Why Can't Robot Fingertips Reuse Each Other's Experience?

Researchers studying robotic touch have found a use for an ordinary consumer-grade 3D printer.

Researchers at Northwestern University turned a 3D printer into a device that repeatedly presses tactile sensors. Their 3D Cal system records camera images alongside measured indentation depths at known positions, providing reference data for calibration.

They collected roughly 36,000 images apiece from two optical tactile sensors, DIGIT and GelSight Mini. A model trained on the images and their measured indentation depths reconstructed the surfaces of previously unseen objects with average errors of about 0.16 and 0.21 millimeters, respectively.¹

Why press a sensor tens of thousands of times?

An optical tactile sensor has a soft material at its tip. Contact deforms the surface, and an internal camera records the change. But the pattern seen by the camera does not itself reveal the indentation depth. The images must first be matched to known depths—much like adding graduations to a ruler.

That process is calibration. Before a robot can interpret what it touches, it must learn what the signals from its own fingertip mean.

But calibration can drift as a sensor is used.

Fingertips change as they collect data

Tactile sensors measure contact by pressing and rubbing against objects. As the material on a fingertip wears or deforms, the same object can produce a different signal.

In 2026, researchers led by F. Richard Cottrell subjected optical tactile sensors to a deliberately harsh repeated-pressing test. The surface membranes of the commercial GelSight Mini and DIGIT sensors ruptured.²

Even without visible damage, a fingertip's mechanical properties may change over time. That makes it important to check whether its calibration remains valid. Researchers have also proposed ways to update calibration by measuring a sensor's mechanical properties without taking it apart.³

Stored data can remain unchanged even as the sensor that produced it changes.

Different sensor designs add another obstacle.

The same touch produces different signals

GelSight-type sensors use a camera to read surface deformation. TacTip tracks small internal pins. uSkin infers contact forces from changes in a magnetic field.

Even when these sensors press the same object with the same force, their outputs differ. A model trained to infer force from one sensor cannot necessarily be applied directly to another.

GenForce, developed by researchers including a team at King's College London, attempts to bridge that gap.⁴

The researchers represented each sensor's contact signal as a pattern of markers and learned to translate these patterns between sensors. A method for estimating force learned on one device could then be transferred to another.

Some paired measurements were still needed to relate the sensors, but far fewer new examples with ground-truth force labels were required. In tests across fingertip materials of different stiffness, the researchers also reduced errors in estimated contact force.

GenForce primarily transfers force estimation. Its developers also demonstrated grasping and slip compensation, but transferring force estimates is not the same as transferring an entire manipulation policy.

To manipulate an object, a robot must decide not only where and how hard to touch it, but also how to move its fingers.

Tactile datasets now contain millions of samples

Introduced in 2009, ImageNet is a large image database that made it possible for researchers to train artificial intelligence models on shared photographs and compare results.

Tactile researchers have also built large datasets.

The Sparsh model learned representations usable across tasks from more than 460,000 tactile images.⁵ TacX combined images, sound, movement and pressure from about one million interactions involving sustained contact.⁶

Released in 2026, AnyTouch 2 learns jointly from different optical tactile sensors. Its ToucHD dataset contains around 2.43 million tactile samples from real and simulated contact.⁷

These totals use different units. Some count images, others interactions, and still others frames or paired tactile–force measurements.

It is no longer enough to describe tactile robotics as a field short of data.

The question is whether data collected with one kind of robot fingertip can be used by another.

From a human hand to a robot hand

Think of someone screwing in a light bulb.

The fingers hold the bulb, adjust pressure to prevent slipping and turn it a little at a time. Changes in contact guide each movement.

Transferring that experience to a robot means bridging two differences. A human tactile glove and a robot sensor produce different signals from the same contact. Human and robot fingers also have different lengths and joint structures.

TactAlign, presented at the 2026 Robotics: Science and Systems conference, connects human and robot tactile signals.⁸

The researchers studied how human and robot hands moved relative to the objects they handled. Those interactions provided approximate correspondences from which the signals could be aligned.

In an experiment involving screwing in a light bulb, the researchers used 20 human demonstrations. They reused a tactile mapping learned from earlier tasks, so no robot demonstrations specific to the light-bulb task were required. The robot succeeded in all ten evaluation attempts reported by the team.

Contact experience from a human hand had helped a differently constructed robot hand perform the task.

TACROSS, described in a preprint posted in October 2026, examined how human tactile data could support robot learning.⁹

The human glove measured changes in electrical resistance when pressed, while the robot sensor measured changes in capacitance.

Rather than matching raw sensor readings, the researchers aligned patterns such as contact onset and changes in pressure.

Across four manipulation tasks, training with 30 robot demonstrations alone produced an average success rate of 51.3%. Adding 150 human demonstrations to the same 30 robot demonstrations and applying the full TACROSS alignment method raised the rate to 91.9%.

The robot succeeded more often without requiring additional robot demonstrations. Human tactile data contributed to learning, while the robot demonstrations still supplied the reference actions.

Linking the signals from different fingertips could therefore reduce the amount of experience that has to be collected directly on a robot.

Ten thousand hours collected without a target robot

TacClaw-Infinity, described in an October 2026 preprint, took a different approach to gathering experience.¹⁰

Researchers used TacClaw, a handheld gripper-like device. When a person grasped and moved an object, the system recorded camera footage, tactile signals and hand movements together.

The researchers reported collecting more than 10,000 hours of manipulation and over one million manipulation episodes without operating a particular target robot. As of 21 October 2026, the public release covered the first 1,000 hours, not the entire reported collection.

Models trained with this material were evaluated on three different robot types. In the reported tests, the robots could stack or sort objects and wipe surfaces without additional task demonstrations gathered on each target robot. Models with tactile information had higher average success rates than those without it.

These results suggest that experience gathered without operating a target robot can help train more than one kind of robot; they do not establish transfer to every robot or task.

But the researchers still had to check whether movements recorded with the handheld device were executable by each target robot.

Understanding contact and executing the movement it calls for are separate problems.

A giant data repository is not enough

Three differences emerge when these studies are considered together.

First, signals from different sensors have to be connected: the same touch is recorded differently depending on the device.

Second, movements must be adapted to each hand. A robot with different joints cannot necessarily reproduce the same motion even if it understands the contact.

Third, sensors themselves change with use. Methods now exist to reduce force-estimation errors caused by differences in fingertip stiffness, but detecting wear and updating calibration during actual work require further testing.

ImageNet mattered not only because it contained many photographs, but because researchers could share the same data and compare the results of learning from it.

Tactile datasets are growing rapidly too. Yet if the fingertip that collected the data differs from the one trying to use it, the signals must be interpreted across sensors and the movements adapted to the new hand.

Success in robotic touch depends not on one enormous data store, but on making the contact experience of one fingertip reusable by another.

Notes

1. Rohan Kota et al., “3D Cal: An Open-Source Software Library for Depth Reconstruction on Vision-Based Tactile Sensors”, IEEE Robotics and Automation Letters 11(5), 6248–6255 (2026). DIGIT contributed 36,630 images and GelSight Mini 36,270. Mean depth-reconstruction errors on unseen objects were 156 and 205 μm.

2. F. Richard Cottrell et al., “A Durable Vision-Based Tactile Fingertip for Robotic Manipulation”, arXiv:2608.24242 (2026). A deliberately accelerated test applied 39.2 N at 45 presses per minute; initial membrane failures occurred after approximately 25–35 minutes under these conditions, not after that amount of normal service life.

3. Can Zhao et al., “In-situ Mechanical Calibration for Vision-based Tactile Sensors”, ICRA (2023). In-situ estimation of Young's modulus and Poisson's ratio without disassembling the sensor.

4. Zhuo Chen et al., “Training tactile sensors to learn force sensing from each other”, Nature Communications 17, 2101 (2026). Tests included GelSight, TacTip and uSkin. Location-paired data represented less than 10% of the force-labeled images in the reported experiments. Across seven silicone materials of different stiffness, mean normal-force errors fell by 30% for hard-to-soft transfer and 16% for soft-to-hard transfer. This was not longitudinal testing of autonomous wear monitoring.

5. Carolina Higuera et al., “Sparsh: Self-supervised touch representations for vision-based tactile sensing”, CoRL 2024, PMLR 270, 885–915 (2025). More than 460,000 tactile images.

6. Carolina Higuera et al., “Tactile Beyond Pixels: Multisensory Touch Representations for Robot Manipulation”, CoRL 2025, PMLR 305, 105–123. Around one million contact-rich interactions collected with Digit 360. The earlier preprint used the name Sparsh-X.

7. Ruoxuan Feng et al., “AnyTouch 2: General Optical Tactile Representation Learning For Dynamic Tactile Perception”, ICLR (2026). ToucHD's 2,426,174 tactile samples comprise 1,118,896 simulated touch frames, 584,842 real-manipulation frames and 722,436 paired tactile–force measurements. They are not independent contact-event counts.

8. Youngsun Wi et al., “TactAlign: Human-to-Robot Policy Transfer via Tactile Alignment”, Robotics: Science and Systems (2026). Preprint. Both the OSMO glove and Xela robot sensors are magnetic-based, but their signal structures and sensing mechanisms differ. The light-bulb task reused alignment learned on previous tasks; 20 human demonstrations were used, with no task-specific robot demonstrations, and all ten evaluations succeeded.

9. Bo Chen et al., “TACROSS: An Efficient and Low-Cost Scalable Human Touch System Across Heterogeneous Tactile Sensors for Dexterous Robot Learning”, October 2026 preprint. Average success across four tasks was 51.3% with 30 robot demonstrations alone and 91.9% with those same demonstrations, 150 human demonstrations and the complete TACROSS method. Removing human touch reduced performance to 73.1%, while removing tactile alignment reduced it to 49.4%. Not all of the gain can be attributed to touch alone.

10. Fucai Zhu et al., “TacClaw-Infinity: A Large-Scale Vision-Tactile-Language-Action Dataset for Dexterous Manipulation”, October 2026 preprint. The authors report more than 10,000 hours, one million episodes and one billion frames. As of 21 October 2026, the public dataset portal announces its first 1,000 hours; collection, publication and training volumes are different. Two policy families showed mean success-rate increases from 50.0% to 74.0% and from 46.0% to 72.7%, respectively. The research repository provides robot replay/kinematic tools but describes policy training and task-level evaluation code as forthcoming.