Target-driven Navigation of a Mobile Robot using an End-to-end Deep Learning Approach

The capabilities demonstrated by deep neural networks have positioned them as viable alternatives to traditional approaches in the fields of autonomous driving and robotics. Classical local motion planning algorithms often involve multiple steps of data preprocessing and can be inherently complex to manage. In this paper, A target-oriented, mapless, end-to-end navigation algorithm is introduced that learns the navigation policy through demonstrations provided by an expert driver. The proposed navigation model relies on laser data and relative target positions to generate the necessary velocity commands, employing a lightweight CNN/DNN network to achieve the desired goal. The experiments demonstrates the proposed navigation model’s capability for real-time navigation within unexplored areas, efficiently avoiding obstacles, and successfully reaching determined goals. The comparison between the expert model and the deep planner indicates the model’s capacity to mimic the expert’s navigation policy. Additionally, the model is able to navigate in dynamic scenarios.