Document Type
Poster
Start Date
6-3-2026 12:30 PM
Description
LiDAR sensors are widely used in autonomous robots and vehicles for mapping, navigation, and obstacle avoidance. Because mapping is a foundational capability in nearly all robotic systems, it must be both accurate and reliable. However, LiDAR frequently fails in environments containing mirrors or highly reflective surfaces, where distorted returns corrupt maps and generate false obstacles. Current solutions typically rely on additional sensors and computationally expensive sensor-fusion pipelines, which is costly and difficult to scale for compact robots with limited onboard resources. This paper presents a machine-learning–based LiDAR mirror detection model that identifies reflective surfaces directly from LiDAR data, reducing the need for auxiliary sensors and complex post-processing. Unlike earlier heuristic-based approaches that attempt to smooth reflections, the proposed method learns to robustly classify mirror-induced artifacts while remaining computationally efficient. The model is deployed and evaluated on a resource-constrained embedded platform, demonstrating the feasibility of real-time mirror detection on microcontroller-class hardware. By enabling more reliable single-sensor mapping, the approach lowers system complexity and frees computational resources for faster mapping and additional autonomous capabilities.
Included in
LiDAR Mirror Detection Model for Autonomous Robotics
LiDAR sensors are widely used in autonomous robots and vehicles for mapping, navigation, and obstacle avoidance. Because mapping is a foundational capability in nearly all robotic systems, it must be both accurate and reliable. However, LiDAR frequently fails in environments containing mirrors or highly reflective surfaces, where distorted returns corrupt maps and generate false obstacles. Current solutions typically rely on additional sensors and computationally expensive sensor-fusion pipelines, which is costly and difficult to scale for compact robots with limited onboard resources. This paper presents a machine-learning–based LiDAR mirror detection model that identifies reflective surfaces directly from LiDAR data, reducing the need for auxiliary sensors and complex post-processing. Unlike earlier heuristic-based approaches that attempt to smooth reflections, the proposed method learns to robustly classify mirror-induced artifacts while remaining computationally efficient. The model is deployed and evaluated on a resource-constrained embedded platform, demonstrating the feasibility of real-time mirror detection on microcontroller-class hardware. By enabling more reliable single-sensor mapping, the approach lowers system complexity and frees computational resources for faster mapping and additional autonomous capabilities.