Object-of-Interest Perception in a Reconfigurable Rolling-Crawling Robot

Area of Science:

• Robotics
• Computer Vision
• Biomimetics

Background:

• Reconfigurable robots require sophisticated perception systems to adapt locomotion to diverse environments.
• The huntsman spider (Cebrenus Rechenburgi) exhibits versatile locomotion, serving as inspiration for robotic development.

Purpose of the Study:

• To develop and validate an object-of-interest perception system for a reconfigurable rolling-crawling robot.
• To enable the robot to identify appropriate locomotion modes (rolling or crawling) based on environmental perception.

Main Methods:

• A reconfigurable robot, Scorpio, with rolling and crawling capabilities was utilized.
• A Pyramid Scene Parsing Network (PSPNet) was trained on a custom dataset distinguishing unobstructed and obstructed paths.
• The framework was evaluated using mean Intersection over Union (mIOU) in offline and real-time tests.

Main Results:

• The trained PSPNet achieved mIOU scores of 72.28% (offline) and 70.63% (online).
• The proposed framework outperformed semantic segmentation benchmarks (HRNet, Deeplabv3) in mIOU and speed.
• The system demonstrated stable robot maneuverability and accurate locomotion mode selection in complex pathways.

Conclusions:

• The developed perception framework effectively guides locomotion mode selection in reconfigurable robots.
• Biomimetic approaches, inspired by spiders, can lead to advanced robotic capabilities.
• The system offers enhanced accuracy and efficiency for robots navigating varied terrains.