Towards Infrastructure-Free Autonomous Robot Navigation: Machine Learning-Based Stereo Vision vs. Motion Capture
Session Number
CMPS 31
Advisor(s)
Daniel Torres and Dr. Pranav Bhounsule, University of Illinois Chicago
Discipline
Computer Science
Start Date
17-4-2025 11:25 AM
End Date
17-4-2025 11:40 AM
Abstract
This research examines the effectiveness of machine learning-based stereo vision as an alternative to traditional motion capture (MoCap) systems for autonomous mobile robot (AMR) navigation. Using the Unitree Go1 quadruped robot, navigation accuracy and efficiency are assessed in environments with varying obstacle densities. MoCap provides highly precise localization data, serving as a baseline for comparison, while the stereo vision system employs depth cameras to classify obstacles and generate real-time path adjustments. The stereo vision pipeline integrates disparity map computation, depth estimation, and obstacle detection to construct a dynamic 3D representation of the environment. By analyzing factors such as navigation precision, computational efficiency, and adaptability to dynamic environments, this study evaluates the feasibility of stereo vision for autonomous navigation without reliance on external tracking infrastructure, as required in MoCap. Preliminary findings indicate that stereo vision-based navigation can successfully handle complex, unstructured environments, demonstrating its potential for broader applications in robotics. The project contributes to advancements in vision- based navigation and pathfinding technologies, offering a scalable and infrastructure-free alternative for autonomous systems operating in real-world conditions.
Towards Infrastructure-Free Autonomous Robot Navigation: Machine Learning-Based Stereo Vision vs. Motion Capture
This research examines the effectiveness of machine learning-based stereo vision as an alternative to traditional motion capture (MoCap) systems for autonomous mobile robot (AMR) navigation. Using the Unitree Go1 quadruped robot, navigation accuracy and efficiency are assessed in environments with varying obstacle densities. MoCap provides highly precise localization data, serving as a baseline for comparison, while the stereo vision system employs depth cameras to classify obstacles and generate real-time path adjustments. The stereo vision pipeline integrates disparity map computation, depth estimation, and obstacle detection to construct a dynamic 3D representation of the environment. By analyzing factors such as navigation precision, computational efficiency, and adaptability to dynamic environments, this study evaluates the feasibility of stereo vision for autonomous navigation without reliance on external tracking infrastructure, as required in MoCap. Preliminary findings indicate that stereo vision-based navigation can successfully handle complex, unstructured environments, demonstrating its potential for broader applications in robotics. The project contributes to advancements in vision- based navigation and pathfinding technologies, offering a scalable and infrastructure-free alternative for autonomous systems operating in real-world conditions.