In this review, I will systematically explore the integration and application of sensors in humanoid robots, highlighting their critical role in enabling advanced mobility and intelligence. The powerful locomotion and high intelligence of humanoid robots stem from their ability to accurately perceive the surrounding world. Sensors, as vital components of the perceptual system in humanoid robots, serve as key links for acquiring environmental information and achieving human-robot integration, playing an indispensable supportive role in human-robot interaction and motion control processes. First, I will comprehensively review the current state of sensor integration in humanoid robots. Then, I will outline the existing problems and challenges faced by various types of sensors, proposing specific solutions to address these issues. Finally, I will offer perspectives on the future development directions of sensor technology for humanoid robots.
The core driving force behind the intelligent development of robots lies in the continuous enhancement of their perceptual and cognitive capabilities, which is key to achieving the leap from simple automation to high intelligence in robots. Sensors, as critical components for information interaction between robots and the external environment, undertake the important mission of perceiving various external information, forming the foundational support for building intelligent robotic systems. In recent years, humanoid robots have become a research hotspot in the robotics field due to their greater versatility and adaptability compared to other robots, enabling smoother human-robot collaboration and role substitution. This has propelled humanoid robots and their associated sensors to the forefront of current research.
Compared to traditional industrial robots primarily used in repetitive, regular manufacturing scenarios, humanoid robots are oriented toward more diverse, complex, and highly uncertain core application scenarios. This difference significantly increases the perceptual requirements for humanoid robots. In addition to traditional sensors such as current sensors and temperature sensors that meet basic operational monitoring needs, high-value sensors crucial for performance enhancement, including force sensors, electronic skin, vision sensors, and inertial sensors, have become indispensable components of the perceptual system in humanoid robots. The modern technological framework of humanoid robots consists of three parts: the “brain,” “cerebellum,” and “limbs,” with sensors serving as key components贯穿 these layers, acting as critical mediators for information perception and human-robot interaction, and playing an irreplaceable supportive role. Among these, the “brain” functions as the central decision-making system of humanoid robots, enabling environmental perception, task planning, and autonomous decision-making through high-performance computing platforms and AI algorithms; the sensor system is a key component for environmental perception in the “brain” of humanoid robots. The “cerebellum” is the core module of the motion control system in humanoid robots, primarily responsible for real-time motion coordination and precise execution; the sensor system provides critical technical support for motion control, balance adjustment, environmental adaptation, and safety assurance in the “cerebellum” of humanoid robots. The “limbs” of humanoid robots are motion execution systems designed to mimic human body structure, covering parts such as the head, torso, and limbs, achieving flexible motion and interaction in complex environments through high-degree-of-freedom joints, precise驱动 devices, and sensor collaboration. Sensors play a core role in the “perception-execution closed loop” within the limbs of humanoid robots, converting information from the physical world into digital signals to drive the limbs for precise, adaptive motion while ensuring safety and interaction capabilities. Thus, sensors in humanoid robot limbs构建 a complete “perception-decision-execution” closed loop, not only enabling high-precision motion control but also endowing robots with the ability to adapt to complex environments and ensure safe interaction.
Compared to other types of robots, the core application scenarios面向 by humanoid robots impose more stringent requirements on the perceptual capabilities of the robots themselves, thereby有力地 promoting the continuous development of sensor technology related to humanoid robots. This review aims to investigate the application and shortcomings of sensors in humanoid robots, discuss the characteristics and main challenges of current sensors in humanoid robots, and展望 future development directions. By comprehensively analyzing domestic and international research成果 and technological status of sensors in humanoid robots, I hope to provide reference and insights for further research and development of humanoid robots.
Principles and Classification
The locomotion and intelligence of humanoid robots originate from their ability to perceive the world. The sensing system, analogous to human sensory systems, provides data support for control decisions in robots through physical signal acquisition and digital conversion. The perceptual capabilities of humanoid robots are divided into two broad categories: proprioception (understanding自身 state) and environmental perception (认识 external world). Proprioception is the core for robots to comprehend their own motion姿态 and force states, forming the basis for precise motion control and safety. Environmental perception endows robots with the ability to understand surrounding space, objects, people, and environmental changes, serving as the foundation for intelligent behavior. Accordingly, sensors used in humanoid robots can generally be classified into internal and external types. Internal sensors are core components for perceiving and monitoring the robot’s own state (e.g., joint angles, speeds, torques), belonging to the “proprioceptive system” of humanoid robots, primarily distributed in key areas such as joints and torso, mainly used for adjusting and controlling robot motion, enabling motion control, balance adjustment, and precise operation in humanoid robots. External sensors are components for perceiving and detecting environmental information and interaction objects external to the robot, equivalent to the “sensory system” of humanoid robots, with main functions including environment recognition, obstacle avoidance navigation, human-robot interaction, etc., enabling robots to automatically correct and adapt based on different environments, such as obstacle avoidance and route planning. The classification of sensors in humanoid robots can be summarized as follows:
A sensor is a device capable of receiving information from a measured object and converting this signal into an electrical signal or other desired signal type through its internal circuitry, serving as the primary means of acquiring environmental and equipment information. The typical structure of a sensor includes sensitive elements,转换 elements, auxiliary power sources, and signal conditioning and conversion circuits. The output voltage \( V \) can often be expressed as a function of the input signal \( S \):
$$ V = f(S) = k \cdot S + C $$
where \( k \) represents the sensitivity and \( C \) is an offset. For linear sensors, this simplifies to \( V = k \cdot S \).
Internal sensors in humanoid robots primarily include force sensors, inertial sensors, and encoders. Internal sensor technology is relatively mature, largely able to沿用 mature solutions from industrial robots and collaborative robots. External sensors mainly include vision, auditory, and tactile sensors, which are current research hotspots in academia and industry, demonstrating significant innovative potential in environmental perception and human-robot interaction scenarios, and are regarded as core directions for future breakthroughs. The classification can be further detailed in the following table:
| Sensor Category | Sub-types | Primary Function in Humanoid Robot |
|---|---|---|
| Internal Sensors | Inertial Sensors (IMU) | Navigation, dynamic stability control |
| Encoders | Precise motion control and positioning | |
| Force/Torque Sensors | Fine force control and motion balance | |
| External Sensors | Vision Sensors | Environmental perception, modeling, navigation |
| Tactile Sensors | Contact detection and dynamic force control | |
| Auditory Sensors | Environmental perception and human-robot interaction |
Current Application Status of Sensors
Internal Sensors
Internal sensors are core components of the motion control feedback system in humanoid robots. In the context of日益增长的精细操作 and柔性 control demands, the need for high-performance sensors has emerged. These high-performance sensors cover types such as inertial sensors, encoders, and force sensors. The core functions and application areas of these sensors in humanoid robots are compared below:
| Sensor Type | Typical Quantity per Humanoid Robot | Application Areas in Humanoid Robot | Core Function |
|---|---|---|---|
| Inertial Sensor (IMU) | 1-2 sets | Pelvis/Thorax | Navigation, positioning, and dynamic stability control |
| Force/Torque Sensor | 2-4 units | Wrist/Ankle | Precise force control and motion balance |
| Encoder | 50+ units | Joints throughout body | Precise motion control and positioning |
Inertial Sensors
In research on gait control for humanoid robots, accurately obtaining姿态 information of each joint and the overall center of gravity position of the robot is a core prerequisite for optimizing gait planning and ensuring运动 stability and coordination. For this purpose, inertial sensors are employed to perform real-time, precise acquisition of multi-dimensional姿态 data of humanoid robots during dynamic motion.
Inertial sensors are foundational for robot motion control and navigation. Classified by measured physical quantity, inertial sensors can be divided into acceleration sensors (accelerometers) and angular velocity sensors (gyroscopes). Integrating accelerometers and gyroscopes yields the basic six-axis Inertial Measurement Unit (IMU). A six-axis IMU measures three-axis angular velocity and acceleration of an object, typically containing three single-axis accelerometers and three single-axis gyroscopes. Some high-end models also integrate magnetometers to form a nine-axis IMU. The IMU plays a role akin to the “cerebellum” in humanoid robots, serving as the core sensor for the humanoid robot to perceive its own motion state. It is usually布置 in the pelvis or thoracic area of the robot to achieve balance, navigation, and coordinated motion. During system initialization, the IMU can output姿态 information, enabling the humanoid robot to achieve stable standing in combination with the center of mass state. During walking, the IMU provides real-time姿态 data feedback, ensuring the humanoid robot moves along a预定路线 and enabling协同 control modules to achieve precise in-place turning based on parameters like heading angle to avoid path deviation. When the humanoid robot requires emergency braking or task中止, the system can respond rapidly to immediately恢复 stable standing状态. Besides Tesla’s Optimus, leading global humanoid robot manufacturers such as Boston Dynamics’ Atlas, Estun Cozmo’s Codroid 02, Zhiyuan Robotics’远征 A1, Ubtech’s Walker X, Unitree’s H1, and Xiaomi’s CyberOne have all integrated IMUs to achieve precise limb motion control.
Encoders
In the application scenarios of humanoid robots, data feedback from encoders plays a pivotal role. It primarily undertakes the task of real-time monitoring and regulating the运动 position of robot joints, thereby ensuring the humanoid robot can achieve precise motion control and positioning. Humanoid robots obtain real-time position and velocity feedback signals for each joint through a large number of encoders. This information is an indispensable input for closed-loop control systems, constituting the “sensory nerves” for the humanoid robot to perceive its own motion state, and is a prerequisite for all precise actions. For example, the motion control module of the Tesla humanoid robot is mainly divided into linear actuator modules, rotary actuator modules, and dexterous hands. Each rotary actuator module has independent output and input encoders, while each linear actuator module is配置 with one encoder. In the dexterous hand module, to meet its complex functional requirements, each coreless motor module is搭载 with one encoder, totaling approximately 40 encoders per humanoid robot.
Encoders have different classification methods. Based on measurement type, encoders can be divided into linear encoders and rotary encoders. According to output signal characteristics, encoders can also be classified as absolute encoders and incremental encoders. From a technical principle perspective, encoders can be categorized into optical encoders, inductive encoders, and magnetic encoders. The performance comparison of these three encoder types is shown in the following table:
| Type | Accuracy | Cost | Environmental Adaptability | Primary Application Scenarios |
|---|---|---|---|---|
| Optical Encoder | Extremely High | High | Weaker | High-precision machine tools, medical robots |
| Inductive Encoder | High | Relatively High | Strong | Outdoor robots |
| Magnetic Encoder | Medium-High | Medium | Extremely Strong | Rotary joints, dexterous hands |
With the development of electromagnetic technology and increasing market demand for抗干扰性 and low cost, inductive encoders and magnetic encoders have begun to be gradually applied in the field of humanoid robots due to their advantages such as small size, light weight, wide operating temperature range, and shock resistance. For instance, humanoid robots like Estun Cozmo’s Codroid 02 and Tesla’s Optimus extensively use magnetic encoders in their joint motors.
Force Sensors
Force sensors, as core components for humanoid robots to understand their own state, directly determine the operational precision, safety, and智能化水平 of the robot. Leveraging force feedback signals, the humanoid robot can perceive in real-time the force state information of its internal joints and parts during motion, and dynamically adjust its actions accordingly to enhance environmental adaptability. Through force control strategies, the humanoid robot can actively plan and execute precise actions based on acquired force information. With the real-time monitoring and feedback capability of force sensors, the humanoid robot demonstrates stronger adaptability in complex and changing environments, and its operational stability and reliability are significantly improved.
Based on measurement dimensions, force sensors can be categorized into: (1) One-dimensional force sensors, used to measure force or torque in a single axis (e.g., force or torque in the z-axis direction), enabling basic force/torque closed-loop control functions. These sensors are often installed at major joints such as hips, knees, ankles, shoulders, and elbows to sense torque/force on joint驱动 shafts or末端, and are key to achieving fine force control,柔性 interaction, and collision detection. They are characterized by low cost (hundreds of RMB level) and high technological maturity. (2) Three-dimensional force sensors, used to detect forces in three axes: Fx, Fy, and Fz, without torque measurement capability. Often used to partially替代 six-dimensional sensors for basic force control scenarios (e.g., dexterous hand grasping). (3) Six-dimensional force/torque sensors, capable of simultaneously measuring three-dimensional force (Fx, Fy, Fz) and three-dimensional moment (Mx, My, Mz), enabling full-space force perception. Six-dimensional force/torque sensors are mostly installed at wrists and ankles for high-precision assembly and dynamic balance control. When installed at the wrist, they can directly perceive interaction forces/moments between the hand and the operated object (grasping force, operational impedance, etc.), serving as the ultimate sensory organ for achieving dexterous operation and precise force control. When installed at the ankle, they can precisely measure the interaction force between the foot and the ground (Ground Reaction Force, GRF), serving as important input for bipedal walking stability (Zero Moment Point, ZMP calculation), gait optimization, terrain adaptation perception, and fall预警.
The ZMP is a key concept for stability in humanoid robot walking and can be calculated from force sensor data. For a humanoid robot, the condition for static stability is that the ZMP lies within the support polygon. The location of the ZMP can be derived from the moments measured by a six-axis force sensor at the foot. If \( \tau_x \) and \( \tau_y \) are the moments about the x and y axes at the sensor, and \( F_z \) is the vertical force, the ZMP coordinates \( (x_{zmp}, y_{zmp}) \) can be approximated as:
$$ x_{zmp} = -\frac{\tau_y}{F_z}, \quad y_{zmp} = \frac{\tau_x}{F_z} $$
Based on measurement principles, force sensors can be divided into strain gauge, piezoelectric, capacitive, photoelectric, and other types. The characteristics of several types of force sensors are compared in the following table:
| Performance | Strain Gauge | Piezoelectric | Capacitive | Photoelectric |
|---|---|---|---|---|
| Accuracy | High | High | Ultra-high | Medium-High |
| Range | Large (1N~10kN) | Large (0.01N~100kN) | Low (suited for微小力) | Medium (limited by optical path) |
| Stability | Requires temperature compensation | Good (wide temperature range) | Susceptible to humidity | 抗电磁干扰 |
| Cost | Low (industrially mature) | High (requires专用电路) | Medium | High (optical components expensive) |
| Static Measurement | Supported (zero-frequency response) | Not supported (dynamic only) | Supported | Supported |
Constrained by factors such as cost and stability, strain gauge-based and their衍生 six-dimensional force sensors remain the mainstream choice for humanoid robots.
External Sensors
Enabling intelligent interaction between humanoid robots and dynamic physical environments essentially involves构建 a perception-motion system with biological similarity. By simulating human vision, proprioception, and hearing through multiple types of sensors, environmental information is converted into digital encodings akin to neural signals. These are then processed by类脑 decision algorithms to ultimately drive the actuators of the humanoid robot to perform adaptive actions, forming a complete “perception-decision-execution” feedback loop. Therefore, external sensors are core components for achieving environmental perception and interaction, typically including vision sensors, tactile sensors, auditory sensors, and others. The core functions and application areas of these external sensors in humanoid robots are compared below:
| Sensor Type | Typical Quantity per Humanoid Robot | Application Areas in Humanoid Robot | Core Function |
|---|---|---|---|
| Vision Sensor | 1 set | Head | Environmental perception, modeling, and navigation |
| Tactile Sensor | 10+ units | Fingers/Palm | Contact detection and dynamic force control |
| Auditory Sensor | 1 set | Head | Environmental perception and human-robot interaction |
Vision Sensors
In the process of humans perceiving external information, the visual system dominates, with approximately 80% of external information acquired through the visual channel. Similarly, machine vision本质上 is about构建类人 visual perception capabilities for robotic systems, capturing light signals reflected from environments and object surfaces through optical sensors, thereby enabling intelligent acquisition and parsing of external information by robots. Vision sensors are typically installed on the robot’s head or other areas requiring visual perception, enabling the humanoid robot to achieve visual functions similar to humans, such as object recognition, navigation, and obstacle avoidance.
As core components for robots to perceive the external environment, the technical principles of vision sensors are primarily based on optical imaging and image processing technologies. They capture object images through photoelectric sensor devices and convert them into digital signals for subsequent image processing and analysis. Based on the dimensionality of image information acquisition and the type of data processed, vision sensors can be classified into 2D vision sensors and 3D vision sensors. 2D vision sensors are mainly used for planar image acquisition (e.g., color, texture recognition), suitable for object presence detection or QR code recognition in simple scenarios, but cannot acquire depth information, leading to reliance on预设 algorithms to compensate in tasks like grasping and obstacle avoidance. 3D vision sensors directly acquire three-dimensional coordinates of objects through structured light, time-of-flight, or multi-view stereo vision techniques, enabling precise grasping, dynamic obstacle avoidance, and environmental modeling. The differences in application between 2D and 3D vision sensors in humanoid robots are significant, mainly reflected in environmental perception, interaction capability, and technological adaptability, as compared in the following table:
| Performance | 2D Vision Sensor | 3D Vision Sensor |
|---|---|---|
| Cost | Low cost, simple equipment | High cost, requires professional equipment |
| Data Type | Planar images, no depth information | Point cloud/depth map,包含 spatial coordinates |
| Algorithm Dependency | High (requires additional algorithms to infer spatial relationships) | Low (directly outputs 3D data) |
| 抗干扰性 | Susceptible to lighting, angle influences | Adapts to complex lighting, dynamic environments, and occlusion |
With the continuous development of intelligent manufacturing and embodied intelligence, facing demands for precise identification of complex objects, high-precision dimensional measurement, and adaptability to complex scenarios in human-robot interaction, traditional 2D vision technology has gradually revealed certain technical limitations in improving measurement accuracy and achieving effective distance measurement. Therefore, 3D vision technology with three-dimensional spatial perception capabilities has seen significant development. 3D vision sensors achieve three-dimensional information acquisition through various technologies, which can be divided into passive and active categories based on their working principles. Passive 3D sensing techniques primarily achieve three-dimensional perception through ambient light or natural radiation from objects without actively emitting light sources, and can be further divided into monocular vision (based on motion or focus) and multi-view vision types. Multi-view vision significantly outperforms monocular solutions in accuracy and stability, but monocular vision technology dominates lightweight applications due to its cost and adaptability advantages. Among these, motion parallax methods are more suitable for dynamic environments, while focus/defocus methods offer significant advantages in static high-precision scenarios. A comparison of passive 3D sensing techniques is shown below:
| Performance | Monocular Vision (Motion Parallax) | Monocular Vision (Focus/Defocus) | Multi-view Vision |
|---|---|---|---|
| Accuracy | Low | Medium | High |
| Hardware Requirements | Single camera, but requires motion capability (e.g., mounted on robot) | Single camera, but requires lens capable of快速精确调焦 | Multiple cameras, require precise calibration to determine relative positions (extrinsic parameters) |
| Advantages | Low hardware cost, simple system structure | Low hardware cost, simple system structure | Works in static environments without motion or focus adjustment, high accuracy and reliability |
| Disadvantages | Must move to measure distance, highly sensitive to motion estimation errors | Extremely small measurement range, susceptible to image noise and texture | High hardware cost, complex calibration; matching困难 in weak texture,重复 texture,强反光 areas |
| Applicable Scenarios | Environmental perception, modeling and navigation | Inspection and grasping | Indoor navigation and grasping |
Passive techniques are developing towards “environmental signal adaptation” and “algorithm noise reduction,” but currently still require supplementation by active techniques for low-light scenarios. Active 3D sensing techniques achieve three-dimensional perception by actively emitting light waves or sound waves, including Structured Light method, Time of Flight (ToF) method, and Laser Triangulation method. A comparison of active 3D sensing techniques is provided below:
| Performance | Structured Light | Time of Flight (ToF) | Laser Triangulation |
|---|---|---|---|
| Accuracy | Millimeter-level at close range (0.1~1mm) | Centimeter-level at medium-far range (1~5cm) | Micrometer-level (industrial grade can reach ±0.01mm) |
| Ranging Range | 0.2~1.2m (limited by light intensity attenuation) | 0.4~5m (advantage for long distance) | 0.01~2m level (dedicated for close range) |
| 抗干扰性 | Suitable for low light, but fails under strong light | Strong resistance to ambient light interference | Susceptible to reflective surfaces |
| Real-time Performance | Medium (requires image processing) | High (direct time measurement) | Medium (requires scanning or point-by-point calculation) |
| Applicable Scenarios | Indoor static measurement | Dynamic long-distance measurement | Static high-precision measurement |
Currently, 3D vision is experiencing rapid market growth in fields like robot navigation and high-precision inspection, but 2D vision remains mainstream due to high maturity and low cost. In the future, as humanoid robots demand more complex interaction, 3D vision sensors, with their high precision and large information capacity, will likely become the mainstream choice for vision schemes in humanoid robots. Currently, mainstream humanoid robots mostly adopt 3D vision schemes. However, due to some inherent defects of 3D vision, multi-sensor synergy and fusion will become another future trend. By combining texture information from RGB-D cameras with geometric data from LiDAR, the reliability of complex environment modeling and the robustness of environmental perception for humanoid robots can be enhanced.
Tactile Sensors
Touch is an important function of human skin and a crucial channel for humans to perceive the external environment and interact dynamically with it. Specialized receptors distributed within the skin can precisely respond to多元 stimuli such as external pressure, deformation, stretch, and temperature changes, enabling humans to identify the physical properties and spatial states of接触 objects. Similarly, a tactile sensor is a type of sensor that mimics human tactile perception capabilities. It can detect and quantify mechanical or thermal stimuli generated during physical contact and convert this information into electrical signals, granting electronic systems the ability to perceive touch. Therefore, tactile sensors have received widespread attention in humanoid robot research. Tactile sensors are primarily applied to the hands, feet, or other parts of the robot that require direct contact with the external world,用于感知 contact force and surface texture of objects, among other information. The development of flexible tactile sensor (electronic skin) technology has enabled robots to achieve finer manipulation and interaction, such as grasping fragile items or operating complex objects.
Tactile perception, due to complex application scenarios, diverse functional requirements, and stringent technical standards, results in complex working mechanisms and a rich variety of sensor types. Similar to force sensors, based on different working principles, tactile sensors can be divided into piezoresistive, piezoelectric, capacitive, inductive, optical, and other types. Their advantages and disadvantages are summarized in the table below, with capacitive, inductive, and piezoelectric types being relatively mature in application.
| Principle | Advantages | Disadvantages |
|---|---|---|
| Piezoresistive | Relatively high sensitivity, strong overload承受能力 | Large volume, high power consumption, susceptible to noise, contact surface fragile |
| Optical | High spatial resolution, less affected by electromagnetic interference | Low linearity under multi-force conditions, poor real-time performance, difficult calibration |
| Capacitive | Large range, good linearity, low cost, high real-time performance | Large size,不易集成化, susceptible to noise, stability issues |
| Inductive | Low manufacturing cost, large measurement range | Low resolution, poor consistency across different contact points |
| Piezoelectric | Wide dynamic range, good durability | Susceptible to thermal response effects |
The development of tactile sensor technology has significantly enhanced the operational dexterity of humanoid robots, thereby promoting the emergence of numerous dexterous hand companies. A comparison of tactile schemes from several common dexterous hand manufacturers is detailed below.
It should be noted that although both force sensors and tactile sensors are core components for humanoid robots to perceive force information, they differ significantly in application scenarios, measurement objects, and other aspects. Force sensors are primarily used to measure macroscopic mechanical parameters (e.g., spatial force/torque), addressing global issues such as dynamic balance control and joint torque adjustment for the humanoid robot. Tactile sensors, however, are crucial for the humanoid robot to understand complex scenes and perform fine manipulation. The performance comparison between the two is as follows:
| Performance Indicator | Force Sensor | Tactile Sensor |
|---|---|---|
| Measurement Object | Spatial vector force/torque (macro mechanical quantity) | Distributed force on contact surface/surface characteristics (micro mechanical quantity + physical attributes) |
| Application Scenario | Joint torque control (joint/wrist/foot) | Surface of dexterous hand or electronic skin覆盖层 |
| Accuracy Requirement | Six-axis force sensor requires ±0.1% FS | Pressure resolution needs to reach 0.01 kPa |
| Technical Difficulty | Complex six-axis force decoupling algorithms, time-consuming calibration | Flexible material deformation affects repeatability |
| Cost & Quantity | High cost, low quantity used | Lower unit cost but high quantity used |
In summary, force sensors focus on enhancing the precision of macroscopic force control, while tactile sensors specialize in achieving细腻 simulation of微观触觉. The two form functionally complementary协同 applications in the field of humanoid robots. Taking a humanoid robot grasping a glass杯 as an example, the force sensor monitors the dynamic balance of the robot’s overall force state in real-time, while the tactile sensor is responsible for精细 regulating the grip force distribution at each contact point.
Auditory Sensors
Hearing is an important way for humans to perceive the external environment, enabling humans to identify sound方位, intensity, and characteristics through声波 frequency (16-20000 Hz). When a humanoid robot is in a非结构化 dynamic environment, the visual perception module may produce perception歧义 due to factors such as sudden changes in lighting intensity, occlusion in dynamic scenes, motion blur, and optical noise interference, leading to reduced accuracy in spatial positioning and object recognition. At this point, it is necessary to introduce auditory sensors. Leveraging their capabilities such as sound source localization and spectral analysis, combined with visual information to form spatiotemporally synchronized跨模态 fusion perception, the accuracy and robustness of the humanoid robot’s environmental perception can be improved.
The hearing system of a robot includes two basic functions: Automatic Speech Recognition (ASR) and sound source localization. Auditory sensors搭载 on humanoid robots are currently primarily based on multi-microphone arrays. The essence of “localization” is to extract the target’s方位 (direction and position) information from the acquired sound signal. For this, the collected signal must contain spatial information of the target, hence requiring two or more sensors配置 at different spatial positions to构成 an array. Multi-microphone arrays employ spatial distribution of multiple microphones (typically 4-8), achieving high-precision sound source localization through beamforming algorithms, with accuracy可达 ±3°, significantly superior to human hearing (±10°~15°). Microphones can be divided into MEMS silicon-based microphones and piezoelectric ceramic types based on materials used; the former has high integration and stable performance, while the latter has high sensitivity. Based on transduction principles, they can be divided into dynamic coil and electret capacitive types, with electret capacitive types having advantages of small size and low cost, sensitivity between 40-60 dB, capable of clearly capturing声音 within normal conversation volume range. Currently, mainstream humanoid robots all use microphone arrays to构建 their hearing systems.
Existing Application Problems and Solutions
Existing Problems
Inertial Sensors
Inertial sensors are important components of the proprioceptive sensors in humanoid robots and are currently the most feasible方案 to assist humanoid robots in achieving bipedal motion. They are of great significance for preventing falls in humanoid robots and generating dynamically stable walking motions. Their basic principle involves using built-in sensors such as accelerometers, gyroscopes, and magnetometers to estimate the position and orientation of the IMU, thereby obtaining the position and orientation of various body parts where the IMU is installed. Their advantage lies in not requiring external cameras to measure the robot’s motion, thus effectively avoiding obstacle occlusion issues. The disadvantage is that built-in sensors like accelerometers and gyroscopes themselves have drift problems; their data accumulate errors over time, leading to some loss of accuracy. The actual application scenarios of humanoid robots present nearly持续 dynamic characteristics, with几乎没有静止状态, and often伴随 impact acceleration effects. This harsh application environment poses极大的 challenges for the stable application of IMUs in humanoid robots.
Encoders
In traditional application scenarios, encoders typically adopt shaft-mounted installation. In this installation mode, the motor shaft, encoder shaft center, and chip shaft center achieve three-axis concentric alignment installation, possessing significant advantages of simple structure and excellent accuracy stability. However, with the development of humanoid robot technology, its urgent demand for miniaturization and lightweight has催生了 the off-axis installation method for encoders. But the off-axis installation method brings unprecedented巨大 challenges to the structural design, accuracy assurance, and抗干扰能力 of encoders. Currently, how to effectively规避 the inherent conflict between encoder miniaturization进程 and reliability requirements, as well as properly resolving the contradiction between accuracy indicators and dynamic response characteristics, has become a critical issue to be攻克.
Force Sensors
Although force sensors play a pivotal role in the complex motion control and precise interaction processes of humanoid robots, being key components to ensure the robot achieves fine manipulation and intelligent perception, their technical implementation层面 still faces many亟待攻克的 challenges. (1) In terms of performance enhancement: How to further improve the sensitivity, accuracy, and stability of sensors to precisely契合 the increasingly stringent and精细化 requirements for force perception in humanoid robots is one of the core issues in current research. (2) In terms of cost control: How to effectively reduce the manufacturing cost of sensors, breaking the桎梏 of high cost on market applications to promote their large-scale deployment and application in broader fields is a key bottleneck for industrial development. (3) In terms of structural design: How to deeply optimize the structure of sensors, significantly improving their reliability and durability, ensuring they can operate stably for long periods in complex, changing, and high-load working environments, is an important foundation for ensuring robot reliability.
Vision Sensors
Vision sensors are an indispensable part of humanoid robot technology, endowing robots with类似人类 “visual perception” capabilities, enabling robots to navigate, recognize, and make decisions in complex environments. Although research and application of vision sensors have made颇为显著的 progress in recent years, they still face many challenges in practical application: (1) Insufficient environmental adaptability: The performance of vision sensors is极易受限 by various external factors, among which dynamic changes in lighting intensity, background complexity, and object occlusion conditions can all significantly干扰 them, thereby affecting their stable and accurate information output. (2) Lack of robustness in recognition accuracy: Although advancements in cutting-edge technologies like deep learning have在一定程度上 promoted the improvement of object recognition accuracy, in some specific scenarios, due to scene complexity and特殊性, existing recognition technologies still难以达到 ideal recognition effects, highlighting the robustness短板 of recognition accuracy in应对复杂情况. (3) Weak data processing capability: Robotic vision systems typically need to process海量的 image data during operation, posing极为严苛的要求 on the system’s computational capabilities. Especially in real-time application scenarios, the system must complete rapid data processing and precise analysis within极短的时间 to make timely decisions. However, currently insufficient data processing capability has become a key problem制约 its development. (4) High cost: High-performance vision sensors often require substantial investment in研发 resources and advanced manufacturing processes, leading to居高不下研发 and manufacturing costs. Excessive costs not only increase the overall difficulty of robot development but also很大程度上阻碍 the promotion and application of robots in broader fields.
Tactile Sensors
Tactile sensors, as a currently备受瞩目的 research hotspot field, although having achieved a series of具有重要价值的研究成果, still face the following challenges in their迈向 batch production and large-scale application进程: (1) There is a难以调和 contradiction between performance and structure in tactile sensors. On one hand, simultaneously achieving high flexibility and high elasticity is极具挑战, and to achieve multiple functions, it is often necessary to fuse multiple different types of sensors, which inevitably negatively affects the轻薄特性 of tactile sensors. On the other hand, tactile sensors need to connect to acquisition circuits, and complex wiring布局 can在一定程度上限制 their degrees of freedom of motion, thereby affecting their performance表现 in practical applications. (2) Material cost and preparation工艺 become key factors制约 large-scale production. Flexible substrate materials with special properties, highly conductive活性 materials, and materials for微加工 generally have较高 market prices. Simultaneously, the production of flexible tactile sensors imposes极高 requirements on the consistency of substrate materials,必须确保 their digital stability and temperature stability meet stringent standards. Taking some materials like graphene and carbon nanotubes used for活性层 as examples, their preparation工艺 is complex and繁琐,这不仅导致材料成本居高不下, but further加剧 cost pressure during large-scale production processes, becoming a bottleneck阻碍 the widespread application of flexible tactile sensors.
Solutions
Compared to consumer-grade and ordinary industrial-grade sensors, sensors for humanoid robots have higher requirements in reliability, stability, durability, and environmental adaptability. Simultaneously, facing diversified scenario application demands, higher standards are提出 for performance requirements. Addressing the current application problems of sensors in humanoid robots, improvements can be made from the following aspects:
Optimizing Sensor Technology: Humanoid robots have extremely high requirements for sensor sensitivity, measurement accuracy, and environmental抗干扰能力 to achieve rapid perception of the external environment and ensure the accuracy and coordination of motion control. Sensitivity can be improved by enhancing sensor materials, optimizing composite structures, etc., combined with algorithms like adaptive filtering and nonlinear compensation to strengthen accuracy and抗干扰能力, meeting the higher demands of humanoid robots for environmental perception.
Reducing Manufacturing Costs: Humanoid robots use a large number of sensors, significantly推高 overall costs. Therefore, reducing sensor cost is of great importance. By adopting new economical and applicable materials, improving production manufacturing工艺流程, and optimizing product design方案等途径, the production cost of sensors can be effectively reduced.
Innovating Structural Design: Comprehensively optimize the structural design of sensors, conducting detailed consideration and improvement from各个环节 such as structural layout, material selection, to connection methods. Through these optimization measures, improve the reliability and durability of sensors, so that even when facing complex and harsh working environments, they can still maintain stable and reliable working performance, ensuring accurate data acquisition and transmission.
Advancing Intelligent Integration: Deeply intelligently integrate sensors with various other sensors, actuators, and other key components,构建 an advanced robotic perception system with adaptive capabilities. This system can automatically adjust perception strategies and parameters based on different working environments and task requirements, further improving the智能化水平 of humanoid robots, enabling them to more autonomously and flexibly应对 various complex situations.
Future Development Directions
High Precision and High Reliability: Maintaining high-precision and high-reliability operation in complex and variable operating environments will become a key direction for the development of sensors in humanoid robots, especially in fields like industrial automation and human-robot collaboration that have极高 requirements for precision and stability. This characteristic is indispensable. Precise and reliable sensors can ensure the efficient and stable operation of industrial production processes, guarantee safety and smoothness during human-robot collaboration, laying a solid foundation for the deep application of humanoid robots in these fields.
Multi-Sensor Fusion and Multi-Modal Perception: When humanoid robots识别 complex scenes, a single sensor often难以全面、准确地获取信息 due to its own limitations. In this context, multi-sensor fusion technology has emerged and gradually become a core驱动力 for humanoid robot technology development. Simultaneously, the sensor system of humanoid robots is developing towards the direction of multi-modal perception, i.e., integrating functions of multiple sensors such as tactile and auditory to构建 a comprehensive perception system. This comprehensive perception capability will endow robots with stronger understanding and应对能力 in complex situations. Taking practical schemes as an example, apart from Tesla adopting a纯视觉方案, most leading companies in the industry tend to融合多种传感器 such as cameras, LiDAR,毫米波 radar, and infrared sensors to enhance the perception accuracy and range of humanoid robots in different environments.
Intelligence and Miniaturization: With the蓬勃 development of artificial intelligence technology, sensors are gradually acquiring more powerful intelligent processing capabilities. Through built-in intelligent algorithms and chips, sensors can perform preliminary analysis and processing of collected data locally, reducing data transmission volume and improving response speed. Simultaneously, at the design层面, miniaturization and low power consumption have become important trends in sensor development to meet the紧凑机身 design requirements of humanoid robots. Miniaturized sensors not only help reduce the overall weight of the robot but also腾出更多空间 for internal layout, facilitating the integration of other functional modules.
Integration and Modularization: To further achieve高度集成 and miniaturization of sensor modules, the trend of integrating sensors with processing units is becoming愈发显著. This integrated design closely combines sensors and data processing units, reducing连接线路 and interfaces between them. This not only enhances the flexibility and机动性 of humanoid robots, enabling them to more自如地完成 various complex actions, but also effectively reduces communication costs among multiple sensors, improving data transmission efficiency and system stability.
From Perception to Intelligent Fusion: The deep integration of large models and embodied intelligence will bring a qualitative leap in the cognitive and decision-making levels of humanoid robots. Leveraging training on海量数据, the sensor system can突破 traditional perception limitations,识别更为复杂的信息, precisely distinguishing object features, attributes, and judging user intentions. Based on these precise perception information, the robot can generate autonomous decisions, achieving a重大转型 from “passive perception” to “intelligent fusion.” This转型标志着 that humanoid robots are developing towards a more intelligent and autonomous direction,有望 playing important roles in more fields in the future.
Concluding Remarks
Through systematic literature review and case analysis, this article has深入探讨 the integration and application of sensors in humanoid robots. First, based on functional dimensions, sensors for humanoid robots were classified, systematically梳理 the working principles and characteristics of sensor types such as vision, force, tactile, inertial, and environmental perception, and对比分析 the differential performance of different sensors in terms of accuracy, response speed, environmental adaptability, etc. Then, a summary was provided on the existing problems and challenges of various sensors in humanoid robots, proposing specific solutions to address these issues. Finally, perspectives were offered on the development directions of sensors for humanoid robots. The advancement of sensor technology is crucial for the evolution of humanoid robots, and continuous innovation in this domain will unlock new potentials for humanoid robots across diverse applications, bringing us closer to truly intelligent and versatile robotic companions.