Summary
DexSkin is a soft, wearable capacitive electronic skin presented as a CoRL 2025 Oral paper that covers nearly the entire surface of gripper fingers, rather than just a small fingertip patch like conventional robotic tactile sensors. This high-coverage sensing allows robots to detect not only fingertip contact but also where on the sides of a finger an object touches—a capability essential for tasks such as in-hand object rotation. The work also introduces a cross-instance calibration scheme: because manufacturing tolerances cause different DexSkin units to produce inconsistent readings, the proposed calibration enables tactile models and data-driven policies (imitation learning, reinforcement learning) trained on one sensor to transfer to another. Online reinforcement learning experiments validate the approach's feasibility. The authors' core insight is that the biggest bottleneck in robotic manipulation is not the 'brain' but the 'skin'—robots need full-surface tactile perception akin to human skin. This article summarizes DexSkin's design, the calibration contribution, and its implications for tactile-driven manipulation.
Human skin provides tactile sensing across the entire body, while robotic tactile sensors typically cover only a small patch of the fingertip. DexSkin, accepted as a CoRL 2025 Oral paper, aims to change this—a soft, wearable capacitive electronic skin that covers nearly the entire surface of gripper fingers.
This means a robot can sense not only fingertip contact, but also *where* on the sides of its fingers an object is touching—critical for tasks like in-hand object rotation. Traditional fingertip-only tactile sensing is like feeling the world blindfolded with just your fingertips: you can touch things, but you don't know where they are in your hand.
A second key contribution is cross-sensor-instance calibration. Different DexSkin instances produce inconsistent readings due to manufacturing tolerances, and the paper proposes a calibration scheme that allows models to transfer between sensors. This means data-driven methods (imitation learning, RL) can be trained on one sensor and deployed on another. Online reinforcement learning experiments also validate the feasibility of the approach.
*Core insight*: The biggest bottleneck in robotic manipulation is not the "brain" but the "skin"—robots need tactile perception distributed across their whole body, just like humans.
Reference: *DexSkin: High-Coverage Conformable Robotic Skin* / CoRL 2025 Oral
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/177619968