Tactile Object Recognition for Humanoid Robots Using New Designed Piezoresistive Tactile Sensor and DCNN

Area of Science:

• Robotics
• Sensor Technology
• Artificial Intelligence

Background:

• Tactile sensor arrays are essential for humanoid robot physical interaction.
• Developing compact, high-resolution tactile sensors is key for advanced robotic applications.

Purpose of the Study:

• To develop a palm-size tactile sensor array for object recognition in humanoid robot hands.
• To evaluate the performance of deep convolutional neural networks (DCNNs) for tactile-based object recognition.

Main Methods:

• A 16x16 pixel tactile sensor array (56.0 mm × 56.0 mm) was fabricated using PCB technology and piezoresistive principles with a conductive polymer composite.
• Tactile images were enhanced using bicubic interpolation and used to train/test 19 DCNNs.
• A multimodal learning approach combining InceptionResNetV2 and XceptionNet was employed.

Main Results:

• The developed tactile sensor achieved high-resolution tactile imaging.
• InceptionResNetV2 demonstrated strong performance with 91.82% accuracy.
• The multimodal learning method achieved a peak recognition rate of 92.73%, further improving with object exploration.

Conclusions:

• The developed tactile sensor array is effective for object recognition in humanoid robots.
• Multimodal deep learning significantly enhances tactile recognition accuracy.
• Object exploration further refines recognition capabilities, paving the way for more sophisticated robotic interaction.