VPNet: Vision and Planning Network for Robotic Navigation

Robotic navigation like human drivers is a key capability for autonomous driving in real, dynamic urban environments. How can a delivery robot reliably navigate to the destination of a new office building with a minimum of priori information? Traditional autonomous driving methods are realized by multiple components such as perception, planning and control. Due to complex assumptions and interdependencies, these methods are difficult to be generalized to different scenarios. More recently, an end-to-end driven approach has emerged and is suitable for new environments by directly learning and exporting the data provided. However, many existing approaches neglect to examine the confidence level of driving behavior and the ability to recover from driving errors. This paper introduces a two-level hierarchical approach, called vision and planning network (VPNet), which integrates vision and path planning. The experimental results suggest that under various weather and lighting conditions, our network can reliably generate trajectories in different urban environments.