Saigan Technology completes a Pre-A++ funding round of tens of millions of yuan, with its iontronic flexible tactile sensor landing in automotive manufacturing

Author | Huang Nan

Editor | Yuan Silai

Hardware Kr has learned that flexible tactile sensor company Saigan Technology has completed a Pre-A++ funding round of tens of millions of yuan, exclusively invested by the Guangzhou Industrial Control Donggao Fund. The funds will be mainly used for R&D of iontronic flexible technology and the iteration and implementation of a new generation of products, to continuously advance cross-industry customer validation and industrialized market expansion, and accelerate penetration into multiple scenarios.

Among them, Guangzhou Industrial Control, as an industrial investor, is simultaneously bringing the team a deterministic entry point for scenario implementation. The first batch of orders has now been finalized and will be implemented in the automotive manufacturing segment. This also means that the application boundaries of tactile sensors are extending outward from the embodied intelligence track, gradually penetrating into diverse scenarios such as industrial quality inspection and safety perception.

Saigan Technology was founded in 2023, with technology originating from the Ultra-Flexible Electronics Laboratory of Southern University of Science and Technology. The company focuses on the R&D, production, and sales of iontronic flexible tactile sensors and robotic electronic skin, and has completed a full-chain technology layout from materials and devices to systems.

Founder and Chief Scientist Guo Chuanfei is a tenured professor in the Department of Materials Science and Engineering at Southern University of Science and Technology and a recipient of the National Science Fund for Distinguished Young Scholars, with deep research in flexible electronics and more than 40 patent applications; CEO Xiong Gengchao has years of industrialization and operational experience within the Haier system; the team brings together more than 20 doctoral and postdoctoral research backbone members.

A dexterous hand equipped with Saigan Technology tactile sensors (Image source/Company)

The position of touch in embodied intelligence has completed a cognitive shift from "optional" to "mandatory." "Last year we still had to explain to customers why robots need touch, but after March this year, this is no longer a question that needs to be argued. The real divergence has shifted to another level: how exactly tactile signals are put to use," said Xiong Gengchao, CEO of Saigan Technology.

The real engineering challenges have only just begun. One fact that must be acknowledged is that touch and vision are not consistent on the time scale; visual update frequencies are usually lower, while tactile feedback requires higher frequencies to support closed-loop control. Simply splicing tactile information into vision-dominated VLA models can easily cause frequency mismatch, and touch may not only fail to provide gains but may instead interfere with original decisions.

Touch is not an auxiliary channel for vision; it requires an independent data stream, processing frequency, and model interface. This means that when tactile signals enter embodied foundation models, it is not as simple as adding an input channel; rather, the problem of compatibility with the visual time scale while preserving the high-frequency characteristics of touch must be solved at the data pipeline level.

To ensure that the tactile signal itself is sufficiently stable and usable, the Saigan Technology team starts from the device structure. Its core iontronic interface capacitance technology generates sensing signals through changes in the contact area between electronic conductors and ionic conductors, enabling simultaneous improvements in sensitivity, range, and linearity that differ from traditional routes.

Xiong Gengchao told Hardware Kr that traditional flexible sensors have long been trapped by the structural contradiction that high sensitivity, large range, and high linearity are difficult to achieve simultaneously, while the iontronic solution can maintain a linear response across the full range from milligram-level perception to hundred-kilogram-level load bearing.

In terms of specific indicators, Saigan's single-point sensors and array electronic skins both feature low detection limits, high sensitivity, large range, and high linearity: the detection limit is as low as about 1 milligram, the response frequency can reach 1 KHz, and the range can maintain a linear response at 1 MPa. This level can still be maintained in the form of a tactile glove, while device thickness and wearing burden are significantly lower than traditional solutions that require bulky structures for isolation.

Iontronic flexible tactile technology (Image source/Company)

"We found that the indicators customers list in requirement documents, such as tangential force, friction, centripetal force, and temperature, are not low, but when they truly enter real interaction, acceptance often first lands on stability and consistency," Xiong Gengchao pointed out. The engineering value of tactile sensors lies not in how high a single peak indicator is, but in whether data across devices and batches are comparable. "If cross-device consistency is insufficient, the larger the data scale, the higher the cleaning and alignment costs on the model side, and the training value is instead diluted."

This performance boundary further propagates to the delivery strategy of the Saigan Technology team. The company has not focused its competition on continuously raising the upper limit of single-point precision, but instead places stability and consistency as constraints in commercial design, optimizing precision performance on the basis of repeatable and calibratable devices.

Accordingly, Saigan Technology delivers in the form of fingertip modules integrating sensors, boards, and packaging, with factory calibration completed, so commercial customers can directly add the tactile module to the body without device assembly, calibration, and debugging. By moving part of system integration and consistency control forward to the device end, it reduces downstream customers' engineering coupling and cross-device data alignment costs.

A more noteworthy change is taking place in the physical coverage of touch. Xiong Gengchao told Hardware Kr that in the past, orders from embodied companies were mainly concentrated on dexterous fingertips. Since the second half of the year, tactile demand has simultaneously emerged for finger pads, palms, arms, and even shoulders, torsos, and soles. The full-size humanoid robot electronic skin system that Saigan Technology will launch in mid-2026 has already extended touch from the fingertips to the whole body, covering pressure perception, collision detection, and gait feedback.

Saigan Technology flexible tactile sensors (Image source/Company)

As flexible tactile technology moves from single-point fingertip sensing to full-body electronic skin deployment, sensors must again confront a whole new set of physical constraints. As the degrees of freedom of dexterous hands continue to increase, the space inside the fingers is continuously being squeezed by motors and transmission structures, leaving increasingly tight installation space for traditional rigid-substrate sensors. At the same time, humanoid bionic forms impose requirements on the sensing layer for curved-surface conformity and stretchability, and signal drift during bending and stretching deformation must remain within a controllable range.

Saigan Technology's flexible skin solution is designed precisely for this set of constraints: it supports direct attachment to curved surfaces, rather than reserving flat slots in a mold and then embedding sensors. This capability means the body structure does not need to reserve separate mounting positions for sensing units, greatly improving the flexibility of body iteration and curved-surface coverage, and enabling rapid deployment of full-body tactile perception across humanoid robots of different configurations.

"Touch, like vision, is becoming a foundational perception channel for embodied intelligence, but the two are not a substitute relationship—they are complementary," said Xiong Gengchao. Vision lets robots see the world; touch lets robots feel the world. In the main line of evolution from VLA to VTLA, the VTLA model is not limited to simply layering touch into the visual architecture; the key is to make the two kinds of information collaborate at the correct frequency and semantic level.

The position Saigan Technology seeks to occupy is that, building on iontronic materials and device processes, it provides system-level solutions ranging from fingertip modules and full-body electronic skin to tactile data streams, while using non-embodied scenarios such as automotive manufacturing and battery safety to validate the engineering boundaries of touch under extreme loads and long-term reliability.

As touch moves from "whether it is useful" to "whether it is easy to use," what determines whether a perception solution can be scaled up and deployed is not just sensitivity figures, but device consistency, delivery cadence, and the depth of coupling with the downstream model layer.

The following is an excerpt from an interview between Hardware Kr and Saigan Technology CEO Xiong Gengchao (edited):

Hardware Kr: As touch moves from the fingertip to the whole body, customer demand is shifting from single-point precision to stability and consistency. How do you understand this change?

Xiong Gengchao: These two directions are actually two sides of the same problem.

The true engineering value of a tactile sensor lies not in the peak precision of a single sampling, but in whether data is comparable across devices and across batches. Once consistency is insufficient, the larger the dataset, the higher the cost of data cleaning and alignment on the model side, and the training value of the data itself is diluted. The industry has already reached consensus on basic metrics, such as a touch-detection capability of 5 to 10 grams and a 30 N range; dexterous hand loads are mostly in the 3 to 5 kg range, and the stability gap between different solutions is significant.

Saigan's approach is to set stability and consistency as preconditions for commercialization: first ensure devices are reproducible and calibratable, then optimize precision. For delivery, it uses fingertip modules integrating sensor, board, and packaging; some products are calibrated at the factory and can be integrated directly by customers, saving them the work of assembly and calibration. By moving integration and consistency control upstream, it reduces downstream engineering coupling and the cost of data alignment across multiple devices.

And when touch extends from the fingertip to the finger pad, palm, shoulder and arm, torso, and sole, sensors face another set of constraints. In shoulder scenarios, for example, the core need is not heavy load, but light-contact recognition in human-robot interaction; moreover, requirements for range, sensitivity, flexibility, and cost vary across the whole body, so the same sensor cannot simply be used to cover everything. On the other hand, the interior space of dexterous hands is squeezed by motors and transmission structures, making rigid-substrate solutions difficult to place; the humanoid body also requires the sensing layer to conform to curved surfaces and be stretchable, with signal drift under deformation kept controllable.

Saigan's flexible skin solution can be attached directly to curved surfaces without reserving mounting slots during mold opening, decoupling the sensing unit from the body structure and making it convenient to rapidly deploy full-body touch on robots of different configurations.

Simply put, stability and consistency determine whether tactile data can be effectively used by models, while flexible skin determines whether tactile perception can be deployed over large areas on the robot body. Together they form the engineering foundation for touch to move from single-point fingertips to full-body electronic skin.

Hardware Kr: In the VTLA model, why can't touch serve merely as an auxiliary input channel for vision, and what is the core significance of hierarchical processing of tactile data?

Xiong Gengchao: Touch and vision differ fundamentally in their perception timing and response logic.

Vision leans toward environmental perception, with a relatively limited acquisition frame rate, serving slower cognitive decisions such as spatial understanding and scene reasoning; touch, however, is contact-based perception, and interaction events are often transient, requiring millisecond-level high-speed feedback to support physical closed-loop control. If tactile signals are treated as an additional input and stuffed into a vision-dominated VLA model, timing mismatch easily occurs: high-frequency tactile events are diluted by synchronization with low-frequency visual frames, which not only fails to leverage the unique interaction value of touch, but the extra noise introduced also interferes with the inference stability of the original model.

So connecting touch to a multimodal model is not about adding an auxiliary signal; it is about building an independently adapted data pipeline that both achieves temporal alignment and can collaborate with the visual information stream in the model's semantic space, while fully preserving the native high-frequency transient features of touch—without flattening the dynamic signals of the contact instant just to match the visual rhythm.

At the delivery level, Saigan's solution, on the one hand, is compatible with customers' existing data interfaces, ensuring smooth access for existing VLA systems; on the other hand, it outputs a standardized tactile data stream oriented to the native VTLA architecture, adapting to a unified multimodal perception framework.

What we clearly see is that downstream demand has become explicit. High-frequency reflex actions requiring immediate response, such as touching, tapping, and slipping, are not suitable for being entirely uploaded to the large model's main brain for inference. Such events have extremely short fault-tolerance windows; only on-device local recognition and local closed-loop control can meet the real-time requirements of robot physical interaction. This has given rise to a hierarchical tactile data processing architecture: low-frequency interaction decisions that require comprehensive understanding of environmental information are left to the VLA large model; high-speed closed-loop logic such as millisecond-level contact reflexes and safety protection is pushed down to the edge for execution.

Touch must first become an independent, controllable information stream with its own temporal characteristics, rather than an auxiliary supplement to vision, in order to form true complementary collaboration with vision in the VTLA system, rather than simple concatenation.