Actuation, sensing, and integration in soft robotics: principles, progress, and future trends

3.1. Tethered actuation

3.1.1. Fluidic actuation

Fluidic actuation, one of the earliest actuation strategies investigated in soft robotics, is inspired by the locomotion of soft-bodied animals such as octopuses and sea cucumbers. It controls deformation by modulating internal fluidic pressure and exhibits advantages such as fast response, tunable output force, simple structure, and low cost^[27,28]. Fluidic actuation in soft robotics primarily includes pneumatic^[29] and hydraulic approaches^[30], which use gas and liquid as the driving medium, respectively.

Pneumatic actuation is the most widely adopted, owing to the low density of the working medium and ready availability. This approach typically employs elastomeric materials (e.g., silicone rubber^[31]) to construct structures with inflatable chambers, while constraining layers made of fibers or fabrics are used to guide and constrain deformations^[32]. Specific configurations, such as cylindrical or folded shapes, enable precise and controllable motions. However, reliance on external pumps increases system bulk, and the nonlinear behavior of elastomeric materials increases control complexity. Addressing these limitations remains a key focus for future research^[33-37].

Traditional implementations typically involve compressed air storage combined with solenoid valves to control airflow. However, pneumatic systems are often constrained by the size of air supply equipment, limiting their deployment in spatially restricted or unstructured environments. To address challenges related to system complexity and control dependency, some studies have proposed structurally driven solutions that do not require electronic control. For instance, one study developed a self-oscillating pneumatic limb system composed of soft tubing, achieving high-frequency motion through structural design^[38]. This system demonstrated the highest motion speed reported to date among pneumatic soft robots, while significantly simplifying the overall structure [Figure 3A]. The configurational diversity of pneumatic actuation continues to expand. For example, a “geometrically constrained contraction–extension” pneumatic artificial muscle was constructed from a single-material folded membrane without additional strain-limiting structures. This actuator can be rapidly fabricated using low-cost additive manufacturing for devices such as bioinspired hands, demonstrating excellent functional integration and manufacturability^[39].

Figure 3. Representative tethered actuation technologies in soft robotics. (A) The tethered robot has four limbs connected to a 3D-printed monolithic body, with four inner coupling channels, scale bars: 1 cm. Adapted with permission from Ref.^[38], Copyright © 2025, AAAS; (B) Upon inflation, an initially flat panel composed of programmed Gaussian cells self-shapes into a complex 3D structure. Adapted with permission from Ref.^[42], Copyright © 2023, AAAS; (C) Hydraulic actuation-enabled soft robotic system with integrated sensing capability. Adapted from Ref.^[44], under CC BY-NC 4.0 license; (D) A holistic design of the biomimetic rigid-soft finger with reduced complexity; (E) MIDI signal comparisons of the music played by the robotic hand, the human hand, and the standard notation. (D) and (E) were adapted with permission from Ref.^[49], Copyright © 2025, AAAS; (F) Schematic diagram and photographs showing a robot navigating through a spiral pipe for inspection, scale bars: 10 mm; Adapted with permission from Ref.^[52], Copyright © 2026, AAAS; (G) Design strategy of weaving LCE fiber soft actuators for multifunctional soft robotics. Adapted from Ref.^[58], under CC BY-NC 4.0 license; (H) Electrothermal-driven amphibious insect-scale robot based on SMA actuators. Reprinted from Ref.^[59], under CC BY-NC-ND 4.0 license. MIDI: Musical instrument digital interface; LCE: liquid crystal elastomer; MCP: metacarpophalangeal; IP: interphalangeal; DE: dielectric elastomer.

In terms of programmable configurations, another study proposed an origami-inspired pneumatic gripper capable of dynamically adjusting its grasping mode. The gripper can switch between bistable and monostable states after fabrication, enabling precise grasping across multiple forms and enhancing adaptability and performance in complex tasks. Such designs overcome the limitations of fixed traditional gripper geometries, providing soft grasping systems with increased flexibility^[40]. Another strategy, termed net rolling-splicing (NRS), rapidly constructs multifunctional pneumatic units by curling and splicing two-dimensional mesh structures, enabling complex functions such as walking, climbing, load carrying, and stable rolling. This approach offers an efficient method for building stackable pneumatic structures^[41]. Moreover, the environmental adaptability of pneumatic actuation can be further enhanced through bioinspired strategies. Inspired by bulliform cell structures in monocot leaves, one study developed a planar panel system with internally designed structures capable of programmable deformation. Under compressive conditions, the panel exhibits simultaneous in-plane deformation and bending, forming specific three-dimensional shell shapes. Using flexible polymers, the system demonstrates good manufacturability, rapid response, and tunable stiffness, offering new approaches for constructing large-area, multi-morphology shape-morphing robots [Figure 3B]^[42].

Hydraulic actuation employs liquid as the driving medium and exploits its incompressibility to achieve advantages in precision control, energy transmission, and output force. Compared with pneumatic actuation, hydraulic systems generally provide faster response and lower energy loss, making them suitable for applications that require high precision or heavy loads. For instance, one study developed a hydraulically actuated bioinspired soft robotic fish using an innovative four-cylinder piston pump and a self-sensing dual-joint actuator. Integrated with a closed-loop control system, the robot demonstrated efficient turning, diving, and long-distance swimming, exhibiting excellent posture control and adaptability in complex environments^[43]. Hydraulic actuation also offers notable benefits in structural integration and control accuracy. Kashyap et al. developed a catheter-deployable soft robotic sensor array driven by hydraulic actuation, in which inflatable soft structures were integrated with stretchable electronics to achieve highly conformable cardiac mapping [Figure 3C]^[44]. Nevertheless, challenges persist, particularly due to the effects of fluid weight on system modeling and control strategies, which are especially critical in tasks involving rapid movements or posture stabilization. Therefore, enhancing miniaturization and integration of hydraulic systems, along with the development of adaptive control algorithms, remains a key focus in current research. With advances in material technologies and fluid control methods, hydraulic actuation has evolved from traditional rigid systems to diverse, highly flexible structures. Its integration with sensing and control systems is becoming increasingly tight, providing essential support for future applications of soft robots in precise manipulation, medical wearables, and underwater exploration.

3.2. Untethered actuation

3.2.1. Magnetic actuation

Magnetic actuation is a remote, non-contact method in which magnetic materials are embedded within the robot, and an external magnetic field induces forces or torques to achieve directional motion and shape control. To enable programmability, the magnetic domain structures of these materials can be erased and reprogrammed by heating above the Curie temperature after fabrication, enhancing structural diversity and functional flexibility. With advances in magnetic material design and magnetic field control, magnetically actuated soft robots can perform complex and flexible motion patterns, including rolling, crawling, undulating, and even flying. This technology offers remote and non-contact control, high compliance, and excellent biocompatibility, making it particularly suitable for operation in confined environments. Given the strong penetration of magnetic fields through human tissue, magnetic actuation is widely regarded as a promising approach for in vivo medical robots and minimally invasive surgical systems^[62-64].

For example, one study employed a microfluidic fabrication platform to construct shape-morphable magnetic robots, enabling precise control of microscale structural configuration and function. These microrobots can be efficiently manipulated and rotated by external magnetic fields, with potential applications as active drug carriers, microreactors, or micromixers^[65]. Another study demonstrated a subcentimeter-scale rotor robot driven by a uniaxial alternating magnetic field, capable of hovering, controlled flight, and impact recovery [Figure 4A and B]. It represents one of the lightest and smallest magnetically driven aerial robots to date, providing a novel design paradigm for wireless aerial robotics^[66].

Figure 4. Representative untethered actuation technologies in soft robotics. (A) Working principle of the untethered subcentimeter-scale flying robot; (B) Optical images for the sustained upward flights. (A) and (B) were adapted from Ref.^[66], under CC BY-NC 4.0 license; (C) Image of a fingertip holding the fabricated Kresling crawler and designed magnetization directions of the four magnetic plates for distributed torques, scale bars: 5 mm; (D) Steering and navigation of the Kresling crawler, scale bars: 5 mm. (C) and (D) were adapted from Ref.^[67], under CC BY-NC 4.0 license; (E) Accordion crawling inside the inclined tube, scale bars: 10 mm. Adapted from Ref.^[71], under CC BY 4.0 license; (F) Schematic of the actuating mechanism for strip-shaped actuators by reversible adsorption and desorption of solvents, scale bars: 2 mm. Adapted from Ref.^[76], under CC BY 4.0 license; (G) Translational motion of the microrobot in a circular microchannel, scale bars: 100 μm. Adapted from Ref.^[81], under CC BY-NC 4.0 license.

In bioinspired structural actuation, a magnetic origami robot inspired by earthworms and inchworms was developed to achieve planar contraction-based crawling. The system consists of bipolar dipole modules constructed from four-unit Kresling origami structures, generating anisotropic torque distributions under a magnetic field for tether-free, multi-degree-of-freedom motion [Figure 4C and D]. Its internal cavity can be utilized for drug storage and release, demonstrating multifunctional potential for minimally invasive medical applications^[67]. To address mobility challenges in complex 3D environments, another study proposed a magnetically actuated soft robot featuring a “peel–load” mechanism. This system combines microstructured adhesive pads with bioinspired adhesive feet, enabling vertical and inverted climbing on complex geometries or soft, wet surfaces (e.g., porcine tissue) through coupled global deformation and magnetic field control. This multimodal adhesion mechanism significantly enhances the utility of soft robots for environmental sensing and minimally invasive interventions^[68].

Despite these advantages, magnetic actuation still faces challenges in practical applications, including non-uniformity in controlled magnetic fields and interference from metallic environments, which can limit deployment in complex scenarios. Nevertheless, magnetically actuated soft robots continue to show broad potential in minimally invasive surgery, in vivo navigation, and targeted therapy, representing a key direction for the development of next-generation intelligent medical devices.

3.3. Comparative analysis of actuation technologies

This section reviews the major actuation strategies used in soft robotics, including fluidic, tendon-driven, electroactive, thermal, magnetic, optical, chemical, and acoustic actuation. These approaches exhibit distinct characteristics in terms of output capability, response speed, controllability, structural complexity, and environmental adaptability. Table 1 summarizes the performance characteristics of representative soft actuation technologies under different stimulation methods, including their key advantages, main limitations, output capability, and response speed.

Although existing actuation technologies have enabled a wide range of deformation modes and locomotion behaviors, no single actuation strategy can simultaneously satisfy all requirements for output force, response speed, energy efficiency, controllability, miniaturization, and environmental adaptability^[83]. Fluidic and tendon-driven systems generally provide large deformation and output force but often rely on external driving equipment, whereas electroactive and thermal actuators offer compact structures and high integration potential while facing challenges related to driving conditions, energy efficiency, and long-term stability^[84]. Untethered actuation approaches, including magnetic, optical, chemical, and acoustic actuation, enable remote and non-contact operation, yet their practical deployment remains constrained by limitations in energy transfer, environmental dependence, and control complexity.

Future developments are expected to move beyond the optimization of individual actuation performance and instead focus on application-oriented co-design of smart materials, mechanical structures, energy supply, sensing feedback, and control strategies. Such integrated approaches will be essential for achieving soft robotic systems with predictable deformation, improved robustness, enhanced adaptability, and autonomous operation.

4.1. Resistive and piezoresistive sensors

Resistive and piezoresistive sensors primarily rely on changes in electrical resistance to detect mechanical stimuli, including force, pressure, strain, and deformation, and represent one of the most widely studied sensing approaches in flexible electronics and soft robotics. When subjected to stretching, compression, or bending, the conductive networks within the sensing material may be disrupted, reconfigured, or brought into closer contact, resulting in measurable resistance variations^[104]. For conductive materials obeying Ohm’s law, the resistance can be expressed as:

where ρ is the material resistivity, L is the length, and A is the cross-sectional area^[105]. According to this relationship, the resistance is proportional to the conductor length and inversely proportional to its cross-sectional area. Flexible resistive and piezoresistive sensors often exhibit nonlinear response behavior owing to structural heterogeneity, conductive network evolution, and microcrack formation during deformation. As the applied pressure increases, different sensing mechanisms may dominate the resistance response, resulting in distinct low-, medium-, and high-pressure regimes with varying sensitivities. To improve sensing performance, various surface microstructures, such as hemispherical, honeycomb, and pyramidal architectures, have been introduced to increase effective contact area and optimize stress distribution, thereby enhancing pressure and strain sensitivity^[106-108].

It should be noted that resistive and piezoresistive sensing are closely related but not identical. In general resistive sensing, the measured resistance change can arise from geometric variations, contact resistance changes, conductive pathway reconstruction, or crack opening and closure during deformation. This mechanism is commonly observed in flexible composite sensors based on conductive fillers, liquid metals, conductive polymers, carbon nanomaterials, or metallic networks. By contrast, piezoresistive sensing more specifically refers to the intrinsic change in the electrical resistivity of a material under mechanical stress or strain, which is particularly evident in semiconductor-based sensing materials. In practical flexible sensors, especially composite or microstructured systems, these two effects often coexist and jointly contribute to the output signal^[109]. Therefore, resistive and piezoresistive sensors are discussed together in this section because both are widely used to convert mechanical deformation into resistance-based electrical signals in soft robotic systems; however, their distinct physical origins should be considered when evaluating sensitivity, linearity, hysteresis, temperature dependence, and long-term stability.

Leveraging these structural optimization strategies, flexible resistive and piezoresistive sensors have achieved significant improvements in sensitivity, detection range, and mechanical adaptability. For example, a wearable piezoresistive sensor constructed from carbon nanofiber and PDMS composites exhibited a linear force sensitivity of approximately 1.82 kN^-1 under forces below 20 N, while maintaining sewability and mechanical adaptability^[110]. With further research, these sensors have expanded from simple pressure detection to multimodal recognition, enabling differentiation between touch, sliding, and other tactile modes [Figure 6A]^[111]. Such capabilities further broaden the applicability of resistive sensing technologies in flexible and soft robotic systems^[112,113].

Figure 6. Representative examples of resistive, capacitive, and piezoelectric sensors for flexible sensing applications. (A) Different sensing signals for tactile and slip sensing of objects are captured using ultrasensitive touch sensors. Adapted with permission from Ref.^[111], Copyright © 2024 Wiley-VCH GmbH; (B) Schematic diagram of the MPSS sensor to be applied in hearing aid. Adapted from Ref.^[114], under CC BY 4.0 license; (C) Fabrication process and photograph of the strain sensor. Adapted from Ref.^[121], under CC BY-NC 4.0 license; (D) Image displaying the fabricated 15 × 15 flexible capacitive pressure sensor array. Reprinted from Ref.^[125], under CC BY-NC-ND 4.0 license; (E) Optical images and physical photos of smart wristbands based on flexible piezoelectric ion sensors. Adapted with permission from Ref.^[131], Copyright © 2025 Wiley-VCH GmbH; (F) Schematic image and digital photographs of the implantation of packaged piezoelectric-device in the thigh and chest areas of rats. Adapted with permission from Ref.^[132], Copyright © 2025 Wiley-VCH GmbH. MPSS: MXene/polyvinyl alcohol sound sensor.

Beyond tactile sensing, resistive and piezoresistive technologies have also been extended to acoustic and vibration sensing. For example, a piezoresistive hearing aid was reported to detect sound pressures as low as 60 dB with a frequency response of 20-4,000 Hz [Figure 6B]^[114]. In addition, sensors incorporating porous inclined hollow-cone microstructures achieved ultrahigh sensitivity and were successfully integrated into artificial tympanic membrane systems for acoustic signal detection^[115]. Recent developments have further demonstrated the capability of resistive sensing platforms to capture voice-induced vibrations and complex human motions, enabling advanced signal recognition and decoding through data-processing algorithms^[116,117]. These advances highlight the versatility of resistive sensing technologies and broaden their potential for intelligent perception systems.

In general, resistive and piezoresistive sensors require relatively simple signal-conditioning circuits and offer excellent design flexibility, making them adaptable to a wide range of sensing requirements. However, they remain subject to limitations such as baseline drift, hysteresis, and dependence on external power supplies, which may affect measurement accuracy and long-term stability. In addition, commonly used sensing materials, including conductive polymers, carbon-based nanomaterials, metallic conductors, and composite materials, are often sensitive to temperature variations, particularly under extreme environmental conditions, introducing additional signal noise and reducing reliability^[118]. Nevertheless, owing to their mature fabrication processes, structural adaptability, and design versatility, resistive and piezoresistive sensors have demonstrated strong capabilities in detecting strain, pressure, and complex deformation states. As sensing materials and structural designs continue to advance, these sensors are expected to remain important components in soft robotics, wearable electronics, and intelligent perception systems, while ongoing efforts focus on improving signal stability, environmental robustness, and long-term reliability.

4.2. Flexible capacitive sensors

Flexible capacitive sensors represent another widely used sensing modality in soft robotics. Their typical structure consists of two flexible electrodes sandwiching a dielectric layer, forming a capacitor structure, most commonly in a parallel-plate configuration. The capacitance C can be expressed as:

where ε₀ is the vacuum permittivity, ε_r is the relative dielectric constant of the dielectric layer, A is the effective electrode area, and d is the separation distance between the two electrodes^[119]. When subjected to external stimuli such as pressure, stretching, or bending, one or more of these parameters change, resulting in measurable variations in capacitance. Capacitive sensors offer fast response, high sensitivity, wide dynamic range, and stable signal output, making them particularly suitable for high-resolution, reliable detection of physical quantities.

Thanks to their structural flexibility, low power consumption, and compatibility with soft materials, capacitive sensors are widely integrated into wearable and bio-integrated platforms. For example, a smart contact lens combining a Kirigami-inspired design with an ionic capacitive sensing array demonstrated the capability of conformal pressure monitoring on curved ocular surfaces^[120]. In medical applications, capacitive strain sensors integrated with communication coils in surgical sutures enabled simultaneous strain sensing and wireless signal transmission, highlighting their potential for minimally invasive monitoring [Figure 6C]^[121]. Furthermore, capacitive sensing has been incorporated into flexible textile-based interfaces, where high spatial resolution and stable signal output support accurate detection of human motions and interaction inputs^[122].

Despite these strengths, capacitive sensors face several challenges. Environmental factors such as temperature, humidity, and electromagnetic interference can affect signal stability^[123]. The proximity of conductive objects may induce false triggers, while sensitivity to both pressure and strain can lead to multi-source signal coupling. Furthermore, achieving flexibility and stretchability requires precise electrode design and encapsulation, increasing integration complexity and cost. To address these limitations, various structural and material innovations have been proposed. For instance, a capacitive sensor based on a polyimide porous layered ionic fiber network with high-adhesion interfaces demonstrated significantly improved signal stability and mechanical durability. The optimized structure achieved a high sensitivity of 156.6 kPa^-1, a broad sensing range of up to 4,000 kPa, and stable operation over more than 150,000 loading cycles^[124]. These results highlight the effectiveness of interface engineering and porous structural design in enhancing the reliability and long-term performance of capacitive sensing systems. Another study identified edge fields as the primary source of capacitive interference and developed a sensor architecture that concentrates the electric field between electrodes, eliminating the need for shielding [Figure 6D]. By introducing a nanoscale electrode-gap structure (~900 nm spacing), crosstalk between adjacent sensing units was significantly reduced, enabling spatial resolution comparable to human tactile perception while mitigating bending-induced signal interference^[125]. In addition to improving sensing performance, recent advances in flexible conductive fabrics, ionic hydrogels, and nanocomposite electrodes have expanded the functional scope of capacitive sensing systems. For example, capacitive sensors incorporating magnetically responsive microstructures can simultaneously detect pressure, proximity, and magnetic stimuli^[126]. Similarly, ionic hydrogel- and nanocomposite-based electrode designs have enabled distributed tactile sensing, deformation monitoring, and multifunctional sensing capabilities, further enhancing the versatility of capacitive sensing technologies.

In summary, flexible capacitive sensors offer several advantages, including high sensitivity, fast response, low power consumption, and excellent structural compatibility. However, challenges such as environmental susceptibility, signal coupling, and fabrication complexity remain. Continued advances in microstructured dielectric layers, electrode materials, and device integration strategies are expected to further improve sensing reliability and multifunctionality, supporting broader applications in soft robotics, wearable electronics, and intelligent sensing systems.

4.3. Piezoelectric sensors

Piezoelectric sensors operate based on the piezoelectric effect, in which certain materials generate electrical charges when subjected to external mechanical stimuli such as pressure, vibration, or deformation. Common piezoelectric materials include piezoelectric ceramics, such as lead zirconate titanate (PZT) and barium titanate (BTO), as well as piezoelectric polymers, including PVDF and its copolymers^[127]. Upon mechanical loading, these materials generate a potential difference proportional to the applied stimulus, producing measurable electrical signals. Because electrical outputs are generated directly from mechanical inputs, piezoelectric sensors can operate in a self-powered manner without continuous external power supply. Their high sensitivity, rapid response, and excellent capability for detecting dynamic mechanical stimuli make them particularly suitable for pressure, vibration, acoustic, and tactile sensing applications^[128-130].

For example, a piezoelectric ionic sensor based on CNT and graphene nanocomposite electrodes exhibited high sensitivity to dynamic deformation and was capable of detecting subtle facial and body motions [Figure 6E]^[131]. In another study, a piezoelectric silk fibroin thin film generated sufficient electrical output under mechanical stimulation, highlighting the potential of piezoelectric materials for self-powered bioelectronic systems [Figure 6F]^[132]. Furthermore, piezoelectric nanofiber-based sensing platforms have been applied to acoustic signal detection, where the combination of high-frequency response and advanced signal-processing algorithms enabled accurate sound-source localization^[133].

However, piezoelectric sensors also exhibit several intrinsic limitations. Because the piezoelectric effect generates transient electrical outputs, continuous measurement of static forces remains challenging. In addition, piezoelectric materials are susceptible to fatigue, and long-term operation may result in sensitivity degradation. Material selection therefore plays a critical role in device performance. Piezoelectric ceramics generally exhibit strong electromechanical coupling and high piezoelectric performance but suffer from intrinsic brittleness and limited mechanical compliance, which restrict their use in highly deformable systems^[134,135]. In contrast, polymer-based piezoelectric materials, such as PVDF and its copolymers, offer superior flexibility, processability, and biocompatibility owing to their molecular dipole orientation mechanisms, making them attractive candidates for next-generation flexible piezoelectric sensors^[136]. Recent work combining PVDF piezoelectric microspheres with a PDMS elastomer matrix produced a flexible composite sensor with a sensitivity of up to 540 mV·N^-1 and a broad response range of 5-40 N, highlighting the advantages of polymer-based piezoelectric materials in achieving both mechanical flexibility and effective electromechanical conversion^[137].

Unlike resistive and capacitive sensors, piezoelectric sensors can directly convert mechanical stimuli into electrical signals, enabling self-powered sensing without continuous external power input. They are particularly well suited for detecting dynamic and high-frequency mechanical stimuli, although their ability to continuously monitor static forces remains inherently limited. To address these challenges, ongoing research focuses on advanced piezoelectric materials, composite structures, and flexible device architectures to improve durability and sensing performance. Owing to their unique combination of self-powered operation, rapid response, and high sensitivity, piezoelectric sensors are expected to remain key components in next-generation flexible sensing systems.

4.4. Triboelectric sensors

Since the introduction of triboelectric nanogenerators (TENGs) in 2012, this technology has injected new momentum into the field of flexible electronics with its unique energy self-supply mechanism and has shown significant advantages in building smart sensing systems without external power supply^[138,139]. TENG operates on the principles of triboelectrification and electrostatic induction: when two materials with differing electron affinities contact and separate, charge transfer occurs at the surfaces, generating a potential difference and an induced current. Its self-powered nature, structural simplicity, and high energy conversion efficiency have facilitated broad applications in biomedical monitoring^[140-145], acoustic detection^[146], and noise suppression^[147,148], energy harvesting^[149-153], and flexible sensing platforms^[154-157].

Owing to their high sensitivity to weak mechanical stimuli, TENG-based sensors have shown great potential in wearable motion monitoring and physiological signal detection^[158,159]. For example, a drum-structured TENG (DS-TENG) was developed to detect subtle pressure variations with high sensitivity, enabling accurate acquisition of biomechanical and physiological signals^[160]. Similarly, flexible TENG sensors based on hierarchically reduced graphene oxide (rGO) demonstrated excellent motion-sensing and deformation-monitoring capability, enabling precise detection of joint movements and other dynamic mechanical stimuli^[161]. TENG-based sensors have also demonstrated strong capabilities in physiological signal monitoring. For example, triboelectric stethoscopes achieved a signal-to-noise ratio of 36 dB and a recognition accuracy of 97%, outperforming conventional piezoelectric counterparts in acoustic signal acquisition [Figure 7A]^[146]. Similarly, a fully nanofiber-based TENG respiratory sensor enabled real-time and long-term monitoring of respiratory activity^[162]. Further studies have also explored TENG-based sensing systems for rehabilitation monitoring, where triboelectric sensing has been shown to support recovery assessment of hand sensory functions after stroke^[163]. Beyond single-parameter sensing, TENG technology has gradually evolved toward multifunctional sensing and intelligent information processing. Flexible sensing patches integrating non-contact triboelectric layers with hydrogel-based pressure and temperature sensors enabled simultaneous detection of pressure, temperature, and proximity signals without cross-interference^[164]. Likewise, TENG-driven smart neckbands provided continuous monitoring of neck motion and deformation^[165]. When coupled with embedded electronics and deep-learning algorithms, triboelectric sensing systems further expanded from signal acquisition to intelligent behavioral analysis, enabling functions such as fall detection, behavior recognition, and non-contact navigation^[166,167].

Figure 7. Representative examples of triboelectric and magnetic sensors for flexible sensing applications. (A) Schematic illustration of cardiac sounds sensing using the triboelectric stethoscope and exploded view diagram of the overall design structure. Adapted with permission from Ref.^[146], Copyright © 2024 Wiley-VCH GmbH; (B) Structure of the triboelectric tactile perception smart finger. Adapted from Ref.^[169], under CC BY-NC 4.0 license; (C) The 3D object tele-perception system based on the bionic electroreceptor matrix (20 × 20 units) and the CNN. Adapted from Ref.^[172], under CC BY 4.0 license; (D) Optical images of the prepared magnetoelectric tactile sensor and the sensor installed at the tip of the robot arm. Adapted from Ref.^[195], under CC BY 4.0 license. CNN: Convolutional neural networks; Al: aluminium; FEP: fluorinated ethylene propylene; PI: polyimide; ABS: acrylonitrile butadiene styrene; OLED: organic light-emitting diode.

Compared with many other flexible sensing technologies, TENGs exhibit exceptional sensitivity to contact electrification signals, enabling the extraction of rich information beyond conventional force and deformation measurements^[168]. This capability allows triboelectric sensors to distinguish material properties, surface textures, contact conditions, and spatial characteristics of external objects. By combining triboelectric sensing with machine-learning algorithms, researchers have further enhanced the ability of TENG-based systems to interpret complex tactile information. For example, an intelligent finger system integrated triboelectric outputs with deep-learning models to extract and classify subtle differences in material-dependent triboelectric signals, enabling rapid identification of complex material properties [Figure 7B]^[169]. Another study reported a bio-inspired finger-shaped triboelectric tactile sensor capable of simultaneously perceiving multidirectional forces and surface-dependent triboelectric signatures, achieving material discrimination among 12 different materials with an accuracy of 98.33%^[170]. These studies demonstrate the potential of triboelectric sensing for advanced tactile information acquisition and intelligent perception.

The application of TENGs in non-contact sensing has further expanded the scope of tactile perception^[171]. For instance, a non-contact triboelectric sensor achieved a remote sensitivity of 14.2 (ΔV/Δd) and a material recognition accuracy of 99.56%, enabling simultaneous extraction of spatial and material information for three-dimensional object reconstruction and material identification [Figure 7C]^[172]. Other studies have demonstrated that non-contact TENGs can reliably capture spatial position, proximity variation, motion trajectory, and dynamic interaction information under continuously changing conditions, highlighting their robustness in dynamic sensing scenarios^[173-177]. These advances indicate that the strength of triboelectric sensing lies not only in identifying static material properties, but also in acquiring rich spatial and dynamic interaction information. As a result, TENG-based sensing systems are evolving from conventional contact detection platforms toward advanced perception systems capable of extracting multidimensional environmental information. Beyond sensing, the self-powered nature of TENGs enables simultaneous energy harvesting and signal generation, distinguishing them from most conventional flexible sensing technologies and providing opportunities for autonomous soft robotic systems^[178,179].

In summary, TENGs have emerged as an important class of flexible sensing technologies owing to their self-powered operation, structural versatility, and high sensitivity. From physiological signal monitoring and advanced tactile perception to non-contact sensing and multimodal information acquisition, TENG-based sensors have demonstrated capabilities that extend beyond conventional flexible sensing approaches. Particularly in soft robotic systems, TENGs enable real-time detection of environmental stimuli and motion-related information while simultaneously harvesting mechanical energy from the surroundings. These unique characteristics provide a promising foundation for the development of next-generation flexible intelligent systems featuring self-powered operation, multimodal perception, and enhanced environmental adaptability.

4.5. Other sensors

In addition to the sensor types discussed above, a variety of other sensing technologies, including optical, magnetic, chemical, and acoustic sensors, have also been explored for soft robotic perception. Among them, optical sensors are particularly attractive due to their high sensitivity, immunity to electromagnetic interference, and capability for distributed sensing. Optical sensors detect external stimuli by monitoring changes in light intensity, phase, wavelength, or transmission characteristics induced by mechanical deformation^[180]. Traditional optical tactile sensing systems often rely on rigid optical configurations. For example, TacTip^[181] and GelSight^[182] employ camera-based imaging to capture deformation of compliant interfaces, enabling high-spatial-resolution tactile sensing. However, their bulky structures and rigid optical components limit integration into highly deformable soft robotic systems. To address these limitations, fiber Bragg grating (FBG) technology has been introduced for strain and pressure sensing in soft robots, enabling distributed and multipoint sensing along a single optical fiber without requiring local electronic components^[183]. Nevertheless, FBG sensors still face challenges including complex fabrication processes, high cost, and limited stretchability, which restrict their broader deployment in soft robotic applications. More recently, flexible and stretchable optical fiber sensors have demonstrated rapid and stable responses, large-deformation compatibility, and good cyclic durability, providing promising alternatives for distributed sensing in highly deformable systems^[184]. Recent studies have further demonstrated that soft optical waveguide sensors combined with intelligent decoding algorithms can reconstruct complex three-dimensional deformations from a limited number of optical signals, highlighting the potential of optical sensing for high-density proprioceptive perception with reduced sensing complexity^[185]. Furthermore, integrating polymer optical fibers with artificial intelligence algorithms has expanded the capabilities of intelligent optical sensing, enabling more efficient interpretation of complex deformation and interaction information^[186,187].

The adoption of rigid–flexible hybrid architectures has further improved the interference resistance and mechanical robustness of optical sensing systems, enhancing their reliability under complex deformation conditions^[188]. In parallel, emerging optical sensing strategies have continued to push the performance limits of conventional systems. For example, exceptional point enhanced optical sensing has demonstrated significantly improved sensitivity, providing a promising route toward high-performance flexible sensing platforms^[189]. Furthermore, by exploiting multi-wavelength encoding mechanisms induced by graded structures in flexible optical fibers, researchers have achieved efficient decoupling of pressure and stress signals while enabling spatially distributed sensing, highlighting the potential of optical sensors for multidimensional information acquisition and high-density perception^[190]. These advantages have also facilitated the adoption of optical sensing technologies in demanding scenarios such as wearable monitoring and biomedical diagnostics^[191-193]. Overall, optical sensing technologies provide unique advantages in high-resolution tactile perception, distributed deformation monitoring, and electromagnetic-interference-free operation. Although challenges remain in terms of stretchability, system complexity, and manufacturing cost, ongoing advances in flexible optical materials, intelligent signal processing, and distributed sensing architectures continue to expand their potential for soft robotic perception and intelligent sensing systems.

Magnetic sensors detect external stimuli by monitoring variations in magnetic field distributions generated by magnetic materials or permanent magnets. Owing to their non-contact sensing capability, high sensitivity, and compatibility with soft structures, magnetic sensing has emerged as an attractive approach for monitoring force, displacement, deformation, and orientation in soft robotic systems. Furthermore, magnetic signals can penetrate non-magnetic encapsulation materials, enabling distributed and remote sensing in complex or confined environments. However, susceptibility to environmental magnetic interference and the complexity of signal processing remain important challenges that limit sensing accuracy and system integration^[194]. Recent advances in flexible magnetic sensing have focused on both structural design and material engineering. For example, self-powered flexible magnetic sensing architectures have been developed to improve device autonomy and system-level integration, providing new opportunities for intelligent sensing platforms [Figure 7D]^[195]. In parallel, bioinspired magnetic cilia structures integrated into flexible magnetic sensing arrays have enabled simultaneous detection of force magnitude and direction, significantly improving multidimensional tactile perception and spatial sensing resolution^[196]. Furthermore, the combination of flexible magnetic materials with amorphous magnetic wires has led to wide-range magnetic sensing structures capable of maintaining low detection limits, high stability, and broad dynamic response ranges, thereby expanding the capability of magnetic sensors for multidimensional and long-range perception^[197]. Despite these advances, magnetic sensors continue to face challenges in interference suppression, signal interpretation efficiency, and scalable manufacturing. Further progress in structural optimization, magnetic material development, and integrated sensing architectures will be essential for improving reliability and practical applicability. Nevertheless, their unique advantages in non-contact sensing, distributed perception, and multidimensional force detection suggest considerable potential for future soft robotic perception and intelligent sensing systems.

Acoustic and chemical sensing technologies have been explored as complementary sensing modalities for soft robotic perception, providing capabilities beyond conventional mechanical stimulus detection^[198]. Acoustic sensors exploit the propagation and reflection characteristics of sound waves to achieve non-contact sensing, offering strong penetration capability and high sensitivity to subtle structural variations. These characteristics make them attractive for applications such as structural state monitoring and remote perception^[199]. However, challenges associated with environmental noise, signal attenuation, and integration with highly deformable materials continue to limit their broader adoption in soft robotic systems. Chemical sensors, in contrast, enable direct detection of environmental parameters such as gas composition, ion concentration, and pH variation, thereby expanding the perception capabilities of soft robots toward chemical and biological environments^[200]. Recent efforts have focused on incorporating flexible functional films, nanomaterials, and bioinspired sensing materials to improve sensitivity, selectivity, and response speed. Nevertheless, achieving simultaneous flexibility, long-term stability, and high integration density remains a significant challenge for practical deployment. Overall, acoustic and chemical sensors provide additional perceptual dimensions beyond force, deformation, and motion sensing, enabling soft robots to acquire information from complex physical, chemical, and biological environments. Although their current applications remain relatively limited compared with mainstream flexible sensing technologies, continued advances in material design, sensing mechanisms, and system-level integration are expected to further expand their role in future intelligent soft robotic systems.

4.6. Multimodal sensing technologies

In addition to single-modality sensing mechanisms, multimodal sensing technologies have attracted increasing attention in soft robotics because they enable simultaneous acquisition of multiple types of physical and environmental information. By integrating different sensing mechanisms, multimodal sensors can detect pressure, strain, bending, torsion, temperature, humidity, proximity, slip, and material-related information within a single device or sensing system^[201]. This capability is particularly important for soft robots operating in unstructured environments, where deformation, contact, and environmental stimuli often occur simultaneously^[202].

Multimodal sensing can be realized through several strategies. One approach is to integrate multiple sensing units with different transduction mechanisms onto a shared flexible substrate. Another approach is to design multifunctional materials or microstructures that generate distinguishable electrical, optical, or magnetic responses to different stimuli^[203]. In addition, signal processing and machine learning methods are increasingly used to decouple overlapping signals and identify complex interaction states. Compared with single-modality sensors, multimodal sensing systems can provide richer perceptual information and improve the robustness of robotic perception in tasks such as object recognition, human–robot interaction, manipulation, locomotion, and confined-space exploration^[204].

Despite these advantages, multimodal sensing still faces several challenges. The simultaneous integration of multiple sensing mechanisms increases structural complexity and fabrication difficulty. Signal coupling among different stimuli can cause cross-sensitivity, making accurate signal decoupling and calibration difficult^[205]. In addition, the increased number of sensing channels may lead to more complicated wiring, higher data acquisition requirements, and greater computational burden. Therefore, future multimodal sensing systems for soft robotics should focus not only on increasing the number of detectable stimuli, but also on improving signal independence, mechanical compatibility, long-term stability, scalable fabrication, and intelligent data interpretation. Table 2 summarizes the representative sensing technologies discussed in this section by comparing their sensing principles, sensing functionalities, advantages, and limitations in soft robotic perception.

4.7. Comparative analysis of sensing technologies

The sensing technologies discussed above have significantly expanded the perceptual capabilities of soft robots, enabling deformation monitoring, tactile perception, environmental recognition, and closed-loop control. However, each sensing mechanism exhibits inherent advantages and limitations, and no single sensing technology can simultaneously achieve high sensitivity, large deformation compatibility, fast response, long-term stability, low complexity, and seamless integration. Resistive and piezoresistive sensors remain widely adopted due to their simple structures, low cost, and compatibility with large-area fabrication, but their performance can be affected by hysteresis, drift, temperature variation, and mechanical fatigue. Capacitive sensors provide fast response, low power consumption, and stable electrical signals, while requiring careful consideration of parasitic capacitance, electromagnetic interference, and structural coupling effects. Piezoelectric and triboelectric sensors offer high sensitivity to dynamic stimuli and, in some cases, self-powered operation, but their applications are limited by static sensing capability, environmental dependence, and signal stability. Optical and magnetic sensing approaches provide advantages in remote perception, distributed sensing, and immunity to electrical interference, although they usually require more complex signal-processing systems and integration strategies^[206].

Therefore, the selection of sensing technologies for soft robotic systems should be determined by specific sensing tasks, deformation characteristics, and operating environments rather than by a single performance metric. For proprioceptive sensing, key requirements include mechanical compatibility, repeatability, low hysteresis, and stable responses under large deformation. For exteroceptive perception, high spatial resolution, sensitivity, response speed, and environmental robustness become more critical^[207]. Meanwhile, multimodal sensing represents an emerging direction for overcoming the limitations of individual sensing mechanisms by combining complementary information from different modalities. However, future developments should focus not only on integrating more sensing elements, but also on achieving effective signal decoupling, scalable sensor networks, intelligent data interpretation, and reliable perception under complex operating conditions. The combination of advanced materials, structural design, multiplexed sensing architectures, and data-driven algorithms will be essential for developing next-generation soft robotic systems with robust and adaptive perception capabilities^[208].

5.1. Proprioception in soft robotics

Proprioception in soft robots is considerably more challenging than in rigid robots due to their theoretically infinite degrees of freedom and the complex deformations that can arise from both internal actuation and external loads. To achieve effective self-sensing, soft robots have integrated various types of sensors, often complemented by computational approaches such as machine learning and deep learning to enhance proprioceptive capabilities^[215]. Here, we review representative sensing strategies that enable proprioception in soft robots.

Among various sensing strategies, resistive sensing has become one of the most widely adopted approaches for proprioception in soft robots owing to its high flexibility, fast response, simple structure, and ease of integration. By embedding resistive sensing elements directly into soft robotic structures, continuous monitoring of strain and deformation can be achieved, providing essential feedback for shape estimation, motion-state perception, and closed-loop control. To improve the capability of soft robots to perceive complex deformations and motion states, a variety of integrated resistive sensing schemes have been developed, many of which employ conductive materials such as CNTs, MXene, and liquid metals to construct flexible sensing structures^[216].

For instance, Banerjee et al. integrated a carbon nanotube-reinforced natural rubber (CNT/NR)-based strain sensor onto a pneumatic soft arm to achieve continuous monitoring of actuator deformation during motion. The sensor enabled real-time perception of complex deformation states and provided proprioceptive feedback for closed-loop control, thereby enhancing motion accuracy during multi-degree-of-freedom operations^[217]. Kim et al. integrated SCP (self-coagulating conductive Pickering)-based conductive sensing structures into soft balloons, grippers, and artificial muscle actuators, enabling synchronous self-sensing of structural deformation during actuation [Figure 8A]^[218]. Wang et al. integrated a MXene/CNT/fluoroelastomer-based strain sensor into a caterpillar-inspired soft robot, enabling continuous shape perception and self-monitoring throughout locomotion. Benefiting from its wide sensing range and high sensitivity, the system provided reliable feedback on structural deformation and body configuration during movement^[219]. Additionally, Wang et al. developed a liquid-metal-based resistive strain sensor and applied it to movement-state monitoring in a crawling soft robot. Owing to its high stretchability and self-healing capability, the sensor continuously tracked large deformations during locomotion and provided real-time feedback on the robot’s movement state^[220].

Figure 8. Application of flexible sensors in the proprioception of soft robot. (A) A silicone rubber four-legged gripper, onto which SCP emulsion traces are printed. Adapted with permission from Ref.^[218], Copyright © 2020, AAAS; (B) Image of the fabricated hand mounted on a robotic arm with each finger. Adapted with permission from Ref.^[229], Copyright © 2016, AAAS; (C) Image of the SLIMS-integrated soft glove with LED light source and electronics module. Adapted with permission from Ref.^[230], Copyright © 2020, AAAS; (D) Robotic skins embed distributed actuation and sensing into a conformable substrate to produce different forms of motion effects. Adapted with permission from Ref.^[234], Copyright © 2018, AAAS; (E) Different deformation modes by self-actuation as well as human inputs and corresponding sensor data. Adapted with permission from Ref.^[235], Copyright © 2020, AAAS; (F) Fiber-based deformation sensors, relaxed and bent states. Adapted from Ref.^[236], under CC BY 4.0 license. SCP: Self-coagulating conductive Pickering; SLIMS: stretchable lightguide for multimodal sensing; LED: light emitting diode.

To further enhance the perception of complex deformation states in soft robots, sensing systems capable of directional strain recognition and curvature estimation have been developed. For example, Mousavi et al. integrated a 3D-printed anisotropic resistive sensor into a soft robotic system to detect and quantify multi-axis bending angles and deformation degrees^[221]. By combining heterogeneous conductive fillers with tailored sensor geometries, the sensor achieved a high directional response ratio of up to 31.4, enabling more accurate discrimination of deformation directions. In addition, Liu et al. developed a helically wound conductive fiber sensor for curvature estimation in soft robotic components such as soft grippers. With a wide strain range and high response linearity, the sensor can accommodate large-curvature deformation and different structural configurations. When integrated into flexible grasping tasks, it enabled continuous posture monitoring and dynamic motion control^[222].

To further simplify the structure and functionality of soft robots, researchers have actively explored the deep integration of sensors with actuators, enabling direct perception of the robot’s state through changes in material resistance. Cao et al. developed a thermally actuated soft actuator^[223], in which internal heating wires serve both as actuators and as self-sensing elements, detecting deformation through resistance changes. This design establishes a closed-loop control mechanism without the need for additional sensors. Similarly, Yang et al. constructed modular actuation units based on shape-changing polymer artificial muscle (SCPAM) wires. The resistance response of the wires during deformation was used to estimate bending angles, and this approach was applied in caterpillar-inspired robots and soft manipulators, achieving direct coupling of sensing and motion at the material level^[224].

To extend proprioception from individual sensing units to robot-level state awareness, researchers have explored distributed and system-level sensing architectures for soft robots. These approaches enable continuous monitoring of robot configurations across multiple locations, thereby improving adaptability and control precision in complex environments. Goldoni et al. integrated a stretchable CNT–silicone sensing system with flexible circuits to achieve wireless monitoring of soft robot motion states^[225]. By tracking strain variations across different segments during movement, the system provided distributed proprioceptive feedback and supported operation in confined or remote environments. Alatorre et al. embedded stretchable ionic-liquid-based resistive sensors into soft robotic arms to enable proprioception of joint angles and curvature in continuum structures^[226]. The system provided reliable state feedback during deformation, supporting closed-loop control. Truby et al. developed a distributed proprioceptive system for a pneumatic soft robotic arm using plasma-modified flexible electronic skins. The system enabled continuous real-time monitoring of actuator shapes and could further predict future deformations, providing enhanced state awareness^[227].

Beyond resistive sensing, alternative sensing strategies have been explored to provide higher-resolution proprioceptive information, particularly for shape reconstruction and complex deformation perception in soft robots. Optical fiber sensors, in particular, have shown significant potential for shape sensing applications. Galloway et al. first embedded a monolithic multi-core fiber optic shape sensor (FOSS) into a fiber-reinforced soft actuator, achieving submillimeter real-time position feedback and enabling detection of actuator shape changes, collision locations, and external object stiffness. Although this sensing scheme operates in an open-loop configuration, it has demonstrated effectiveness in complex morphology monitoring tasks^[228]. In 2016, the Shepherd group integrated stretchable optical waveguides into fiber-reinforced soft prostheses, constructing a proprioceptive system based on changes in light transmission. The optical strain sensors exhibit high compliance, low hysteresis, and stable output, allowing real-time monitoring of soft structure bending, stretching, and applied forces. They were successfully applied to prosthetic perception tasks, such as discriminating the shape and firmness of different tomatoes, enabling human-like active sensing and interactive control [Figure 8B]^[229]. Building on this work, the team further developed a stretchable distributed fiber optic sensing system (stretchable DFOS) in 2020, incorporating continuous or discrete chromatic patterns. By analyzing light color and intensity, the sensor can simultaneously detect bending and pressure at different finger joints, enabling real-time decoupling of multiple positions and multi-modal deformations [Figure 8C]^[230]. Yang et al. developed a plant-inspired soft helical gripper that performs automatic wrapping under a single pneumatic input. This system integrates a high-birefringence optical fiber torsion sensor to provide feedback on wrapping angle and external disturbances, achieving a sensitivity of up to 0.03 nm. Such performance supports applications in small-object manipulation and operations within confined spaces^[231].

Beyond resistive and optical sensing approaches, other sensing mechanisms have also been employed to enhance proprioception in soft robots. These systems provide complementary capabilities, including vibration perception, shape reconstruction, and embedded state awareness. Wang et al. proposed a 0.1 mm-thick Kirigami-structured triboelectric e-skin, fabricated using screen printing and liquid metal particles (LMP), which can conform to the soft robot surface to passively monitor vibration signals generated by its own motion^[232]. Hu et al. reported a high-resolution shape reconstruction technique for soft robots based on stretchable capacitive sensors. Integration of this sensor with machine learning techniques enables accurate reconstruction of soft gripper geometry^[233]. Booth et al. from Yale University designed a deformable sheet embedding both actuators and high-deformation capacitive strain sensors, which can be integrated onto any soft object surface to achieve proprioception and realize various motion modes such as grasping, locomotion, and wearable applications [Figure 8D]^[234].

To further enhance the capability of soft robots to perceive complex deformation states, multimodal sensing systems have been developed to simultaneously capture and decouple multiple types of proprioceptive information. For example, Kim et al. developed a multi-modal soft sensor integrating optical, microfluidic, and piezoresistive mechanisms. This system can decouple multiple deformation types, including stretching, bending, and compression, while maintaining a compact form, and can recognize eight different deformation modes with an accuracy exceeding 95% using neural network algorithms. Such sensors not only improve robot state discrimination but also provide possible pathways for flexible human–machine interface applications [Figure 8E]^[235]. Moreover, Lin et al. developed a hybrid sensing network composed of electronic-free fiber-optic sensors and microfluidic pressure sensors, combined with digital cameras and dye-responsive units to achieve remote reading of soft robot states. This system has enabled visual monitoring of bending, twisting, and pressure states in elastic fingers, representing an exploration into low-power, highly integrated proprioception [Figure 8F]^[236].

5.2. Exteroception in soft robotics

Although soft robots can passively adapt to object interactions through their compliant bodies, passive mechanical adaptation alone is insufficient for perceiving external environments and performing precise manipulation tasks. Therefore, exteroception plays a critical role in enabling soft robots to actively acquire information about their surroundings. Tactile perception allows soft robots to detect contact forces, identify object shapes and surface textures, and recognize environmental interactions in real time^[237,238]. Such capabilities provide essential information for tasks ranging from object grasping and manipulation to environmental exploration and navigation^[239,240]. To achieve exteroception, flexible sensing systems are commonly integrated into soft robotic structures, enabling continuous monitoring of interactions between the robot and its environment and providing key sensory feedback for closed-loop control and intelligent decision-making.

In tactile systems based on single-modality sensors, various actuation–sensing co-design strategies have been developed to endow soft robots with environmental perception capabilities. For example, Gu et al. developed a bioinspired pneumatically actuated soft prosthetic hand capable of perceiving fingertip contact pressure during grasping tasks^[241]. By integrating capacitive tactile sensors with an electrical stimulation feedback pathway, the system provided real-time tactile information to users and enabled a preliminary closed loop for tactile perception and interactive response. Inspired by lobster tails, Li et al. developed a soft pneumatic rehabilitation actuator capable of perceiving tensile forces during human–robot interaction through integrated triboelectric sensing. The system enabled force monitoring, spasm assessment, and rehabilitation assistance, demonstrating the potential of tactile perception for interactive rehabilitation applications [Figure 9A]^[242]. Alshawabkeh et al. developed a capacitive proximity tactile sensor that can conform to the free motion of a soft robot. The sensor enabled simultaneous perception of nearby objects and contact forces, allowing the robot to identify different materials and interact more effectively with unstructured environments^[243]. To further enrich environmental perception, Li et al. integrated a flexible piezoresistive sensor based on dual-size microspheres (DMS) into a soft gripper. The system was capable of simultaneously perceiving pressure and temperature stimuli, enabling automatic contact detection and responsive grasping behaviors. Such multimodal tactile feedback supported interactive manipulation tasks, including object grasping and environmental interaction^[244].

Figure 9. Representative exteroceptive sensing systems for soft robots. (A) Bioinspired TENG-SPA design and performance evaluation. Adapted with permission from Ref.^[242], Copyright © 2025 Wiley-VCH GmbH; (B) The soft robotic hand identifies the material and roughness of different objects. Adapted with permission from Ref.^[248], Copyright © 2023 Wiley-VCH GmbH; (C) Schematic of the mechanical gripper mounted with three sets of single-electrode TENGs and tactile sensors. Adapted with permission from Ref.^[249], Copyright © 2025 Wiley-VCH GmbH; (D) Experimental setup of texture discrimination using the hybrid finger. Adapted from Ref.^[250], under CC BY 4.0 license; (E) A depiction of the principle of touchless human-soft robot interaction based on flexible smart skin. Adapted from Ref.^[251], under CC BY 4.0 license. TENG: Triboelectric nanogenerator; SPA: soft pneumatic actuator.

In addition, Lu et al. developed a self-powered tactile sensing system for soft grippers to monitor gripping forces during fruit and vegetable harvesting. By providing real-time feedback on grasping forces, the system enabled adaptive manipulation and non-destructive handling of delicate agricultural products^[245]. Exteroceptive sensing has also been extended to underwater soft robotic systems. For example, triboelectric sensing units have been integrated into octopus-inspired dorsal membranes and tentacle structures, enabling simultaneous underwater grasping and environmental perception^[246]. This design allowed soft robots to interact with surrounding objects while continuously acquiring information about the external environment, demonstrating the potential of integrated sensing for underwater operations.

To further enhance environmental perception in complex scenarios, researchers have explored multimodal sensing strategies that enable soft robots to simultaneously acquire different types of external information. For example, Zhou et al. developed an intelligent soft gripper integrating a photothermal actuator with triboelectric sensing functions. The system was capable of simultaneously perceiving temperature and material characteristics while performing grasping tasks. By combining multimodal sensing information with neural-network-based analysis, the robot achieved accurate identification of different targets, demonstrating enhanced environmental awareness and object recognition capability^[247].

Liu et al. developed a bimodal self-powered sensor combining triboelectric and giant magnetoelastic (GME) effects. When integrated into a soft gripper, the sensor enabled simultaneous detection of non-contact and tactile interactions, allowing the robot to identify object shapes, materials, and surface textures with high accuracy (97%), demonstrating enhanced multimodal environmental perception [Figure 9B]^[248]. Similarly, Zhou et al. proposed a multifunctional soft gripper system integrating capacitive and triboelectric sensing mechanisms. This system achieved high-precision perception of subtle grasping motions and material properties, supporting accurate object manipulation and differentiation in complex environments [Figure 9C]^[249]. Furthermore, to overcome the limitations of fully soft structures in terms of gripping force and sensing accuracy, Sankar et al. developed a rigid–soft hybrid robotic hand integrating piezoresistive and piezoelectric sensors. By combining the precision of rigid mechanisms with the compliance of soft structures, the system enhanced texture recognition and dexterous grasping capabilities [Figure 9D]^[250]. Additionally, Liu et al. designed a flexible bimodal smart skin (FBSS) based on triboelectric and piezoresistive sensing principles. This system could simultaneously detect proximity and contact, enabling contactless teaching and manipulation of soft robotic arms [Figure 9E]^[251].

5.3. Integration of proprioception and exteroception sensing

As discussed above, proprioception enables soft robots to perceive their internal states, including deformation, posture, and motion, whereas exteroception provides information about external objects and environmental interactions. Although significant progress has been achieved in both sensing modalities, practical robotic tasks often require the simultaneous acquisition and interpretation of both internal and external information. In real-world scenarios, proprioceptive and exteroceptive information are intrinsically coupled. For example, during grasping, a soft robot must not only perceive its own deformation and configuration but also identify contact locations, interaction forces, and object properties. Similarly, in navigation and environmental exploration tasks, the robot needs to continuously monitor its body configuration while sensing surrounding obstacles and environmental conditions. Therefore, relying on a single sensing modality is often insufficient to fully characterize the robot’s dynamic behavior and interaction states. To address this challenge, researchers have increasingly explored integrated sensing systems that combine proprioception and exteroception within the same robotic platform. Such systems enable soft robots to simultaneously perceive their internal states and external environments, providing richer sensory information for adaptive control, intelligent decision-making, and complex task execution^[252-255].

For example, Shen et al. developed a conductive polymer hydrogel strain sensor and integrated it into a soft robotic gripper. By monitoring strain variations during grasping, the system could simultaneously estimate the gripper’s bending state and identify the size of grasped objects [Figure 10A]^[256]. Wang et al. proposed an expected–actual perception–action framework and demonstrated it on a soft continuum robot equipped with resistive sensors. By comparing expected and measured shapes, the system could distinguish whether deformation originated from internal actuation or external contact [Figure 10B]^[257]. Homberg et al. developed a soft robotic gripper integrating bending and force sensors to simultaneously monitor gripper configuration and object interaction. The system enabled robust grasping and recognition of objects with uncertain geometries and poses [Figure 10C]^[258]. To simultaneously perceive robot motion and environmental stimuli, Lai et al. developed a self-powered active robotic skin capable of detecting proximity, contact, and pressure information without requiring an external power source. Integrated with soft actuators, the system enabled active perception and responsive interaction, demonstrating potential for applications such as medical monitoring, infant care, and robotic palpation [Figure 10D]^[259]. Similarly, to simultaneously monitor robot posture and object characteristics during manipulation, Xu et al. proposed a Kirigami-structured sensing system that could estimate gripper bending angles while identifying the size of grasped objects. Through the coordinated arrangement of multiple sensing units, the system achieved high-precision posture monitoring and object recognition^[260]. Shu et al. developed a sensing electronic skin based on a differential piezoelectric matrix for soft robots. The system reconstructs the surface shape with high precision and fast response during motion, providing proprioceptive information. Combined with machine learning algorithms, it can also recognize different terrains and obstacles, extending to exteroceptive perception and supporting autonomous task execution in complex environments [Figure 10E]^[261].

Figure 10. Representative single-modality sensing systems for integrated proprioception and exteroception in soft robots. (A) Soft gripper with integrated hydrogel strain sensor for bending monitoring. Adapted with permission from Ref.^[256], Copyright © 2025 Wiley-VCH GmbH; (B) Proprioception and simultaneous contact detection during motion. Reprinted from Ref.^[257], under CC BY-NC-ND 4.0 license; (C) Soft robotic hand performing object grasping. Adapted from Ref.^[258], under CC BY 4.0 license; (D) Image of conscious gripper to hold the doll’s hand. Adapted with permission from Ref.^[259], Copyright © 2018 Wiley-VCH GmbH; (E) Illustration of the SSES-mediated soft robot assembly. Adapted with permission from Ref.^[261], Copyright © 2023 Wiley-VCH GmbH. SSES: Shape-sensing electronic skin; PEDOT:PSS: poly(3,4- ethylenedioxythiophene):poly(styrenesulfonate); PVA: poly(vinyl alcohol); PVDF: polyvinylidene fluoride.

The studies discussed above demonstrate that, through appropriate sensing architectures and integration strategies, a single sensing modality can simultaneously provide information about both robot states and environmental interactions. Such systems offer a simple and effective route toward integrated proprioception and exteroception in soft robots. However, these approaches often rely on specific sensing mechanisms, structural designs, or task scenarios. In complex environments, it remains challenging to reliably distinguish between internal actuation-induced responses and externally induced stimuli, which may limit sensing accuracy, robustness, and general applicability. To address these limitations, recent research has increasingly focused on multimodal sensing systems that combine complementary sensing mechanisms. By simultaneously acquiring and fusing multiple sources of information, these systems enable more comprehensive and reliable perception of both robot states and environmental conditions during task execution.

For example, Zhou et al. developed a soft gripper capable of simultaneously perceiving object interaction and gripper deformation during grasping. By combining multimodal sensing information with deep-learning-based analysis, the system achieved highly accurate object recognition under both room-temperature and freezing conditions. Furthermore, the gripper could detect object slippage and automatically adjust the grasping force in real time, enabling more stable and adaptive manipulation^[262]. Similarly, Diao et al. developed a multimodal perception system that simultaneously monitored grasping force and gripper bending state during operation. By continuously tracking the interaction between the gripper and the target object, the system could identify the minimum force required for stable grasping and prevent slippage of fragile objects, thereby improving manipulation reliability^[263].

Wang et al. developed an intelligent soft robotic system capable of simultaneously perceiving robot deformation and object interaction during manipulation. By fusing information from multiple sensing modalities, the system achieved reliable bending perception and adaptive grasping control, while enabling cross-modal object recognition through feature-level information integration. Such multimodal perception significantly enhanced the robot’s ability to understand and interact with target objects^[264]. Similarly, Jin et al. developed a smart soft gripper capable of simultaneously acquiring continuous motion information and tactile feedback during operation. By combining multimodal sensory information with machine-learning-based analysis, the system achieved accurate object recognition while maintaining continuous awareness of the gripper state and contact conditions, reaching a recognition accuracy of 98.1% [Figure 11A]^[265].

Figure 11. Representative multimodal sensing systems for integrated proprioception and exteroception in soft robots. (A) Enhanced object recognition system via machine learning technology. Adapted from Ref.^[265], under CC BY 4.0 license; (B) In vivo validation of a soft robotic thera-gripper for epicardial sensing and electrical stimulation (E-stim). Adapted from Ref.^[266], under CC BY 4.0 license; (C) Two flexible sensors detect the contact events and mechanical deformation of the soft actuator. Adapted from Ref.^[267], under CC BY 4.0 license; (D) Demonstration of the sensors integrated with the soft robot for motion and material recognition. Adapted with permission from Ref.^[268], Copyright © 2023 Wiley-VCH GmbH.

Beyond object manipulation tasks, multimodal sensing has also been explored in soft robotic systems for biomedical applications. For example, Zhang et al. developed a soft robotic platform capable of simultaneously perceiving multiple environmental and physiological stimuli, including touch, pressure, temperature, and pH, while maintaining adaptive actuation capability. By integrating multimodal perception with wireless motion functions, the system supported a variety of medical applications, including blood pressure monitoring, bladder volume measurement, digestive system monitoring, drug delivery, and cardiac therapy, demonstrating the broad potential of multimodal sensing in medical soft robotics [Figure 11B]^[266].

To further improve the perception capabilities of soft robots in unstructured environments, recent studies have explored higher-dimensional multimodal sensing. Sun et al. demonstrated a soft robotic system capable of simultaneously sensing strain, temperature, and mechanical stimuli, and, with deep learning analysis, accurately recognized posture changes including bending, twisting, and stretching, as well as thermal events [Figure 11C]^[267]. Dai et al. developed a self-healing multimodal sensing system that monitors strain, temperature, and material properties, enabling real-time characterization of both robot posture and environmental state when integrated into a pneumatic quadruped robot [Figure 11D]^[268]. To enhance positioning accuracy and overcome limitations of camera- or optical-sensor-based guidance, Chengkuo Lee’s team implemented a multimodal sensing system combining ultrasonic and triboelectric signals, achieving integrated perception of objects position, size, hardness, and material, with deep learning algorithms enabling nearly 100% recognition accuracy^[269]. Similar learning-based perception frameworks have also been explored in robotic systems, where heterogeneous feature fusion networks improve the extraction and integration of multi-source sensory information for accurate robotic perception and manipulation^[270].

Despite notable progress in integrating actuation and sensing within soft robotic systems, most developments remain at the proof-of-concept stage. Critical challenges - including system stability, long-term durability, and reliable operation in real-world environments - still need to be fully addressed. Achieving highly integrated designs, improving energy efficiency, and ensuring robust performance in unstructured settings remain key barriers to wider application. Looking forward, the convergence of multifunctional materials, advanced manufacturing strategies, and intelligent control algorithms is expected to further advance the integration of actuation and sensing, enabling more coordinated, autonomous, and adaptive soft robotic systems.

5.4. Comparative analysis of integrated actuation–sensing strategies

The integration of actuation and sensing represents an important transition from conventional soft robots with separated functional modules toward adaptive robotic systems capable of perception and feedback control. Existing integration strategies can generally be classified into modular integration, embedded integration, and intrinsic actuation–sensing coupling based on the degree of functional coupling between actuators and sensors^[271]. Modular integration provides a practical approach by combining mature actuators and flexible sensors, enabling relatively straightforward fabrication and system implementation. However, the mechanical mismatch between sensing components and soft structures, interface reliability, and additional wiring complexity remain major challenges^[272]. Embedded integration improves structural compactness and mechanical compatibility by incorporating sensing elements within soft bodies, but fabrication complexity and sensor replacement remain limiting factors^[273].

Intrinsic actuation–sensing coupling represents a more advanced direction, where the actuator itself generates sensing signals through changes in electrical, optical, magnetic, or mechanical properties during deformation^[274]. Such approaches can reduce system complexity and provide more direct feedback information, but they require precise material design, signal decoupling, and long-term stability. In addition, the increasing demand for adaptive soft robotic systems has promoted the development of multimodal integration, where multiple sensing mechanisms are combined with data-driven algorithms to achieve comprehensive perception and adaptive responses.

Overall, integrated actuation–sensing systems have progressed from simple combinations of independent functional modules toward more compact and intrinsically coupled architectures. However, achieving reliable perception–action loops remains challenging due to issues including mechanical compatibility, signal decoupling, fabrication scalability, and long-term stability. These challenges highlight the need for coordinated optimization of materials, structures, sensing mechanisms, and control strategies, which will be further discussed in the Outlook section.