ECT Applications in Robotic Embedded Sensing

TL;DR

ECT sensor integrated into robotic gripper enables object detection and localization under variable-spacing conditions through dynamic calibration framework and XGBoost prediction model—providing a non-invasive, real-time sensing solution for robotic grasping.

---

I. Background: Challenges in Robotic Sensing

🤖 Limitations of Traditional Capacitive Sensors

Capacitive sensors are widely used in robotic sensing due to their unique advantages:

Capacitive sensor advantages:
├─ Fast response
├─ Easy calibration
├─ Can detect metallic and non-metallic objects
├─ Can distinguish different materials
└─ Relatively low cost

But traditional applications have limitations:

Limitation	Description	Impact
Distance-only	Only measures distance to objects	Cannot determine if object is centered
Single-point measurement	Limited information	Cannot plan optimal grasping path
Fixed geometry	Fixed sensor spacing	Difficult to adapt to dynamic scenarios

💡 ECT Opportunity

Electrical Capacitance Tomography (ECT) offers a new solution:

ECT advantages:
├─ Full-field imaging (not single-point)
├─ Visualizes permittivity distribution
├─ Non-invasive measurement
└─ Real-time image reconstruction

Challenges:
├─ Traditional ECT sensors have fixed geometry
├─ Cannot adapt to variable-spacing scenarios
└─ Dynamic imaging requirements difficult to meet

---

II. Innovation: Variable-Spacing Parallel-Plate ECT Sensor

🔧 Sensor Design

Tianjin University research team designed a parallel-plate ECT sensor integrated into a robotic gripper:

Basic structure:
├─ Two parallel plates (simulating gripper jaws)
├─ 6 electrodes per plate, 12 measuring electrodes total
├─ Electrode size: 40mm × 80mm
├─ PCB size: 335mm × 110mm
└─ Plate spacing D variable (changes with gripper opening)

Electrode naming:
├─ Plate A: A1, A2, A3, A4, A5, A6
└─ Plate B: B1, B2, B3, B4, B5, B6

Working principle:

Capacitance measurement → Permittivity distribution → Object detection and localization

Gripper spacing D:
├─ Minimum: w = 40mm
├─ Maximum: 3w = 120mm
└─ Continuously variable

🎯 Comparison with Traditional ECT

Feature	Traditional Industrial ECT	Variable-Spacing Robotic ECT
Geometry	Fixed circular/square	Parallel plates, variable spacing
Sensing region	Enclosed, fixed	Open, variable
Sensitivity distribution	Constant	Changes dynamically with spacing
Applications	Pipeline/vessel monitoring	Robotic grasping sensing
Calibration needs	One-time calibration	Dynamic calibration framework

---

III. Core Challenges: Technical Problems from Variable Spacing

📊 Challenge 1: Dynamic Sensitivity Distribution

Problem: When gripper spacing changes, electric field distribution and sensitivity matrix change accordingly.

Fixed spacing assumption:
├─ Traditional ECT assumes fixed sensor geometry
├─ Sensitivity matrix is constant
└─ Calibration parameters unchanged

Variable-spacing reality:
├─ Electric field line distribution changes with D
├─ Sensitivity matrix changes dynamically
└─ Calibration parameters need updating

Electric field characteristics:

1. Open-field divergence:
└─ Field lines extend beyond geometric boundaries, affected by environment

2. Distance-dependent field density:
├─ D small: Field concentrated, uniform sensitivity
├─ D large: Field divergent, center sensitivity decreases
└─ Single sensitivity matrix cannot meet requirements

🔧 Challenge 2: Full-Field Calibration Efficiency

Problem: Traditional methods require full-field calibration at each spacing, which is time-consuming and impractical.

Traditional calibration process:
├─ Set spacing D
├─ Fill entire sensing area
├─ Measure full-field capacitance
├─ Change spacing, repeat above
└─ Extremely inefficient

Actual needs:
├─ Gripper spacing changes continuously
├─ Cannot repeat calibration at each spacing
└─ Need more efficient solution

---

IV. Solution: Dynamic Calibration and Reconstruction Framework

🧮 Solution 1: Sensitivity Matrix Interpolation

Idea: Precompute sensitivity matrices at multiple discrete spacings, interpolate during runtime to obtain current spacing sensitivity.

Implementation steps:
1. Select discrete spacing points
   ├─ Range: w to 3w
   ├─ Step: 0.2w
   └─ 11 points total

2. FEM calculation
   ├─ Use COMSOL for modeling
   ├─ Calculate field distribution at each spacing
   └─ Generate corresponding sensitivity matrix

3. Interpolation function
   ├─ Runtime based on current D
   ├─ Interpolate to calculate sensitivity matrix
   └─ Real-time adaptation to geometric changes

Performance comparison:

Method	Computation time	Relative error
Direct FEM recalculation	40.68 sec	-
Linear interpolation	0.34 sec	below 3.4%
Advantage	119× faster	Sufficient accuracy

🤖 Solution 2: XGBoost Capacitance Prediction Model

Idea: Only one reference calibration needed, use machine learning to predict full-field capacitance at other spacings.

Model architecture:
Input features X (67-dim):
├─ Gripper spacing D (1 dim)
└─ Full-field capacitance at reference spacing Ch (66 dims)

Prediction target Y (66-dim):
└─ Full-field capacitance Ch at target spacing

XGBoost parameters:
├─ Number of regression trees: 200
├─ Maximum tree depth: 5
├─ Learning rate: 0.1
└─ L2 regularization: 1.0

Performance metrics:

Model	MSE	RMSE	MAE	R²
XGBoost	9.455×10⁻⁸	3.075×10⁻⁴	1.539×10⁻⁴	0.9908
Multiple Linear Regression	2.995×10⁻⁵	5.473×10⁻³	2.119×10⁻³	0.9419
Random Forest	3.980×10⁻⁶	1.995×10⁻³	1.067×10⁻³	0.9714

Practical application:

• Only one full-field calibration at reference spacing (50mm) needed
• Model prediction time: 36.72ms/sample
• Suitable for real-time sensing applications

✅ Validation Results: Cross-Validation

Leave-one-distance-out cross-validation results:

• MAE: 0.0006672 (log scale)
• Conclusion: Model successfully learned physical relationship between distance and capacitance, not data memorization

---

V. Experimental Validation

🧪 Experimental Setup

Test scenarios:
├─ Stratified targets (different material layers)
├─ Single object targets
└─ Dual object targets

Spacing tests:
├─ D = 50mm (closer)
└─ D = 95mm (farther)

Reconstruction algorithms:
├─ LBP (Linear Back Projection)
├─ Landweber (50 iterations)
└─ Tikhonov regularization

📊 Experimental Results

1. Stratified Target Imaging

Position	D	CC	SSIM	Description
Near PCB	95mm	0.5484	0.3661	Clearer boundaries
Center region	95mm	0.5095	0.2256	Center blurred, edges preserved

2. Single/Dual Object Detection (50mm)

Target	CC	SSIM	Position error
Single object (center)	0.8787	0.8741	0.3082
Dual objects (opposite sides)	0.8647	0.6797	0.6737

3. Single/Dual Object Detection (95mm)

Target	CC	SSIM	Position error
Single object (center)	0.8360	0.9027	0.9621
Dual objects (opposite sides)	0.6221	0.7424	0.9644

📈 Key Findings

1. Image quality decreases with increased spacing
   ├─ CC value: 0.97 at w → 0.86 at 3w
   └─ Sensitivity decreases in center region

2. Center vs edge position
   ├─ Center position: Better reconstruction quality
   └─ Edge position: Higher sensitivity, clearer imaging

3. Practical validation
   ├─ Can detect object presence
   ├─ Can locate dielectric targets
   ├─ Can reconstruct approximate contours
   └─ Meets robotic grasping requirements

⏱️ Computational Efficiency

Single frame processing time: 385.60 ms

Meets requirements:
├─ Near-real-time robotic sensing
├─ Dynamic monitoring during grasping
└─ Practical application feasible

---

VI. Technical Advantages and Application Prospects

✅ Technical Advantages

1. Non-invasive sensing
├─ No object contact needed
├─ Doesn't affect grasping process
└─ Safe and reliable

2. Full-field information
├─ Obtain object distribution
├─ Determine if object is centered
└─ Plan optimal grasping path

3. Adaptive capability
├─ Dynamically adapt to spacing changes
├─ Real-time sensitivity updates
└─ No repeated calibration needed

4. Computational efficiency
├─ Fast interpolation
├─ Quick model prediction
└─ Meets real-time requirements

🚀 Application Prospects

Short-term applications:
├─ Industrial robotic grasping sensing
├─ Object detection and localization
└─ Grasping process monitoring

Medium-term development:
├─ Optimize shielding structure design
├─ Extend to complex object shapes
└─ Integrate into closed-loop grasping control

Long-term directions:
├─ Multi-modal sensing fusion
├─ Dexterous manipulation applications
└─ General robotic sensing platform

⚠️ Current Limitations

1. Image quality limitations
├─ Quality decreases with increased spacing
├─ Limited spatial resolution
└─ Contour accuracy needs improvement

2. Open-field interference
├─ Electric field susceptible to environment
├─ Needs shielding optimization
└─ Stability requires improvement

3. Application scope
├─ Limited permittivity range
├─ Complex shape recognition difficult
└─ Quantitative accuracy needs improvement

---

VII. Connection with Tianjin Youyi Technology

🔬 Research Translation

This paper comes from Tianjin University Electrical Imaging Research Group, sharing the same origin as Tianjin Youyi Technology:

Technical connection:
├─ Research team: Tianjin University Electrical Imaging Research Group
├─ Core technologies: ECT sensor design, image reconstruction algorithms
├─ Application extension: From industrial monitoring to robotic sensing
└─ Product translation: TJUECT series can leverage related technologies

💼 Commercial Value

Potential application areas:
├─ Industrial robots: Grasping sensing
├─ Service robots: Environmental sensing
├─ Special robots: Underwater, space, etc.
└─ Research and education: Robotic sensing teaching

Market opportunities:
├─ Growing demand for robotic sensing technology
├─ Non-visual sensing solutions needed
└─ Domestic sensor market gap

---

VIII. Frequently Asked Questions

Q1: Why not use vision sensors?

A: ECT provides information difficult for vision to obtain:

Vision sensor limitations:
├─ Affected by lighting conditions
├─ Transparent/translucent objects hard to identify
├─ Serious occlusion problems
└─ Depth information needs complex algorithms

ECT advantages:
├─ Unaffected by lighting
├─ Can detect permittivity differences
├─ Penetrating measurement
└─ Direct depth information

Complementary, not replacement.

Q2: Does the XGBoost model need retraining?

A: No. Train model once:

Training phase:
├─ Use FEM simulation data
├─ Cover entire working range (w-3w)
└─ Model has strong generalization

Practical application:
├─ Only one reference calibration needed
├─ Model predicts other spacings
└─ No repeated training needed

Q3: Can this technology be commercialized?

A: Has commercial potential, but needs further development:

Current status:
├─ Laboratory prototype validation
├─ Core algorithms proven
└─ Application scenarios clear

Commercialization needs:
├─ Product design (miniaturization, integration)
├─ Cost control
├─ Reliability improvement
└─ Real-scenario validation

Q4: How does it compare to traditional force sensors?

A: Complementary relationship, providing different types of information:

Force sensors:
├─ Measure contact force
├─ Judge grasping stability
└─ Prevent object damage

ECT sensors:
├─ Pre-contact sensing
├─ Object position detection
└─ Material discrimination

Combined use:
└─ Enable smarter grasping control

---

IX. Summary

📚 Core Contributions

1. Variable-spacing ECT sensor design
   └─ First realization of robotic gripper integration

2. Dynamic calibration framework
   └─ Sensitivity matrix interpolation + XGBoost prediction

3. Experimental validation
   └─ Proved technical feasibility

4. Application prospects
   └─ New solution for robotic sensing

🎯 Tianjin Youyi Perspective

As the technology translation platform of Tianjin University team, Tianjin Youyi continuously focuses on:

Research directions:
├─ ECT technology innovation
├─ Multi-modal sensing fusion
├─ Robot integrated applications
└─ Industry-academia-research integration

Product planning:
├─ Industrial ECT systems (TJUECT)
├─ Customized solutions
└─ Emerging application exploration

---

📞 Contact Us

If you’re interested in ECT applications in robotic sensing, or need technical solution discussion:

Tianjin Youyi Technology Co., Ltd.

📧 Email: sales@iptomo.com 📱 Phone: +86-22-87057001 🌐 Website: https://iptomo.com 📍 Address: Tianjin, China

---

🔍 Further Reading

• ECT Technology Guide
• What are Soft Field and Ill-Posedness
• ECT Applications in Gas-Solid Two-Phase Flow
• Electrical Tomography vs Ray Tomography

---

📖 References

Based on Tianjin University research paper:

Cui, Z., Wang, S., Xu, J., & Li, D. (2026). Use of Electrical Capacitance Tomography in Robot Embedded Sensing. IEEE Sensors Journal.

Acknowledgment: Thanks to Tianjin University Electrical Imaging Research Group for their innovative research work.

---

💡 Upcoming Articles

Robotic Sensing Series:

1. ✅ ECT Applications in Robotic Embedded Sensing (this article)
2. 📝 Advances in Impedance Tactile Sensors
3. 📝 Multi-Modal Robotic Sensing Fusion
4. 📝 Vision-Tactile Fusion in Robotic Grasping

Follow our official account for more technical articles!