Practical implementation of Depth Anything V2 as a LiDAR alternative in robotics navigation

This paper presents a systematic evaluation of monocular depth estimation (MDE) using Depth Anything V2 as an alternative to LiDAR for real-time obstacle perception in mobile robot navigation. We investigate the performance of an MDE-based perception pipeline integrated into a Nav2 navigation stack on a differential-drive robot, comparing it to a LiDAR baseline across three structured indoor courses. Qualitative analyses reveal that MDE offers superior spatial resolution, successfully capturing small obstacles often missed by sparsely sampled LiDAR beams. However, challenges such as localisation inaccuracies at long range and edge-induced artefacts are observed, particularly when the camera is forward-facing. These issues are significantly mitigated when using a high camera view angle, which reduces artefact streaking and improves localisation accuracy. Quantitative results show that the MDE system achieves comparable navigation performance to LiDAR, with similar travel distances, path consistency, and replanning deviation, despite greater variability in obstacle localization. The robot successfully achieved collision-free operation in all trials, demonstrating that MDE is a viable modality for local planning in structured environments. This work highlights the potential of vision-based depth estimation to complement or replace LiDAR in resource-constrained robotics applications and identifies key directions for improving its reliability and deployment efficiency.