{"id":392,"date":"2026-06-09T19:38:52","date_gmt":"2026-06-09T11:38:52","guid":{"rendered":"https:\/\/www.modu-tech.cn\/index.php\/2026\/06\/09\/tactile-sensing-supports-embodied-intelligence\/"},"modified":"2026-08-10T13:24:26","modified_gmt":"2026-08-10T05:24:26","slug":"tactile-sensing-supports-embodied-intelligence","status":"publish","type":"post","link":"https:\/\/www.modu-tech.cn\/en\/tactile-sensing-supports-embodied-intelligence\/","title":{"rendered":"Gao Libo, Chief Scientist at Modulus Technology, was invited to participate in the GAIR LIVE online roundtable"},"content":{"rendered":"
\n
\n

ICRA 2026 sent a clear signal: Robotics is transitioning from a \u201cvision-dominated\u201d era to a new phase of \u201cvision-touch integration.\u201d<\/span><\/p>\n<\/section>\n<\/section>\n

\n
\n

Over the past decade, deep learning and computer vision have given robots \u201ceyes.\u201d However, when a robotic arm approaches an object and enters the microscopic range of direct contact, the visual paradigm begins to fail\u2014fingers block the line of sight, light and shadow constantly shift, and critical information\u2014such as whether the object is slipping or how much force to apply\u2014can no longer be obtained simply by \u201clooking.\u201d<\/span><\/p>\n<\/section>\n<\/section>\n

\n
\n

This distance is very short\u2014perhaps only one centimeter. But it is precisely this one centimeter that has become the bottleneck preventing many dexterous maneuvers from being successfully executed.<\/span><\/p>\n<\/section>\n<\/section>\n

\n
\n

Professor Gao Libo, Chief Scientist and Co-founder of Modulus Technology, was invited to attendGAIR LIVE<\/span>An online roundtable discussion provided a systematic overview of the technical approaches and industrial prospects for tactile perception. The following is a recap of the key points from the discussion.<\/span><\/p>\n

\"\"<\/p>\n<\/section>\n

\n
\n
\n
\n
\n
\n

Why Is the Sense of Touch So Important?<\/span><\/strong><\/p>\n

Professor Gao Libo believes that the statement \u201ctactile perception is the key\u201d should be interpreted with greater caution. It is not that vision is no longer important, but rather that when robots shift from \u201cobserving the world\u201d to \u201cchanging the world,\u201d relying solely on vision is no longer sufficient.<\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n

\n

Gripping soft objects, inserting precision parts, unscrewing bottle caps, picking up fine needles\u2014the success of these tasks depends not on \u201cwhat is seen,\u201d but on: Has contact been made? Is the contact force appropriate? Is the object slipping or deforming? How much force should be applied next?<\/span><\/p>\n<\/section>\n

\n

Most of this information cannot be captured by sight. More importantly, by the time sight finally detects an error, the consequences are often irreversible.<\/span><\/p>\n<\/section>\n

\n
\n

Professor Gao summarized it as follows: Vision is responsible for \u201cseeing\u201d and \u201canticipating,\u201d while touch is responsible for \u201cverifying\u201d and \u201cstabilizing.\u201d Any movement without feedback is an irreversible mistake.<\/span><\/strong><\/p>\n<\/section>\n<\/section>\n

\n

This means that the realm of AI learning is shifting\u2014from the virtual space of images, language, and mathematical symbols to the physical feedback of touch, friction, and deformation under force. This is not simply a matter of adding a sensor; it is a shift in the way we perceive the world.<\/span><\/span><\/p>\n<\/section>\n

\n

As for why the sense of touch is receiving unprecedented attention today, Professor Gao believes that three conditions have now come together.<\/span><\/p>\n<\/section>\n

\n
\n

First, flexible materials andMicro- and Nanofabrication Processes<\/span>With significant progress, the performance of domestically produced flexible sensors is now on par with that of foreign ones;<\/span><\/p>\n

Second, the demand for real-world physical interaction data among large models has grown dramatically, making haptic data an indispensable source;<\/span><\/p>\n

Third, as dexterous robots begin to be deployed in real-world operational scenarios, sensors must transition from laboratory specifications to system-level reliability.<\/span><\/p>\n

\n<\/section>\n<\/section>\n

\n
\n
\n
\n
\n
\n

What is the \u201cphysical limit\u201d of electronic skin?<\/span><\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n

<\/section>\n
<\/section>\n<\/section>\n
\n

When discussing the technical challenges of electronic skin, Professor Gao Libo candidly admitted that,<\/span>The greatest challenge lies not simply in increasing the number of pixels, but in ensuring that thousands of sensing units maintain their individual performance and stability even under complex mechanical deformation.<\/strong><\/span><\/p>\n<\/section>\n

\n

He identified three sets of core contradictions.<\/span><\/p>\n<\/section>\n

\n
\n

First, the trade-off between sensitivity and measurement range. Many transducers are extremely sensitive in the low-pressure range but are prone to saturation when subjected to extreme pressures.<\/span><\/p>\n

Second, the trade-off between flexibility and stability. Electronic skin needs to be thin, soft, and conformable, but the softer the material, the greater the creep, hysteresis,Mechanical Fatigue<\/span>...and the more pronounced their sensitivity to temperature and humidity becomes\u2014compared to silicon-based sensors, maintaining baseline stability over the long term is the primary challenge for flexible devices.<\/span><\/p>\n

Third, the conflict between single-point and array consistency. It is not difficult to achieve high performance with a single unit in a laboratory setting, but when thousands of units are integrated, packaging stress, manufacturing errors, and baseline drift are amplified to a level that cannot be ignored.<\/span><\/p>\n<\/section>\n<\/section>\n

\n

Another obstacle that must be overcome is crosstalk.<\/span>Electrical crosstalk manifests as parasitic capacitance and row-column scan interference; mechanical crosstalk, however, is more subtle\u2014when soft materials are subjected to pressure, stress is transmitted to the surrounding areas, causing fluctuations in the signals of adjacent cells. It is difficult to completely resolve this issue through back-end algorithmic correction alone.<\/span><\/p>\n<\/section>\n

\n

In response to the above challenges,<\/span>Modulus Technology has adopted a strategy that integrates materials, circuits, and algorithms.<\/span><\/strong>One of the key technologies is the introduction of a special \u201cspacer layer\u201d packaging structure, which ensures signal isolation between each sensing unit.<\/span>Currently, Modulus Technology has achieved nearly 400 integrated units per square centimeter and is working toward its goal of 1,000 units.<\/span><\/strong><\/p>\n<\/section>\n

\n

Professor Gao also emphasized that the industry should not focus solely on \u201chow many data points have been collected,\u201d but rather on whether it can generate a real, continuous, and repeatable spatiotemporal pressure field. The evaluation criteria for electronic skin are shifting from a single metric\u2014resolution\u2014toward high fidelity, low crosstalk, and consistency in large-scale mass production.<\/span><\/span><\/p>\n

\n<\/section>\n

\n
\n
\n
\n
\n
\n

How is tactile data processed?<\/span><\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n

<\/section>\n
<\/section>\n<\/section>\n
\n

This is an issue that is easily overlooked but extremely challenging from an engineering perspective.<\/span><\/p>\n<\/section>\n

\n

There is a fundamental difference between tactile and visual signals: Tactile data contains a large number of high-frequency transient signals\u2014such as minute slippage at the moment of contact, localized mechanical impacts, or vibrations\u2014that may last only a few milliseconds but are critical to grasping stability and safety control.<\/span><\/p>\n<\/section>\n

\n

When thousands of sensing units in a large-scale flexible array simultaneously output high-frequency raw waveforms and transmit them all to the CPU,Data Bandwidth<\/span>, communication latency, power consumption, and computational load will all pose significant challenges.<\/span><\/p>\n<\/section>\n

\n

Professor Gao Libo proposed that a reasonable architecture should be layered: the skin layer handles \u201crapid reflexes,\u201d while the central system handles \u201chigher-level understanding.\u201d Specifically, preliminary feature extraction is performed at the sensor end\u2014using lightweight algorithms to determine whether contact has occurred, locate the center of contact, and detect abnormal pressure gradients. The central system no longer receives raw voltage or resistance values, but rather \u201ctactile events\u201d that are close to the task\u2019s semantic meaning. This layered design elevates tactile signals from the \u201cmeasurement layer\u201d to the \u201csemantic layer.\u201d<\/span><\/p>\n<\/section>\n

\n

The haptic sensing system currently under development by Modulus Technology follows this approach: synchronously capturing flexibleVisual and Tactile Perception<\/span>, such as fabric pressure and hand movements; pressure and position features are first extracted locally, and these features are then fed, along with visual, joint motion, and proprioceptive information, into the higher-level model.<\/span><\/p>\n<\/section>\n

\n

Professor Gao further elaborated on the design philosophy of functional layering. The fingertips and high-frequency operation zones handle precision control tasks and require strong local computing power; large body areas primarily focus on force detection and safety responses, emphasizing low-power, low-precision solutions; and the central nervous system is responsible for object recognition, strategy generation, and cross-modal understanding. Each part performs its own specific function, rather than adopting a one-size-fits-all approach.<\/span><\/p>\n<\/section>\n

\n

\u201cOnly when tactile sensors no longer output simple waveforms, but instead output state variables and event information that robots can directly utilize, do they truly become part of embodied intelligence.\u201d<\/span><\/span><\/strong>Professor Gao said.<\/span><\/span><\/p>\n<\/section>\n

\n
\n
\n
<\/section>\n<\/section>\n<\/section>\n
\n
\n
\n
\n
\n

Industry Outlook: Opportunities and Challenges Coexist<\/span><\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n

<\/section>\n
<\/section>\n<\/section>\n
\n

Looking ahead to the next 3 to 5 years, Professor Gao Libo predicts that,<\/span>Tactile perception will move from the \u201cproof-of-concept phase\u201d to the \u201cscenario selection and large-scale validation phase.\u201d<\/span><\/strong>The focus of competition is no longer limited to a contest of sensitivity or resolution, but rather centers on which system can operate stably over the long term in real-world applications while offering acceptable maintenance and replacement costs.<\/span><\/p>\n<\/section>\n

\n

In terms of technical approaches, he believes there will not be a situation where a single solution dominates the market; instead, technology combinations will emerge based on specific application scenarios: fingertip operations require high precision, large areas of skin prioritize safety, and industrial assembly emphasizes durability. Optical, piezoresistive, capacitive, ion-electronic, and MEMS solutions will coexist for the long term.<\/span><\/p>\n<\/section>\n

\n
\n

China\u2019s strength lies in its complete industrial chain\u2014ranging from materials, molds, and flexible circuits to finished products\u2014combined with a wide range of application scenarios (3C, logistics, new energy, and service robots), resulting in extremely rapid engineering iteration.<\/span><\/strong><\/p>\n<\/section>\n<\/section>\n

\n

However, the shortcomings cannot be ignored either.<\/span><\/strong>In terms of dedicated chips, most systems currently still rely on discrete circuit boards for data acquisition; in the future, low-noise, multi-channel application-specific integrated circuits (ASICs) will need to be developed to achieve the integrated fusion of sensors and circuitry. Regarding high-end materials and processes, breakthroughs are still needed in highly stable sensitive materials, wear- and contamination-resistant packaging materials, and reliable interconnection processes for complex curved surfaces. In terms of standardization, the industry currently lacks unified standards and urgently needs to establish an automated calibration system and lifespan prediction models to form a complete closed-loop system spanning haptic hardware, data, model training, and interfaces.<\/span><\/p>\n<\/section>\n

\n

Regarding the hotly debated topic of \u201cin-house developed chips\u201d in the industry, Professor Gao\u2019s stance is pragmatic and clear: dedicated haptic chips are undoubtedly the way forward, but at this stage\u2014where various sensing approaches have yet to converge and technological iterations are occurring at a rapid pace\u2014general-purpose chips and programmable solutions offer greater flexibility. Only when application scenarios and production volumes stabilize will the advantages of dedicated chips in terms of power consumption, size, and real-time performance truly become apparent.<\/span><\/p>\n

\n<\/section>\n

\n
\n
\n
\n
\n
\n

In Closing<\/span><\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n

<\/section>\n
<\/section>\n<\/section>\n
\n

Only when vision and touch are deeply integrated under the large-model paradigm will the gap between robots and the physical world truly be narrowed to \u201cthe last centimeter.\u201d<\/span><\/p>\n<\/section>\n

\n

Breaking through this \u201cfinal centimeter\u201d is precisely the area where Modulus Technology continues to focus its efforts\u2014from material innovation to algorithm architecture, and from individual breakthroughs to system integration\u2014ensuring that tactile sensing is no longer confined to sensor spec sheets, but becomes a core capability that enables robots to perceive the physical world and perform dexterous tasks.<\/span><\/strong><\/p>\n<\/section>\n<\/section>","protected":false},"excerpt":{"rendered":"

ICRA 2026 sent a clear signal: robotics is transitioning from a \u201cvision-dominated\u201d era to a new phase of \u201cvision-touch integration.\u201d Over the past decade, deep learning and computing [\u2026]<\/p>","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2],"tags":[],"class_list":["post-392","post","type-post","status-publish","format-standard","hentry","category-company-news"],"_links":{"self":[{"href":"https:\/\/www.modu-tech.cn\/en\/wp-json\/wp\/v2\/posts\/392","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.modu-tech.cn\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.modu-tech.cn\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.modu-tech.cn\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.modu-tech.cn\/en\/wp-json\/wp\/v2\/comments?post=392"}],"version-history":[{"count":0,"href":"https:\/\/www.modu-tech.cn\/en\/wp-json\/wp\/v2\/posts\/392\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.modu-tech.cn\/en\/wp-json\/wp\/v2\/media?parent=392"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.modu-tech.cn\/en\/wp-json\/wp\/v2\/categories?post=392"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.modu-tech.cn\/en\/wp-json\/wp\/v2\/tags?post=392"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}