Who Builds the Sensors? Eight Profiles of the Visuotactile Field

The commercial spinout from the Adelson Lab at MIT, in many ways the company that defines the high-resolution end of the visuotactile market. GelSight Inc. has been shipping research and industrial-grade tactile sensors for over a decade and is the supplier most academic groups default to when buying off-the-shelf hardware.

Beyond robotics research, GelSight sells into industrial inspection — using the same sensing principle to image surface defects on tyres, banknotes, ballistics evidence and printed-circuit-board solder joints. That dual market gives the company a more stable revenue base than most pure robotics suppliers.

The research organisation behind DIGIT (2020), DIGIT 360 (2024) and ReSkin (2021), and through that the single biggest force in lowering the cost of tactile-sensing research over the last six years. Meta AI Research's strategy has consistently been open-source release: CAD, BOM, firmware and supporting libraries (PyTouch, Tacto) are public under permissive licenses.

The lab does not sell sensors directly; commercialisation goes through partner manufacturers. The strategic motivation is clear: make tactile sensing widely available so that the surrounding ML stack (which Meta does sell, indirectly via its compute and AI ecosystem) becomes the constraint.

NYU's robotics groups have become a centre of gravity for the magnetic-skin branch of visuotactile sensing. AnySkin, released in 2024, was designed to address the calibration and durability problems of earlier magnetic skin designs — the skin slips on like a fingertip cover and is intended to be effectively interchangeable across instances of the same robot hand.

The group is also active in dexterous-manipulation policy learning, often using its own AnySkin sensors as the tactile front-end. Several recent papers explore the question of how cheaply replaceable tactile skin changes the design space of long-horizon learning.

The home of TacTip and the European centre of gravity for marker-based optical tactile sensors. Nathan Lepora's group has produced a long line of TacTip variants — flat, dome, finger, even brush-like — and a body of work on tactile servoing, where the controller closes the loop directly on tactile features rather than reconstructed forces.

TacTip designs are open-source (CAD and BOM public) but also distributed in assembled form via Bristol partners. The lab plays a strong role in the European tactile-sensing community and in the Horizon Europe-funded robotics consortia.

Toyota Research Institute is the source of the Soft-Bubble sensor and a long line of work on robotic grocery handling, dish unloading and fragile-object manipulation. Where most visuotactile work miniaturises the sensor, TRI took the opposite design choice with Soft-Bubble: a large, inflated, transparent membrane with an internal camera, optimised for compliant whole-finger contact rather than fingertip precision.

TRI does not commercialise its sensors. Soft-Bubble's main impact has been to influence subsequent academic compliant-fingertip designs and to provide a reference benchmark for fragile-grasp tasks.

Sanctuary AI's Phoenix humanoid is among the most-discussed examples of a commercial humanoid platform whose hand design centres explicitly on tactile feedback. The company has publicly emphasised that its hands integrate multi-modal tactile sensing — including visuotactile and capacitive elements — for fine manipulation tasks beyond what a vision-only humanoid can do.

Sanctuary is included here not because it sells sensors but because it represents the kind of integrator that drives downstream demand for the rest of the field. Apptronik, Figure and the major Chinese humanoid programmes occupy a similar position with varying degrees of public disclosure.

StretchSense is not a visuotactile vendor in the strict sense, but its capacitive stretch sensors and mocap gloves cover the closely related niche of soft, wearable tactile sensing. The company's gloves are widely used in entertainment-industry motion capture and increasingly in robotics teleoperation, where they provide finger-pose data that complements robot-side visuotactile feedback during demonstration recording.

Inclusion here reflects the practical reality that visuotactile sensing is rarely used in isolation — demonstration data and robot-side tactile streams typically co-exist in modern manipulation pipelines.

SynTouch's BioTac was, for over a decade, the premium reference for multi-modal tactile sensing in research robotics — combining electrode arrays in a conductive fluid, vibration sensing and thermal flow into a single fingertip. It is not a visuotactile sensor (no internal camera) but is included for context: the bar SynTouch set on multi-modal tactile sensing is the bar contemporary visuotactile designs measure themselves against.

BioTac sensors have been used in prosthetics research, surgical robotics and dexterous-manipulation papers since the early 2010s. Public availability and pricing have varied over the years; current status should be checked directly with the company.