Abstract
While robotic guide systems offer a scalable alternative to guide dogs, existing solutions often prioritize navigation over human comfort, leading to jerky movements and user anxiety. This paper presents a novel quadrupedal guide robot system designed to treat the human user as an active participant. By combining a custom-built controllable elastic traction device with a force-based motion planner, the system ensures smooth, predictable guidance. Experimental results demonstrate a significant improvement in the Relative Comfort Index (RCI) and a total elimination of uncomfortable force spikes compared to traditional inelastic leash systems.
Introduction
Guide dogs are highly effective for the visually impaired but are limited by high training costs and long lead times. Robotic alternatives, ranging from wheeled platforms to quadrupedal robots, have emerged to fill this gap. However, a critical bottleneck remains: human comfort.
Most existing robotic guides use rigid handles or inelastic leashes. Rigid handles transmit every vibration and sudden mechanical stop directly to the user, while inelastic leashes often oscillate between being slack and suddenly taut. These systems typically ignore the human’s physiological response, treating the person as a passive object to be towed. This leads to imprecise guidance and physical discomfort. This research seeks to bridge the gap between robotic agility and human-centric interaction by focusing on the force interface between the robot and the user.
System Architecture and Hardware Design
The proposed system is deployed on a Unitree Laikago quadruped robot, augmented with specialized sensors and a novel traction mechanism.
Controllable Traction Device
Unlike a simple leash, the system features a custom device consisting of an elastic rope connected to a DC motor, a string reel, and a force sensor. By adjusting the length of the string based on Hooke’s Law ($F = K\Delta l$), the robot can precisely control the magnitude of the traction force applied to the user. The elastic rope acts as a physical low-pass filter, absorbing high-frequency jitters from the robot’s gait before they reach the human hand.
Sensing and Localization
The robot utilizes a 2D LiDAR for environmental mapping and localization. To track the human user, an RGB-D camera is mounted on a 1-degree-of-freedom gimbal, providing a stable field of view regardless of the robot's orientation. An Unscented Kalman Filter (UKF) fuses these inputs to estimate the joint state of the robot and the human, handling the inherent uncertainty of human movement.
Planning and Control Framework
The core of the system is a hierarchical control architecture that prioritizes the interaction force as the primary interface for guidance.
Force-Based Human Motion Model
Human gait is modeled as a discrete state machine (standing vs. walking). The researchers utilized minimum-jerk theory to predict transitions; the robot understands that a human requires a specific "ramp-up" in force to start walking comfortably, rather than a sudden yank.
Two-Stage MPC Planning
- Human Motion Planner: A Model Predictive Control (MPC) framework plans the optimal force (magnitude and direction) required to guide the human along a collision-free path.
- Robot Motion Planner: A second MPC generates specific velocity commands for the robot to maintain the desired leash tension and direction, effectively decoupling the human-centric guidance from the robot's mechanical constraints.
Key Results and Evaluation
The system was tested in narrow indoor corridors against three baselines: an inelastic rope, an elastic rope without active control, and the proposed "Planned" approach.
- Relative Comfort Index (RCI): The proposed system achieved a perfect score of 1.0, significantly outperforming the inelastic rope baseline (0.1082).
- Force Spikes: The duration of force exceeding the comfort threshold ($t_{F > F_{max}}$) was reduced to 0 seconds in the proposed system, compared to over 17 seconds for the inelastic setup.
- Fluency: The number of state changes ($N_{ch}$)—representing instances where the human had to stop or start due to confusing cues—dropped from 20 in the inelastic setup to just 6 with the proposed planner.
Conclusion
By treating force as a planned interaction rather than a byproduct of movement, this research demonstrates that quadrupedal robots can provide a natural and comfortable guidance experience. Future work will focus on expanding the system's capabilities to handle dynamic obstacles in crowded spaces and navigating complex outdoor environments and multi-floor buildings.
Summary updated