Introduction
Animal guide dogs provide invaluable assistance to the visually impaired, yet their high training costs (exceeding $50,000) and long waiting periods mean only about 2% of blind individuals in the U.S. can access them. Previous robotic solutions often fail to replicate the nuanced human-animal team dynamic, treating navigation as a purely autonomous task rather than a collaborative effort. This research introduces a quadruped guide dog robot designed to bridge this gap by focusing on haptic feedback and handler-aware path planning.
Methodology
The system development began with qualitative research involving interviews with handlers and trainers to understand the "unit" philosophy: the handler manages high-level cognitive mapping (knowing the destination), while the dog handles local obstacle avoidance.
Hardware and Haptic Interface
The platform utilizes a Unitree Go1 quadruped equipped with a custom rigid, adjustable harness. Unlike soft leashes, a rigid connection provides instantaneous haptic feedback regarding the robot's speed, heading, and sudden stops. To calibrate the robot's motor controllers for natural interaction, researchers measured pulling forces from real guide dogs during blindfold walking sessions.
Semantic-Aware Path Planning
The robot employs two RGB cameras and a MobileNetV2-based neural network for real-time semantic segmentation at 20 Hz. This allows the robot to distinguish traversable floor space from obstacles without the computational overhead of 3D LiDAR SLAM.
A key innovation is the "Collaborative Trajectory Roll-out." The planner evaluates 40 potential paths using "cost boxes" that account for the physical width of the handler standing beside the robot. If a path is clear for the robot but would cause the handler to collide with an obstacle, the robot proactively adjusts its trajectory to clear the way for the human.
Results and Discussion
In indoor testing across 105m and 90m paths, the robot successfully guided blindfolded users at speeds of 0.7 m/s, matching natural human walking gaits. The system demonstrated robustness in narrow hallways and successfully navigated around dynamic obstacles like pedestrians and construction zones without requiring the handler to slow down.
While effective indoors, future work will address the complexities of outdoor environments, including variable lighting and curb detection. Additionally, while the quadruped is physically capable of climbing stairs, integrating vertical navigation into the collaborative framework remains a priority for achieving full animal-level utility.
Summary updated