Real-Time Navigation of Unmanned Ground Vehicles in Complex Terrains With Enhanced Perception and Memory-Guided Strategies

Accurate navigation of unmanned ground vehicles (UGVs) across challenging outdoor terrains demands precise maneuvering amidst diverse surface characteristics, while ensuring collision avoidance. Conventional navigation methodologies often struggle in dynamic environments due to their reliance on pre-mapped data and limitations in real-time multi-sensory data assimilation. To this end, this study proposes a methodology that integrates a diverse array of sensory inputs, including pose data, images, and point clouds, to enhance UGVs' situational awareness and decision-making capabilities. Central to our innovation is the development of a multi-modal reinforcement learning framework, which enhances UGVs' perceptual and decision-making ability. This framework incorporates a lattice-based motion planning algorithm, meticulously calibrated to optimize action selection while respecting UGVs' kinematic constraints. Additionally, a novel dual-training paradigm is introduced, combining curriculum learning and modal separation techniques to address the complexities of multi-modal learning. A notable contribution is the strategic integration of Long Short-Term Memory (LSTM) algorithms, which mitigate information decay and preserve essential navigational strategies over extended periods. The fusion of advanced perception and memory-guided strategies establishes a new standard for autonomous UGV navigation across diverse and unpredictable terrains.