Temporal deep neural network for tactile sensing in artificial finger pulp skin

Area of Science:

• Robotics and Artificial Intelligence
• Materials Science and Engineering
• Biomedical Engineering

Background:

• Tactile perception is crucial for artificial finger dexterity in grasping and touch.
• Conventional methods struggle to utilize temporal data for accurate tactile demodulation.
• Precise sensing of deformation position and applied force is a key challenge.

Purpose of the Study:

• To develop an advanced tactile sensing system for artificial finger pulp skin.
• To improve the accuracy of tactile demodulation by leveraging temporal data correlations.
• To jointly reconstruct pressing position and force using a novel neural network architecture.

Main Methods:

• Utilized quasi-distributed Fiber Bragg Gratings (FBGs) integrated into artificial finger pulp.
• Developed a two-stage hybrid LSTM-Transformer neural network (TSH-LTNN) for data analysis.
• Trained the network on temporal data variations across three consecutive time steps.

Main Results:

• The TSH-LTNN achieved high accuracy in position (0.2331 mm MAE, R²=0.9971) and force (0.303 N MAE, R²=0.9829) prediction.
• Demonstrated significant improvements over Random Forest, with a 66.06% reduction in position MAE and 31.94% in force MAE.
• Confirmed precise, stable, and real-time pressure-state demodulation.

Conclusions:

• The proposed FBG-based tactile sensing system with TSH-LTNN significantly enhances tactile perception accuracy.
• The hybrid LSTM-Transformer network effectively captures both short-term and long-range temporal dependencies.
• The system shows strong potential for high-precision haptic feedback applications in robotics and beyond.