بهبود برنامه‌ريزي ماموريت براي وسايل‌نقليه زميني بدون سرنشين در زمين‌هاي ناهموار و ناشناخته

Deep Reinforcement Learning , Unmanned Ground Vehicles , Multi-head Attention , Cross-Attention , Rough an‎d Unknown Terrains

Enhancing Mission Planning for Unmanned Ground Vehicles in Rough an‎d unknown Terrains

مهندسي معماري كامپيوتر

Autonomous navigation of autonomous ground vehicles, as a key area of mobile robotics an‎d intelligent robotics, is considered one of the fundamental challenges in autonomous systems due to the complexity of terrain geometry, the presence of sensory uncertainties, the variability of environmental conditions, an‎d the lack of regular structure. In such environments, classical map-based path planning methods, as well as many conventional deep reinforcement learning approaches, suffer from insufficient stability an‎d efficiency owing to their inability to effectively integrate multimodal sensory data, their high sensitivity to noise, an‎d their limited capability in modeling temporal dependencies. In this research, a distributional deep reinforcement learning framework based on Rainbow DQN is proposed for safe an‎d stable local navigation of robotic autonomous ground vehicles in rough terrains. The proposed architecture is designed to extract rich an‎d meaningful representations from heterogeneous sensory inputs, including LiDAR scans, inertial measurements, relative goal angle, relative displacement, an‎d a local elevation map. In this context, a Multi-Head Attention mechanism is employed to model interactions among different sensory tokens, enabling the robotic agent to adaptively focus on the most informative features relevant to decision-making. To enhance spatial perception of the terrain, a spatial attention module is applied to the environmental elevation map, allowing the navigation system to identify hazardous regions an‎d critical terrain irregularities. Furthermore, through the use of non-spatial attention, the relative importance of numerical an‎d sensory inputs is dynamically adjusted. For effective fusion of spatial an‎d non-spatial information, a Cross-Attention mechanism is designed between the elevation map representation an‎d other sensory inputs, which improves decision-making coherence an‎d reduces the system’s reliance on noisy observations. To model temporal dependencies an‎d address the problem of partial observability in robotic navigation, the outputs of the attention modules are fed into an LSTM network with a limited temporal window. Moreover, by leveraging C51-based distributional learning an‎d prioritized experience replay, learning stability an‎d the accuracy of expected return estimation are significantly improved. The proposed method is eva‎luated in the Gazebo simulation environment within the ROS framework, an‎d the results demonstrate that the presented architecture outperforms the baseline method an‎d attention-free variants in terms of learning stability, reduced collision rates, improved path quality, an‎d enhanced navigation safety of autonomous ground vehicles. In addition, qualitative analysis of spatial an‎d non-spatial attention maps increases the interpretability of the robotic agent’s decisions an‎d indicates that the model can meaningfully focus on critical terrain features. The findings of this research suggest that integrating multi-head an‎d cross-attention mechanisms into deep reinforcement learning can play an effective role in the development of intelligent navigation systems for off-road environments.

2

159084