Deep Reinforcement Learning for Safe Local Planning of a Ground Vehicle in Unknown Rough Terrain

Safe unmanned ground vehicle navigation in unknown rough terrain is crucial for various tasks such as exploration, search and rescue and agriculture. Offline global planning is often not possible when operating in harsh, unknown environments, and therefore, online local planning must be used. Most online rough terrain local planners require heavy computational resources, used for optimal trajectory searching and estimating vehicle orientation in positions within the range of the sensors. In this work, we present a deep reinforcement learning approach for local planning in unknown rough terrain with zero-range to local-range sensing, achieving superior results compared to potential fields or local motion planning search spaces methods. Our approach includes reward shaping which provides a dense reward signal. We incorporate self-attention modules into our deep reinforcement learning architecture in order to increase the explainability of the learnt policy. The attention modules provide insight regarding the relative importance of sensed inputs during training and planning. We extend and validate our approach in a dynamic simulation, demonstrating successful safe local planning in environments with a continuous terrain and a variety of discrete obstacles. By adding the geometric transformation between two successive timesteps and the corresponding action as inputs, our architecture is able to navigate on surfaces with different levels of friction. Reinforcement learning, autonomous vehicle navigation, motion and path planning.