To improve robot relocalization accuracy in both static and dynamic environments, we introduce a novel network, FusedNet, which incorporates a cross-attention to fuse global and local image features for end-to-end relocalization. This approach relies solely on a monocular camera sensor that is fixed on the mobile robot, and directly predicts the absolute pose from the input RGB image. Additionally, we have collected a mobile robot relocalization dataset, termed moBotReloc, consisting of dynamic large-scale scenes, using the Unity 3D simulation platform and a real mobile robot. Through extensive experiments on 7Scenes and moBotReloc, we demonstrate that FusedNet achieves significant accuracy in 6-DoF camera relocalization in static scenes, and exhibits superior relocalization performance in dynamic large-scale scenes for mobile robot applications, outperforming existing end-to-end methods that rely solely on a single global or local feature.
1 January 2017 in JOURNAL OF SHENZHEN UNIVERSITY SCIENCE AND ENGINEERING
Robot navigation in densely crowd environments poses significant challenges due to the crucial need for pedestrian safety and comfort. Interpreting and predicting pedestrian actions is particularly difficult for robots, resulting in suboptimal performance by traditional navigation methods. Recent studies have shown that deep reinforcement learning can address these issues effectively; however, most research focuses solely on current pedestrian positions. This work introduces a novel contrastive sampling approach tailored for robot navigation, which leverages historical data of pedestrian locations to anticipate future movement sequences and generate negative samples. Named Social Predictive Negative Contrastive Estimation (SP-NCE), our method enhances safety and efficiency during robotic navigation, enabling robots to better avoid pedestrians while successfully reaching their destinations. Notably, SP-NCE exhibits superior data learning capabilities in imitation learning, especially with smaller datasets. For instance, when trained and tested on an expert demonstration dataset of 1k size, our approach achieves a 7% higher navigation success rate compared to existing contrastive learning method Social-NCE, and concurrently reduces collision probabilities with pedestrians by 7%.