Advances in materials and actuation strategies for wearable medical robotics: optimization for biomedical applications

2.1 Modeling the human motion system for wearable robotics design and optimization

A comprehensive understanding of the human motion system, including its structural, kinematic, dynamic, energetic, and control mechanisms, serves as the foundation for designing and optimizing wearable robots. In modeling this system, higher model complexity generally results in greater anatomical fidelity but also significantly increases computational demands. Thus, researchers must carefully balance model accuracy with computational efficiency when selecting an appropriate modeling approach.

One of the simplest representations is the torque-driven human motion system model, such as the 17 torque actuators (T1, T2, …, T17) of the upper limb and trunk, which is frequently used in structural design and optimal control simulations of wearable robots^[19]. While this model enables rapid computation, its biological realism is low, as it does not accurately reflect the true structure of the human motion system.

At the opposite end of the spectrum, the muscle finite element model-driven human motion system provides high-fidelity simulations by incorporating finite element analysis to model muscle function. This approach enables a highly detailed representation of muscle–tendon interactions, allowing for precise characterization of muscle strain, stress distribution, and energy dissipation under various loading conditions. However, this method requires significant computational resources and is often impractical for real-time applications, limiting its use in the development of assistive wearable technologies^{[20–21–22]}.

A middle-ground approach utilizes Hill-type muscle models in combination with 3-dimensional skeletal models, making it the most commonly adopted framework for human motion analysis and wearable robot development. This hybrid methodology enables an effective balance between computational efficiency and biological accuracy. In this method, the human motion system is represented as a multi-rigid-body structure actuated by Hill-type muscles, incorporating muscle contraction dynamics while maintaining relatively low computational overhead. Several well-established models exist within this framework, including:

• A musculoskeletal model designed for simulating cycling and high-speed running^[23].

• A customized musculoskeletal model optimized for analyzing squatting mechanics^[24].

• The TLEM 2.0 dataset, which provides detailed musculoskeletal geometry for lower-limb modeling^[25].

• The full-body musculoskeletal model is widely used for human gait analysis (illustrated in Figure 1)^[26].

Since the human motion system is typically modeled as a multi-rigid-body system actuated by various types of actuators, its kinematic and dynamic behavior can be described using the fundamental principles of multi-rigid-body system mechanics. However, given the large number of interconnected rigid bodies, traditional 3-dimensional vector-based methods—commonly used in robotics—result in complex and cumbersome calculations^{[27–28–29]}.

To streamline these computations, researchers have adopted a 6-dimensional spatial coordinate system to represent the motion variables of each rigid body. The corresponding dynamic equations in 6-dimensional coordinates are then derived, facilitating efficient inverse and forward dynamics analysis of the human motion system. This formulation enhances computational efficiency by reducing redundant calculations associated with large-scale multibody dynamics.

To further improve efficiency, the recursive Newton–Euler method and chain-body algorithm have been employed to solve dynamic equations, significantly reducing computational burdens associated with large-scale musculoskeletal simulations^[30,31]. These advancements provide significant potential for real-time motion analysis and optimization in wearable robotics applications.

Future research should prioritize the integration of real-time sensor feedback into these models to facilitate adaptive control mechanisms in wearable robotics. Additionally, exploring machine learning-based surrogate modeling techniques could provide opportunities for expediting inverse dynamics analysis, leveraging data-driven approaches to approximate complex biomechanical behavior with reduced computational cost. Leveraging high-fidelity musculoskeletal models in conjunction with real-time physiological data could enable the development of predictive frameworks capable of dynamically modifying robotic assistance based on user-specific movement patterns and fatigue levels. Achieving this level of adaptability will be essential for optimizing user comfort, enhancing performance, and ensuring the long-term practicality of wearable robotic devices.

2.2 Energy consumption modeling of the human motion system

Energy consumption modeling of the human motion system encompasses 2 primary components:

• Generalized power, work, and energy modeling: This method involves computing generalized power by multiplying generalized velocity and generalized force, followed by integration. It provides a mathematical representation of how energy is transferred within the musculoskeletal system, which is critical for analyzing mechanical efficiency and energy expenditure during movement^[32].

• Muscle metabolic energy consumption modeling: This approach is inherently more complex as it accounts for biochemical energy expenditure. In 2003, Umberger at Arizona State University introduced a model for estimating human muscle energy consumption^[33]. This model was later refined by Uchida and colleagues at Stanford University in 2016, improving its accuracy in quantifying metabolic energy expenditure^[34].

Muscle energy modeling is particularly important in wearable robotics, as energy-efficient design directly impacts the endurance and usability of assistive devices. Various metabolic cost functions have been developed to estimate energy expenditure during movement, including oxygen consumption rates, ATP hydrolysis models, and electromyographic (EMG) signal-based approaches^{[32–33–34]}. These models allow for more precise energy optimization strategies in wearable robots, ensuring that assistance is provided with minimal metabolic burden to the user.

Further advancements in energy modeling involve integrating real-time physiological measurements, such as heart rate variability and respiratory rate, into predictive frameworks. By continuously monitoring metabolic markers, wearable robots can dynamically adjust their actuation strategies to optimize energy efficiency based on user fatigue levels and exertion states.

2.3 Modeling the human motion control system

The human motion system is a highly redundant system, meaning that multiple muscle activation patterns can produce the same motion. This redundancy makes it extremely challenging to infer neural control signals and muscle excitations from observed movements. Addressing this challenge requires integrating optimization and control methodologies.

Figure 2 summarizes the tendon–muscle dynamics we adopt: muscles are represented as path actuators that exert force along specified lines of action; antagonistic sets of actuators combine to generate the net joint moments that drive motion. Following Millard et al.^[35], each muscle–tendon unit includes a passive elastic tendon in series with a muscle modeled by an active contractile element, a passive elastic element, and a viscous damper, with muscle force governed jointly by neural activation dynamics and contraction dynamics.

Early approaches relied on dynamic optimization techniques, where muscle excitation levels were treated as optimization variables, and the objective was to minimize the discrepancy between simulated and experimental human motion data^{[36–37–38]}. However, when incorporating numerous muscles with complex excitation patterns, solving this optimization problem becomes computationally infeasible due to the large number of state equations requiring full numerical integration.

To mitigate this challenge, researchers have explored 2 primary strategies:

• Reducing the number of muscles included in the model to decrease computational complexity while maintaining biomechanical realism^[36,39].

• Simplifying muscle control signals to limit the number of optimization variables required for simulation^[38].

Despite these simplifications, dynamic optimization techniques still demand extensive computational resources, often requiring thousands of iterations of state equation integration. To alleviate this computational burden, inverse approaches have been proposed. Yamaguchi and colleagues^[38] developed an inverse method to directly solve for muscle forces, though this approach does not inherently account for muscle properties and requires additional optimization steps to resolve muscle redundancy. Kaplan introduced a second-order dynamic optimization method, utilizing discrete state equations to enhance computational efficiency. However, this method is difficult to implement as it requires manually deriving first- and second-order derivatives of state equations concerning control variables^[37].

A more recent and computationally efficient technique is computed muscle control (CMC), which has been widely adopted in biomechanical simulations. Compared with traditional dynamic optimization approaches, CMC offers significant performance advantages by integrating feedforward and feedback control strategies with static optimization algorithms. Unlike dynamic optimization, CMC requires only a single integration of the model’s state equations while still enabling accurate tracking of experimental kinematics. Thelen and colleagues^[40,41] at Stanford University have been instrumental in advancing CMC research, using it to validate human motion simulations—such as walking—against experimental data

Beyond CMC, researchers have also explored direct collocation optimization algorithms to model human muscle control. This method has been successfully implemented by Lee and Umberger^[42] and De Groote and colleagues^[43]. However, while direct collocation enables the simultaneous optimization of large-scale models, the approach remains computationally demanding due to the substantial number of optimization variables involved.

Future research in this area should explore hybrid control strategies that integrate physiological feedback into computational models, allowing for adaptive and context-aware motion control in wearable robotics. Combining control theory with neural network-based predictive models may further enhance real-time motion assistance, paving the way for more intuitive and responsive robotic systems.

3.1 Simulation-based optimization for wearable robots

To overcome the limitations of physical prototyping and human experimentation, simulation-based design, also known as virtual prototyping, has emerged as a powerful alternative for wearable robot development. This approach leverages computational simulations to iteratively optimize wearable robots, significantly improving design efficiency and reducing costs. Virtual prototyping enables rapid quantitative analysis of numerous design configurations across diverse simulated scenarios, minimizing the need for physical prototypes and extensive human trials.

By employing physics-based simulations, designers can systematically test different actuation strategies, material compositions, and biomechanical interactions, ensuring that wearable robots meet performance objectives while maintaining user safety and comfort. Advanced simulation environments allow researchers to integrate realistic human musculoskeletal models, control strategies, and interaction forces to refine design parameters before fabrication^[46].

Several examples illustrate the advantages of simulation-based design, including:

• Wearable robots optimized through simulation to assist in load-bearing walking^[52].

• Simulated ideal wearable robots designed to enhance running performance^[53].

• Simulation-based wearable robots aimed at reducing metabolic cost during walking in elderly individuals^[54].

Early applications of simulation-based design primarily focused on wearable robots for passive users, where the wearer’s initiative was minimally considered. For instance, early studies optimized the design and control of hand exoskeletons^[55] and developed leg exoskeletons for rehabilitation training^[56]. However, as wearable robotics technology has evolved, the target user population has expanded, and human–robot interaction has become increasingly integral. This is particularly relevant for flexible wearable robots, where the user’s active engagement must be incorporated into the design process^[57]. Consequently, research efforts have shifted toward HITL optimization methods for wearable robots^{[46,58–59–60]} as illustrated in Figure 3.

Building on these advances, recent simulation-to-hardware pipelines have demonstrated the ability to train and deploy exoskeleton assistance strategies without HITL tuning, achieving substantial metabolic reductions^[61]. Integrated frameworks combining physical modeling with AI-driven policy learning have been shown to shape high-performance wearable robots for both restoration and enhancement of human function^[62]. In upper-limb applications, personalized machine learning-based controllers improved impaired arm function across diverse user populations^[63], while novel control formulations based on regulating virtual energy have emerged for coordinated human–exoskeleton locomotion^[64].

3.2 HITL optimization for wearable robots

The HITL optimization approach treats the human body and the wearable robot as a coupled system, integrating human biomechanics and robotic dynamics to define optimization objectives and constraints. Optimization algorithms are then employed to identify the optimal design parameters for the wearable robot. Since this method involves solving a complex, nonlinear coupled dynamics problem, computational demands are substantial, making the optimization process highly challenging.

Unlike traditional optimization approaches, HITL optimization continuously adapts to real-time physiological feedback, enabling wearable robots to dynamically adjust their assistance levels based on user movement patterns and energy expenditure. This adaptability makes HITL particularly well-suited for rehabilitation applications, as assistive devices can be personalized for each user’s specific needs and motor capabilities.

To improve computational efficiency, researchers have increasingly drawn inspiration from machine learning-based optimization techniques. For example, Felt et al.^[65] examined 3 different optimization algorithms—steady-state cost mapping, instantaneous cost mapping, and instantaneous cost gradient search—to optimize the design of active prostheses, orthoses, and exoskeletons based on dynamic estimation and response surface modeling. However, in many cases, the exact relationship between control parameters and metabolic costs remains unknown, leading to risks of overfitting and bias in optimization results^[66,67].

A more robust alternative is Bayesian optimization, which extends traditional response surface methodologies by employing nonparametric regression models and principled data selection strategies^[68,69]. Ding et al.^[46] and Kim et al.^[70] successfully applied Bayesian optimization to the HITL design of flexible wearable robots, demonstrating its effectiveness in refining control strategies. However, most of the existing HITL optimization studies based on Bayesian optimization have focused exclusively on optimizing control schemes for specific wearable robots, relying on partial human trials. These studies do not comprehensively integrate mechanical configuration and material properties into the optimization process.

Future work in HITL optimization should explore multiobjective optimization strategies that balance biomechanical efficiency, user comfort, and energy expenditure. Additionally, integrating physiological markers such as EMG signals, heart rate variability, and metabolic cost estimations into HITL optimization frameworks could further enhance adaptability and robustness. A more comprehensive framework incorporating both real-time and predictive optimization strategies would enable wearable robots to provide truly personalized and adaptive assistance.

3.3 Future directions: generalized HITL optimization

To fully leverage the potential of HITL optimization, it is essential to develop a generalized HITL framework that can simultaneously optimize:

• Material properties, ensuring that wearable robots are lightweight, durable, and adaptable.

• Geometric configurations, refining the robot’s structure for enhanced comfort and biomechanical compatibility.

• Control parameters, improving efficiency and adaptability across diverse user populations.

A truly generalized HITL optimization method would allow for the simultaneous fine-tuning of these 3 aspects, tailoring wearable robot designs to meet the specific needs of different users. This holistic approach would enhance the adaptability, efficiency, and user comfort of wearable robotic systems, paving the way for next-generation assistive and augmentative technologies.

Beyond conventional biomechanical metrics, future research should explore integrating cognitive and neuromuscular factors into HITL optimization. By incorporating neural control models and brain–computer interface (BCI) technologies, wearable robots could be designed to respond more intuitively to user intent, further enhancing user-robot synergy.

Additionally, advancing real-time adaptive learning algorithms within the HITL framework will allow wearable robots to refine their assistance strategies over time. By continuously learning from user movement patterns and physiological signals, assistive devices can evolve dynamically, improving long-term rehabilitation outcomes and enhancing daily usability for individuals with mobility impairments.

Future studies should also focus on developing standardized HITL optimization protocols to ensure reproducibility and cross-study comparisons. A unified approach to optimization in wearable robotics would accelerate progress in this field, fostering more effective and widely accessible robotic assistance solutions.

4.1 Model-free and model-based control approaches

To achieve accurate control of robots driven by flexible actuators, roboticists have explored both model-free and model-based control strategies. The choice between these 2 approaches largely depends on the specific application requirements, computational constraints, and the degree of modeling accuracy needed.

4.2 Advancements in model-based control compensation strategies

To improve model-based control techniques, researchers have pursued advancements in 2 key areas: enhancing modeling techniques for McKibben-type pneumatic artificial muscles and developing compensation strategies for model-based control.

4.3 Model predictive control for pneumatic muscle-driven joints

Model predictive control (MPC)^[95] is a powerful framework for handling multi-input, multioutput systems with constraints, enabling real-time predictions to adjust control actions dynamically. Its ability to explicitly incorporate actuator limits, safety constraints, and performance objectives makes it well-suited for antagonistic joints driven by McKibben-type pneumatic artificial muscles. However, pneumatic muscle-driven joints exhibit strong nonlinearities, hysteresis, and time-varying parameters, which impose stringent model accuracy requirements^[96] and can degrade predictive performance.

4.3.4 Implementation loop

A typical implementation loop involves:

• [1] Sensing: joint angle, velocity, and chamber pressure (inertial measurement unit (IMU), encoders, pressure transducers).

• [2] State estimation: Extended Kalman filter or disturbance observer.

• [3] Prediction: update model with latest parameters.

• [4] Optimization: solve the quadratic program (QP) for $\mathrm{\Delta }{u}_k$.

• [5] Actuation: send updated valve or pump commands.

Figure 5 illustrates this process, integrating trajectory and constraint inputs, prediction modeling, optimization, actuation, sensing, and state estimation into a real-time control loop.

4.4 AI-driven approaches for MPC optimization

With the advancements in AI, many data-driven MPC techniques have been developed to address the model uncertainty challenges associated with pneumatic muscle-driven antagonistic joints. Some notable AI-driven MPC approaches include:

• Neural network-based MPC for piezoelectric actuators, which learns actuator dynamics from historical data^[105].

• Reinforcement learning-based MPC controllers for robotic arms, which continuously refine control policies through trial-and-error learning^[106].

• Echo state network-based MPC for flexible exoskeletons, which leverages reservoir computing to model complex system dynamics efficiently^[99].

Beyond these examples, recent AI-assisted MPC work has landed directly in wearable contexts. Laguerre-based MPC has been deployed on an upper-limb rehabilitation robot with low computation and strong tracking performance, making real-time implementation practical^[97]. For lower-limb devices, disturbance-rejection MPC augmented with an extended-state observer improves robustness to human-induced uncertainties^[107], and MPC has also been used clinically motivated for tremor suppression in a wrist exoskeleton^[108]. Related formulations include robust tube-based nonlinear model predictive control (NMPC) demonstrated on a hybrid knee exoskeleton^[109] and Koopman-based MPC variants tailored to tremor dynamics with online model updating^[110]. On the learning side, deep models embedded inside MPC can replace hard-to-identify dynamics while retaining constraint handling^[111]; similarly, data-driven predictive control frameworks have been proposed to synthesize gaits directly from measured trajectories without explicit plant models^[112]. Complementary to these model-based approaches, sim-to-real reinforcement learning shows how learned assistance policies can transfer to hardware and reduce metabolic cost^[61], with musculoskeletal-model reinforcement learning (RL) providing convergent evidence for robust locomotion controllers^[113]. In parallel, HITL optimization with machine learning surrogates personalizes controller parameters in real time^[114], including sample-efficient preference-based tuning for ankle assistance^[115], EMG-informed HIL on a portable hip exoskeleton^[116], and preferential multiattribute Bayesian optimization for exoskeleton personalization^[117]. User-specific policy learning via Gaussian processes offers a low-data path to personalization and motivates the Gaussian process regression (GPR)-MPC integration discussed next^[118].

Although these methods show promise in learning nonlinear dynamics, they require extensive training data and computational resources. Moreover, training these models is complex, often necessitating the collection of large datasets to achieve optimal performance.

4.5 GPR for MPC in antagonistic joint control

To overcome the challenges posed by deep learning-based control methods, GPR has emerged as a promising alternative. Unlike neural networks, GPR requires minimal prior data and is capable of approximating a wide range of nonlinear functions^[68,119,120].

In regression form, given training data $\mathcal{D}=\{X,y\}$ and a kernel function $k(\cdot ,\cdot )$, the GPR predictive distribution for a test input $x_{\ast }$ is:

where $K$ is the Gram matrix of training points, $k_{\ast }=k(X,x_{\ast })$, and ${\sigma }_n^2$ is the observation noise variance. In GPR-MPC, $\mu \left(x_{\ast }\right)$ models system dynamics (eg, joint torque response), while ${\sigma }^2\left(x_{\ast }\right)$ informs robust constraint handling in the MPC optimization problem^[68,121].

GPR has been successfully implemented for modeling the dynamics of various systems, including:

• Internal combustion engine charging cycles^[122].

• Chemical pH neutralization systems^[123].

• Autonomous robotic ship navigation^[124].

• Wheeled unmanned vehicles^[121,125].

• Lower-limb exoskeletons enabling rapid adaptation to individual biomechanical patterns^[118].

The integration of GPR and MPC presents a compelling strategy for precise control of McKibben-type pneumatic muscle-driven antagonistic joints. By adaptively modeling system uncertainties and feeding both predicted means and uncertainty bounds into the MPC optimizer, this hybrid approach can improve control accuracy while reducing reliance on extensive precollected training datasets.

5.1 Global contributions to wearable robotics: institutions, research initiatives, and applications

Wearable robots, designed to support, restore, or enhance human mobility, have gained significant attention from leading research institutions worldwide. Several prominent organizations—including NASA, DARPA, the BioRobotics Institute at the Sant’Anna School of Advanced Studies, Harvard University, the SPEXO team under the European Union, the Aspire Create team in the United Kingdom, and Stanford University—have made substantial contributions to advancing wearable robotic technology. Their research covers diverse areas such as human–robot interaction, flexible actuation mechanisms, control methodologies, and optimization techniques aimed at improving human movement efficiency.

5.1.1 NASA’s Johnson Space Center

NASA’s Johnson Space Center has been at the forefront of developing wearable robotic exoskeletons, primarily to prevent muscle atrophy and bone loss in astronauts during prolonged space missions. As illustrated in Figure 6, these exoskeletons function as alternatives to traditional, bulky exercise equipment on space stations, thus optimizing space utilization and astronaut health^[126]. To enhance astronauts’ ability to manipulate objects in microgravity, NASA has also engineered space gloves that improve grip strength. Furthermore, the center has designed wearable assistive devices with independently controlled shoulder and elbow joints, which enhance upper-limb dexterity in confined environments^[127]. NASA’s long-term goal is to adapt these technologies for civilian applications, such as reducing fatigue and repetitive stress injuries, assisting individuals with mobility impairments, and augmenting human performance^[128].

5.1.3 The BioRobotics Institute’s Wearable Robotics Laboratory

The BioRobotics Institute at the Sant’Anna School of Advanced Studies in Pisa, Italy, is one of the pioneering research centers in wearable robotics^{[130–131–132]}. The institute houses 10 specialized laboratories and 8 primary research domains. It has developed numerous wearable robotic prototypes, including human–machine-integrated prosthetic hands, exoskeletons for upper and lower limb assistance, and wearable robotic footwear^{[133–134–135]}, as illustrated in Figure 7.

5.1.4 Harvard University’s Biodesign Laboratory

Harvard University’s Biodesign Laboratory specializes in the development of flexible wearable robots using functional textiles^[14,16]. Since 2008, the lab has contributed significantly to wearable robotics research, publishing seminal studies on the metabolic cost reduction through soft exosuits^{[136–137–138–139]} and wearable robots for gait rehabilitation in stroke patients^[140].

This lab has designed and developed flexible wearable robots targeting different populations using functional textiles as the main material, and has published dozens of classic articles since 2008^[137,138]. The flexible wearable robots that reduce metabolism during walking or running and the flexible wearable robots used for gait recovery in stroke patients have been published in the journals Science, Science Robotics, and Science Translational Medicine. As shown in Figure 8, the wearable robot designed by the team is characterized by using Bowden cable drive. Its main advantage is that the wearer’s joints are not constrained by external rigid structures, and it is lightweight, just like ordinary clothing, which can interact with the human body more naturally. The main research directions of the laboratory are functional textiles, lightweight and effective drivers, wearable sensors, intuitive and robust control methods, etc. Recent representative advances from this lab and collaborators include a study showing soft robotic apparel averting freezing of gait in Parkinson disease^[141], a field evaluation where an active back exosuit reduced back muscle activation during an hour-long order-picking task^[142], and an inpatient feasibility study that updated the Lab’s hip flexion exosuit for therapist-driven stroke rehabilitation^[143].

5.1.5 The SPEXO project

The SPEXO project is an EU Horizon 2020-funded initiative dedicated to developing wearable robotic solutions for individuals suffering from back pain. This multidisciplinary team consists of medical professionals and engineers from multiple European countries. They have successfully developed a hybrid wearable system that integrates both passive and active components^{[144–145–146]} as shown in Figure 9.

5.1.6 The Aspire Create Team

The Aspire Create Team, formed in collaboration between Aspire Charity, University College London, and the National Orthopaedic Hospital, was established in 2014 with the aim of improving the quality of life for individuals with spinal cord injuries. As depicted in Figure 10, the team has developed assistive technologies, including cough assist machines and smart wheelchairs for rehabilitation^[147,148], with a focus on human–machine interaction and shared-control systems; recent outputs extend this line of work by presenting a modular smart-wheelchair architecture with built-in assistive features^[149], demonstrating a time-of-flight-based shared-control “table-docking” method on a prototype power wheelchair^[150], and outlining an equitable shared-control framework that personalizes assistance policies for assistive robotics^[151].

5.2 Advancements in wearable robot actuators: from rigid exoskeletons to flexible TPU-based systems

The research efforts of the aforementioned institutions demonstrate the broad and diverse applications of wearable robots. These systems are employed not only to reduce physical strain in workers, facilitate rehabilitation for stroke patients, improve the athletic ability of athletes, and enhance movement efficiency by lowering metabolic energy consumption, but also in areas such as sports performance enhancement, military operations, aerospace exploration, and assistive technologies. Given this wide scope, further research into the technical challenges and optimization strategies of wearable robots is crucial for expanding their capabilities and applicability across multiple fields.

Early wearable robotic systems were predominantly rigid exoskeletons^[171,172], incorporating motor-driven actuators, linkage mechanisms, transmission mechanisms, and clutch spring systems^{[173–174–175]}. Although rigid exoskeletons can provide effective movement assistance, their large weight and limited mechanical degrees of freedom constraints often interfere with the user’s natural biomechanics movement behavior^[176,177]. Additionally, these systems increase user’s metabolic energy expenditure^[178,179], making prolonged use physically demanding. A notable limitation of the rigid exoskeletons is seen in knee joint devices, where the fixed rotational axis of the exoskeleton fails to align with the dynamic, shifting axis of the human knee during various movement^[179]. This misalignment can lead to skin irritation, increased joint loading, and long-term joint degradation, elevating the risk of arthritis for wearers^[180,181]. Additionally, the high cost, complex structural design, and difficulty in control further hinder the widespread adoption of rigid exoskeletons^[171,178]. Consequently, there is a strong demand for lightweight, cost-effective, and easily controllable flexible wearable robots.

Flexible actuators have become a key focus in wearable robotics research^{[182–183–184]}, with scientists increasingly exploring functional materials for actuator development. Some promising materials include shape-memory alloys (SMAs), which can alter their shape in response to temperature or magnetic field changes^[185], and electroactive polymers, which generate torque through high-voltage stimulation^[186]. However, both material types require specialized operating conditions, such as high temperatures or high voltages, which raise safety concerns for wearable applications. Another widely used flexible actuator is the Bowden cable system, which has been successfully integrated into wearable robots. However, due to the linear motion mechanism of cables, this type of wearable robot requires additional rigid components, such as metal frames and gear conversion mechanisms to convert linear forces into rotational motion, thereby achieving rotational motion assistance for joints^[46,139,187].

Pneumatic actuators have attracted a lot of interest in the field of wearable robots due to their characteristics, such as small mass, low cost, skin friendliness, and deformability^{[188–189–190]}. Among them, the McKibben-type artificial muscles have been widely studied for their ability to generate high force outputs. However, their application in wearable robotics remains limited, as they still require linear-to-rotational motion conversion mechanisms^[191,192]. To address this limitation, researchers have modified the McKibben-type actuators to develop the fiber-reinforced pneumatic actuators, which can directly produce bending motion without the need for rigid motion conversion mechanisms^[8]. Despite their advantages, both the McKibben-type and the fiber-reinforced actuators depend on high gas pressure drive systems, introducing potential safety concerns in wearable applications. Additionally, other researchers have proposed rigid-flexible hybrid bending actuators^[193,194], but their complex manufacturing process limits their practicality for widespread adoption.

TPU has emerged as a promising material for flexible actuators, widely used in airbags and food packaging due to its lightweight, cost-effective, and low-pressure operation. TPU actuators present significant potential in wearable robotics, as they are easily shaped and can provide assistance safely at low pressure. Recent studies have explored TPU-based actuators for wearable applications.

5.2.2 Applications of TPU-based actuators in wearable robotics

As shown in Figure 11A, Natividad et al. developed a reconfigurable pneumatic bending actuator with replaceable inflation modules that can achieve bending motion assistance at various angles by adjusting the number of modules. However, its output force is insufficient to provide sufficient motion assistance for wearable robots^[197]. Meanwhile, Nesler et al.^[198] introduced a self-intersecting pneumatic actuator capable of generating joint torque to assist human movement, but its airflow efficiency during inflation and exhaust processes is low at large bending angles (Figure 11B).

Some researchers have applied TPU-based actuators to wearable robots. As illustrated in Figure 11C, Natividad and Yeow^[199] designed a TPU fiber beam-driven shoulder wearable robot for the rehabilitation of patients with cerebral palsy. Similarly, as shown in Figure 11D, O’Neill et al.^[200] developed a TPU-actuated soft shoulder wearable robot, but the 2 air chambers at the far end of the soft wearable robot provided minimal torque assistance

Besides that, the TPU-based actuators have been investigated for lower-limb wearable applications^[201]. As depicted in Figure 12, Sridar et al.^[202,203] designed a knee exosuit driven by the TPU actuator, which could provide knee joint assistance torque to the wearer during walking as needed. However, the assistance torque (4.4 Nm with 2 actuators) generated by this TPU actuator is limited, which restricts its ability to support other joint movements effectively. Fang et al.^[201] proposed a novel design of foldable pneumatic bending actuators (FPBAs), which could produce torque by the inflation or interactive compression of their interconnect air chambers with no airflow restriction occurring at any bending angle. A knee exosuit equipped with the FPBAs was also designed and has been used in the area of knee rehabilitation training.

To fully leverage TPU-based pneumatic actuators in wearable robotics, further refinements in structure and mechanical properties are required. Future research should focus on enhancing actuation efficiency, increasing output torque, and improving airflow control, ensuring that TPU actuators can meet the functional demands of wearable robotic applications.

Beyond actuator performance, sustained operation without tethered power remains a major challenge for wearable robots. Energy autonomy is increasingly addressed through advanced harvesting technologies. Fully autonomous exoskeleton prototypes integrate multisource energy capture and storage^[204]. At the joint level, self-powered and self-sensing knee negative-energy harvesters recover energy from gait without impeding motion^[205]. Recent reviews^[206] and hybrid electromagnetic–triboelectric harvesters^[207] highlight the potential of combining modalities to extend untethered operation.

Despite significant progress in this area, many challenges remain in achieving widespread adoption and optimized functionality. One key avenue for future research is the integration of TPU actuators with advanced sensor networks to improve real-time adaptability. By incorporating force, pressure, and motion sensors, TPU-based actuators could respond dynamically to user movements, enhancing their ability to provide personalized assistance. Additionally, developing closed-loop control strategies that incorporate real-time biofeedback data could significantly improve the effectiveness of wearable robots for rehabilitation and assistive applications.

Another promising area for research is the optimization of TPU actuator designs to improve force output and efficiency. Current TPU actuators still struggle with limited torque generation, preventing them from replacing bulkier actuation mechanisms in wearable robots. Future work should explore new structural designs, such as multichamber inflation systems and variable stiffness configurations, to increase torque while maintaining flexibility. Enhancing material formulations to improve TPU durability and elasticity would also contribute to more reliable long-term performance.

Energy efficiency is another critical factor influencing the feasibility of TPU-based wearable robots for real-world applications. Future research should focus on developing energy-efficient drive systems, including low-power pneumatic pumps and optimized air compression mechanisms. Additionally, exploring hybrid power solutions, such as integrating TPU actuators with energy-harvesting technologies like piezoelectric or triboelectric generators, could extend operational time without requiring frequent recharging.

Expanding the range of applications for TPU-based wearable robots is also a crucial consideration for future studies. While TPU actuators have primarily been explored for rehabilitation and assistive devices, they could also be applied to sports performance enhancement, industrial exoskeletons, and space exploration. The development of TPU-based wearables for enhancing athletic performance, preventing workplace injuries, and supporting astronauts in microgravity environments could further push the boundaries of this technology.

In conclusion, TPU-based actuators present a highly promising alternative to traditional rigid and hybrid actuation mechanisms in wearable robotics. However, further advancements in material properties, structural optimization, real-time sensing, and energy efficiency are necessary to unlock their full potential. By addressing these challenges, researchers can develop next-generation wearable robots that are lightweight, comfortable, and highly adaptable, ultimately enhancing human mobility, rehabilitation, and overall quality of life.

To place these developments in a broader context, the following section compares TPU-based actuators with other common actuation methods used in wearable robots. This quantitative overview highlights the trade-offs in force output, weight, power requirements, and controllability, providing a framework for understanding where TPU actuators excel and where further improvement is needed.

5.3 Summary of advances in materials and actuation

Advances in wearable medical robotics are strongly driven by innovations in both material systems and actuation strategies. Material choice affects comfort, safety, and durability, while actuator selection determines achievable force, control precision, and efficiency. Optimal designs often require coordinated selection of materials and actuators to meet specific clinical and assistive needs.

Figure 13 provides a schematic overview of commonly used material classes and actuation methods, illustrating their relationships, performance ranges, and example applications. Table 2 summarizes key material categories, and Table 3 compares actuation strategies, outlining their advantages, limitations, and representative devices, with multiple literature sources for each. Together, these resources offer a consolidated reference to guide design decisions and highlight trends toward hybrid, sensor-integrated, and portable systems.