Continual Learning with Evaluation for Motion Planning in Unstructured Environments

Learning-based motion planning methods for unmanned ground vehicles (UGV) have shown significant advantages in terms of real-time performance and adaptability. However, their performances are usually dependent on the quality of demonstration data. Thus, even with the increased amount of training data, the model performance may still degrade due to imperfect demonstrations within the increased data. Moreover, newly collected demonstrations usually require to be trained with historical data to avoiding forgetting previous knowledge, which further hinders the efficiency in updating and improving the model. To tackle these problems, this paper introduces a continual learning framework for UGV motion planning in unstructured environments, enabling model learning from the optimal demonstrations without forgetting previously learned knowledge. Within this framework, an evaluation-based training data filtering method is implemented to filter out poorly performing demonstration data, thus preventing model performance degradation caused by imperfect demonstrations. Furthermore, a continual learning algorithm was employed to achieve the continuous and efficient improvement of the motion planning model with accumulative data. Additionally, this study collects real-world data from diverse unstructured scenes for training and evaluating the proposed framework. The results indicate that, using the same training data, the motion planning model based on the proposed continual learning framework achieves a 33.1% improvement in efficiency and a 28.1% improvement in smoothness compared to the model obtained through traditional imitation learning.