robot navigation in sequential environments

Purpose Current approaches to mobile robot navigation primarily focus on optimizing performance for specific tasks, frequently overlooking the adaptability of navigation strategies in sequential environments. The purpose of this paper is to present continual imitation learning benchmark (CILB), a benchmark for continual imitation learning (CIL) in multimodal robot navigation, aimed at enhancing the robot’s capability to navigate in dynamic and sequential settings. Design/methodology/approach CILB comprises a multimodal dataset, a navigation framework based on CIL and five evaluation metrics designed to assess CIL-based robot navigation. First, this study constructed a dataset containing 12,000 samples by gathering images, point clouds and odometry data from diverse scenarios. Then, the study introduced an end-to-end navigation framework that incorporates momentum updates to mitigate catastrophic forgetting across various environments, providing a baseline for CIL-based robot navigation. In addition, the study defines five specific evaluation metrics to measure continuous navigation performance across different scenarios. Findings The experimental results demonstrate that the proposed benchmark significantly enhances multimodal continual robot navigation in sequential environments. This implies that CILB can effectively assess and improve the adaptability and stability of robots operating in dynamic scenarios. Originality/value The CILB proposed in this paper provides an innovative evaluation benchmark for the field of robot navigation. The accompanying dataset and framework establish a solid groundwork for advancing research on CIL methods and provide valuable references for applications within related domains. The dataset is available at https://vsislab.github.io/CIL/.