Hybrid robot navigation: Integrating monocular depth estimation and visual odometry for efficient navigation on low-resource hardware

Abstract

Robotic navigation is a complex task requiring accurate localization, environmental perception, path planning, and control of actuators. Traditional navigation systems rely on pre-built maps or map building techniques such as simultaneous localization and mapping (SLAM). However, these approaches unnecessarily map the entire environment, including all objects and obstacles, making them computationally intensive and slow, particularly on resource-constrained devices. While mapless navigation methods address some of these issues they are often too impulse-based, lacking reliance on planning. Recent advances in deep learning have provided solutions to many navigation paradigms. In particular, Monocular Depth Estimation (MDE) enables the use of a single camera for depth estimation, offering a cost-effective alternative to selective mapping. While these approaches effectively address navigation challenges, they still face issues related to scalability and computational efficiency. This paper proposes a novel hybrid approach to robot navigation that combines map-building techniques from classical visual odometry (VO) with maples techniques that uses deep learning-based MDE. The system employs an object detection model to identify target locations and estimate travel distances, while the MiDaS MDE model provides relative depth to detect the nearest obstacle and navigable gaps after image segmentation removes floor and ceiling areas, enhancing the robot's perception of free spaces. Wheel odometry (WO) and VO determine the robot's position and its metric distance from detected nearest obstacle. An instantaneous Grid map is then formed with robot’s position, navigable gap, nearest obstacle and the goal location. Path planning is conducted using a modified A-star (A*) algorithm, followed by path execution with a Proportional Integral Derivative (PID) controller. The system’s performance is evaluated at both the modular level and the final system level using various metrics, such as Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and inference time for depth estimation, and navigation success rate across different robot speeds for final navigation performance. Additionally, a Friedman statistical test is conducted to validate the results. Experimental results show that the proposed approach reduces memory and computational demands, enabling real-world navigation on low-resource hardware. To our knowledge, this is the first integration of MDE-based mapless navigation with VO-based map-building, presenting a novel direction for research.