Wenyuan Chen

Due to their high compliance, environmental adaptability, and inherent safety in human–robot interaction, soft robots have demonstrated considerable potential, including medical rehabilitation, compliant grasping, and operations in complex environments. Unlike conventional robots relying on rigid structures and discrete joint motions, soft robots use flexible materials such as elastomers and hydrogels to achieve motion and functionality through continuous deformation. Despite providing excellent adaptability, these compliant materials also introduce complex behaviors, including continuous morphological changes, material nonlinearities, time-varying properties, and environmental coupling, posing substantial challenges to modeling, control, and motion planning. In recent years, artificial intelligence (AI), through nonlinear representation, data-driven learning, adaptive optimization, and autonomous decision-making, has provided new technical pathways for addressing these challenges. Unlike previous reviews organized primarily by AI algorithm type, this work narratively reviews 39 studies and organizes AI-driven approaches according to the soft robotic system implementation process, encompassing modeling, control, and motion planning. Specifically, in modeling, neural networks, Koopman representations, and physics–data fusion methods have enhanced the representation of complex dynamical relationships. In control, learning-augmented control and reinforcement learning have reduced reliance on accurate system models and improved adaptability under uncertainty. In motion planning, learning from demonstration, reinforcement learning, and intelligent optimization methods have strengthened the capability for autonomous motion generation under complex constraints. Finally, this paper summarizes the current state of the field and discusses future trends in AI-driven soft robotics, emphasizing stronger coordination across functional modules. Closed-loop integration of prediction, decision-making, control, and feedback updates may further advance soft robots from stage-specific intelligence toward system-level autonomy.

While soft hand exoskeleton robots have approached human-level dexterity in terms of degrees of freedom, precise control methods for fine motor movements remain a significant challenge. Surface electromyography (sEMG) is widely employed in gesture recognition to enable patients to independently control a soft hand exoskeleton. However, individual finger control remains challenging through sEMG-based control due to the complexity of decoupling synergistic muscle activities. In this study, we propose a muscle fiber-based ultrasound perception strategy for fine hand motion recognition and soft hand exoskeleton control. Ultrasound imaging enables non-invasive visualization of forearm muscle morphology and provides information associated with underlying muscle-fiber activity. By reconstructing muscle morphology from ultrasound images, biologically relevant muscle-fiber features are extracted and fused to characterize fine hand movements. A lightweight Random Forest classifier is subsequently employed to map these biologically informed features to discrete hand actions, providing a computationally efficient recognition module for real-time control. To the best of our knowledge, publicly available ultrasound image datasets specifically designed for fine hand gesture recognition in rehabilitation applications remain limited. In the experiments, a dataset containing 21 hand gestures based on muscle ultrasound images was constructed to evaluate the proposed method. All data were collected from healthy participants as a preliminary proof-of-concept investigation. The results show that the proposed approach achieves an average recognition accuracy of 95.24% across three subjects in finger motion recognition. This preliminary study demonstrates the potential of machine learning-based ultrasound perception for improving fine hand gesture recognition and providing an intuitive control interface for soft hand exoskeletons, thereby enhancing their applicability in hand rehabilitation scenarios.

The rapid evolution of next-generation electronics urgently demands high-performance functional materials. Two-dimensional (2D) semiconductors, characterized by tunable bandgaps, magnetic properties, and excellent optical and electronic properties, hold significant potential for applications in nanoelectronic devices, magnetic storage, and optoelectronics. However, the high computational cost of traditional Density Functional Theory (DFT) severely restricts large-scale high-throughput screening. Meanwhile, problems such as insufficient datasets and non-uniform data quality remain prevalent. Against this background, machine learning (ML), which captures intricate nonlinear correlations and accelerates the discovery of novel materials, has emerged as an efficient technical approach. This review systematically summarizes recent advances in ML-driven property prediction for 2D semiconductors. It first elaborates the fundamental properties and classifications of 2D semiconductors, and then compares traditional computational simulations with ML algorithms, clarifying the distinct advantages of data-driven approaches. Subsequently, this work focuses on the latest progress in predicting critical properties, including bandgap, magnetism, and other physical characteristics. For bandgap prediction, classical algorithms such as random forests are compared with deep learning models represented by graph neural networks. The results demonstrate that deep learning performs much better in low-data regimes and complex material systems. For magnetic property prediction, the impact of feature engineering strategies on model accuracy and efficiency is systematically analyzed. In addition, the research progress of other physical property prediction tasks is briefly summarized. Finally, future research directions for machine learning, including standardized materials databases, physics-informed machine learning, multimodal modeling, and the integration of machine learning with experimental and theoretical methods, are outlined to address challenges in data quality, model interpretability, and cross-system generalization ability. This work aims to provide a systematic theoretical foundation and methodological guidance for research on two-dimensional semiconductor materials assisted by machine learning.

Accurate and prompt speed grasp intention recognition is crucial in online human-robot interaction (HRI). However, dynamic grasping relying on complete motion for high recognition accuracy will lead to an unavoidable delay in real-time prediction. To address this issue, we propose a Spatial Gating and Temporal Weighting Early Grasp Prediction (STEGP) method that utilizes incomplete dynamic grasping data from sliding windows to reduce the time delay for reliable robot teleoperation. The proposed method comprises a synergy-based feature extraction module, a spatial gating classification module, and a time-decay weighting fusion prediction module. The spatial-temporal mechanism with gating units effectively classifies sequential movements, achieving performance comparable to that of Transformers but being much easier to implement. Integrating a time-decay weighting frame enables reliable early prediction even with incomplete data. gMLP is chosen for the classification of hand dynamic grasping because of its high accuracy, realizing 93.83% accuracy for 33 grasping categories. The prediction tests demonstrated 85.4% accuracy, with the first 25% grasp completion across 28 subjects. Online robotic teleoperation grasp experiments achieved a 57.4% reduction in time delay and a 93.3% success rate.

Fish can adaptively adjust their body kinematics and swimming modes by sensing to realize optimal propulsion. However, most soft robotic fish have an unchangeable swimming mode through simple structure design, making them difficult to adapt to dynamic and complex fluid environments. Here, inspired by the multiple muscle synergy and lateral line sensing function of fish, we developed a soft robotic fish with multiple actuating units and embedded sensing elements. By collaboratively controlling the amplitude and phase of excitation from the multiple flexible actuating units, the soft robotic fish can successfully realize various swimming modes very similar to those of natural fish. Additionally, the embedded flexible sensing elements enable the robotic fish to sense the swimming state and the surrounding fluid environment in real time. The multiple actuation and embedded sensing allow the soft robotic fish to adaptively switch to an optimal swimming mode in a certain fluid environment. The multi-mode swimming and perception capabilities proposed in this work not only make soft robotic fish more intelligent and adaptable to complex fluid environments, but also contribute to the future implementation of autonomous control capabilities for robotic fish.

It remains a formidable challenge to accurately recognize motion intentions of patients thus to control hand exoskeletons according to their volition. Current methods primarily focus on recognition of limited patient’s motion intentions, with the purpose of controlling preconfigured gestures of a hand exoskeleton for grasping objects. These methods exhibit a marked shortfall when encountering scenarios that are unexpected or not designed in advance, such as non-preprogrammed hand movements and object manipulation tasks. To tackle this issue, large language model (LLM) and speech recognition technology are employed in this study to allow the patient to control a hand exoskeleton at will. In particular, two LLMs are tailored to formulate codes of either generating non-preprogrammed gestures or dealing with unencountered objects. Additionally, an incremental learning framework is proposed to enable patients to perform both predefined and non-predefined operation tasks by integrating a natural language parser with the two LLM-based learners. The natural language parser can directly control the hand exoskeleton to perform predefined operations tasks from prestored command set, while the LLM-based learners can incrementally expand the control command set so as to enhance adaptability of the hand exoskeleton to complex activities over daily use. This study is a pioneering work in the field of hand exoskeletons, which will revolutionize the way to control hand exoskeletons. Furthermore, the proposed framework can be easily generalized to any other robots by modifying the prompt of customized LLMs, which provides a new idea to achieve autonomous learning in robotics. Note to Practitioners —The motivation of this article is to tackle the challenge of intention recognition for performing activities of daily living (ADLs) by stroke patients using a multi-degree of freedom hand exoskeleton. Existing methods for intention recognition so far can only be used for several tasks that are predefined in advance, thus none of them allow patients to control the hand exoskeleton completely at will. To surpass this limitation, an LLM-enabled incremental learning framework that integrates a hand exoskeleton controller with Large Language Model (LLM) is proposed and validated in this study. The framework offers patients an intuitive interface via voice interaction and enables patients to perform not only predefined operation tasks by the hand exoskeleton controller but also non-predefined ones that can be learned from the LLM. As a result, the hand exoskeleton controller continues to learn from the LLM, therefore is gradually able to perform all tasks in daily life. This pioneering study paves a new way in building patient-controlled hand exoskeletons with autonomous intelligence that can deal with non-predefined operation tasks in unstructured environments.

It remains a formidable challenge to accurately recognize patients’ motion intentions to flexibly control hand exoskeletons. Current methods primarily focus on recognition of limited patient’s motion intentions, with the purpose of controlling preconfigured gestures of a hand exoskeleton for grasping objects. These methods exhibit a marked shortfall when encountering scenarios that are unexpected or not designed in ad-vance, such as non-preprogrammed hand movements and object manipulation tasks. This paper proposes a large language model (LLM)-enabled incremental learning framework for controlling hand exoskeletons. The framework enables patients to perform both predefined and non-predefined operation tasks with hand exoskeletons through a hand exoskeleton controller and LLM-based learners via voice interaction. Specifically, the framework embeds LLMs as expert modules that are capable of inferring appropriate hand motions and formulating corresponding control commands in accordance with human experiences to fulfill tasks that are unknown to be completed by the hand exoskeleton controller. At the same time, the hand exoskeleton controller incrementally learns to perform these non-predefined tasks by updating the control command set. Therefore, with the frame-work, the hand exoskeleton can incrementally expand its control command set so as to enhance patients’ adaptability to complex activities over daily use until the hand exoskeleton controller no longer needs to consult the LLM-based expert modules. This study is a pioneering work in the field of hand exoskeletons, which will revolutionize the way to control hand exoskeletons by patients.

Hand exoskeletons are wearable devices that can provide outer kinematic coupling with human hands and thus assist movement of human fingers. However, conventional rigid hand exoskeletons are characterized by their bulky and complex structures, which are often incompatible with human finger joints and restrict finger’s natural motion. This paper reports an underactuated base-to-distal hand exoskeleton that provides adaptive grasping assistance. An underactuated 8-bar base-to-distal linkage driven by a cable is used to flex and extend fingers and it applies force only to the distal phalanges of fingers, which not only makes the hand exoskeleton adapt to different sizes of fingers but also allows all phalanges to naturally accommodate the geometry of the objects to be grasped. The kinematic model of the 8-bar linkage is derived in order to generate desired hand ges-tures. A five-finger hand exoskeleton with active flexion/extension (F/E) for all fingers and active abduction/adduction (Ab/Ad) for the thumb is assembled and then tested on a healthy subject and a stroke survivor. Experimental results show that the hand exoskeleton can generate sufficient fingertip force for regular tasks. The hand exoskeleton enables the healthy participant and the stroke survivor to achieve 90% and 52% of their passive range of F/E motions respectively. In addition, the stroke survivor can accomplish various training tasks, such as grasping, pinching and writing, with the assistance of the hand exoskeleton. These results demonstrate that the underactuated base-to-distal hand exoskeleton can be an effective device for rehabilitation training or daily-life assistance for patients with a hemiparetic hand.

Personalized control strategies can improve the overall performance of exoskeleton robots by providing adaptive assistance to different individuals. However, current personalized strategies require accurate models or large amounts of training data, which are not suitable for soft exoskeletons due to human-robot interaction uncertainties and low data acquisition efficiency. This paper presents a framework of Model-Agnostic Personalized Knowledge Adaptation (MPKA) consisting of two modules of Soft-to-Rigid Kinematics Model (SRKM) and Data-Driven Personalized Adaptation (DDPA) to generate a personalized strategy. Specifically, the first SRKM module is used to simplify and convert uncertain and time-varying HRI model to a certain and fixed model. The second DDPA module is used to accelerate model adaptation to new individuals using meta-learning with less data, based on the learned knowledge from the previous subjects. The experiments were conducted on three different sized platforms, four healthy and one post-stroke subjects, respectively. The results demonstrate that the MPKA can generate personalized models for new individuals more rapidly with less data than a pre-trained neural network, and the personalized models can obtain better control accuracy by over 62% than mechanism-based mathematical model. The proposed method has the potential to provide personalized rehabilitation strategies, which can improve effectiveness of rehabilitation.

A significant number of stroke patients are permanently left with a hemiparetic upper limb after the poststroke six-month golden recovery period, resulting in a drastic decline in their quality of life. This study develops a novel foot-controlled hand/forearm exoskeleton that enables patients with hemiparetic hands and forearms to restore their voluntary activities of daily living. Patients can accomplish dexterous hand/arm manipulation on their own with the assistance of a foot-controlled hand/forearm exoskeleton by utilizing foot movements on the unaffected side as command signals. The proposed foot-controlled exoskeleton was first tested on a stroke patient with a chronic hemiparetic upper limb. The testing results showed that the forearm exoskeleton can assist the patient in achieving approximately 107°of voluntary forearm rotation with a static control error less than 1.7°, whereas the hand exoskeleton can assist the patient in realizing at least six different voluntary hand gestures with a success rate of 100%. Further experiments involving more patients demonstrated that the foot-controlled hand/forearm exoskeleton can help patients in restoring some of the voluntary activities of daily living with their paretic upper limb, such as picking up food to eat and opening water bottles to drink, and etc. This research implies that the foot-controlled hand/forearm exoskeleton is a viable way to restore the upper limb activities of stroke patients with chronic hemiparesis.