⚡ electrical engineering

Pneumatic-Tomographic Tactile Skin for Multicontact Localization and Force Estimation

This paper presents a dual-channel pneumatic-tomographic tactile skin that combines an electrical impedance tomography layer for robust multicontact localization with a pneumatic pressure layer for stable force estimation, utilizing a location-aware correction framework to achieve high-accuracy force sensing across diverse contact scenarios without requiring complex machine learning pipelines.

Original authors: Haofeng Chen, Jiri Kubik, Bedrich Himmel, Matej Hoffmann, Hyosang Lee

Published 2026-08-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Haofeng Chen, Jiri Kubik, Bedrich Himmel, Matej Hoffmann, Hyosang Lee

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a robot that can not only feel a gentle touch but also understand exactly where it is being touched and how hard it is being pressed, all across its entire body. For machines to interact safely and naturally with humans, they need this kind of sensitive "skin." While humans have billions of tiny sensors in their skin that tell the brain about pressure and location simultaneously, giving robots this ability has been a difficult engineering puzzle. Traditional approaches often rely on dense grids of individual sensors, which are expensive to build, hard to wire up, and prone to interference. A more promising alternative involves using electrical signals to map the inside of a material, a technique known as electrical impedance tomography. This method can locate where a touch happens using very few wires, but it struggles to tell exactly how hard the touch is, often giving inconsistent readings depending on where the touch occurs.

To solve this problem, a team of researchers has developed a new type of robotic skin that combines two different sensing methods into a single, flexible layer. Their design pairs the electrical mapping layer with a thin, air-filled layer that acts like a pressure sensor. The electrical layer is excellent at pinpointing the location of a touch, even if multiple things are touching the robot at once. The air-filled layer, meanwhile, is very good at measuring the total force of the touch, but it cannot tell where that force is coming from. By fusing these two signals, the researchers created a system that knows both where a touch is and how hard it is, without needing complex wiring or massive amounts of training data.

The core of this innovation lies in how the two layers work together. The electrical layer sends a small current through a porous foam soaked in a conductive liquid. When something presses on the skin, the liquid is squeezed, changing the electrical resistance in that specific spot. This change allows the system to reconstruct an image of where the contact is happening. However, this electrical method is notoriously tricky when it comes to measuring force; the signal strength varies wildly depending on whether the touch is in the center of the sensor or near the edge. To fix this, the researchers added a pneumatic layer, which is essentially a sealed shell filled with air. When the skin is pressed, the air inside compresses, and a sensor measures the change in pressure. This pressure change correlates very strongly with the total force applied, regardless of where the touch happens.

The challenge was that the air layer, while reliable for measuring total force, still had slight variations in sensitivity across its surface, much like the electrical layer. To overcome this, the team developed a clever calibration method. They pressed the sensor at many different spots with known weights and used the electrical layer to record exactly where each touch occurred. They then used this data to create a digital map that corrects the pressure readings based on location. This map acts like a guide, telling the system how to adjust the force reading depending on whether the touch is near a stiff edge or in the flexible center. Once this correction is applied, the system can take the total force measured by the air layer and split it among the different touch points identified by the electrical layer.

The results of this approach were tested extensively. In experiments where a single object pressed on the sensor, the new system estimated the force with a high degree of accuracy, making errors of less than 0.6 Newtons on average. This was a significant improvement over using the electrical method alone, which produced errors more than twice as large. The system also proved robust when tested with objects of different sizes, from small 10-millimeter tips to larger 25-millimeter ones, showing that the calibration held up even when the contact area changed. Perhaps most impressively, the system handled multiple touches at the same time. In scenarios where two or three objects pressed on the skin simultaneously, the system could still distinguish the individual locations and estimate the force of each touch separately. Without the location-aware correction, the system would often misjudge the force of individual touches in these complex situations, but with the correction, the error dropped by nearly 40 percent.

This work demonstrates that combining a location-finding sensor with a force-measuring sensor creates a more capable and practical tool than either could be on its own. The researchers found that by letting the air layer handle the heavy lifting of force measurement and using the electrical layer simply to find the "where," they could avoid the need for massive datasets or complicated machine learning models that other systems often require. The resulting skin is simple to build, flexible enough to wrap around a robot's body, and capable of providing the detailed feedback necessary for safe human-robot interaction. While there are still limitations, such as slight delays in response and challenges with very large or overlapping touches, the study confirms that this dual-channel approach offers a scalable and effective path forward for giving robots a true sense of touch.

Technical Summary: Pneumatic-Tomographic Tactile Skin for Multicontact Localization and Force Estimation

Problem Statement
Large-area tactile sensors are essential for safe and effective human-robot interaction, yet existing solutions face significant trade-offs. Dense sensor arrays offer high resolution but suffer from high manufacturing costs, wiring complexity, and electromagnetic interference. Reconstruction-based methods, specifically Electrical Impedance Tomography (EIT), provide a cost-effective alternative for large-area sensing using few electrodes. However, EIT-based tactile skins suffer from nonuniform sensitivity distributions caused by electrode placement and boundary effects. This nonuniformity severely limits the accuracy of contact force estimation, particularly in multi-contact scenarios where reliable calibration is difficult to achieve without extensive datasets or complex machine learning pipelines. While pneumatic pressure sensors offer stable force estimation, they lack the spatial resolution to localize contacts within a large area.

Methodology
The authors propose a dual-channel tactile skin that synergistically integrates an EIT layer with a pneumatic pressure layer to overcome the limitations of each individual modality.

Hardware Architecture:
- EIT Layer: Constructed from polyurethane (PU) foam infused with ionic liquid and surrounded by 16 silver electrodes. This layer is responsible for spatial localization.
- Pneumatic Pressure Layer: A thin shell made of thermoplastic elastomer (TPE) with a gyroid infill pattern, connected to a pressure sensor. This layer provides a scalar measurement of total contact force.
- Integration: The EIT layer is embedded into a slot within the pneumatic shell, creating a compact, flexible, and large-area sensor.
Signal Processing and Calibration Framework:
- EIT Reconstruction: Boundary voltages are measured and reconstructed into a conductivity change image ( $\Delta\sigma$ ) using a Tikhonov/Newton's One-Step Error Reconstructor (NOSER).
- ROI Segmentation: The reconstructed image is segmented into Regions of Interest (ROIs) using Otsu's method and morphological operations. For each ROI, the system extracts the intensity-weighted centroid (contact location) and the sum of conductivity changes.
- Force Distribution: The pneumatic layer measures the total pressure change ( $\Delta P$ ). This total pressure is distributed across the identified ROIs proportionally to their respective conductivity sums.
- Location-Aware Correction: To address the mild nonuniform sensitivity of the pneumatic layer (caused by edge stiffness), the authors introduce a calibration framework. Using single-contact data, they learn smooth spatial fields for gain ( $g(x,y)$ ) and offset ( $b(x,y)$ ). These fields are modeled as quadratic functions of the contact coordinates derived from the EIT layer.
- Final Estimation: The distributed pressure for each contact is converted to force using the learned global pressure-force mapping, then refined by the location-aware gain and offset fields specific to that contact's centroid.

Key Contributions

Tomographic–Pneumatic Dual-Channel Framework: A novel architecture that assigns complementary roles: EIT handles contact localization (centroids and ROIs), while the pneumatic layer handles force estimation. This separation avoids the need for EIT to perform force estimation directly.
Location-Aware Correction Fields: A calibration method that learns spatial gain and offset fields from a single session. This corrects for the pneumatic layer's spatial sensitivity variations, enabling consistent force estimation across the entire surface without requiring dense, multi-point calibration for every location.
Scalable Fabrication: The design utilizes standard materials (PU foam, ionic liquid, TPE) and simple fabrication techniques (3D printing, encapsulation), making it suitable for large-area robotic skins.

Experimental Results
The system was evaluated on single-contact and multi-contact scenarios using various indenter sizes (10, 20, and 25 mm) and loads.

Localization: The system achieved an average position error of $4.4 \pm 1.9$ mm, comparable to EIT-only systems.
Single-Contact Force Estimation:
- The proposed Pneumatic-LocAware method achieved a Root Mean Square Error (RMSE) of 0.59 N across 10–25 mm indenters.
- This represents a 35–60% reduction in error compared to EIT-only approaches (1.45–1.48 N) and a significant improvement over the uncorrected pneumatic baseline (0.91 N).
- The method demonstrated generalization to indenter sizes not used during calibration (10 mm and 20 mm), though performance slightly degraded for 25 mm indenters at higher loads due to boundary interactions.
Multi-Contact Force Estimation:
- In experiments involving 2 and 3 simultaneous contacts, the per-contact RMSE decreased from 0.91 N (uncorrected pneumatic) to 0.55 N (proposed method), a 39.6% improvement.
- The system maintained robust performance across diverse configurations (central, edge, clustered), though accuracy decreased slightly when contacts overlapped significantly.

Significance and Claims
The paper claims that this dual-channel design provides a practical, scalable, and easy-to-calibrate solution for large-area robotic skins. By leveraging the pneumatic layer for force estimation and the EIT layer for localization, the method avoids the large datasets, complex calibration setups, and heavy machine-learning pipelines often required by EIT-only approaches. The authors highlight that the pneumatic pressure layer offers a more reliable basis for force estimation than EIT conductivity sums, which suffer from inherent nonuniform sensitivity. Furthermore, the use of a pneumatic layer aligns with industrial safety standards (similar to AIRSKIN), suggesting potential for certified human-robot collaboration. The work demonstrates that accurate multi-contact force estimation is achievable with minimal calibration requirements by fusing complementary sensing modalities.

More like this