End-to-End Learning for Visual Navigation of Forest Environments

Off-road navigation in forest environments is a challenging problem field robotics. Rovers are required to infer their traversability over priori unknown and dynamically changing terrain using noisy onboard sensors. The compounded for small-sized rovers, such as that of swarm. Their size-proportional low-viewpoint affords them restricted view navigation, which may be partially occluded by vegetation. Hand-crafted features, typically employed analysis, often brittle fail discriminate obstacles varying lighting weather conditions. We design low-cost system tailored rovers self-learned features. MobileNet-V1 MobileNet-V2 models, trained following an end-to-end learning approach, deployed steer mobile platform, with human-in-the-loop, towards traversable paths while avoiding obstacles. Receiving 128 × 96 pixel RGB image from monocular camera input, the algorithm running on Raspberry Pi 4, exhibited robustness motion blur, low lighting, shadows high-contrast It was able successfully navigate total 3 km real-world comprising shrubs, dense bushes, tall grass, fallen branches, tree trunks, standing trees, five different conditions four times day.