On May 27, Shanghai NeoteAI Intelligent Technology Co., Ltd. (hereinafter referred to as "NeoteAI") announced the completion of its angel-funding round totaling nearly RMB 100 million. This round was co-led by Shanghai Science & Technology Venture Capital (Group) Co., Ltd. (hereinafter referred to as "Shanghai STVC Group") under Shanghai State-owned Capital Investment Co., Ltd. (SSCI) and Fudan Innovation & Entrepreneurship, with participation from Ceyuan Venture Capital Fund under Shanghai STVC Group. Following the funding, NeoteAI will continue to expand the scale of tactile data collection and accelerate the training and validation of tactile-based embodied models in real-world operational tasks.
NeoteAI originated from the Institute for Trustworthy Embodied AI, Fudan University. As a major outcome of the strategic cooperation between Fudan University and Jing'an District, NeoteAI received strong incubation support from its inception by the Science, Technology, and Economy Commission of Jing'an District, in collaboration with Shanghai Shibei Hi-Tech (Group) Co., Ltd. It has secured strategic funding from Jing'an District and special funds for high-quality industrial development from the Shanghai Municipal Commission of Economy and Informatization. Furthermore, the Science and Technology Commission of Shanghai Municipality provided special funding for the R&D of its core visuo-tactile sensors, accelerating the company's technological breakthroughs.
At this critical stage, as embodied intelligence transitions from technical validation to industrial application, this startup is making its breakthrough in tactile sensing.
Robots don't need more eyes; they need a sense of touch.
What is tactile-based embodied intelligence? Over the past two years, resources and research in embodied intelligence have been overwhelmingly concentrated on visual perception, focusing on teaching robots to see the world, including recognizing cups, screws, and even aligning with USB connectors. But when it comes to actual manipulation, things often go wrong. A robot might line up a plug perfectly but fail to master the force required for insertion; it might pick up a garment but remain oblivious to the fabric slipping or where the tension is too high.
Contrast this with how humans perform the same tasks: as our fingers touch the port edge, our wrists naturally micro-adjust based on resistance feedback; when we pick up a paper cup, our fingertips modulate pressure based on the cup's deformation; when smoothing fabric, our hands dynamically shift as we sense changes in tension. These judgments are almost subconscious—what we commonly call a "sense of touch." This is, precisely, the missing link in today's robotic perception.
Whenever a robot interacts with the physical world, tactile information becomes an indispensable core element of perception. NeoteAI believes that embodied intelligence is transitioning from a vision-centric paradigm to a visuo-tactile bimodal perception framework. While vision handles global positioning and semantic understanding, tactile perception provides physical feedback and enables dynamic adjustments upon contact.
Their approach elevates a fundamental physical problem to the core of robot training: Robots must go beyond seeing the world—they need to feel it, and in that moment of contact, recognize what they've touched, assess if it's right, and decide what to do next.
However, deploying tactile perception is far from simply adding a sensor. If tactile information remains limited to isolated hardware readings, it cannot be effectively transformed into robotic manipulation capabilities. NeoteAI's primary objective is to convert tactile perception into standardized, trainable data and deeply embed it across the entire pipeline for model training and inference.
At NeoteAI, the core team comes from the Institute for Trustworthy Embodied AI, Fudan University, and brings a deep-seated culture of industry-academia-research integration.
CEO Zhao Shihao earned his bachelor's and master's degrees from Fudan University and his PhD from the University of Hong Kong. He served as a core researcher at Microsoft Research and Alibaba's Tongyi Lab, with deep experience in the R&D of cutting-edge models, spanning video world models and generative models. Chief Scientist Wu Zuxuan, Vice President of the Institute for Trustworthy Embodied AI, formerly worked at Meta, with extensive expertise in core areas such as video and multimodal models. COO Dong Daoguo is a multidisciplinary talent with both academic and industry experience. With nearly 20 years in the industry, he was the chief architect of the first-generation Honor Magic released by Huawei. He is currently a research fellow at the Institute for Trustworthy Embodied AI, playing a key role in driving and de-risking the commercialization of the company's technologies.
Building on the team's deep expertise in multimodal models, NeoteAI has established three core pillars: visuo-tactile sensors, a high-precision data acquisition platform, and a large tactile-embodied model, forming a complete technical closed loop.
Capturing tactile signals first
NeoteAI's first step in deploying its technology is the high-precision acquisition of contact information from robotic end-effectors. As the inaugural landmark achievement of industry-academia-research integration with Fudan University, the company's independently developed visuo-tactile sensors are compatible with a wide range of end-effectors, from industrial grippers to dexterous hands, and can accurately capture multi-dimensional physical data during contact, including force, slip, deformation, and boundary contours.
The core of this technical approach lies in the visuo-tactile perception paradigm. Traditional piezoresistive and capacitive tactile sensors are typically restricted to providing single-point force feedback and are unable to fully reconstruct critical information about the contact interface, such as geometric contours, slip directions, surface textures, and deformation boundaries. More importantly, the performance ceiling for such solutions is essentially fixed at the hardware manufacturing stage, making it difficult to enhance capabilities through subsequent algorithmic iterations.
In contrast, visuo-tactile technology is built on a novel perception paradigm, offering two core advantages:
First, it offers significantly higher information density, enabling comprehensive contact characterization that traditional sensors cannot provide. Second, its output data format is highly compatible with visual data, facilitating seamless integration with existing Transformer architectures. This compatibility dramatically lowers the technical barrier to incorporating the tactile modality into pre-existing embodied models.
Building a "Tactile Data Factory" to enable robotic haptic memory
However, sensor breakthroughs constitute only the initial step toward tactile embodied intelligence. Unlocking the true potential of tactile perception requires the support of large-scale, high-quality tactile interaction data.
To address the industry-wide pain point of tactile data scarcity, NeoteAI has established a specialized tactile embodied data collection center exceeding 1,000 square meters.
Data collection is centered on precision manipulation scenarios. These include: contact-based operations (e.g., USB plugging, screw driving, and RAM installation), tasks demanding high-precision force control (e.g., wiring harness assembly, thin-walled container grasping, and elastic component handling), and the manipulation of deformable objects (e.g., fabric smoothing, paper folding, and tape application).
In its approach to data scale, NeoteAI employs a cross-industry comparative perspective. While current data volumes in the embodied AI field are in the tens of millions, language models start at billions or even tens of billions. Consequently, the company has prioritized large-scale data accumulation at this stage, following a development path of "scaling up first, optimizing costs later." At present, all collected data is prioritized for internal model training. Once the data system and collection processes mature, the company will move on to explore commercial models such as data services.
How tactile data makes robots smarter?
The ultimate value of data lies in its deep integration into model training and inference pipelines, where it is translated into the robot's operational capabilities.
NeoteAI aims to integrate tactile sensing into pre-trained large embodied models (Vision-Touch-Language-Action (VTLA) and Tactile World Model). By combining this integration with reinforcement learning pathways that incorporate tactile modalities, the company seeks to systematically build haptic-enabled embodied intelligence and achieve significant breakthroughs in multiple precision manipulation tasks.
Traditional Vision-Language-Action (VLA) models rely on visual and linguistic inputs to generate actions, but they often struggle with delicate operations due to perceptual blind spots.
By integrating tactile (touch/force) feedback into its proprietary VTLA (Vision-Touch-Language-Action) architecture, NeoteAI enables the real-time perception of contact states, such as clamping, slippage, positioning, and deformation, thereby precisely guiding task execution. It is much like a person plugging in a charging cable with its eyes closed, relying entirely on haptic feedback: if misaligned, it retracts slightly, adjusts the angle, and lightens the touch.
World models are required to learn the causal relationships between actions and resulting environmental changes. However, vision-only approaches struggle with fine-grained tasks such as handling flexible materials or precision assembly. NeoteAI's tactile world model addresses this lack of physical information, thereby significantly boosting the success rate of delicate operations. In reinforcement learning, touch serves as the signal for real-time action correction. The model leverages real-time touch signals, such as abnormal resistance, unstable clamping, or slip, to adjust its next actions.
Validating it in the factory first
Commercially, NeoteAI's first stop is the factory. The rationale is straightforward: factory tasks are clearly structured, their outcomes are quantifiable, generalization demands are lower than in domestic settings, and the benefits of tactile feedback are more readily validated.
Wire harness assembly, RAM insertion/removal, flexible material handling, and home textile operations are typical examples of fine-grained operational scenarios. These tasks have long relied on human workers, not because visual positioning is insufficient, but because the state transitions after contact are too complex to model.
Shanghai's robust industrial ecosystem provides ideal conditions for the deployment of this technology. The region's highly clustered industries, such as automotive, 3C electronics, and home textiles, involve numerous fine-grained operational scenarios. This aligns perfectly with NeoteAI's strategic focus on data collection and technical validation. Currently, the company has secured Proof of Concept (POC) orders across several industrial niches.
A strategic opening for latecomers
NeoteAI maintains a clear competitive outlook. The team's analysis, from the perspective of a latecomer, is that the vision-only approach, after years of rapid advancement, has now entered a phase of resource-intensive competition. As R&D costs continue to soar, the inherent limitations of this approach in fine-grained operations have also become increasingly prominent.
Merely following the pioneers' path rarely leads to breakthroughs. Sustainable competitive advantage is built by identifying unresolved, core industry challenges and introducing novel technical variables.
NeoteAI's strategic vision is to elevate tactile sensing from an optional modality to a standard, essential feature for robots. Task understanding is merely the entry ticket for embodied intelligence; the true competitive watershed in the next phase will be the ability to make real-time adjustments during physical interaction to ensure reliable task completion.