Physics-Driven Learning Framework for Tomographic Tactile Sensing.

Area of Science:

• Biomedical Engineering
• Electrical Engineering
• Computer Science

Background:

• Electrical impedance tomography (EIT) offers potential for large-area tactile sensing but suffers from artifacts due to its nonlinear inverse problem.
• Existing deep learning methods often lack physical plausibility and generalization capabilities.

Purpose of the Study:

• To develop a physics-driven deep reconstruction framework (PhyDNN) for EIT tactile sensing.
• To improve the accuracy, physical plausibility, and generalization of EIT reconstructions by embedding the forward model into the learning objective.

Main Methods:

• Developed PhyDNN, a framework integrating the EIT forward model into the deep learning objective.
• Created a differentiable forward-operator network for efficient, physics-guided training.
• Validated the method through extensive simulations and real-world experiments on a 16-electrode soft sensor.

Main Results:

• PhyDNN demonstrated superior performance compared to NOSER, TV, and standard deep neural networks (DNNs).
• The framework achieved higher accuracy in reconstructing contact shape and location.
• Reconstructions exhibited fewer artifacts, sharper boundaries, and improved quantitative scores.

Conclusions:

• PhyDNN effectively addresses the limitations of traditional EIT reconstruction methods.
• The physics-driven approach enhances both the accuracy and physical interpretability of tomographic tactile sensing.
• This framework shows significant promise for high-quality EIT-based tactile sensing applications.