SciGroveBeta
Robotics

FELT: Generating Tactile Signals from Vision for Visuo-Tactile Manipulation

Zinan Li, Yiyang Ling

Featured July 25, 2026

Original Paper

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

A new AI system called FELT learns to imagine how things feel just by looking at them, helping robots do tricky tasks like picking up cups without needing real touch sensors.

In depth
The paper introduces FELT, a framework that synthesizes per-finger pressure tactile images from RGB observations, addressing the scarcity of tactile data in robotic manipulation. It leverages a frozen visual encoder and a lightweight query decoder with separate branches for each finger, respecting the physical sensor topology. This allows for augmenting vision-only datasets with generated tactile information, enabling sensor-free deployment of visuo-tactile policies.

Key Takeaways

  • 1
    FELT synthesizes physically consistent per-finger tactile images and latent features from RGB input, reducing the need for real tactile sensors during policy training and deployment.
  • 2
    The architecture incorporates sensor-topology-aware design, using separate decoder branches for each finger and a gated cross-panel exchange module to model asymmetric contact patterns.
  • 3
    Generated tactile signals and latent tactile features significantly improve policy success rates in contact-rich manipulation tasks compared to vision-only baselines, often matching or exceeding real tactile sensing performance in low-data regimes.

Conceptual Flow

HIGH LEVEL
1
Methodology: How FELT Generates Touch from Sight

The system looks at a picture, then uses a smart brain to guess what the robot's fingers would feel, making separate guesses for each finger.

Robot Camera View
Imagine Touch
Left Finger Feel
Right Finger Feel
2
Results: Better Robot Performance

By imagining touch, the robot can do difficult tasks much better, sometimes even as well as if it had real touch sensors.

Robot Sees Only
Robot Sees + Imagines Touch
Robot Sees + Real Touch
Do Tricky Task
Low Success
High Success
High Success

This breakdown was generated by SciGrove. Get the same analysis — intuition, storyboard, peer review, a runnable prototype and a glossary — on any paper you upload or paste a DOI for.