Vision-based tactile sensors face overvaluation, says Quantum Bit analysis
A Quantum Bit analysis argues that vision-based tactile sensors, despite their popularity, face low technical barriers, poor durability, and unresolved algorithmic issues, making their investment value questionable.
The analysis traces the technology to the mature vision industry, noting that VBTS still relies on small CMOS camera modules, which are readily available from suppliers. Hardware prototyping is extremely simple, requiring only a miniature camera, a light source, a transparent elastomer, and marker dots. The algorithms used in the field are largely borrowed from computer vision, such as ResNet, U-Net, and optical flow, with few original contributions to tactile physics. This “borrowing” accelerates paper output but means that the domain has made almost no independent algorithmic breakthroughs.
Low barriers and high hardware reusability have attracted dozens of startups worldwide, all developing products based on the same principles and open-source algorithms. The resulting sensors show negligible differences in core performance. According to the report, the sector remains fragmented, with no player yet possessing technological representativeness or market dominance. From an investor’s perspective, a track with non-proprietary core technology and an easily copied business model offers poor long-term return predictability.
The report then details the engineering obstacles for industrial deployment. Because of the minimum focal length constraint, lens-based VBTS modules are typically thicker than 10mm, making them difficult to integrate into small dexterous hands. The rigid form cannot conform to curved surfaces, so the technology can only serve as a local fingertip sensor, not as a distributed electronic skin for whole-body sensing. When extending to a dual-hand, five-finger configuration with nearly 30 modules, the system requires over 160 TOPS of computing power, an external server, and power consumption exceeding a hundred watts. This breaks the millisecond-level latency required for compliant control, and the dense wiring cannot pass through the narrow finger roots, creating a critical bottleneck.
The article also highlights a material trade-off. The sensitivity of VBTS depends on the softness of the elastic body: softer materials produce clearer deformation images but age quickly; harder materials last longer but lose the ability to capture fine texture. The report calls this a zero-sum game, noting that industrial environments require durability that directly conflicts with the optical sensing principle, leading to continuous performance drift and poor long-term reliability.
Finally, the report argues that the core of dexterous manipulation is force, not texture. VBTS compresses three-dimensional deformation into a two-dimensional image, which is an ill-posed inverse problem. The effective spatial resolution is not equal to the pixel count; it is constrained by viscoelastic hysteresis, the multi-solution nature of inverse algorithms, and the limits of force decoupling. These issues cannot be eliminated by increasing pixel numbers.
The article concludes that the popularity of VBTS is built on visual algorithm spillover and low hardware thresholds rather than a fundamental breakthrough in tactile measurement. It recommends that investors direct capital to players who can truly achieve full-stack core technology and overcome physical limitations, implying that the current VBTS route is overvalued.