FibTac combines pneumatic gripping and tactile sensing

Research news — paper published August 31, 2026; first reviewed September 11; source review updated November 4, 2026

FibTac is a pneumatic gripper that uses carbon fibers both to manipulate objects and to sense their interaction with the environment. Researchers at Purdue University describe the system in a peer-reviewed paper published in npj Robotics on August 31, 2026. Its useful distinction is the shared mechanical structure: the fibers do the grasping, while an internal camera reads their movement as tactile information. Read the paper.

Explore tactile sensing for the broader comparison.

The laboratory results cover chess-piece identification, liquid classification, granular-material classification, underwater object recognition, and water-flow estimation. These are separate experiments with different training sets and conditions. A single accuracy number would obscure what the gripper has actually demonstrated.

How a fiber gripper senses touch

Carbon fibers are embedded in silicone inside a waterproof housing. Pneumatic pressure deforms the silicone membrane, translating and reorienting the rigid fibers to grasp or release an object; the fibers themselves are not the compliant bending element. The camera observes fiber-tip motion, which also changes when an object or surrounding medium resists that movement. Learned models use these image sequences for classification or regression.

This combines actuation and sensing in a compact contact structure. It also makes the signal depend on how the robot moves: liquid tests use active grasp–release cycles, whereas the granular tests move the gripper vertically with passive fibers. The sensing protocol is part of the measurement, not just a detail of the demonstration.

The authors report a payload up to 186 g and more than 3,000 actuation cycles without observable damage under their tested conditions. These establish a laboratory operating range; they do not establish an industrial lifetime or a payload rating across arbitrary objects. The paper’s results and methods describe the tested hardware and procedures.

What the classification results measure

The following values are author-reported results from the paper. Clips, held-out samples, and online manipulation trials are kept separate because they are different evaluation units.

The chess experiment’s 20 online trials should not be substituted for the larger clip-based training and validation collection. Likewise, the underwater result does not imply that the gripper can identify unseen marine objects. These distinctions matter when comparing FibTac with other robotic gripper tactile sensors.

Why motion and medium matter

For liquids, the gripper alternates grasp and release every second, collecting the resistance pattern over two cycles. For granular media, the robot moves vertically with a 4 cm amplitude and a 1.4-second period without pneumatic actuation. These motions actively expose material-dependent signals to the fiber array.

The flow experiment adds a different task: regression against measured water-flow conditions. The collection uses seven angles from 0° to 90° in 15° increments, 20 flow levels, and 12 clips per condition, allocated as nine training, one validation, and two test clips. The reference flow-meter values are volumetric flow in L/min; they should not be described as water velocity in m/s. See the flow-sensing methods.

For an engineering reader, this makes FibTac a candidate for studying coupled grasping and active tactile perception in wet or compliant environments. A comparison should match motion, fluid or material, camera settings, object set, and train/test separation before attributing a result solely to the sensor design.

What can be inspected or reproduced today?

The authors’ FibTac repository provides evaluation notebooks for five tasks and links external data and checkpoints. The November 4 root and README review still describes validation/test assets rather than a complete training-data release. RoboSkin reviewed the repository and its instructions but did not download the external data or rerun the models.

The paper and repository point to different Google Drive folders. Both are source-provided links; their contents and equivalence were not independently verified in this review. A reproducer should first confirm that the selected notebook, checkpoint, and evaluation split belong together. The November 4 GitHub root has no license file and its license metadata is empty, so the paper’s open-access status should not be treated as a software or data license.

Limits to carry into a comparison

The paper discusses practical constraints, including camera-cable fragility, the tradeoff between gripping strength and fine sensing, and generalization to unseen objects or environments. Those constraints are consequential for a combined actuator and sensor: changing stiffness or fiber geometry may improve one function while changing the other.

RoboSkin did not conduct these laboratory experiments. The article synthesizes the authors’ paper and public repository, and the cover is an illustration. For alternative contact structures, see soft robotic skin and the tactile sensor benchmark guide. For a different approach that reconstructs contact geometry from sparse electrical measurements, compare TacPrint’s controlled grasping results.

Sensing does not automatically close the pneumatic control loop

FibTac’s active sensing uses timed pneumatic commands and predefined robot waypoints. In the chess example, classification selects a predefined destination; it does not establish a controller that continuously adjusts grasp pressure from measured contact. The camera’s 60 fps is an acquisition setting, not evidence of a 60 Hz robot feedback loop. Control and experimental methods.

There is also a model-selection boundary in the flow experiment: the methods select the model using the lowest test-set MAE. That score therefore should not be presented as performance on an untouched final holdout. A new comparison should reserve an additional evaluation set, freeze model selection first and preserve the grouping by angle, flow condition and capture sequence. This is a proposed evaluation design; RoboSkin has not rerun the model.

For a new gripper or medium, record commanded pressure, robot motion, camera timing, fiber geometry and the actual outcome alongside the predicted class or flow. Recheck those relationships after replacing the membrane or fibers. Include unfamiliar objects and an explicit rejection decision instead of assuming every input belongs to a known class. The paper itself reports unreliable weight estimation, so successful classification must not be converted into a calibrated weight or contact-force claim. See the contact-evidence comparison for the separate tests needed before using a sensor output in control.

Sources

• Athar et al.: FibTac, npj Robotics, August 31, 2026
• Purdue MARS Lab research listing
• FibTac evaluation repository, reviewed revision

*Source review updated November 4, 2026. Original publication dates are retained. Experiments are attributed to their authors; RoboSkin did not reproduce the hardware, execute the models or inspect downloaded experimental payloads.*