Gongxin Li

Because lesions associated with digestive system diseases are often deeply seated, embedded within complex luminal milieus, and protected by substantial local delivery barriers, conventional diagnostic and therapeutic modalities remain constrained in their targeting capability, minimal invasiveness, and precision. Microrobots, leveraging micro-scale motion, active navigation, programmable controllability, and theranostic integration, open a new avenue for the precise diagnosis and treatment of digestive system diseases. Here, we review the major actuation modalities and structural designs of microrobots and discuss recent advances in their use across the gastrointestinal and hepatopancreatobiliary systems, focusing on targeted drug delivery, biospecimen harvesting, lesion detection, and interventional treatment. We further discuss the major obstacles to progress in this field, including robust operation in complex in vivo settings, real-time imaging and closed-loop feedback control, biosafety, degradability, and eventual clinical implementation. With continued advances in high-performance materials, multimodal actuation, intelligent control, and convergence with endoscopic and medical imaging technologies, microrobots are poised to accelerate the transition of digestive disease care toward precision, intelligence, and minimally invasive intervention.

Precise continuous fingertip force estimation based on surface electromyography (sEMG) is crucial for human-robot interaction and prosthetic control. However, traditional deep learning models, such as conventional CNN-based models, are constrained by limited receptive fields, focusing mainly on local features while failing to capture the long-range global dependencies required for dynamic force estimation. In addition, substantial inter-individual variability further limits practical deployment. To address these challenges, this paper proposes sEMG-Conformer (Convolution-Transformer hybrid architecture) that integrates local spatiotemporal features and global temporal dependencies for robust force estimation. Specifically, a hierarchical convolution module encodes noise-filtered local patterns, while a self-attention mechanism models global correlations across the sequence. Furthermore, a parallel multi-task prediction head jointly performs force-presence detection and force-magnitude regression, evaluated by accuracy and NRMSE, respectively. To tackle inter-subject variability, we implement a physiology-inspired, layer-selective transfer learning strategy that freezes the CNN encoder and fine-tunes the Conformer and decoder/head modules. Experimental results on a 12-gesture sEMG-force dataset show that sEMG-Conformer achieves 95.29\% force-presence detection accuracy and 6.46\% NRMSE, outperforming representative traditional and deep learning baselines. Crucially, the framework transfers from a laboratory pressure array to a 16~Hz wearable tactile glove, enabling real-world 3D grasping. These findings highlight its potential for practical, user-adaptive myoelectric interfaces. The code is available at \url{https://anonymous.4open.science/r/main_code-74B3}.

Background: Soft robots utilised for intestinal diagnosis and therapy use, compliant materials and novel actuation were employed to address the limitations of rigid endoscopic tools. Methods: This paper systematically reviewed the literature (2000–present) on robotic systems explicitly developed for intestinal applications, organising findings across actuation, materials, fabrication and clinical use. Ethical approval: Not applicable because this study is a literature review and involved no human participants or animal experiments. Results: Fluidic, shape‐memory alloy, electroactive polymer and magnetic actuation strategies were classified and compared by operating principles, advantages and limitations. Common material platforms (e.g., silicones and hydrogels) were analysed for flexibility, biocompatibility, responsiveness and fabrication methods (casting, 3D printing and shape deposition manufacturing) and were assessed for producing functional, complex structures. Applications clustered around examination, targeted drug delivery and minimally invasive interventions, with designs grouped by function and morphology. Conclusions: Intestinal constraints (peristalsis, variable lumen geometry and biosafety) remain key barriers; interdisciplinary integration is central to clinical translation.

Untethered magnetically actuated soft microrobots are promising for biomedical and industrial tasks, yet their practical deployment remains limited by poor adaptability to heterogeneous environments and insufficient motion controllability. This paper presents a soft magnetic microrobot, termed the Bio-mimetic Mudskipper Bot (BM-bot), inspired by the morphology and crawling mechanism of amphibious mudskippers. The BM-bot consists of a compliant body and a pair of soft-magnetic half-wheel feet, enabling crawling locomotion driven solely by magnetic torque–induced foot rotation without onboard actuation. A vision-based pose estimation framework integrating temporal differencing and adaptive histogram equalization is developed for real-time localization, achieving a mean position error of approximately 0.3 mm across multiple experimental conditions. For motion control, a Stepwise Dynamic Compensation (SDC) strategy is introduced, which decouples local motion execution from global path planning and performs discrete heading corrections based on real-time visual feedback. This control scheme allows the BM-bot to maintain a median trajectory deviation of 0.33 mm during long-duration path-following tasks in dynamic environ-ments. Experimental results demonstrate that the proposed robot and control framework enable stable and repeatable locomotion across diverse terrains, including granular media, soil, shallow water, constrained maze environments, and ex vivo soft and dynamic biological tissues. The presented system provides a practical solution for achieving robust, high-precision locomotion in magnetically actuated soft microrobots under unstructured, compliant, and dynamically changing conditions.

Magnetically driven microrobots operating in intestinal environments face two major challenges: difficulty in traversing low-height confined spaces and limited local visibility caused by mucosal obstruction. To address these issues, this study proposes a gradient-field-based force-driven control method for a mudskipper-inspired magnetic microrobot. By establishing the mapping among coil current, magnetic field, and magnetic force at the robot working point, and by solving the control input through singular value decomposition and linear programming, effective magnetic-force output along a desired direction was achieved. On this basis, two representative force-driven motions were designed. The first was a translational mode based on pulsed magnetic-force actuation for stable navigation in low-height confined spaces. The second was a lifting mode based on continuous loading and gradual adjustment of the magnetic-force upper bound to locally lift a flexible “mucosa-like” membrane, thereby simulating intestinal mucosal elevation and local visual field expansion. Experimental results showed that the robot could stably pass through narrow tunnels and effectively lift an overlying flexible membrane under vertical magnetic-force actuation. The proposed method extends both the locomotion capability and the local interaction capability of the mudskipper-inspired magnetic microrobot, and demonstrates a feasible proof-of-concept approach for confined-space navigation and localized manipulation in intestinal applications.

Saccharomyces cerevisiae (S. cerevisiae) can be used to treat diarrhea and the diseases associated with malaria, but it may cause invasive infections in individuals with weakened immune systems. While antibiotics are available for treating these infections, they can also produce side effects in the body. Additionally, macrophages in the human immune system can engulf infected S. cerevisiae, but they have some limitations, including a relatively weak active chemotactic response and a prolonged engulfing process. To address this, this paper proposes a novel method to enhance the phagocytosis of S. cerevisiae by macrophages using optically induced dielectrophoresis (ODEP). ODEP is a cellular micromanipulation technique that directs phagocytes towards S. cerevisiae cells, significantly improving phagocytosis efficiency. Furthermore, the optical electrodes created by ODEP expedite the phagocytosis process. To validate this approach, the optimal operating parameters for ODEP were determined through a combination of numerical simulations and experiments, enabling the swift and precise capture of phagocytes targeting S. cerevisiae. A comparative experiment was conducted to assess macrophage phagocytosis of S. cerevisiae with and without the aid of optical electrodes. Results showed that the time required for macrophages to engulf S. cerevisiae was reduced by 50%, highlighting a promising method for the early prevention or accelerated treatment of invasive S. cerevisiae fungal infections.

As robotic systems transition from structured industrial settings to dynamic, human-centric environments, the demand for natural and intuitive human-robot interaction (HRI) has become paramount. Given its inherent advantages in expressiveness, intuitiveness, and universality, Hand Gesture Recognition (HGR) has emerged as a cornerstone technology for effective non-verbal communication, enabling applications from precise robotic control to skill learning. Distinct from prior surveys that often focus narrowly on algorithmic advancements, this paper provides a holistic, three-tiered analysis encompassing hardware, algorithms, and applications, with a unique emphasis on the practicalities of system implementation. A primary contribution is our in-depth analysis of sensing hardware. We move beyond theoretical principles to critically examine the essential software interfaces (SDKs/APIs) and data characteristics that fundamentally govern real-world deployment—a dimension frequently overlooked. Furthermore, we establish a quantitative performance baseline by presenting representative State-of-the-Art (SOTA) accuracies for major sensing modalities on benchmark datasets. This serves as a crucial reference for technology selection. Finally, to address the challenges of robustness and semantic ambiguity in HGR, we propose a forward-looking vision where multi-modal perception is integrated with Large Language Models (LLMs) to elevate gesture analysis from pattern recognition to genuine intent understanding.

Cell manipulation using optically induced dielectrophoresis (ODEP) in a microfluidic system has drawn much attention due to its simplicity and being damage-free at the cellular level. Additionally, to improve its manipulation efficiency and accuracy, automatic manipulation methods have been applied in the ODEP system. However, the current automatic manipulation methods of ODEP rarely consider the impact of non-manipulated targets on cell manipulation, thereby reducing the operating efficiency and accuracy. Here, we propose a new, to our knowledge, automatic manipulation method of ODEP based on a path planning algorithm of the improved A-star. First, the maximum influence range of ODEP force generated by the virtual electrode was investigated by a numerical simulation, and it was also taken as the limit to expand the scope of the infeasible region in path planning to avoid the impact of the non-operational target on manipulation accuracy. Then, an improved A-star algorithm with target range constraints was proposed to optimize the manipulation path and improve the operation efficiency. Finally, experiments on cell separation were also carried out to validate the feasibility of the proposed automatic manipulation method. This work provides an automated method to improve the accuracy of ODEP manipulation, which is of great significance to further promote the application of ODEP in cell manipulation.