PhD Dissertation Proposal: Hochul Hwang, Human-Centered Quadruped Robot for Navigation Assistance for Blind and Low-Vision Individuals
Content
Speaker:
Abstract:
Independent mobility is a fundamental challenge for blind and low-vision (BLV) individuals. While animal guide dogs provide highly effective mobility assistance, their availability is severely limited due to extensive training requirements and high costs. With the vision of providing a scalable alternative solution using robots, nearly half a century of efforts have been made to create navigation assistant robots for BLV individuals. In particular, recent progress in quadruped robots has attracted attention as a potential counterpart to animal guide dogs; however, deploying these systems in complex human environment requires seamless integration of user-centered design, robust perception, and reliable navigation. This thesis presents a comprehensive framework for developing an autonomous quadruped guide dog robot capable of safely navigating real-world environments while aligning with the practical needs of BLV users.
The research is structured around five core contributions:
1. User-Centered Foundations: An in-depth evaluation of handler-guide dog interactions to establish the fundamental design requirements and behavioral expectations for robotic mobility aids.
2. System Integration & Locomotion: The development of a foundational legged robotic architecture capable of safety-critical tasks, including dynamic obstacle avoidance, silent walking, and stable stair climbing, validated through interactive studies with BLV participants.
3. Vision-only Robust Route Following: The implementation of an outdoor visual teach-and-repeat framework, enabling the robot to reliably traverse complex pedestrian routes from demonstrations without GPS, LiDAR, or dense maps, evaluated by real-world navigation trials with BLV users and a trainer.
4. Safety-Critical Perception: The integration of advanced visual reasoning for high-stakes environments, including the evaluation of Vision-Language Models for safe street crossing and the deployment of robust tactile paving segmentation models trained on large-scale synthetic datasets.
5. (Proposed work) Responsive Path Deviation and Recovery: A proposal for extending the visual teach-and-repeat framework to support flexible, user-driven interruptions. This section outlines methods to enable the robot to handle intermittent requests and robustly re-engage the navigation path, allowing for dynamic deviations without compromising route fidelity.
Ultimately, this thesis bridges the gap between human-robot interaction and embodied AI, demonstrating a viable path toward deploying reliable, intelligent guide robots.
Advisor:
Donghyun Kim