1.  How Vision-Based Tactile Sensing Works

VBT sensors replicate — in miniature — the photographic workflow. A soft elastomer gel coated with a reflective or marker layer is pressed against an object. Internal LEDs illuminate the contact zone, and a small camera captures the resulting light-and-shadow pattern as the gel deforms. An image-processing algorithm then reconstructs the full tactile state: normal force to millinewton resolution, shear force vector, slip direction, contact pose and surface microgeometry.

The sensor hardware comprises four functional modules: a contact module (the gel layer), an illumination module (typically multi-colour LEDs), an image-capture module (miniature camera), and an information-processing module (embedded compute running the reconstruction algorithm). The optical signal path makes the device inherently immune to electromagnetic interference — a meaningful advantage in welding stations, motor-assembly lines and nuclear environments where capacitive and resistive sensors perform poorly.

Three limitations remain active engineering challenges. First, the optics, illumination path, and gel layer together produce a sensor that is physically larger than competing technologies — a critical constraint for fingertip integration. Second, flexible gel materials wear and contaminate, degrading image quality over extended duty cycles. Third, continuous image reconstruction is compute-intensive, adding latency and on-board processing requirements.

1.1  Six Technical Milestones

The technology has matured through a series of research advances spanning roughly four decades:

Table 1: Key VBT milestones. Sources: academic literature; Kaiyuan Securities research (August 2025

2.  Why Tesla's Dexterous Hand Is Likely to Use VBT

Tesla's Gen2 Optimus dexterous hand carries 22 degrees of freedom across both hands (11 per hand), with tactile sensors installed across all fingers. The robot's broader perception architecture is built on a camera-first / vision-only philosophy, a strategy carried over from the Full Self-Driving programme, and extended to embodied AI through end-to-end VLA models trained on visual and action data.

Four attributes make VBT the most natural fit for that architecture:

(i)  High-dimensional contact perception

Most conventional tactile sensors measure only normal (perpendicular) force. VBT sensors simultaneously resolve normal force, shear force, slip vector, object pose, surface texture, and hardness — a dimensionality approaching the human fingertip. This enables manipulation capabilities that are impossible with pressure-array sensors alone, including stable egg-grasping (balancing grip force against shell fragility in real time) and multi-object dynamic transfer with mid-air regrasp corrections.

(ii)  Native AI compatibility

VBT output is an image. That is not an analogy: the sensor literally produces JPEG-compatible frames that can be concatenated with RGB camera feeds in a standard tensor pipeline. State-of-the-art VLA models — which accept image, language and action-trajectory inputs — can therefore ingest tactile data without any modal-conversion layer. Training data generation, domain adaptation and model fine-tuning all proceed with existing computer-vision toolchains.

(iii)  Cost and integration economics

GelSight's commercial VBT sensor (the DIGIT, co-developed with Meta) lists at USD 300–350 per unit — a fraction of the price of a six-axis force-torque wrist sensor. Integration time from unboxing to first data is approximately five minutes; the DIGIT SDK supports ROS, Python and C++ across standard Linux distributions, with documented compatibility across several major robot platforms.

(iv)  Environmental robustness

Optical signal propagation is immune to the interference sources that degrade electrical-signal sensors: temperature fluctuation (which causes resistance drift in piezoresistive types), surrounding conductive objects (which distort edge-capacitance in capacitive types), and electromagnetic fields (relevant in welding, motor assembly and medical-imaging environments). VBT sensors have been demonstrated operating reliably in metallurgical high-heat zones, cryogenic logistics warehouses, welding bays, motor-assembly lines and proximity to nuclear plant equipment.

3.  Tactile Sensor Technology Comparison

Six principal sensing modalities are commercially deployed or in active development for robotic tactile applications. The table below summarises operating principles, key trade-offs and typical deployment contexts:

Table 2: Tactile sensor modality comparison. Compiled from multiple sources including Kaiyuan Securities (August 2025) and published literature.

4.  Global Competitive Landscape

The VBT market is led by US incumbents but is being actively contested by a cohort of Chinese start-ups. The following table covers the ten most relevant organisations, including two non-VBT companies included for competitive context:

Table 3: Global VBT and multimodal tactile sensing companies. Sources: company websites, Kaiyuan Securities (August 2025), public filings. Note: SynTouch is included for historical context; the company ceased trading circa 2022.