Using 3D Convolutional Neural Networks for Tactile Object Recognition with Robotic Palpation.

Area of Science:

• Robotics
• Artificial Intelligence
• Sensor Technology

Background:

• Robotic manipulation requires sophisticated perception systems.
• Current tactile sensing methods often lack detailed information about object properties.
• Active exploration is crucial for robots to gather comprehensive sensory data.

Purpose of the Study:

• To present a novel active tactile perception method for robots.
• To leverage 3D neural networks for enhanced object classification using tactile data.
• To explore the utility of 3D tactile tensors derived from robotic palpation.

Main Methods:

• Developed a robotic gripper with a high-resolution tactile sensor.
• Implemented a haptic exploratory procedure involving robotic palpation and varying grasp forces.
• Represented tactile information as 3D tactile tensors capturing pressure variations.
• Utilized a 3D Convolutional Neural Network (3D CNN), termed 3D TactNet, for object classification.

Main Results:

• The proposed method effectively captures both external shape and internal features of objects.
• 3D TactNet demonstrated superior performance in object recognition compared to other methods.
• The 3D CNN approach achieved higher recognition rates with reduced training data requirements.

Conclusions:

• Active tactile perception using 3D neural networks offers a powerful approach for robotic object recognition.
• The 3D tactile tensor representation effectively encodes rich information for tactile sensing.
• This method advances robotic capabilities in understanding and interacting with objects through touch.