Mobile Robot Evaluating Traversability Based on Self-Supervised Learning

Mobile Robot Evaluating Traversability via Self-Supervised Learning of Elevation Map Features

This technology is a self-supervised mobile robot that converts 3D point cloud data into grid-based elevation maps, extracts multiple terrain features, and generates an AI model to determine traversability through a self-learning algorithm.

Existing manual labeling methods are costly, simulation data often differs from real-world environments, and simple threshold-based rules struggle to provide precise traversability assessments in complex urban settings.

This technology proposes a method that initializes positive samples from previous driving trajectories and negative samples from grids exceeding thresholds, then iteratively refines the model by reclassifying data based on the classifier's inference probability. This allows the model to improve its accuracy autonomously without human manual labeling. It can be applied to outdoor delivery and patrol robots, providing an economical solution that adapts to new environments without the need for separate data collection.

‍

‍

Key Features:

• Elevation map generation unit that creates grid-based elevation maps using LiDAR point cloud data
• Feature extraction unit that extracts multiple types of feature values, such as slope and roughness, for each grid from the generated elevation map
• Dataset generation unit that creates labeled and unlabeled datasets based on feature values for labeling
• Self-learning unit that generates an AI model for traversability assessment through self-supervised learning using the two datasets