1380-GRM Traversabiliy-Aware Semantic Navigation for Quadruped Robots

Location

https://campus.kennesaw.edu/colleges-departments/ccse/events/computing-showcase/fa26-cday-program.php

Document Type

Event

Start Date

20-11-2026 2:00 PM

Description

Standard ROS 2 navigation stacks rely on geometric occupancy grids that treat all obstacles equally. As a result, low-profile objects such as cables, mats, or small indoor clutter are frequently flagged as impassable barriers, forcing unnecessary detours and severely underutilizing a quadruped robot’s physical capability to step over minor obstacles. To address this limitation, and building upon the Unitree Go2 Edu ROS 2 architecture, we integrated a modular, vision-based Semantic Traversability Classifier. This perception layer fuses live LiDAR height-bands with YOLO11s detections and a Vision-Language Model (VLM) to distinguish between solid barriers and steppable clutter. Evaluated in ROS 2 / Gazebo across 10 paired trials per mode, the system significantly improved foot clearance over low obstacles (53% vs. 26% baseline) with zero unsafe semantic labels. A costmap backstop also prevented 100% of collisions on camera-missed objects, validating the architecture for upcoming physical Go2 hardware testing.

This document is currently not available here.

Share

COinS
 
Nov 20th, 2:00 PM

1380-GRM Traversabiliy-Aware Semantic Navigation for Quadruped Robots

https://campus.kennesaw.edu/colleges-departments/ccse/events/computing-showcase/fa26-cday-program.php

Standard ROS 2 navigation stacks rely on geometric occupancy grids that treat all obstacles equally. As a result, low-profile objects such as cables, mats, or small indoor clutter are frequently flagged as impassable barriers, forcing unnecessary detours and severely underutilizing a quadruped robot’s physical capability to step over minor obstacles. To address this limitation, and building upon the Unitree Go2 Edu ROS 2 architecture, we integrated a modular, vision-based Semantic Traversability Classifier. This perception layer fuses live LiDAR height-bands with YOLO11s detections and a Vision-Language Model (VLM) to distinguish between solid barriers and steppable clutter. Evaluated in ROS 2 / Gazebo across 10 paired trials per mode, the system significantly improved foot clearance over low obstacles (53% vs. 26% baseline) with zero unsafe semantic labels. A costmap backstop also prevented 100% of collisions on camera-missed objects, validating the architecture for upcoming physical Go2 hardware testing.