By Theodore Li 
By the Latest Robot News editorial team. Adapted from a report originally published by Jijia Juexing (机甲觉醒) on 26 October 2026.
Embodied AI has spent two years arguing about the brain and the body. The next contested surface is skin.
MoSense (模感科技), a Shanghai-registered company building flexible multimodal sensors and electronic skin, has closed an angel-plus round in the hundred-million RMB range. CDH Investments and Oasis Capital led; Unitree-adjacent investors and robotics supply-chain funds followed, including funds linked to the embodied AI companies that will eventually buy the sensors. Existing backers Sequoia China, Gaorong Ventures and AgiBot participated again — the same three that put several tens of millions of RMB into the company’s angel round in July 2026.
Two rounds in four months is not unusual in 2026. What is unusual is the investor mix: a sensor company raising from the robot makers it sells to. That is the tell. Whole-body tactile is moving from research demo to bill of materials.
Why touch became the bottleneck
Cameras tell a robot where an object is. They do not tell it whether the grip has closed, whether the object is sliding, or how much force is being applied at the contact patch. Every manipulation failure that survives good vision fails at contact.
The physical argument is simple. A vision-only policy can plan a grasp that looks correct from three camera angles and still crush a carton, drop a glass, or miss that the foot has landed on gravel rather than concrete. Contact state is not observable from outside the contact.
MoSense’s bet is that this gap needs hardware at many points on the body, not one sensor at the fingertip: chest, sole, palm, finger pad, and the floor the robot walks on.
| Body location | What the robot needs to know | Consequence of missing it |
|---|---|---|
| Chest / torso | push, press, collision from a human | unsafe human-robot interaction, no compliant yielding |
| Sole | pressure distribution, slip, friction, temperature | falls, unstable gait, no terrain adaptation |
| Palm / grip | grip force, slip tendency, object weight | crushed or dropped objects |
| Fingertip | micro-contact, 3-axis force, texture | no fine assembly, no surface identification |
| Floor | contact position, trajectory, load field | gait data cannot be validated |
The team: HKUST, Imperial, Zhejiang
The founding group is unusually academic for a hardware startup, and unusually commercial for an academic one.
Yan Zhaoxu (严朝旭), co-founder and CEO, holds an interdisciplinary PhD in microelectronics from HKUST, with work spanning RF electromagnetic fields, advanced electronic devices and multimodal sensing. He has published in IEEE TCPMT, Electronics, ICRA and APLM, worked on focused electromagnetic wireless charging, and has run industry research, strategy and financing at technology and financial firms.
Zhou Hang (周航), co-founder and CTO, took his undergraduate degree at Imperial College London and a robotics PhD at HKUST, working on embodied intelligence, autonomous driving and multimodal sensor fusion. He was a senior algorithm researcher at Desay SV in Singapore, on perception for autonomous driving.
Yang Mujun (杨穆君), co-founder and CFO, holds a Zhejiang University undergraduate degree, a Hong Kong University master’s and an HKUST PhD in advanced materials, and has worked at Industrial Securities International, High-Flyer, ByteDance and CITIC Securities.
The chief scientist is Professor Wen Weijia (温维佳), chair professor at HKUST and dean of the HKUST (Guangzhou) Function Hub. He holds a PhD in condensed matter physics from the Institute of Physics, Chinese Academy of Sciences, and was a postdoc at UCLA. He received the National Science Fund for Distinguished Young Scholars (overseas) in 2000, led the project on giant electrorheological fluid structure and physical properties that won a Second Prize of the National Natural Science Award in 2014, has published more than 400 papers with over 18,000 citations (h-index 65), and holds 19 US and more than 40 Chinese patents.
That is a materials-and-devices bench sitting next to a robotics-algorithms bench. Electronic skin fails on exactly the seam between those two disciplines: a sensing material that works in a lab coupon often drifts, hysteresis-loops or delaminates once it is a curved pad on a moving joint.
The product line: chest, sole, hand, floor
Rather than sell one sensor, MoSense has laid out a body map.
MoChest is a chest-mounted tactile system for humanoids and bionic platforms. A flexible curved structure conforms to the ribcage; a magnetic induction array reads field changes on contact or approach, and the algorithm resolves force distribution, indentation depth and three-axis force vectors. Use cases are push, press and collision detection for dynamic obstacle avoidance, compliant interaction and care-giving scenarios.
MoSenseFoot Wireless targets bipeds, exoskeletons and bionic platforms: a high-density Hall sensor array acquires three-dimensional magnetic field changes and temperature-compensated data from the sole, transmitted wirelessly. The output is pressure distribution, contact direction and region identification. A fuller version adds multi-channel ranging at toe, heel and forefoot, monitoring pressure, friction, temperature and slip, with low-power modules, abrasion protection and wireless charging for gravel, sand and water.
MoGrip / MoTip / MoPad split the hand into three problems. MoGrip reads multi-dimensional force fields on the contact surface at high sampling rate and detects slip tendency so a gripper can adjust force or pose. MoTip handles micro-contact, three-axis force, friction and slip detection for fine-part grasping and texture recognition. MoPad covers the finger belly and larger grasp areas, reading pressure distribution and three-axis vectors so grip force follows object shape, size and weight.
MoCarpet takes tactile out of the robot: a carpet-form magnetic tactile array with 45 magnetic sensing channels, reconstructing discrete measurements into 2D pressure and direction fields and a 3D deformation field. It is a gait-validation and human-robot interaction research instrument as much as a product.
MoTact and MoWorld: from signal to decision
Hardware produces field data. Something has to turn that into force, displacement, slip and material estimates that a controller can consume. MoSense calls this layer MoTact — effectively a tactile cerebellum — and claims sparse super-resolution, complex topology integration, drift and magnetic-interference rejection, and kHz-class array sampling.
Above that sits MoWorld, a tactile world-action model under development: fuse high-frequency multi-dimensional touch into the policy so the robot corrects its judgment during execution, using touch as a physical prior to narrow the sim-to-real gap.
The honest reading: the hardware claims are specific and checkable; the model claims are directional and not yet independently validated. Task success rates in real scenes are the number that will decide this.
What to watch
Three things determine whether this round converts into a category position.
Cost per square centimetre. Electronic skin only reaches a humanoid’s whole body if the BOM allows it. A chest pad and two soles at aerospace prices is a demo; at consumer-electronics prices it is a product line.
Durability under real duty cycles. A sole sensor on a walking robot sees millions of impact cycles. Lab cycle counts of “hundreds of thousands” are a starting point, not a proof.
Whether the model layer ships. Sensor companies that stay sensor companies get valued on unit economics. Sensor companies that ship a perception stack get valued on the robot’s capability. The second is what the investors in this round are paying for.
For the component supply chain, the signal is straightforward: whole-body tactile is being specified now, in the same design cycles that are locking in joint modules, reducers and thermal architecture. Once a humanoid’s mechanical envelope is frozen around a chest sensor and a sole stack, the next generation of that platform inherits the choice.
Source: Jijia Juexing (机甲觉醒), 26 October 2026; company materials from MoSense (模感科技); supporting reporting by HardTech (硬氪). Figures as reported by the source and not independently verified by Latest Robot News.
Sarah Xiao