21 October 2025 in Measurement Science and Technology
With the widespread deployment of mobile robots in complex task scenarios, enhancing their ability to achieve efficient autonomous navigation in unknown environments has become a major challenge in the field of intelligent robotics. Traditional deep reinforcement learning methods often struggle to capture and model long-term spatio-temporal dependencies, resulting in unstable navigation policies and poor generalization in unseen environments. To address these issues, this paper proposes a vision-based navigation framework that integrates a spatio-temporal adaptive convolutional attention model (SACAM) and an asynchronous meta-variational policy algorithm (AMPA). SACAM enhances the agent’s ability to capture long-term dependencies through self-supervised learning, improving adaptability in complex settings. Meanwhile, AMPA employs an asynchronous update mechanism and a meta-variational approach to handle policy uncertainty, enabling robust decision-making and rapid adaptation. Experimental results in the AI2-THOR simulation environment demonstrate that the proposed method outperforms several advanced baselines in terms of success rate, path efficiency, and generalization capability, showcasing its robust performance and adaptability in challenging navigation tasks.
30 December 2025 in Measurement Science and Technology
Multimodal perception capabilities are crucial for enhancing autonomous navigation performance in mobile robots based on deep reinforcement learning (DRL). However, existing multimodal perception-based DRL navigation methods face two main challenges. One is insufficient environmental representation, the other challenge is low sample utilization efficiency during experience replay. To address these issues, this study proposes the Transformer-convolutional neural network (CNN) feature fusion with ratio-weighted hindsight experience replay (TCFusion-RWHER) framework. This framework employs TCFusion to fuse global spatial correlations and local geometric features from depth images, then performs multimodal fusion with CNN-extracted pseudo-LiDAR data to enhance environmental representation. The RWHER mechanism further improves sample efficiency and accelerates policy convergence. Experimental results indicate that the proposed method achieves navigation success rates of 97.2%, 86.6%, and 82.3% in simple, complex, and dynamic scenarios, respectively, improving by up to 40.1 percentage points over baseline methods in dynamic environments with a zero timeout rate in simple scenarios. Furthermore, the framework successfully transfers from simulation to real-world scenarios, validating its excellent generalization capability and robustness.