Advanced Obstacle Avoidance in Walking Aids: Overcoming the Challenges of Low-Lying Obstacles with Deep Learning

An autonomous walking aid is a type of Autonomous Mobile Robots (AMRs) designed to navigate and perform tasks independently. In this paper, we introduce an innovative obstacle avoidance approach for walking aids, with a particular emphasis on avoiding non-stereotypical obstacles such as water puddles. Traditional methods relying on LiDAR, depth sensors, or cameras often struggle with low-lying obstacles due to challenges in detecting reflective surfaces and interpreting complex visual data. To overcome these limitations, we developed two deep learning-based network architectures: the Single-Image Action Network (SIAN) and the Sequence-to-Action Network with Feedback Network (SAN-FN). The first architecture processes a single image to generate an action sequence for obstacle avoidance, while the second architecture takes an image sequence and feedback from wheel velocities to infer real-time action responses. In addition, we employed Gazebo to create a 3D simulation environment and gather training data for deep learning networks, significantly reducing the cost and effort associated with data collection. The first approach achieved a 100<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup> success rate in simulated obstacle avoidance tests involving puddles, while the second approach demonstrated a 76% success rate. This study highlights the significant potential of deep learning in enhancing the obstacle avoidance capabilities of walking aids, particularly in the context of challenging obstacles like water puddles.