Dynamic Path Planning, Mapping, and Navigation for Autonomous Non-Holonomic Differential Drive Robot : Assessing scalable and predictable navigation on a custom low-cost mobile robot platform

Sammanfattning

This thesis evaluates how well the ROS2 NavFn/DWA navigation stack enables a custom-built, low-cost non-holonomic differential drive robot to navigate a dynamic indoor environment. While impressive humanoid systems have been demonstrated in recent years, they remain costly and complex to operate. This does not diminish their potential, but it underscores the importance of also advancing non-humanoid robotic designs, which can offer simpler, more reliable, and more accessible solutions for many real-world tasks. Tasks such as loading a coffee brewer or dishwasher continue to challenge most budget-friendly autonomous systems. This project aims to bridge this gap by focusing on robot autonomy on a mid-sized robot (around 1 m3 in size), using real-time 2D LiDAR sensor data for navigation, without attempting to replicate human anatomy or relying on AI for decision-making. Importantly, this approach avoids the black-box nature that real-time artificial intelligence decision-making might entail, ensuring a transparent system dealing with uncertainty in a predictable way. The system was thoroughly tested and evaluated in both simulated and real-world environments. Hardware and software integration included establishing communication between microcontrollers and the onboard computer, enabling motor control and sensor data processing. Extensive testing validated motor encoders, control systems, and LiDAR functionality, providing a solid foundation for reliable navigation. The main navigation test involved three waypoints, designed to assess path efficiency and adaptability in two scenarios: when the pre-scanned 2D map matches the environment, and when obstacle avoidance is required to navigate around new obstacles (moved table and the introduction of a person). Results revealed the robot’s ability to navigate in static, pre-scanned environments and adapt to dynamic obstacles that were not previously scanned. However, the tests also highlighted the trade-off between efficiency and adaptability that can arise, as well as the need for further parameter tuning to improve efficiency, particularly with respect to angular velocity overshooting. The navigation stack used NavFn as the global planner, set to compute the shortest path using Dijkstra’s algorithm based on a costmap generated from the pre-scanned map, and the Dynamic Window Approach (DWA) as the local planner, demonstrated effective real-time adaptation while highliting the challenge of balancing efficiency with adaptability. The evaluation is based on a single real-world run after simulation tuning so the findings are not broadly generalizable but reflect performance under the tested conditions after simulation. Future work will focus on hyperparameter optimization using machine learning, leveraging both the initially planned path and incremental updates during navigation. This work provides a robust starting point for future efforts, paving the way for automating routine tasks in home and office environments, improving quality of life, and promoting sustainability through energy-efficient navigation.