Using a Variable-Friction Robot Hand to Determine Proprioceptive Features for Object Classification During...

Area of Science:

• Robotics
• Machine Learning
• Haptics

Background:

• Within-hand manipulation (WIHM) involves complex gripping, sliding, and pivoting actions.
• Object re-orientation during WIHM generates rich haptic information through sensory organ interactions.
• Proprioceptive data from robotic actuators can be leveraged for object classification.

Purpose of the Study:

• To investigate the use of proprioceptive data from variable friction (VF) robotic fingers for object classification during rolling WIHM.
• To identify key features from actuator data that are most effective for distinguishing between different geometric objects.
• To evaluate the classification accuracy using a reduced set of important features compared to a comprehensive feature set.

Main Methods:

• Utilized VF robotic fingers to perform rolling WIHM on various objects.
• Recorded proprioceptive actuator data (position and load profiles) during manipulations.
• Computed 66 general features based on gradient changes in actuator data.
• Employed an Extra Trees classifier for object classification and feature importance ranking.

Main Results:

• Achieved 86% classification accuracy for distinguishing six geometric objects using only the six most important 'Key Features'.
• Using all 66 computed features resulted in a classification accuracy of 89.8%.
• The Extra Trees classifier effectively ranked feature importance, enabling the selection of a concise, high-performing feature set.

Conclusions:

• Proprioceptive data from robotic fingers, specifically actuator position and load profiles, can effectively classify objects during rolling within-hand manipulation.
• A reduced set of key features derived from this data can achieve high classification accuracy, simplifying the recognition process.
• This approach offers a promising method for object recognition in robotic systems without relying on tactile sensors.