Single-Grasp Object Classification and Feature Extraction with Simple Robot Hands and Tactile Sensors.

Area of Science:

• Robotics
• Artificial Intelligence
• Sensor Technology

Background:

• Traditional robotic object identification relies on complex grippers and exploratory procedures (EPs).
• Recent findings suggest tactile property inference is possible from brief, non-exploratory motions, termed 'haptic glance'.

Purpose of the Study:

• To implement tactile object identification and feature extraction using data from a single, unplanned grasp.
• To evaluate cooperating machine learning and parametric schemes for object property estimation.

Main Methods:

• Utilized a simple, underactuated robot hand with inexpensive barometric pressure sensors.
• Implemented two cooperating schemes: random forests (machine learning) and parametric property estimation.
• Data included actuator positions and force sensor values from a single grasp.

Main Results:

• The developed schemes achieved tactile object identification without requiring object exploration, re-grasping, or force modulation.
• Collaborative operation of the schemes synergistically improved overall identification results.
• The method demonstrated effectiveness for arbitrary object start positions and orientations.

Conclusions:

• The proposed approach enables practical robotic grasping by integrating tactile identification without adding time or manipulation overhead.
• This technique offers a more efficient alternative to classical robotic tactile identification methods.