🤖 machine learning

FlexiTac: A Low-Cost, Open-Source, Scalable Tactile Sensing Solution for Robotic Systems

The work presents FlexiTac, a low-cost, open-source, and scalable piezoresistive tactile sensing system featuring flexible sensor mats and compact evaluation electronics that enables real-time data acquisition and supports advanced tactile learning pipelines for diverse robotic platforms.

Original authors: Binghao Huang, Yunzhu Li

Published 2026-05-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Binghao Huang, Yunzhu Li

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 you are trying to teach a robot to pick up a fragile egg or thread a needle. Currently, most robots rely heavily on their "eyes" (cameras). Yet eyes can be deceived by shadows, or they cannot see what happens when the robot's hand holds something. Humans, however, possess a secret superpower: our skin. We can feel whether something is slipping, how hard we are pressing, and exactly where an object touches our fingers—all without looking.

FlexiTac is a new, open-source project designed to give robots this same superpower of touch, in a way that is low-cost, easy to manufacture, and works on almost any robot hand.

Here is a simple breakdown of how it works, using everyday analogies:

1. The "Smart Skin" (The Sensor Pad)

Imagine the FlexiTac sensor pad as a high-tech, flexible bandage you can stick onto a robot's finger.

How it is made: It is built like a "club sandwich" of three thin layers. The top and bottom layers are flexible circuit boards (like the wires in a remote control), and the middle layer is a special black film called Velostat.
The Magic: When you squeeze the sandwich, the middle film changes its electrical resistance (it becomes "easier" or "harder" for electricity to flow). The robot's computer reads this change and knows exactly how hard it is pressing.
The Upgrade: In older versions, people had to hand-sew tiny wires to make this work, which was slow and messy. FlexiTac uses pre-made circuit boards with tiny, milled holes. It is like switching from hand-knitting a sweater to using a knitting machine—it is faster, more consistent, and you can produce hundreds quickly.
Flexibility: Because it is so thin (less than 1 mm) and flexible, it can wrap around a rigid metal robot finger or a soft, squeezable gripper without needing to rebuild the robot.

2. The "Translator" (The Readout Electronics)

The sensor pad generates a large amount of data (hundreds of tiny pressure points), but the robot's main memory does not speak "sensor language."

The Task: The readout electronics act like a translator or a traffic cop. They take the messy signals from the sensor pad, organize them, and send a clean, synchronized data stream 100 times per second to the robot's computer.
Simplicity: It uses very common, inexpensive electronic components (like an Arduino, a tiny, affordable circuit board). This keeps the entire system low-cost—about 30 dollars per sensor—so researchers can buy many without breaking the budget.

3. What Can Robots Do With This?

The study shows that FlexiTac is not just a gadget; it is a tool that helps robots learn better. Here are three main ways:

Seeing Through Touch (Visuo-tactile Fusion):
Imagine a robot trying to pick up a cup in a dark room. Its eyes are useless. But with FlexiTac, the robot can "feel" the shape and texture of the cup. The system combines what the robot sees (3D vision) with what it feels (tactile points) into a single, unified map. It is like giving the robot a "sixth sense" that fills in the gaps when its eyes are blocked.
Learning From Humans (Cross-Embodiment):
Imagine a human wearing a special glove with FlexiTac sensors, picking up objects and showing a robot how to do it. Because the robot has the exact same type of sensors on its hand, it can perfectly understand the human's movements. It is like two people speaking the same language; the robot can copy the human's "feeling" and apply it to its own body, even if the robot looks different.
Training in a Video Game (Real-to-Sim-to-Real):
Training robots in the real world is slow and risky (they might break things). Usually, we train them in computer simulations. But simulating "touch" is difficult because real skin is complex.
FlexiTac makes this easier because it measures simple pressure (like a scale) rather than complex images. The researchers developed a simulation that perfectly mimics the pressure sensors of FlexiTac. You can train the robot in a virtual world with these simulated "feelings" and then transfer that ability to the real robot with very little adjustment. It is like practicing a piano piece on a silent keyboard before playing it on a real piano.

Summary

FlexiTac is a "plug-and-play" solution. It is a low-cost, open-source method to give robots a sense of touch that:

Is Affordable: About 30 dollars per unit.
Is Easy to Manufacture: Uses simple, repeatable production.
Is Versatile: Works with hard metal robots, soft grippers, and even wearable gloves for humans.
Is Intelligent: Helps robots learn complex tasks by combining vision, touch, and computer simulations.

The goal is not just to build a sensor, but to make it so simple and affordable that every robotics lab can give its robots "skin," enabling them to handle fragile objects and learn much more effectively from human demonstrations.

1. Problem Statement

Despite rapid advances in robotic manipulation and learning-based control, robotic force grippers still lack the dexterity and sensory feedback of human skin. Current limitations include:

Sensory Gap: Robots often lack the high-frequency tactile feedback necessary for tasks such as grasping fragile objects, detecting slip motions, or performing touch-rich manipulations without continuous visual monitoring.
Cost and Complexity: Existing high-performance tactile sensors are frequently expensive, proprietary, or difficult to manufacture and integrate into diverse robotic platforms (e.g., soft grippers, multi-fingered hands).
Scalability: There is a lack of standardized, open-source solutions that enable large-scale data collection and easy deployment across various robotic embodiments, hindering the development of robust visuo-tactile learning pipelines.

2. Methodology

The authors propose FlexiTac, a modular, open-source tactile sensing system designed as a "plug-and-play" solution. The methodology is divided into hardware design, manufacturing, and system integration.

A. Hardware Design

FlexiTac consists of two main components:

Flexible Sensor Pads:
- Principle: Based on a piezoresistive sensor matrix, where mechanical pressure alters the electrical resistance of a conductive film.
- Structure: A sealed three-layer laminate stack: FPC (Flexible Printed Circuit) – Velostat (piezoresistive film) – FPC.
- Electrode Innovation (FlexiTac V2): Unlike previous designs using manually aligned conductive threads, FlexiTac V2 uses FPCs with patterned copper openings that serve directly as electrodes. This creates an orthogonal grid whose intersections define "Taxels" (tactile pixels).
- Geometry: The pads are thin (<1 mm) and lightweight. The design includes narrow slots between electrodes to increase compliance (adaptability to curved surfaces) and concentrate deformation to improve sensitivity. Stiffeners are added to maintain structural integrity without compromising flexibility.
- Configurations: Modular sizes support various resolutions (e.g., 8×16, 12×32, 32×32) to adapt to different grippers.
Readout Electronics:
- Components: Built with widely available, low-cost components: an Arduino Nano (MCU), CD74HC4067 multiplexers, and SN74HC595 shift registers.
- Function: The board scans the sensor matrix and streams synchronized data via a serial interface at 100 Hz to a host computer.
- Scalability: The design is backward-compatible; a single board can support lower-resolution pads by utilizing a subset of channels.

B. Manufacturing Process

Fabrication: The process is designed for high throughput and repeatability. A desktop cutting machine (Silhouette Cameo 5) is used to cut Velostat and laminating films to ensure consistent dimensions.
Assembly: The three-layer stack is encapsulated using laminating films. The entire process for a single pad takes approximately 5 minutes.
Cost: A complete unit costs approximately $30. The sensor pad (FPCs + Velostat) costs about $3.55 at scale, and the readout PCB costs about $3.10 (excluding the approximately $25 Arduino Nano).

C. System Integration and Learning Pipelines

The paper demonstrates the utility of FlexiTac in three advanced learning scenarios:

3D Visuo-Tactile Fusion: Tactile signals are mapped into 3D space using robot proprioception and forward kinematics. These "tactile points" are fused with visual point clouds (from RGB-D cameras) to create a unified 3D representation for contact-aware decision-making (e.g., using diffusion policies).
Cross-Embodiment Skill Transfer: FlexiTac is deployed on both a wearable human data-collection glove and a robotic arm (xArm). The common tactile interface allows strategies trained on human demonstrations (for capturing natural manipulation) to be transferred directly to the robot, as the tactile signal format remains consistent.
Real-to-Sim-to-Real Pipeline: The matrix-based pressure sensing is modeled in a GPU-parallel tactile simulator.
- Simulation: Each taxel is modeled as a contact point interacting with the Signed Distance Field (SDF) of an object using a Kelvin–Voigt penalty model (spring-damper).
- Calibration: A simple calibration procedure adjusts normal stiffness ( $k_n$ ) and damping ( $k_d$ ) to fit real force-response curves, enabling effective pre- and post-adjustment in simulation before deployment on real hardware.

3. Key Contributions

Open-Source Platform: Introduction of FlexiTac, a low-cost ($30), open-source tactile sensing solution integrating flexible pads, compact electronics, and unified software interfaces.
Manufacturing Innovation: A novel FPC-based electrode design that eliminates manual thread alignment, significantly improving manufacturing throughput, repeatability, and scalability.
Versatile Deployment: Demonstration of the sensor on diverse platforms, including rigid grippers (Robotiq, xArm), soft grippers, dual-arm systems (Aloha), and wearable human data-collection devices.
Learning Pipeline Support: Validation of FlexiTac in modern robotic learning workflows, particularly by enabling 3D visuo-tactile fusion, cross-embodiment transfer, and Sim-to-Real post-adjustment with calibrated tactile simulation.

4. Results

Hardware Performance: The sensors successfully deliver dense tactile signals (up to 32×32 resolution) at 100 Hz. The design proves robust on both rigid and soft interfaces without significant mechanical redesign.
Deployment Success: The system was successfully integrated into complex setups, including a semi-humanoid robot, a wearable tactile glove, and a dual-arm manipulation system, underscoring its adaptability and modularity.
Learning Outcomes:
- Visuo-Tactile: The 3D fusion approach enables strategies to reason over contact intensity and spatial relationships, improving performance on occluded or touch-rich tasks.
- Cross-Embodiment: Strategies learned with the wearable device from human demonstrations were successfully transferred to the xArm robot, leveraging the consistent tactile interface.
- Simulation: The calibrated Sim-to-Real pipeline enabled large-scale RL post-adjustment for assembly tasks (e.g., nut and bolt), with normalized real and simulated tactile histograms closely overlapping.

5. Significance

FlexiTac addresses a critical bottleneck in robotic research: the high cost and complexity of capturing dense tactile data. By lowering entry barriers, it enables:

Scalable Data Collection: Researchers can easily deploy sensors on many robots to collect large-scale tactile datasets.
Reproducibility: The open-source nature and simple manufacturing process allow other labs to replicate the hardware and compare results directly.
Advanced Learning: It fosters the development of contact-aware strategies that combine vision and touch, moving beyond purely visual manipulation.
Simulation Alignment: The structured nature of piezoresistive signals makes them easier to simulate than optical tactile sensors, bridging the gap between simulation and reality for reinforcement learning.

In summary, FlexiTac offers a practical, democratized foundation for advancing dexterous manipulation in robotics through accessible hardware and integrated learning pipelines.