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

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