Sensors & Perception: How China's Robots See and Feel

China's Sensor Market Overview

China's overall sensor market reached ¥406.1 billion ($55.6B) in 2024, growing 11.4% year-over-year. The market is projected to reach ¥579.3 billion by 2027, implying a compound annual growth rate of 12.6%. Within this, pressure sensors account for the largest share at 17%, followed by temperature/humidity sensors at 8%, and gas sensors at 2%, with the remaining 73% spread across dozens of other sensor types.

What makes this market particularly interesting for the robotics supply chain is that embodied intelligence — humanoid robots and other AI-driven machines — is emerging as a major new demand driver on top of the traditional automotive, home appliance, and industrial automation segments. Several sensor categories that were mature, slow-growth markets are now seeing renewed investment precisely because of the robotics opportunity.

The Full Sensor Suite: What One Humanoid Robot Needs

To put the sensor opportunity in concrete terms, CAICT's analysis of the Tesla Optimus provides a detailed bill of sensors for a single humanoid robot:

That's 40+ individual sensor units per robot. At the industry's target of tens of thousands of robots per year, this creates enormous volume demand — especially for the more specialized types like six-axis F/T sensors and MEMS tactile arrays that are still supply-constrained today.

The overall trend in humanoid robot sensors, per CAICT: multi-dimensional, high-precision, high-integration, high-flexibility. The two hottest R&D areas right now are high-dimensional force/torque sensors and high-dimensional tactile sensors (electronic skin).

Force Sensors & Market Leaders

Force and torque sensors are among the most critical — and most expensive — components in a humanoid robot. They allow a robot's joints and hands to detect resistance, adjust grip, and avoid damaging delicate objects or people. According to industry analysis, force sensors account for roughly 11% of total robot component value, making them the fourth-largest cost item after lead screws, frameless torque motors, and reducers.

The global force/torque sensor market reached $2.86 billion in 2024 and is projected to grow to $4.27 billion by 2031 (5.9% CAGR). China's share stood at $510 million (17.8% of global) in 2024 and is expected to reach $950 million by 2031, increasing its global share to 22.3%. The leading Chinese players in torque sensors include ATI (which maintains a presence through imports), Xinjingcheng (鑫精诚), Kunwei Technology (坤维科技), and Yuli Instruments (宇立仪器).

Demand is driven by industrial controls, automotive, transportation, medical devices, robotics, semiconductor manufacturing, aerospace, and new energy sectors — but humanoid robots and new energy are the fastest-growing applications by far.

Joint Torque Sensors

Knee, ankle, and hip joints require torque sensors to achieve compliant, safe motion — especially when interacting with humans or uneven terrain. Based on Tesla's AI Day 2022 specifications, a humanoid robot with 28 degrees of freedom requires one torque sensor per rotary joint and one push/pull force sensor plus one position sensor per linear joint — meaning each robot needs dozens of force sensing components. Current torque sensor specifications for humanoid joints range from 50 to 1,500Nm, with two main technology approaches: MEMS silicon strain gauges combined with glass micro-fusing, and traditional metal foil strain gauges.

Fingertip Force Sensors

Dexterous hands like the LinkerBot O6 and Yuansheng Apex Hand require miniaturized force sensors at the fingertip level to enable tasks such as picking up eggs, handling smartphones, or threading a needle. These sensors must measure forces as low as 0.1N (the weight of a business card) while surviving grip forces up to 30N — a 300:1 dynamic range in a package small enough to fit inside a fingertip (typically under 10mm diameter). The cost trajectory has been dramatic: imported fingertip sensors that cost ¥100,000+ five years ago can now be sourced domestically for under ¥200, enabling the sub-¥10,000 dexterous hands that are transforming the market.

Six-Axis F/T Sensors & the Humanoid Robot Opportunity

Six-axis force/torque sensors are the highest-value, highest-growth segment within the force sensor category, and the one most directly tied to the humanoid robot opportunity. These sensors measure forces and torques simultaneously across all three spatial axes (Fx, Fy, Fz, Tx, Ty, Tz), enabling robots to perform force-controlled tasks like polishing, assembly, and delicate manipulation.

The global six-axis F/T sensor market was just $230 million in 2023 but is projected to explode to $2.3 billion by 2030 — a compound annual growth rate of 40.5%. China's segment is growing even faster at 47.5% CAGR, expected to reach $620 million by 2030. This growth is almost entirely driven by the humanoid robot production ramp.

Technology Approaches for Six-Axis F/T

Chinese companies are pursuing two main technology paths for six-axis force/torque sensors. The first uses traditional metal foil strain gauges — a mature technology that is faster to bring to market and is already in sample testing with robot makers. The second uses MEMS silicon strain gauges combined with glass micro-fusing processes, which promises higher precision and better miniaturization but is still in the R&D phase. Some companies, like Ampron (安培龙), are pursuing both paths simultaneously to hedge technology risk.

Domestic Substitution Dynamics

The competitive landscape for force/torque sensors in China mirrors a pattern seen across the robotics supply chain: foreign incumbents (primarily ATI from the US) dominate the high-end market, while Chinese entrants compete on price, customization speed, and willingness to co-develop with domestic robot makers. Companies transitioning into this space from adjacent sensor markets — like Ampron, which brings decades of ceramics, MEMS, and pressure sensor expertise — have a cost structure advantage because they can share manufacturing infrastructure and materials knowledge across product lines.

Tactile Sensing Technology

Electronic skin — also called e-skin or tactile skin — is a distributed sensing layer that can cover a robot's hands, arms, and body to detect pressure, temperature, and texture across a wide surface area. Human skin is our largest sensory organ, covering roughly 1.5–2 square meters — and electronic skin aims to replicate that capability for machines. Tactile perception encompasses contact, pressure, force, slip, and temperature sensing, reflecting not just whether a robot is touching something but also revealing the object's position, shape, stiffness, softness, texture, and thermal conductivity. While Japan and the US have led in academic research, several Chinese startups are now producing commercially viable tactile sensor arrays for integration into humanoid robots. Tactile sensing is widely regarded as the single biggest bottleneck for dexterous robot hands — the component that will determine whether hands can cross from demonstration to real-world deployment.

Five Competing Technology Approaches

The technology landscape for tactile sensors has not yet converged, with five main approaches competing:

Piezoresistive sensors use elastomer materials whose electrical resistance changes under applied pressure, converting contact pressure into electrical signals. They are cheap and easy to fabricate at scale, but are plagued by hysteresis and signal drift over time.

Capacitive sensors measure changes in the relative position of two electrode plates under external force, causing a measurable shift in capacitance. They offer high sensitivity and low temperature drift, but require careful shielding from parasitic capacitance. This is the approach used by Tashan Technology (他山科技) and US-based Pressure Profile Systems (PPS).

Inductive sensors use electromagnetic induction to convert pressure into changes in a coil's self-inductance and mutual inductance, then translate these magnetic field changes into electrical signals via a magnetic circuit. This is the approach behind Pacini Perception's (帕西尼) 6D Hall-effect array technology, which can detect 15 distinct types of tactile perception.

Piezoelectric sensors rely on the piezoelectric effect — when a piezoelectric material is deformed by force, its surface generates an electrical charge that can be measured. These sensors excel at detecting dynamic forces and vibrations but struggle with static pressure measurement.

Optical/photoelectric sensors detect changes in light reflection intensity, wavelength, or frequency to measure deformation. The GelSight approach, developed at MIT, is the best-known example — a camera captures the deformation of an elastomer surface, enabling rich 3D force and texture reconstruction. However, these sensors remain bulky and computationally intensive.

Among Chinese companies, Pacini Perception Technology (帕西尼感知科技) — named after the Pacinian corpuscles in human skin — stands out with the DexH13 GEN2, which combines multi-dimensional tactile sensing with AI vision in a single hand system. Yuansheng Intelligent has taken a different approach, developing its own proprietary e-skin integrated directly into the Apex Hand.

E-Skin vs. Discrete Sensor Arrays

Two deployment architectures are competing for adoption. Discrete sensor arrays place individual sensors at key contact points — fingertips, palm centers, and grip surfaces. This is cheaper and simpler to integrate, but leaves gaps in coverage that can cause missed contacts or dropped objects. Continuous e-skin aims to mimic human dermis with a flexible sensing layer covering the entire hand or body surface. This provides richer spatial data and more natural force distribution detection, but adds significant cost, wiring complexity, and signal processing overhead. Most production hands today use discrete arrays; continuous e-skin remains largely a research-stage technology with a few notable exceptions like Hysense Technology's (超珀科技) flexible tactile products.

Core Competencies: Materials & Algorithms

Electronic skin is fundamentally a marriage of advanced materials and electronics. Building a flexible tactile sensor with high sensitivity, stretchability, and fast response requires innovations in three material layers: the substrate (what the sensor sits on), the active sensing layer (what actually changes under force), and the electrode layer (what carries the signal). Soft functional materials with excellent electrical and mechanical properties form the core. On top of the hardware, data acquisition and intelligent algorithms are equally critical — they are the key to translating raw sensor data into the kind of real-time tactile understanding that humanoid robots need for fine manipulation tasks.

Competitive Landscape

The high-end tactile sensor market is still dominated by overseas companies: Pressure Profile Systems (PPS, US — capacitive, launched Robotact in 2024 for dexterous hand integration), Tekscan (Germany), Sensor Products and Interlink (US), Nitta (Japan), and GelSight (US, optical). On the Chinese side, publicly listed players include Hanwei Technology (汉威科技, 300007.SZ — via subsidiary Suzhou Nengstda, focused on flexible micro-nano sensing), Fulai New Materials (福莱新材, 605488.SH — functional composite coating materials), and Riying Electronics (日盈电子, 603286.SH). Among private Chinese companies, Pacini (帕西尼) leads on shipment volume with its 6D Hall-array technology, followed by Tashan Technology (他山科技, AI tactile sensing chips, capacitive approach) and Moxian Technology (墨现科技, pressure-sensing approach).

Vision Systems

Humanoid robots require multiple overlapping vision modalities to operate safely and effectively: RGB cameras for object recognition, depth cameras (structured light or time-of-flight) for 3D spatial mapping, and wide-angle fisheye lenses for peripheral awareness. Lidar, historically the domain of autonomous vehicles, is also beginning to appear in high-end humanoid platforms for reliable long-range obstacle detection.

The camera and sensor configurations used in China's leading humanoid platforms reflect a layered approach: RGB for recognition, depth for spatial awareness, and lidar for long-range safety. Vision also plays a foundational role in VLA (Vision-Language-Action) models — the dominant AI architecture for embodied intelligence — where camera feeds serve as the robot's primary observation channel. Chinese suppliers like Orbbec (奥比中光) and Hesai (禾赛) are gaining traction as alternatives to Intel RealSense and Stereolabs.

The Role of Vision in VLA Models

Vision-Language-Action (VLA) models treat visual input as the robot's primary observation channel, making camera quality a direct determinant of AI performance. The resolution, frame rate, and latency of camera feeds directly affect how well a model like Galbot's GraspVLA or AgiBot's GO-1 can generalize to new objects and environments. Modern VLA training pipelines typically require at least 720p resolution at 30fps with sub-50ms end-to-end latency from photon to motor command. Higher-resolution feeds improve zero-shot generalization to novel objects, but increase the compute burden on edge inference hardware — creating a practical tradeoff between visual quality and real-time control responsiveness. For more on how VLA architectures work and who is leading the race, see our Embodied AI deep dive.

Market Landscape & Key Takeaways

China's sensor ecosystem for humanoid robots is at an early but rapidly maturing stage. Foreign brands still dominate at the high end — particularly for precision force/torque sensors — but Chinese suppliers are gaining ground through aggressive pricing and willingness to co-develop custom solutions with robot makers. Key battlegrounds include miniaturization, waterproofing, and the integration of sensing with onboard processing.

• Force/torque sensors: $2.86B global market (2024), China at $510M (17.8% share, rising to 22.3% by 2031). ATI (US) and Robotous (Korea) dominate high-end; Chinese entrants Xinjingcheng, Kunwei, and Yuli competing on cost and customization
• Six-axis F/T sensors: The breakout segment — $230M (2023) projected to reach $2.3B by 2030 (40.5% CAGR). China segment growing at 47.5% CAGR. Driven almost entirely by humanoid robot demand
• Tactile skin / e-skin: Projected $5.32B market by 2029 (18% CAGR). Five technology approaches competing (piezoresistive, capacitive, inductive, piezoelectric, optical) with no convergence yet. Overseas leaders (PPS, Tekscan, Nitta, GelSight) dominate high-end; Chinese players gaining ground — Pacini Perception leads on shipments, Hanwei Technology (via Nengstda) and Fulai New Materials entering as listed players. Coverage expanding from fingertips to full-body deployment
• Vision systems: China has world-class lidar suppliers (Hesai, RoboSense); depth cameras maturing with Orbbec
• Platform companies entering robotics: Established sensor makers like Ampron (安培龙, ¥940M revenue) are leveraging existing ceramics + MEMS expertise to enter the force sensor market, creating a new competitive dynamic
• IMUs and proprioceptive sensors: Largely commoditized; domestic supply is strong