Mapless Navigation With Safety-Enhanced Imitation Learning

摘要

Mapless navigation is a popular approach for guiding a robot in an unknown environment. However, current learning-based methods for mapless navigation cannot guarantee safe exploration during training. In this article, we aim to safely train a mapless navigation policy with 2-D LiDAR inputs using imitation learning (IL). We propose a novel IL training scheme based on the dataset aggregation with extra safety enhancement. In order to avoid collisions during training, we design a controlled exploration strategy for demonstration data collection. In this strategy, the output of the novice and expert policy are fused adaptively to control the robot. In order to improve the obstacle avoidance capability of the trained policy, we design a safety-enhanced loss function for network training. This proposed loss function focuses on the obstacle avoidance by prioritizing the training samples collected in different states. In addition, inspired by the symmetrical pattern of the navigation task, we propose a new data augmentation technique, which increases the training speed significantly. In both simulation and real-world experiments, our method demonstrates the remarkable performance in terms of training safety and navigation success rate compared with other IL-based approaches.