TacGBA tactile upgrade.

PHYSICAL INTELLIGENCE · ROBOT LEARNING

TacGooseBumpsHacking tactile sensors
to feel friction.

For learning contact-rich manipulation

Wenjie Li1,*Binyu Yang1,*Yuxin Chen1Ambrose Wang2Masayoshi Tomizuka1,†

1 University of California, Berkeley2 Saratoga High School

* Equal contribution   ·   † Corresponding author

VideoPaperCode Coming soonBibTeX
SCROLL TO DISCOVER
01 / TL;DR

Cabinet bumpers.
Shear-aware touch.

Stick cabinet bumpers onto any existing normal-only tactile sensor. Shear (tangential forces) tilts the domes and redistributes pressure, encoding tangential cues for robot learning.

TacGB concept, two sensor integrations, data-collection pipelines, and four manipulation tasksView full figure +
One passive interface principle, evaluated across two tactile sensors and four contact-rich manipulation tasks.
2 sensors

Tachin

Commercial / off-the-shelf
Capacitive

FlexiTac

Lab-made / open-source
Piezoresistive
2 grippers & pipelines

Rigid gripper · Compliant gripper

SpaceMouse teleoperation · Hand-held iPhUMI
4 tasks

Insert · Twist-lock · Place · Draw

From discrete contact to continuous interaction

02 / THE VIDEO

See touch differently.

From a passive surface
to better contact-rich manipulation.

Research video

YouTube video coming soon
Paper ↗

03 / HOW IT WORKS

Shear becomes a pattern
the policy can learn.

01

Add the surface.

A thin array of soft domes adheres to an existing normal-only tactile sensor.

02

Let mechanics encode shear.

Tangential loading tilts the domes. Pressure increases along the leading side and decreases along the trailing side.

03

Learn directly from pressure.

The policy uses the resulting pressure maps, without reconstructing a calibrated shear-force vector.

Qualitative static-load demonstration on FlexiTac. A tangential load is applied to a stationary 500 g mass.

Inspect the pressure maps +
Before, after, and difference pressure maps with and without TacGBView full figure +
Before/after maps reveal directional redistribution with TacGB.

CUT. ATTACH. LEARN.

An upgrade that starts
with cabinet bumpers.

Cut an off-the-shelf sheet to fit, or mold domes for a target geometry. Both versions attach directly to the sensor surface, without modifying its electronics or aligning individual domes to individual taxels.

Integration with commercial capacitive Tachin and open-source piezoresistive FlexiTac sensors.

Explore the design and finite-element analysis +
Dome designs, finite-element strain fields, and integration procedureView full figure +
Under shear, dome tilt redistributes the normal response. The paper’s finite-element illustration magnifies deformation for visibility.

04 / EXPERIMENTS

Better decisions
at the point of contact.

— w/o TacGB— w/ TacGB

Same policy family. Same pressure-map input format. A different mechanical tactile interface.

Representative rollouts · playback speeds are annotated in the video.

Recognize alignment.
Know when to stop.

Faceplate contact, socket entry, and full seating can look nearly identical. TacGB makes their load transitions more visible in the tactile input.

95%insertion success
w/o TacGB75%
15 / 20 trials
w/ TacGB95%
19 / 20 trials

Off-the-shelf TacGB. The molded version achieves 18/20 (90%). Mean completion time for successful trials: 29.3 s without TacGB, 27.1 s off-the-shelf, and 26.7 s molded.

View the full USB results +
USB insertion setup, observations, success rates, substage survival, risky behaviors, and completion timeView full figure +
USB insertion: all variants are evaluated on the same 20 initial configurations.

Representative rollouts · playback speeds are annotated in the video.

Push. Twist.
Then release.

A bayonet lightbulb must be mechanically locked before the gripper lets go. Merely lighting up is not enough: without the twist lock, the bulb springs back out.

96%overall insertion success
w/o TacGB60%
15 / 25 trials
w/ TacGB96%
24 / 25 trials

20 nominal trials plus 5 altered-condition trials per sensing condition. Nominal: 13/20 to 20/20. Altered socket height and free rotation: 2/5 to 4/5.

View the full lightbulb insertion results +
Bayonet lightbulb setup, tactile dynamics, initializations, and locking successView full figure +
Lightbulb insertion: gains concentrate around alignment and twist locking.

Representative rollouts · playback speeds are annotated in the video.

Place gently.
Release with support.

Both policies complete every trial, but binary success hides how the egg lands. TacGB helps the policy transfer support to the cup before release.

−66%mean peak cup-sensor readout
w/o TacGB0.77
mean peak readout
w/ TacGB0.26
mean peak readout

25/25 binary successes in both conditions, using five eggs per condition. Mean contact-transfer duration rises from 0.99 s to 3.84 s.

View the full egg results +
Egg transfer setup, observations, raw eggs, cup-sensor traces, and peak-versus-duration resultsView full figure +
Egg transfer: lower peak readouts and longer support transfer indicate gentler placement.

Representative rollouts · playback speeds are annotated in the video.

Maintain contact.
Stay on the line.

Drawing needs sustained tangential interaction in two directions. TacGB helps keep the marker on the board while reducing lateral wandering.

+66.5%relative ink coverage
−34.5%mean perpendicular deviation

20 rollouts per sensing condition. Coverage is inked length divided by projected stroke length. Straightness is measured by mean perpendicular deviation from the endpoint chord.

View the full drawing results +
Whiteboard drawing setup, observations, trajectories, ink coverage, and straightnessView full figure +
Whiteboard drawing: coverage improves while mean perpendicular deviation decreases by 34.5%.

BibTeX

@misc{li2026tacgoosebumpstacgbretrofittingnormalonly,
  title={TacGooseBumps (TacGB): Retrofitting Normal-Only Tactile Sensors with Shear Encoding for Learning Contact-Rich Manipulation},
  author={Wenjie Li and Binyu Yang and Yuxin Chen and Ambrose Wang and Masayoshi Tomizuka},
  year={2026},
  eprint={2609.34006},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2609.34006},
}
Download .bib