MonoNav*(right) collides because of erroneous 3D reconstruction and a simplistic planning approach whereas ours(left) successfully navigates to the goal position whilst avoiding obstacles.
*Simon, N., Majumdar, A. (2024). MonoNav: MAV Navigation via Monocular Depth Estimation and Reconstruction. In: Ang Jr, M.H., Khatib, O. (eds) Experimental Robotics. ISER 2023. Springer Proceedings in Advanced Robotics, vol 30. Springer, Cham. https://doi.org/10.1007/978-3-031-63596-0_37
While NoMaD*(right)-predicted trajectories tend to steer away from the obstacle, they lack sufficient precision, resulting in a collision whereas our method(left) explicitly incorporates depth uncertainty into risk-aware trajectory optimization, enabling successful navigation even in complex scenes.
*A. Sridhar, D. Shah, C. Glossop and S. Levine, "NoMaD: Goal Masked Diffusion Policies for Navigation and Exploration," 2024 IEEE International Conference on Robotics and Automation (ICRA), Yokohama, Japan, 2024, pp. 63-70, doi: 10.1109/ICRA57147.2024.10610665.
CARE*(right) applies collision correction on NoMaD trajectories. Noise in the estimated depth/pcd results in erroneous correction and leads to collision in narrow spaces
*Kim, J., Sim, J., Kim, W., Sycara, K. P., and Nam, C., "CARE: Enhancing Safety of Visual Navigation through Collision Avoidance via Repulsive Estimation.", In Proceedings of the 9th Annual Conference on Robot Learning (CoRL), 2025.
ROS Navigation Stack(ROSNAV)(right) uses global planning followed by Dynamic Window Approach (DWA) for local control on costmaps built from estimated point clouds. This classical approach fails mainly due to its inability to handle noise in the estimated depth. Offsets between estimated and true point clouds lead to false free space and collisions(top right), while noisy estimates inflate costmaps near goals, causing planning failures(bottom right).