From imaging enhancement to surgical autonomy: the evolutionary role of nanotechnology in AI-driven robotic surgery

Abstract

Robot-assisted minimally invasive surgery provides stable visualisation and manipulation, but direct access to local tissue state remains limited. Lesion boundaries, mechanics, molecular activity and microenvironmental changes are often inferred from conventional feedback. This review evaluates how nano-enabled imaging, flexible and local sensing, and micro/nanorobotic intervention may extend signals and localised actions in robot-assisted procedures. We organise the evidence along an information-flow framework: signal acquisition, processing or computational interpretation, task-relevant state estimation, conversion into a decision variable or control-relevant input, and feedback-guided robotic action. Processing may involve calibration, signal processing, computational modelling, conventional machine learning, deep learning or artificial intelligence (AI)-based methods. The evidence reviewed is strongest for localised signal acquisition and supervised intervention. Selected forms of computational interpretation have been demonstrated, and localisation- or motion-derived geometric states enter feedback controllers in some systems. By contrast, biochemical, spectral, tactile and physiological states are usually displayed, analysed offline or returned through human-facing feedback; they rarely modify an online surgical controller. Closed-loop navigation in phantoms, surrogate-device studies and constrained navigation autonomy therefore represent partial control capabilities, not clinically relevant surgical autonomy. The principal missing transitions are robust validation of task-relevant tissue states, uncertainty-aware conversion of those states into control inputs, and safe action within realistic surgical workflows. Near-term clinical use is more likely to centre on bounded nano-enabled sensing or intervention modules embedded within supervised procedures. Progress towards bounded task autonomy will require reproducibility, material safety, workflow compatibility, explicit human supervision and prospective evidence of patient and task benefits.

Introduction

Robot-assisted minimally invasive surgery provides stable visualisation, articulated manipulation and motion scaling in anatomically constrained procedures [1]. These capabilities improve access and execution, but the system still depends largely on camera images, instrument kinematics and surgeon interpretation [2]. Conventional feedback may not directly resolve lesion boundaries, local mechanics, molecular activity or changing microenvironmental cues. This limitation concerns access to task-relevant tissue states, rather than the robot’s ability to execute commanded motion.

At the field level, a recent review of robot-assisted surgery for prostate cancer described a staged progression from clinical adoption towards increasingly intelligent systems. It also emphasised that future development remains conditional on clinical, engineering and societal challenges [3].

Nanotechnology is relevant because it can place imaging, spectral acquisition, flexible sensing, microscale access and localised intervention closer to disease-associated biological scales [4–9]. Urological malignancies illustrate biological heterogeneity beyond conventional surgical vision alone. In clear cell renal cell carcinoma, pyroptosis-regulator patterns have been associated with distinct immune-microenvironment profiles [10]. CircPPP1CB has separately been implicated in bladder-cancer progression through the miR-1307-3p/SMG1 axis [11]. These studies neither establish intraoperative sensing targets nor show nano-enabled detection during surgery. They instead illustrate molecular and microenvironmental states that may be clinically relevant but are not directly resolved by standard visual feedback.

Cross-scale complementarity provides a more realistic system context than device substitution. Conventional surgical robots can support access, visualisation, planning and macroscopic manipulation. Nano-enabled interfaces and micro/nanorobots may add localised signal acquisition, restricted anatomical access or focal intervention [7, 12–14]. These components could be coupled within selected workflows, but the literature does not establish that a unified cross-scale architecture is inevitable or clinically mature.

Computation links measured signals to potential actions, but the methods and endpoints differ. Calibration and signal processing can convert raw electrical, optical or mechanical outputs into usable measurements. Physical modelling can estimate dynamics, while conventional machine learning or deep learning can support classification, localisation or prediction where appropriate [9, 14, 15]. State estimation, path planning and feedback control are separate functions that require independent validation. Not every signal requires artificial intelligence (AI), and an interpreted output does not become a control-relevant input unless it changes a decision or command. In this review, ‘AI-driven’ denotes the broader analytical trajectory through which computational methods may contribute to signal interpretation, state estimation or control. It does not imply that every reviewed platform uses AI or that clinically relevant surgical autonomy has been achieved.

This functional definition also distinguishes the present review from adjacent reviews that offer complementary organisational perspectives on microrobotic development and translation. Bozuyuk et al. examined the clinical translation of mobile microrobots and associated fabrication, actuation, imaging and biological constraints [16]. Ju et al. organised micro/nanorobot development across materials, propulsion, collective behaviour, intelligence, applications, manufacturing and regulation [17]. Ceylan et al. proposed a clinical-value and technology-readiness framework that progresses from unmet need to adoption [18]. The present review uses a different, but compatible, classification axis centred on information flow through a robot-assisted procedure.

For each representative system, we assess which signal is acquired, how it is processed or interpreted and which task-relevant state is estimated. We then determine whether that state becomes a decision variable or control-relevant input and modifies feedback-guided action. This analytical framework explicitly distinguishes demonstrated links from missing transitions. It also recognises that not every system uses AI, and that automation or navigation autonomy does not by itself constitute surgical autonomy. Evidence maturity, safety, workflow compatibility, human supervision and clinical relevance are treated as cross-cutting dimensions. Section  3 examines acquisition, local intervention and interpretation; Sect.  4 evaluates state-to-control links; and Sect.  5 considers translational barriers and bounded task autonomy.

Literature search and analytical framework

Literature search and source selection

A literature search was conducted in PubMed and the Web of Science Core Collection, with no initial publication-date restriction. In PubMed, Medical Subject Headings (MeSH) were combined, where applicable, with title and abstract terms; corresponding topic terms were used in Web of Science. The initial search retrieved 472 records, comprising 383 from PubMed and 89 from Web of Science. Records were imported into and deduplicated in Zotero. One author performed the initial screening; the supervising author reviewed the decisions and resolved any disagreement.

Peer-reviewed original studies, relevant reviews, and roadmap or regulatory-context articles were retained when they directly addressed nano-enabled imaging or sensing, micro/nanorobotic intervention, computational interpretation, feedback-guided control, autonomy-supporting capabilities, or clinical translation relevant to robot-assisted procedures. Duplicate records, articles without direct thematic relevance, non-biomedical robotic studies without a clear transferable surgical context, commentary without a distinct evidence contribution, and records providing insufficient information for thematic synthesis were excluded. Reference lists of relevant reviews and representative original studies were also checked manually.

A targeted supplementary update was conducted on 1 August 2026 to identify recent evidence on surgical foundation models, surgical-video foundation models, vision–language models, surgical scene or workflow understanding, and surgical-robot autonomy classifications. Four additional sources were retained, resulting in a final structured narrative corpus of 63 cited sources.

Analytical framework

The evidence was synthesised using a qualitative information-flow framework: signal acquisition → processing or computational interpretation → task-relevant state estimation → decision or control-relevant input → feedback-guided robotic action. This framework was used to distinguish demonstrated links from missing transitions rather than as a formal quantitative methodology. Tables 1 and 2 provide structured technology comparisons, Fig. 1 maps demonstrated and incomplete information-flow links, and Fig. 2 summarises translational and evidence milestones.

Table 1.

Representative nano-enabled imaging and local sensing interfaces relevant to robotic intervention

Study	Platform and readout principle	Measured target or measurand	Processing or interpretation	Experimental model	Key quantitative result	Principal advantage	Principal limitation	Evidence stage
Wang et al. (2026) [21]	Magnetic biopsy microrobot with Au-nanospike SERS interface	Microrobot position; H₂O₂; retrieved tissue class	Image guidance; SERS calibration; CNN classification	Rabbit airway; simulated lesion; ex vivo human tissue	50 µM H₂O₂ read as 54 µM; ≈94.3% five-class accuracy (70:30 split)	Integrated navigation, sampling, spectral readout and tissue classification	Ex vivo post-retrieval classification; no classifier-to-controller feedback	Preclinical proof of concept
Xu et al. (2022) [41]	Soft CNT/PDMS capacitive tactile sensor with six-channel readout	Pressure and loading direction	Capacitance normalisation; naive Bayes/SVM/ANN	Benchtop loading across angular and pressure classes	0.306 ± 0.001 kPa− 1 below 10 kPa; 99.83% naive Bayes/SVM, 98% ANN	Omnidirectional loading-state classification from multichannel tactile signals	No tissue model, surgical robot or control integration	Benchtop proof of concept
Hu et al. (2025) [42]	Two-photon-polymerised microspring piezoresistive force sensor	Microforce; fabrication printability as a separate target	Force calibration/finite-element modelling; radial-basis-function SVM for printability	Benchtop fabrication and testing; 222 fabrication samples	Gauge factor 5.65 over 0–300 µN; SVM 92.66% test and 95.56% validation accuracy	Microscale force sensing in conductive photosensitive resin	Manufacturing ML does not interpret force; no tissue or control loop	Benchtop sensor/fabrication proof of concept
Guo et al. (2020) [37]	Stretchable Cu–EGaIn electronic skin with CT and electrical readout	External marker position; conductor state	CT inspection; electrical characterisation; no AI	Benchtop marker tests; rabbit surface-marker demonstration	2-, 3-, 5- and 10-mm markers visible; 1 mm not visible at 120 kV	Flexible, conformal radiopaque external localisation cue	External marker only; no robotic navigation or feedback control	Benchtop/small-animal feasibility
Yan et al. (2018) [43]	Self-powered 3D piezoelectric sensor array	Pressure and physiological-event patterns	Waveform inspection and calibration; no AI	Benchtop loading and human wearable demonstrations	0.19 V kPa− 1 sensitivity; <16-ms response time	Thin, self-powered and mechanically compliant sensing	No robotic system, tissue task or feedback controller	Benchtop/wearable feasibility
Han et al. (2022) [22]	Vision-based microfluidic pressure sensor with RGB deformation readout	Grip/contact force	SIFT feature extraction; displacement-to-force calibration	Benchtop loading and hand-device demonstrations	0–35 N calibration with < 4% error to 25 N; ≈4-s recovery	Visual force readout without complex electronics	Slow recovery and rehabilitation context; no robotic feedback	Benchtop/rehabilitation
Fu et al. (2025) [23]	Magnetically steered catheter with triaxial FBG force sensing	Axial and lateral contact forces	FBG calibration and triaxial force reconstruction; no AI	Ex vivo porcine kidney; bronchial and stomach-polyp phantoms	494.86, 1338.99 and 623.35 pm N⁻¹ sensitivities; 2.08%, 4.81% and 1.09% RMSE	Integrated steering, force sensing, vision and biopsy	Force did not modify steering/biopsy; no in vivo evidence	Benchtop/ex vivo
Hou et al. (2024) [24]	Endoscopic forceps with three-axis piezoresistive MEMS sensor	Three-axis contact force	Bridge calibration; least-squares force reconstruction; no AI	Calibration; fresh ex vivo porcine stomach	0–1.2 N axial and ± 0.6 N lateral ranges; 1.15%, 2.43% and 1.18% full-scale errors	Miniaturised triaxial sensing in endoscopic forceps	Operator-directed measurement; no robot-controller integration	Benchtop/ex vivo
Othman and Qasaimeh (2026) [25]	Laparoscopic grasper with soft microfluidic force sensors	Handle actuation and jaw contact forces	Resistance-to-force calibration and real-time readout; no AI	Benchtop grasper; PDMS tissue-like model	65-ms response, 80-ms recovery and 8.9% hysteresis	Separates operator actuation from jaw contact load	Tissue-like model; ageing requires recalibration; no automatic control	Benchtop proof of concept
Li et al. (2025) [26]	Multimodal proximity–tactile sensor with haptic/digital-twin feedback	Distance, contact force and simulated-tissue class	Rule-based switching; one-dimensional CNN	Benchtop robot; five simulated tissues (100 samples)	99.05% ultrasonic and 96.47% capacitive accuracy; 91.6% CNN accuracy	Spans non-contact approach to tactile contact with human feedback	Simulated tissues; CNN class remained outside controller	Benchtop proof of concept

Note: Results are not directly comparable because measurands, input ranges, readout mechanisms and experimental models differ

Table 2.

Representative micro/nanorobotic platforms and feedback-guided systems

Study	Platform and actuation	Localisation or feedback variable	Localised task and robotic action	State or control-relevant input	Experimental model	Key quantitative result	Human supervision	Principal limitation	Evidence stage and autonomy relevance
Wang et al. (2026) [21]	Magnetically actuated conical-helix biopsy microrobot	Bronchoscopic and fluoroscopic guidance; post-retrieval output outside control	Drilling-assisted tissue sampling and retrieval	Position guided navigation; SERS/CNN output remained outside control	Rabbit airway; simulated lesion; ex vivo human tissue	90% repeatability within 100-µm radius; 100% within 200 μm	Externally supervised image-guided actuation	No classifier-to-controller feedback	Preclinical; no autonomous control
Han et al. (2024) [36]	Bioresorbable acoustic hydrogel microrobot with ultrasound propulsion and magnetic steering	Ultrasound-derived location	Bladder navigation and localised 5-fluorouracil delivery	Operator-selected ultrasound and steering; no image-to-actuator input	In vitro; orthotopic mouse bladder-tumour model	≈ 93% bioluminescence reduction by day 14 after four treatments (n = 3 per group)	Experimenter-controlled navigation and treatment schedule	Propulsion, imaging, steering and therapy externally orchestrated	Small-animal preclinical; no autonomous treatment
Barroso et al. (2015) [34]	Optically assembled Bacillus subtilis–zeolite L biohybrid with bacterium-driven motility	Microscope trajectory tracking; no online feedback	In vitro assembly and propulsive motility	No task-relevant state or control input	In vitro optical assembly and tracking	≈ 2-µm zeolite; ≈45 μm in 6.6 s (≈ 7 μm s− 1)	Human optical assembly and observation	No tissue intervention, sensing or feedback-guided task	In vitro proof of concept; no autonomy relevance
Wei et al. (2026) [35]	Shape-memory rod-to-helix micro-transformer with magnetic propulsion and magnetothermal recovery	Optically observed position/configuration; no closed-loop use	Simulated-vessel navigation, helix transformation and thrombus fragmentation	Externally selected phase; preset magnetic/heating commands	Zero-flow, fully occluded simulated vessel	92.11% recovery; 148 ± 8 s switching time; 52.7% clearance	Externally commanded actuation sequence	Heating requirement and simulated vessel; no sensing–control loop	Benchtop; externally commanded switching, no autonomy
Wang et al. (2024) [45]	Magnetically actuated tPA nanorobot swarm with catheter/balloon assistance	Fluoroscopic swarm visibility; procedural recanalisation assessment	Deployment, local thrombolysis and retrieval	Operator imaging, catheter/balloon manipulation and magnetic commands	In vitro vessel; ex vivo human placenta; rat and rabbit models	Complete lysis in 18 min with 42-fold dose reduction; ≈80% ex vivo retrieval	Operator controlled	Separate experimental settings; incomplete sample-size reporting; no automatic state-to-control pathway	Benchtop/ex vivo/preclinical; targeted thrombolysis is not autonomy
Liu et al. (2024) [46]	Thermally triggered untethered microgrippers with magnetic manoeuvring	Observed open/folded material state; no feedback localisation	Ex vivo urinary-tract biopsy and specimen retrieval	Temperature exposure plus operator catheter/magnet commands	Ex vivo porcine ureter and renal pelvis	≈ 37 °C folding in ≈ 10 min; 79.5%, 87% and 88% deployment	Operator deployed, manoeuvred and retrieved grippers	No in vivo or feedback-guided task evidence	Ex vivo; autonomous folding, not autonomous biopsy
Yang et al. (2015) [29]	AFM nanorobotic probe; position/velocity/force commands	Probe position and interaction force	Live-cell puncture and intermediate-filament nanodissection	Operator-defined trajectory with position/force commands	In vitro live keratinocytes	≈ 1.9 ± 0.5 nN penetration force; ≈18 nN maximum cutting force	Operator-selected trajectory	Cell model; no clinical workflow or safety framework	In vitro operator-directed feedback; no surgical autonomy
Song et al. (2012) [30]	AFM nanorobot with imaging, manipulation and force/position control	Tip position, interaction force and local imaging	Cell imaging, force interrogation, manipulation and local delivery	Operator-selected task and force/position/trajectory commands	In vitro cell system	One frame s− 1 for a 500 × 500 nm compressive-sensing scan	Operator-controlled haptic interface	Loop latency NR; no surgical workflow or safety constraints	In vitro operator-directed feedback; no surgical autonomy
Martel (2007) [27]	MRI-gradient-propelled 1.5-mm ferromagnetic sphere surrogate	MRI position and path error at 24 Hz	Closed-loop endovascular path traversal	Position/path state converted into gradient-field commands	Living swine carotid; clinical 1.5-T MRI	24-Hz control; 10 round trips at ≈ 10 cm s− 1 in a 25-kg swine	Human preplanning/oversight; automatic real-time loop	Navigation surrogate; no therapy, tissue-state input or fallback framework	Large-animal navigation; automated under preplanning, no task autonomy
Martel et al. (2008) [28]	MRI-gradient-propelled sphere with automated waypoint control	MRI position, waypoint state, path error and timing	Automatic waypoint progression and trajectory correction	Position/waypoint state converted into MRI-gradient commands	Living swine carotid; clinical MRI	41-ms feedback cycle; 13 cm s− 1 with a 10-mm waypoint radius	Human planning/oversight outside real-time loop	Surrogate device; no therapy, tissue-state input or bounded surgical task	Large-animal navigation; automated under preplanning, no task autonomy
Harduf et al. (2018) [51]	Flexible superparamagnetic microswimmer; open-loop magnetic actuation	Optical trajectory/speed/direction; no online correction	Open-loop swimming and stability-regime characterisation	Model-informed parameter choice; no controller input	Benchtop fluid experiment and nonlinear-dynamics model	Direction transition at β ≈ 2; ≈1.52–1.69 μm s− 1 at 3 Hz, β = 2.5	Human-selected magnetic-field parameters	No biological model, closed-loop controller or task-state signal	Benchtop model/experiment; model-level relevance only
Hu et al. (2024) [52]	Helical micro/nanorobot with open-loop magnetic propulsion	Optical displacement/velocity for offline model validation	Open-loop propulsion and offline locomotion prediction	Predicted motion for design; no online control input	Benchtop fluid tests and fluid–structure-interaction simulation	0.05–15.94% prediction errors; 612.54 μm s− 1 predicted at 8 Hz	Human-selected design and actuation conditions	No biological model or online controller	Benchtop open-loop validation; model-level relevance only
Yang et al. (2013) [31]	AFM nanorobot with fluorescence microscopy and force control	AFM position/force; Ca²⁺ and stiffness outputs	Repeated nanoindentation with β-cell mechanical/Ca²⁺ monitoring	Position/force commands; biological states outside controller	In vitro RINm5f β-cell culture	≈ 7.0 kPa baseline stiffness; ≈14.8 kPa after 40 mM glucose	Operator-supervised AFM protocol	Cell model; biological states do not drive action	In vitro mechanistic evidence; no tissue-state-driven action
Lang et al. (2026) [19]	Magnetic hydrogel microrobot with swimming and rolling gaits	Ultrasound-derived position, velocity and orientation with Kalman filtering	Closed-loop trajectory tracking and adaptive obstacle traversal	Geometric state/path error to MPC/PID with rule-based gait switching	Tissue-mimicking phantom	≈ 90% frames with ≤ 3-mm localisation error; <4-mm representative obstacle-test path error	Experimenter-specified goal with human oversight	Millimetre scale; no tissue intervention, animal or safety evidence	Phantom closed loop; bounded navigation, not surgical autonomy
Zhao et al. (2026) [56]	Magnetic microrobot/swarm with physics-constrained learned navigation	Vision-derived robot, target and obstacle state	Target reaching and obstacle avoidance	Physics-constrained networks/planners generate magnetic-field parameters	Simulation and optical benchtop tests; no biological model	23 targets in 20 min/15 obstacles; ≈300-µm swarm traversed 40 mm in 17 min	Human demonstrations and target selection	Near-idealised perception; no biological sensing, tissue intervention or therapy	Navigation autonomy only; no surgical autonomy

Review-level, simulation, theoretical and non-biomedical context sources [50, 53–55] are discussed separately in Sect. 4.2 and 4.3and are not presented here as direct biomedical experimental evidence

Fig. 1.

Evidence boundaries linking nano-enabled signal acquisition to feedback-guided robotic action. a, Representative imaging, sensing, biochemical-readout and micro/nanorobotic interfaces. b, Corresponding geometric, tissue-contact and biochemical or spectral states. c, Solid arrows indicate demonstrated links, whereas dashed arrows indicate incomplete or indirect transitions. Geometric states have entered bounded feedback control in selected phantom or benchtop systems [19, 20], whereas tissue-contact and biochemical outputs generally remain operator-facing or outside the online controller [21–26]. d, Current evidence does not establish a continuous pathway from nano-enabled tissue-state estimation to clinically relevant surgical autonomy. Navigation autonomy is not surgical autonomy; human supervision, override and safe-stop remain necessary. All insets are schematic. MPC, model-predictive control; PID, proportional–integral–derivative control

Fig. 2.

Translational barriers and evidence milestones towards clinically relevant task autonomy. a, Imaging and sensing barriers and the milestone of a validated task-relevant state. b, Micro/nanorobotic platform barriers and the milestone of safe localised action. c, Computation and control barriers and the milestone of a validated control-relevant input. d, Conditional roadmap from supervised modules to state-informed decisions and bounded task autonomy. Representative studies illustrate localised intervention, feedback-guided navigation and model-based control at different evidence stages [27–33]. Dashed arrows indicate proposed evidence milestones rather than a demonstrated or inevitable progression. Advancement remains conditional on safety, reproducibility, workflow compatibility, prospective clinical value and retained human supervision, fallback, takeover and safe-stop. All insets are schematic

Nano-enabled signal acquisition and localised intervention

Nano-enabled and microscale technologies contribute distinct sensing and intervention functions relevant to robot-assisted procedures. Imaging interfaces provide spatial or spectral signals, local sensors transduce mechanical or physiological events, and micro/nanorobotic platforms perform localised tasks. Signal processing and computational methods can then convert selected outputs into interpretable estimates. These functions are assessed as partial information-flow links at different levels of evidence maturity, rather than as a completed pathway to surgical autonomy. Figure 3 provides an overview of representative anatomical targets, wireless actuation modalities, and the supervised functional sequence linking signal acquisition to feedback-guided localised action.

Fig. 3.

Representative target sites, actuation modalities and supervised functional pipeline of medical microrobots. a, Representative anatomical targets and localised tasks, including delivery, sampling, sensing or imaging and micromanipulation. b, Major wireless actuation modalities and their principal advantages and limitations. c, Conceptual cross-study sequence from sensing or imaging through signal processing, localisation, task planning and actuation to feedback-guided localised action. Representative evidence includes biohybrid assembly, task-mode switching, MRI-guided navigation, bladder microrobots and image-guided biopsy with post-retrieval SERS/CNN analysis [21, 27, 28, 34–36]. These capabilities were demonstrated in separate studies and do not represent an integrated autonomous clinical system. All insets are schematic and do not depict patient data or original experimental images. 3D, three-dimensional; CNN, convolutional neural network; MRI, magnetic resonance imaging; SERS, surface-enhanced Raman scattering

Imaging-derived localisation and spectral signal acquisition

Imaging-derived cues can support localisation, whereas spectroscopy can characterise material collected at a target. The studies considered here illustrate these roles through an external radiopaque marker and a biopsy microrobot integrating image-guided navigation, sampling and post-retrieval spectral analysis [21, 37].

Guo et al. developed a stretchable Cu–EGaIn interface whose high-density liquid-metal component is radiopaque on computed tomography (CT). In phantom tests with otherwise matched scan parameters, markers with diameters of 2, 3, 5 and 10 mm were visible at 80 and 120 kV. A 1-mm marker was not visible at 120 kV, whereas a 10-mm surface marker was visualised in a rabbit at 80 kV [37]. The work establishes a conformal CT-visible spatial cue from benchtop to small-animal feasibility, without demonstrating lesion targeting or robotic navigation integration.

Wang et al. reported a magnetically actuated conical-helix biopsy microrobot with Au nanospikes for local sampling and surface-enhanced Raman scattering (SERS) analysis [21]. Bronchoscopy and fluoroscopy guided navigation in rabbit airways, while sampling and sensing were assessed in benchtop, ex vivo porcine-lung and ex vivo human-tissue settings. Sampling repeatability was 90% within a 100-µm target-zone radius and 100% within a 200-µm radius. For a simulated lesion containing 50 µM H₂O₂, SERS calibration yielded 54 µM. A convolutional neural network (CNN) classified healthy tissue and four lung-cancer tissue categories with approximately 94.3% accuracy using a 70:30 training/test split [21]. This sequence links image-guided access, sampling and post-retrieval interpretation, but the classification output did not control navigation or sampling.

Together, these studies illustrate distinct information-flow links at different levels of evidence maturity [21, 37]. The marker supports spatial signal acquisition, whereas the biopsy platform reaches a validated task-relevant tissue-state estimate under ex vivo conditions. Neither converts the resulting output into a decision variable or control-relevant input.

More recent work extends imaging-derived localisation from external cues to microrobot visibility during intervention. Han et al. developed a two-photon-polymerised bioresorbable acoustic microrobot containing Fe₃O₄ nanoparticles and 5-fluorouracil within a hydrogel body [36]. Focused ultrasound drove an oscillating trapped microbubble for acoustic propulsion and ultrasound contrast, whereas an external magnetic field provided steering. Hydrolysis supported subsequent material degradation, and therapeutic delivery was evaluated separately in an orthotopic mouse bladder-tumour model. After four treatments, tumour bioluminescence decreased by approximately 93% by day 14 in the treatment group (n = 3 per group) [36]. The study therefore combines propulsion, visibility, magnetic guidance, biodegradation and localised delivery at the platform level, but these functions were externally orchestrated and do not demonstrate autonomous treatment.

Flow-sensitive imaging provides two experimentally distinct pathways for intravascular microrobot localisation. In the Doppler study, Wang et al. used colour Doppler ultrasound to detect flow perturbations from a magnetically actuated Fe₃O₄ microswarm [38]. Dynamic-region extraction and Kalman filtering produced a mean localisation error of 0.34 ± 0.19 mm (n = 10) under the reported conditions. Navigation was then demonstrated in an ex vivo porcine coronary artery [38]. In the speckle-contrast study, Wang et al. tracked a Fe₃O₄@SiO₂ microswarm through tissue, an ex vivo human placenta and a surgically exposed rat femoral vein [39]. Localisation errors of 0.54 ± 0.32 and 0.62 ± 0.29 mm were reported under two tested conditions [39]. Both studies used external magnetic actuation under direct human supervision. They represent separate imaging-guided navigation demonstrations, not an integrated or autonomous vascular pathway.

Flexible and local sensing interfaces

Flexible and local sensing interfaces can be grouped by transduction or readout principle and by measurand. At the broader materials level, review evidence on graphene-based strain and pressure sensors links device structure to sensing performance, but does not establish surgical force feedback or controller integration [40]. The studies reviewed include capacitive pressure and direction sensing, piezoresistive microforce measurement, piezoelectric monitoring of pressure and physiological motion, and vision-based microfluidic force estimation [22, 41–43]. This classification separates signal production from the inference of a task-relevant state.

Xu et al. constructed a fully soft six-channel capacitive sensor using carbon-nanotube/polydimethylsiloxane (CNT/PDMS) micro-spine electrodes and a hemispheric dielectric structure [41]. Under benchtop loading, sensitivity was 0.306 ± 0.001 kPa^{− 1} below 10 kPa, with a reported range of 2.55 Pa to 160 kPa. The sensor also withstood more than 3,200 cycles at 16 kPa. Gaussian naive Bayes and Gaussian-kernel support vector machine (SVM) classifiers each achieved 99.83% accuracy, while an artificial neural network achieved 98% for the tested loading classes [41]. The study therefore demonstrates benchtop classification of tactile loading states, although it includes no tissue model, surgical robot or control loop.

Hu et al. used two-photon polymerisation to fabricate a piezoresistive microspring force sensor from an MXene-containing conductive photosensitive resin [42]. Benchtop force–resistance characterisation covered 0–300 µN and produced a gauge factor of 5.65. A separate radial-basis-function SVM predicted resin polymerisability from composition and laser-processing parameters. Accuracy was 92.66% in testing and 95.56% in validation across 222 samples [42]. Here, machine learning predicts fabrication printability rather than interpreting the measured force signal. The device remains a benchtop microsensor proof of concept without a task-relevant tissue or system-state estimate.

Yan et al. reported an approximately 80-µm-thick cellular sensor array containing piezoelectric nanoparticles that generated electrical outputs under pressure and physiological motion [43]. Benchtop characterisation yielded a pressure sensitivity of 0.19 V kPa^{− 1}, a response time below 16 ms and operation for more than 4,500 loading cycles. Wearable feasibility demonstrations included heartbeat and eye-movement monitoring outside surgical settings [43]. Direct waveform inspection, rather than machine learning, provided the reported interpretation pathway. The device establishes a self-powered, mechanically compliant sensing interface but was not tested surgically or connected to a robotic controller.

Han et al. used a soft microfluidic “Smart Skin” in which contact deformation displaced coloured liquid recorded by a camera [22]. Scale-invariant feature transform (SIFT) processing located the channel, and calibration converted liquid displacement into force. The system was calibrated over 0–35 N and estimated forces up to 25 N with an error below 4%. Its recovery time was approximately 4 s in benchtop and hand-device demonstrations [22]. The resulting output is a camera-derived force estimate from microfluidic displacement. The recovery time and rehabilitation setting limit its present relevance to real-time surgical feedback.

Instrument-integrated force sensing can place a task-relevant mechanical signal directly at a steerable surgical interface. Fu et al. combined magnetic catheter steering, endoscopic and biopsy channels, and triaxial fibre Bragg grating (FBG) sensing [23]. The three axes showed sensitivities of 494.86, 1338.99 and 623.35 pm N⁻¹, with corresponding root-mean-square errors of 2.08%, 4.81% and 1.09% in calibration. Ex vivo porcine-kidney palpation further produced force differences between the tested regions [23]. The platform integrates palpation and biopsy functions, but the measured force was presented for operator interpretation and did not automatically modify catheter steering or biopsy.

Two other instrument-level designs use different transduction mechanisms and evidence models. Hou et al. integrated a three-axis piezoresistive microelectromechanical-system (MEMS) sensor into 3.5-mm endoscopic forceps [24]. The reported ranges were 0–1.2 N axially and ± 0.6 N laterally, with full-scale errors of 1.15%, 2.43% and 1.18%. Testing was subsequently extended to an ex vivo porcine-stomach model [24]. Othman and Qasaimeh instead used PDMS microchannels filled with Galinstan to separate handle actuation from jaw contact force. The benchtop sensor showed a 65-ms response, 80-ms recovery and 8.9% hysteresis, and was demonstrated on a tissue-like PDMS model [25]. These systems extend local force acquisition, but neither force display nor calibrated readout was incorporated into automatic robotic control.

Multimodal acquisition can span the transition from non-contact approach to physical tissue contact. Li et al. combined ultrasonic long-range proximity sensing, capacitive close-range sensing and triboelectric tactile sensing with a vibration wristband and digital-twin interface [26]. Accuracy was 99.05% across the tested 2–14-cm ultrasonic range and 96.47% below 1.6 cm for capacitive sensing. Triboelectric sensitivity was 0.241 V N⁻¹ from 0.5 to 5 N. A one-dimensional convolutional neural network classified five simulated tissue classes with 91.6% leave-one-out accuracy across 100 samples [26]. Distance and contact cues supported human feedback, whereas the simulated-tissue classification remained outside the robot controller. The study therefore does not establish tissue-state-driven robotic action under closed-loop control.

The reported performance metrics are not directly comparable because the measurands, loading ranges, readout mechanisms and test models differ [22, 41–43]. Collectively, the studies establish signal acquisition and, in selected cases, task-relevant loading, force or physiological estimates under reported conditions. None of these systems incorporated its sensor output into a surgical robot controller, leaving the state-to-control transition untested.

The readout principles, measured targets, interpretation methods and evidence stages of these representative imaging and local sensing interfaces are summarised in Table 1.

System components, actuation mechanisms, and localised tasks of representative micro/nanorobots

To make system components and operating principles explicit, we adapt the system-level organisation used by Iacovacci et al. [44]. Their framework relates microrobot size and design to the working environment, actuation, in vivo tracking, biomedical task and translational constraints. Each representative configuration is therefore described through five system elements. These are (i) microrobot body or functional structure; (ii) actuation or energy-transfer mechanism; and (iii) localisation, tracking or feedback method. We also consider (iv) task interface and localised action; and (v) external command, human supervision or closed-loop control. The broader wireless-actuation taxonomy includes magnetic fields, ultrasound, light and hybrid mechanisms [44]. The original studies considered below represent selected configurations rather than every category.

Wang et al. [21] developed a conical-helix microrobot with an Au-nanospike interface for tissue and biomarker sampling and SERS analysis. A rotating magnetic field supplied actuation, while bronchoscopy and fluoroscopy provided localisation and procedural guidance. The task interface coupled drilling-assisted biopsy to post-retrieval SERS analysis. External guidance governed navigation, sampling and retrieval, and the convolutional neural network classifier operated after retrieval, outside the microrobot controller. Navigation was evaluated in an in vivo rabbit model, whereas human-tissue classification was ex vivo. The study therefore combined image-guided access, localised sampling and post-retrieval analysis within a preclinical workflow, but did not convert the classifier output into microrobot motion or sampling commands [21].

Barroso et al. [34] used a bacterium–zeolite biohybrid comprising Bacillus subtilis coupled to an approximately 2-µm zeolite-L microcrystal. Holographic optical trapping assembled the body, after which bacterium-driven motility provided propulsion. Localisation was limited to direct microscopic observation in vitro, without an independent tracking or feedback channel. The task interface was the transport of the abiotic particle. The assembled object travelled approximately 45 μm in 6.6 s, corresponding to an average velocity of about 7 μm s^{− 1}. Operation remained experimenter initiated and observed. These results demonstrate optical assembly and biohybrid motility, but not tissue intervention, a feedback-guided task or surgical control [34].

Wei et al. [35] used a 15-mm Fe₃O₄-containing shape-memory-polymer body that moved as a rod and recovered a helical task mode. A magnetic-field gradient supplied navigation, an alternating field produced magnetothermal recovery, and a rotating field drove mechanical fragmentation. The simulated-vessel experiment did not incorporate sensing-based localisation or feedback. The task interface comprised externally commanded rod-to-helix switching followed by thrombus fragmentation. Shape recovery reached 92.11%, with a structural switching time of 148 ± 8 s, and reported clearance was 52.7% in a zero-flow, completely occluded simulated vessel. Supervisory authority remained external to the platform throughout all reported task stages. The study supports commanded task-mode switching in PVC-vessel and bionic-blood models, but not sensing-based closed-loop control or clinical thrombectomy [35].

Two studies [29, 30] used tethered atomic force microscopy (AFM) probes as instrumented end effectors, with scanner actuation and force/position sensing. Local AFM images and interaction-force signals supplied instrument-level localisation and feedback. The task interface combined imaging and manipulation along operator-defined trajectories. In the first study, the probe punctured and dissected intermediate filaments in live keratinocytes. Penetration force was approximately 1.9 ± 0.5 nN, and maximum cutting force was approximately 18 nN. Young’s modulus decreased from 36.1 ± 3.5 to 18.4 ± 1.9 kPa (mean ± SEM; n = 200) [29]. In the second study, compressive-sensing reconstruction increased the reported imaging rate to one frame per second for a 500 × 500 nm scan. This reconstruction output provided visual information during the reported manipulation task [30]. In both systems, the operator selected the trajectory and retained supervisory authority. The in vitro feedback loop remained at the instrument level and did not constitute an autonomous surgical workflow.

Endovascular thrombolysis illustrates how several intervention components can be combined without establishing autonomous control. Wang et al. formed a magnetic colloidal swarm from Fe₃O₄@mesoporous-SiO₂ nanorobots anchored with tissue plasminogen activator (tPA) [45]. Catheter and balloon assistance, fluoroscopic localisation and an external robotic permanent magnet supported deployment and manipulation. In a 1.5-mm in vitro vessel, 15 µg of anchored tPA on 1 mg of nanorobots achieved complete lysis in 18 min. The authors reported a 42-fold dose reduction and an approximately 20-fold higher thrombolysis rate under those conditions. Separate ex vivo human-placenta experiments reported a retrieval rate of approximately 80%. Rat femoral-vein experiments reported a mean recanalisation time of approximately 37 min, while a distinct rabbit carotid experiment demonstrated fluoroscopy-guided motion [45]. These settings support deployment, localisation, thrombolysis and retrieval as complementary evidence blocks. The workflow remained operator directed and should not be described as autonomous surgery.

Localised sampling can alternatively exploit a stimuli-responsive material transition. Liu et al. fabricated six-arm untethered microgrippers that folded from approximately 717 μm to approximately 200 μm when the paraffin trigger softened near physiological temperature [46]. Folding occurred at approximately 37 °C within about 10 min, while catheter deployment, handheld-magnet manoeuvring and retrieval were evaluated separately in an ex vivo porcine ureter and renal pelvis. Passage/deployment rates were 79.5%, 87% and 88% for the three tested catheter configurations, and retrieved samples supported H&E and GATA3 histopathological assessment [46]. Temperature-triggered folding is an autonomous material response; the biopsy workflow itself remained operator directed and was not an autonomous task.

The comparison exposes distinct system boundaries. Actuation energy came from external magnetic fields in two systems [21, 35], from bacterium-driven motility after optical assembly in another [34], and from tethered AFM scanners in two instrument-level platforms [29, 30]. Localisation or interaction was measured through bronchoscopy/fluoroscopy, direct microscopic observation without feedback, no sensing-based feedback, or AFM force/position and local imaging, respectively. Biological target engagement included drilling-assisted sampling, simulated thrombus fragmentation and cellular nanodissection, whereas the biohybrid platform [34] involved no tissue intervention. Supervisory authority remained with an external operator or experimental command path in every configuration [21, 29, 30, 34, 35].

None of these systems combined local tissue sensing, validated task-state estimation and autonomous closed-loop adaptation within a surgical workflow. The evidence instead spans guided access with off-controller analysis, open-loop biohybrid motility, externally commanded transformation and operator-directed AFM manipulation. These are localised task capabilities with different feedback boundaries, not demonstrations of autonomous surgical control [21, 29, 30, 34, 35].

Signal processing and computational or AI-based interpretation

Interpretation methods span preprocessing and calibration, feature extraction, signal processing, conventional machine learning, deep learning and direct non-AI inference [21, 22, 41–43]. These approaches should not be collapsed into a universal AI layer. The key distinction is whether a raw signal becomes a processed output, a task-relevant state, and ultimately a decision variable or control-relevant input.

Non-AI pathways are evident in direct and image-based sensing. The piezoelectric array associated waveform patterns with pressure or physiological events, whereas Smart Skin used SIFT processing and calibration to estimate force [22, 43]. Both systems converted raw acquisition into a task-relevant estimate, but neither supplied that estimate to a robotic controller.

Conventional machine learning served different purposes across the studies. Xu et al. [41] reported that naive Bayes and SVM classifiers converted six-channel capacitance patterns into loading-condition classes with algorithm-specific accuracies of 99.83%. The artificial neural network achieved 98% under the same reported task [41]. By contrast, Hu et al. [42] reported that the SVM predicted material polymerisability from fabrication parameters. It neither interpreted the force signal nor inferred a tissue or system state.

Deep learning was applied after SERS acquisition and spectral preprocessing to classify retrieved human tissue into five categories with approximately 94.3% accuracy [21]. This output represents a task-relevant tissue classification, but not an intraoperative decision variable. It did not trigger, adapt or terminate navigation, drilling or sampling.

Recent studies further clarify the boundary between computational interpretation and online controller integration. In the multimodal sensor study, Li et al. used rule-based modality switching and a one-dimensional convolutional neural network [26]. These methods produced distance and contact estimates and simulated-tissue classifications. The outputs either reached the human interface or remained outside robotic control. By contrast, Lang et al. used inter-frame ultrasound processing and correlated-noise Kalman filtering to estimate geometric state [19]. This state was supplied online to model-predictive and proportional–integral–derivative controllers. Computation can therefore terminate at an operator-facing estimate or enter a feedback loop. Neither pathway implies that AI is a universal requirement across the systems reviewed.

Recent foundation-model and vision–language research broadens the interpretation layer beyond task-specific classifiers. SurgVISTA uses large-scale self-supervised surgical-video pretraining to learn spatiotemporal representations for downstream surgical-video understanding [47]. SurgVLP aligns surgical video with transcript-derived language for retrieval, temporal grounding, captioning and zero-shot tool, phase and action-triplet recognition [48]. ORQA integrates visual, audio and structured operating-room inputs for question answering, workflow prediction and safety-monitoring outputs that may support cognitive or operator-facing assistance [49]. These studies mainly address perception, representation learning, information retrieval, workflow recognition and human-facing support; none converts a nano-enabled tissue state into an online robotic-control input or establishes surgical autonomy.

The studies reviewed demonstrate several acquisition-to-interpretation and acquisition-to-state links using AI and non-AI methods [21, 22, 41–43]. However, no sensor-derived tissue, tactile, spectral or physiological state entered a surgical robot controller. This missing state-to-control transition separates the component capabilities described here from the feedback architectures assessed in Sect. 4.

From task-relevant state estimation to feedback-guided robotic action

Section 3 showed how selected platforms acquire, process and interpret signals, sometimes yielding task-relevant states. Section  4 asks whether those states become decision variables or control-relevant inputs, enter feedback loops and modify robotic action. We therefore separate state estimation from closed-loop action and compare experimental maturity, update characteristics and human supervision. For each feedback example, the analysis distinguishes the measured variable, inferred state, controller or rule and resulting action.

Task-relevant state estimation, decision variables, and control-relevant inputs

For this review, we distinguish a measured or inferred task state from information that actively enters decision-making or control. A task-relevant state is therefore more specific than an unprocessed measurement. It becomes a decision variable when used to select, modify or terminate an action. It becomes a control-relevant input when supplied to a controller or update rule. A state may still inform an operator without changing a robotic command. This convention prevents displayed classes, force estimates or physiological states from being conflated with closed-loop use.

Yang et al. combined AFM force–displacement measurements with fluorescence imaging to monitor mechanical and Ca²⁺ responses in insulinoma β-cells [31]. Hertz-model fitting yielded stiffness estimates, while ImageJ quantified fluorescence intensity. Baseline stiffness was approximately 7.0 kPa; stimulation with 40, 20 and 16 mM glucose produced values of approximately 14.8, 13.8 and 12.0 kPa, respectively. These measurements defined cellular mechanical and functional states in vitro. The AFM system also regulated probe trajectory and force, but stiffness and Ca²⁺ remained monitored outputs rather than controller inputs [31].

Other examples in Sect. 3 produced post-retrieval tissue classes, pressure and loading-direction classes, physiological-event states and grip-force estimates [21, 22, 41, 43]. These outputs arose from SERS interpretation, capacitive classification, direct piezoelectric readout or image-based calibration, rather than a common computational pathway. They were displayed or reported rather than used to adapt robotic action. Across the studies reviewed, no sensor-derived task state was directly used as a decision variable or control-relevant input [21, 22, 31, 41, 43]. State-estimation examples instead coexist with separate feedback systems driven mainly by position, trajectory or instrument force.

The newer sensing studies broaden the candidate task states while preserving this distinction. Fu et al. and Hou et al. produced triaxial interaction-force estimates with catheter and forceps systems [23, 24]. Othman and Qasaimeh separated handle actuation from jaw contact force using a soft microfluidic grasper [25]. Li et al. estimated proximity, contact and simulated-tissue class with a multimodal system [26]. These outputs were displayed to the operator, converted into haptic feedback or processed as a benchtop classification result. None was demonstrated as an automatic surgical decision variable or control-relevant input.

Feedback architectures based on localisation, motion and force

The feedback examples fall into localisation or trajectory loops and instrument force or position loops. In both groups, controller inputs are geometric or mechanical. Their update rates reflect distinct sensing and control processes and are not directly comparable across platforms. Tissue-state control would instead require a biochemical, spectral, tactile or cellular-state estimate to modify robotic action.

Martel used magnetic resonance imaging (MRI)-derived position to estimate device location relative to a planned path [27]. A computer-controlled loop translated position error into MRI-gradient propulsion commands for a 1.5-mm ferromagnetic sphere in the carotid artery of a living 25-kg swine using clinical 1.5-T MRI. At 24 Hz, the sphere completed ten preplanned round trips at an average speed of approximately 10 cm s^{− 1}. The loop operated automatically after path specification, but it remained a navigation surrogate without therapy, tissue-state input or a task-level safety and fallback framework [27].

Martel et al. subsequently combined path planning, tracking, event timing and propulsion in a waypoint-based protocol in the same large-animal setting [28]. MRI position and timing defined waypoint state, while a position controller and event scheduler generated gradient commands. A 41-ms feedback cycle, including a 22-ms tracking phase, supported 24-Hz operation. The 1.5-mm sphere reached 13 cm s^{− 1} using a 10-mm waypoint radius. Navigation was automatic within the preplanned protocol, with human planning and oversight outside the real-time loop. The experiment addressed geometric navigation, not therapy, tissue-responsive action or a bounded autonomous surgical task [28].

At the instrument scale, Yang et al. used AFM position and interaction force as feedback variables during live-keratinocyte nanodissection [29]. The operator selected the trajectory, while position, velocity and force commands guided puncture and cutting. The study did not report a numerical controller update rate or latency. This operator-directed in vitro system linked instrument feedback to nanomanipulation [29]. The measured stiffness change remained an outcome rather than a controller input.

Song et al. similarly integrated AFM tip position, interaction force and local imaging with an operator-controlled haptic interface [30]. The in vitro platform supported imaging, force interrogation, manipulation and local delivery. Compressive-sensing reconstruction provided one frame per second for a 500 × 500 nm scan, although force/position-loop latency was not reported. The operator selected and directed the task, leaving the system outside a surgical workflow [30]. For β-cells, AFM loading occurred at 1 Hz with a 5.98 μm s^{− 1} tip speed, while fluorescence images were acquired at one frame per second [31]. The operator-supervised protocol used these rates for acquisition and repeated manipulation, not to feed stiffness or Ca²⁺ states back into the AFM command.

As a non-biomedical analogy, Hu et al. coupled six-axis visual metrology to a proportional–integral controller and piezoelectric actuators in a compliant stage [50]. Translational resolutions were 5, 8 and 10 nm, and rotational resolutions were 10, 10 and 20 µrad under benchtop conditions. The study reported neither controller update rate nor the real-time operator role. The architecture illustrates pose-error servo control but contains no biological model, micro/nanorobot or surgical task [50].

Across the direct biomedical examples, position, path, timing and force supported feedback-guided navigation or instrument action [27–31]. Biochemical, spectral, tactile and tissue-state estimates were not used for online control. Geometric correction is therefore distinct from tissue-state control, while operator-directed AFM manipulation remains distinct from clinically relevant task autonomy.

Lang et al. provide a direct geometric state-to-control example using an approximately 8-mm magnetic hydrogel microrobot in a tissue-mimicking phantom [19]. B-mode ultrasound frames underwent inter-frame differencing and correlated-noise extended Kalman filtering. Path planning was coupled to model-predictive control (MPC) for swimming, proportional–integral–derivative (PID) control for rolling and rule-based adaptive gait switching. Approximately 90% of analysed frames had a localisation error of ≤ 3 mm. In a representative unseen-obstacle test, overall path error remained below 4 mm while the robot switched gait to bypass the obstacle [19]. The estimated geometric state therefore entered the controller and modified action. The evidence remains a millimetre-scale phantom proof of concept without tissue intervention, animal validation or surgical autonomy. The Doppler and speckle-contrast studies occupy a separate evidence boundary. Their imaging supported externally supervised magnetic navigation, but the derived position was not demonstrated as an automatic actuator input [38, 39].

Controller architecture must therefore be interpreted together with its sensing source, experimental model and level of human supervision.

Model-based robustness, dynamic regimes and model-level coordination

Dynamic models can identify regimes in which motion should remain predictable. Their control relevance, however, depends on whether the estimated state is updated online and changes a command. Agreement between prediction and experiment supports model validity, but does not by itself demonstrate feedback control. The studies considered here span experimentally tested dynamic-regime models, open-loop model validation and simulation-only controllers.

Harduf et al. combined optical trajectory measurements with a nonlinear parametric-excitation model of a flexible superparamagnetic microswimmer [51]. Here, β = B_y/B_x denotes the magnetic-field amplitude ratio. The model related this ratio and actuation frequency to swimming direction and stability regime. Under benchtop conditions, the direction transition occurred around β ≈ 2; at 3 Hz and β = 2.5, the maximum speed was approximately 1.52–1.69 μm s^{− 1}. Optical measurements did not correct motion online, and no controller selected the field parameters [51].

Hu et al. developed a coupled fluid–structure interaction model for corkscrew motion in general helical micro/nanorobots and compared predictions with benchtop fluid experiments [52]. Across the Table 1 cases, displacement-prediction errors ranged from 0.05% to 15.94%. In one 8-Hz case, the model predicted 612.54 μm s^{− 1}, with a 5.03% displacement error. These values were specific to the tested helical designs, frequencies and fluid conditions. Magnetic actuation remained open loop, and measured trajectories assessed the model rather than updating motion online. The model may inform controller design but is not an online controller demonstration [52].

Belharet et al. presented a review-level MRI navigation architecture with generalised predictive-control simulations under modelled pulsatile flow and noise [53]. Simulated position and path error generated trajectory-correction commands at a 50-ms sampling interval. Two-dimensional root-mean-square errors were 1.0480 and 1.2306 pixels without and with noise, respectively. These simulation results are not directly comparable with the large-animal navigation metrics reported in two original studies [27, 28] because the variables, models and evaluation scales differ. As review-level simulation examples, they do not provide direct in vivo controller performance [53].

Model-level coordination occupies a more indirect position relative to online biomedical control. Lu et al. reviewed magnetic micro/nanorobot swarm design and control, including magnetic, hydrodynamic and boundary interactions [54]. The review identified closed-loop coordination and clinical translation as continuing challenges rather than reporting an original experiment. Kotsuka and Hori analysed stability in a theoretical molecular-communication system containing 100 model agents [55]. At a diffusion coefficient of 10 μm² min^{− 1}, the system converged to a spatially homogeneous equilibrium with a 70-µm communication length. It did not converge when the communication length was 10 μm. The artificial-cell reaction network contained no fabricated swarm, biomedical task or robotic actuation [55]. These studies provide review-level and theoretical context but no biomedical swarm task execution.

Across these studies, models yield predicted trajectories, velocities, stability regimes or population states [51–55]. Such outputs may guide design or parameter selection, but they differ from online state estimates converted into control-relevant inputs during biomedical tasks.

Representative micro/nanorobotic platforms are compared in Table 2 by actuation, feedback variables, control-relevant inputs, human supervision and evidence boundaries.

Current limits of autonomy-supporting capabilities

We use “autonomy-supporting capability” for a bounded task in which a task-relevant state influences action selection or adaptation. The capability must operate under defined human supervision and safety constraints. This differs from automated motion, which executes specified commands, and feedback-guided action, which corrects commands using measured error. Clinically relevant task autonomy also requires an appropriate surgical workflow and model, with explicit interruption and fallback provisions. This analytical boundary supports comparison of the examples and their missing transitions; it is not a universal autonomy taxonomy.

To anchor this descriptor, we use the six-level framework proposed by Yang et al. [56]. It comprises Level 0, no autonomy; Level 1, robot assistance; Level 2, task autonomy; Level 3, conditional autonomy; Level 4, high autonomy; and Level 5, full autonomy. Sensing or operator-display systems do not independently raise the autonomy level, and teleoperation remains at Level 0 unless the robot provides qualifying assistance. Pre-programmed or bounded navigation may exhibit selected task-autonomy characteristics when a human initiates the task, but geometric closed-loop navigation is not tissue-state-driven surgical autonomy. The heterogeneous evidence reviewed does not support assigning high or full clinical autonomy to any nano-enabled system. We therefore retain ‘autonomy-supporting capability’ as an evidence descriptor for component functions that could contribute to a bounded autonomous task without implying that the complete task-autonomy criteria have been met.

The studies reviewed establish selected acquisition-to-state links for tissue class, tactile loading, physiological events, grip force and cellular response [21, 22, 31, 41, 43]. They also establish feedback-guided action for MRI localisation and waypoint navigation and for operator-directed AFM position or force control [27–31]. These groups remain disconnected because no biochemical, spectral, tactile or tissue-state estimate enters an online surgical controller. The airway-biopsy platform combined navigation, sampling, retrieval and post-retrieval classification, but its SERS/CNN output did not modify navigation or sampling [21]. Similarly, MRI motion depended on geometric position, whereas AFM action depended on instrument position or force [27–31].

The studies by Lang et al. and Zhao et al. extend evidence from human-mediated image guidance to automatic geometric control, while remaining outside tissue-state-driven surgery. Lang et al. demonstrate ultrasound-based feedback and adaptive navigation in a tissue-mimicking phantom [19]. Zhao et al. combine physics-constrained monotonic neural networks, reinforcement learning and planning for target reaching and obstacle avoidance in simulation and optical benchtop experiments [20]. In one physical 15-obstacle test, the system reached 23 targets over 20 min without a reported collision. Separately, an approximately 300-µm swarm traversed a 40-mm tortuous channel in 17 min [20]. Human-selected targets and expert demonstrations remained part of the framework. The study included no tissue intervention, biological sensing, therapy or animal evidence.

Maturity ranges from benchtop and in vitro experiments to ex vivo tissue analysis, small-animal navigation and large-animal navigation with surrogate devices [21, 27–31]. Human-tissue classification was ex vivo, whereas the large-animal studies addressed navigation rather than therapeutic action. Across these studies, no complete pathway connects nano-enabled signal acquisition, interpretation and task-state estimation to autonomous surgical action. No study reports a bounded autonomous task with human oversight, fallback behaviour, safety constraints, workflow compatibility and clinically relevant performance.

The synthesis therefore separates demonstrated component links from missing transitions. Local acquisition, selected state estimation, feedback-guided navigation and operator-assisted nanomanipulation have been shown under specific conditions and at different maturity levels. Using a local tissue or system state to alter surgical action within a supervised, safety-bounded workflow remains untested in the studies reviewed. Figure 1 maps the demonstrated and missing information-flow links between nano-enabled signal acquisition, task-relevant state estimation, controller integration and feedback-guided robotic action.

Challenges, translational bottlenecks, and future perspectives

The translational barriers identified in the studies reviewed differ across technology groups. Imaging and sensing interfaces are limited mainly by signal stability, calibration and workflow-compatible readout. Micro/nanorobotic platforms face additional constraints involving actuation, retrieval, material fate and task repeatability. Computational methods require robust and clinically interpretable states, whereas feedback systems require control-relevant inputs that modify action within defined safety boundaries. These distinctions support a target-group analysis before shared clinical and governance issues are considered. Accordingly, this section identifies where each platform currently stops along the information-flow pathway and which experiment could test the next transition.

Imaging-derived localisation and local sensing interfaces

The most mature capability in this group is the acquisition of measurable spatial, spectral or mechanical signals under defined conditions. An external Cu–EGaIn marker provided a CT-visible spatial cue without robotic integration. By contrast, the biopsy microrobot combined image-guided access with post-retrieval SERS analysis and tissue classification [21, 37]. Flexible and local interfaces also demonstrated benchtop tactile classification, microforce measurement, self-powered physiological monitoring and image-based force estimation [22, 41–43]. These studies establish signal acquisition and selected interpretation, but not a common pathway to robotic control. Their performance metrics are not directly comparable because the measurands, readout mechanisms and experimental models differ. Evidence of a measurable output therefore cannot substitute for validation of the state that the output is intended to represent.

The principal bottlenecks are signal reproducibility and the stability of the inferred state across changing conditions. The tactile sensor classified multichannel loading patterns but was tested without tissue or a surgical robot [41]. The micro-spring force sensor remained a benchtop platform, and its machine-learning model predicted fabrication printability rather than interpreting force [42]. The wearable interface [43] responded rapidly, but its physiological demonstrations were outside surgery. Conversely, the approximately 4-s recovery of the vision-based sensor [22] limits its relevance to rapidly changing control. The stretchable marker [37] provides an external radiopaque cue, not lesion localisation or robotic navigation. The post-retrieval classifier likewise remains outside microrobot navigation and sampling control [21]. Across these platforms, tissue variability, repeated deformation and imaging conditions could alter the meaning of a calibrated signal before it reaches a controller.

Review-level analyses reinforce that imaging compatibility is a system constraint rather than an isolated detector specification. Mobile-microrobot translation requires alignment of penetration depth, spatial and temporal resolution, field of view, clinical anatomy, actuation and control [16, 18]. The complete system also requires reliable localisation, deployment and recovery from visibility loss. New imaging studies add Doppler, speckle-contrast and ultrasound-visible microrobot evidence [36, 38, 39]. Their distinct experimental models and supervision paths prevent direct performance ranking.

A useful next step would be to integrate one validated sensor with a supervised robotic task in an anatomically relevant model. Validation could combine repeated deployment and anatomical motion with concise reporting of calibration drift, latency and uncertainty. A defined failure threshold could prevent an unstable estimate from entering control. The inferred state could then inform one bounded action under operator oversight. This design would test the unresolved state-to-control transition without treating improved sensing as evidence of autonomy.

Micro/nanorobotic platforms and localised intervention

Localised access, movement and manipulation are more mature than sensing-guided task adaptation. The reviewed examples include supervised airway biopsy, optically assembled biohybrid motility, externally triggered transformation in a simulated vessel and operator-directed AFM manipulation of living cells [21, 29, 30, 34, 35]. In this evidence set, the biopsy platform performed tissue classification only after retrieval [21]. One platform demonstrated assembly and motility without reconfiguration or intervention [34]. A separate system used external commands and magnetothermal heating to switch task mode in a simulated vessel [35]. The AFM platforms remained operator-directed in vitro systems using instrument position or force feedback [29, 30]. These evidence stages differ in biological realism and task scope, so movement or transformation alone cannot indicate surgical readiness.

Manufacturing and materials evidence remains similarly specific to each platform. Wu et al. [32] reported an oil-free route for controllable alginate/gelatin methacryloyl (GelMA) microparticle production, short in vitro cell-viability testing and magnetically driven movement on an ex vivo patella. Greater Fe₃O₄ loading improved magnetic response but could affect fabrication stability and biocompatibility, while the apparatus limited motion range [32]. This evidence applies to one fabrication route and does not establish systemic safety, sterilisation compatibility or reproducibility for other platforms. A review identifies immunogenicity, biodegradation, biodistribution, retrieval and cross-laboratory reproducibility as connected translational concerns for micro/nanorobotic drug delivery [57]. These concerns depend on route, material composition and intended residence time. They should therefore be evaluated at whole-platform level rather than extrapolated from one component or short compatibility test.

The adjacent roadmaps similarly connect manufacturing robustness to actuation scale, delivery route, imaging compatibility, retrieval and material fate [16–18]. High-throughput production and batch consistency are therefore insufficient without route-specific deployment and contingency plans. For temporary platforms, retrieval and biodegradation represent alternative strategies with different risks: degradation products require characterisation, whereas retained or retrievable devices require reliable localisation and recovery [16, 18].

A more informative translational milestone would be reproducible completion of a bounded task under anatomically realistic disturbances. Future validation should prioritise batch consistency, realistic flow, task repeatability, material fate and recovery from localisation or actuation failure. Sterilisation, storage and deployment could be incorporated as platform-specific variables rather than generic checklist items. These tests could combine route-specific safety measures with defined supervision and safe termination. Localised actuation is the most mature capability in this group. Reliable inference of task success and its conversion into corrective action remain the missing information-flow transition.

Computational interpretation and task-relevant state estimation

Computational interpretation is heterogeneous and should not be treated as synonymous with AI. Direct calibration and conventional signal processing produced physiological or force estimates [22, 43]. Conventional machine learning converted capacitive patterns into tactile loading states [41]. Deep learning separately classified post-retrieval SERS spectra [21]. By contrast, the SVM [42] addressed fabrication printability rather than the sensor output. Physical and computational models identified swimming regimes or predicted helical locomotion in two original studies [51, 52]. AFM analysis generated cellular stiffness and Ca²⁺ states in one study [31]. Each method therefore requires evaluation against the task state it claims to estimate, not against a generic category of computational sophistication.

The broader technology roadmap by Ju et al. places conventional AI alongside onboard and off-board intelligence, physical intelligence and collective behaviour [17]. This taxonomy provides useful context, but it does not establish that a reviewed platform has sensed a biological state or used that state online. For the present review, the decisive evidence remains the demonstrated transition from a measured signal to a validated task state and, separately, from that state to action.

The strongest evidence concerns offline or benchtop interpretation under source-specific calibration. Evidence supporting generalisation and online use remains considerably less developed. The classification studies do not establish performance across independent institutions, prospective patient samples or changing acquisition hardware [21, 41]. The vision-based sensor [22] has a recovery time that limits rapid control relevance. The model outputs in two original modelling studies [51, 52] may inform parameter selection but were not used for online correction. In the AFM study [31], Ca²⁺ and stiffness remained monitored outputs outside the AFM controller. A displayed state is also not automatically a decision variable or control-relevant input. These boundaries make uncertainty, calibration transfer, domain shift and synchronisation central to evaluating any task-relevant state.

One testable design would validate a task-relevant state prospectively using independent data and a prespecified uncertainty measure. It could compare raw-signal display, interpreted-state display and one narrowly bounded robotic adaptation under supervision. Reporting latency, calibration failure and responses to uncertain inputs would clarify whether interpretation adds task value. This comparison would also reveal whether errors are detectable before they influence action. The unresolved transition is the validated use of an inferred state in action selection, rather than further improvement in standalone classification performance.

Feedback-guided control and autonomy-supporting capabilities

Geometric and instrument-level feedback provide the clearest current control evidence. MRI-derived position and path error drove gradient commands for surrogate devices in a living swine carotid artery [27, 28]. Human path planning and oversight remained outside these real-time loops. AFM systems used probe position or interaction force during operator-directed cellular manipulation [29–31]. These studies demonstrate feedback-guided action, but their control variables were geometric or mechanical rather than biochemical, spectral or tissue-state estimates. Geometric correction can improve navigation without establishing that the system recognises task success or tissue condition. These examples therefore represent bounded control components rather than clinically relevant task autonomy.

The studies by Lang et al. and Zhao et al. sharpen the boundary within this control evidence. Lang et al. close a geometric ultrasound-to-controller loop in a tissue-mimicking phantom, whereas Zhao et al. constrain learned magnetic navigation in simulation and optical benchtop environments [19, 20]. Both go beyond displayed state estimates, yet neither detects tissue condition, selects a therapeutic response or completes a supervised surgical task.

Several related sources remain one step further from online control. One original study [51] identified a dynamic stability regime without online correction, while another original study [52] validated locomotion modelling under open-loop actuation. One source [53] presents review-level simulation examples, another source [54] is a swarm-control review, and a separate source [55] presents theoretical non-biomedical stability analysis [51–55]. These sources may inform controller design but do not demonstrate autonomous biomedical task execution. Ren et al. [33] extend the available control evidence through image processing, path planning and PID feedback for magnetic lipiodol microdroplets. Their closed-loop tests used two-dimensional mazes, a printed vascular network and digital subtraction angiography (DSA) vascular images, with operator-specified start and end points [33]. The authors also identified stronger actuation and three-dimensional validation as remaining needs. Separate rabbit experiments demonstrated DSA-visible magnet-guided motion, not in vivo AI-controlled navigation [33]. The in vivo and closed-loop results therefore represent complementary but unconnected evidence blocks.

A useful next step would be to challenge a bounded controller in an anatomically relevant moving environment while retaining explicit human authority. One focused evaluation could combine localisation uncertainty, signal loss and physiological motion with predefined takeover, fallback and safe-stop responses. Control decisions could remain traceable to the measured state and update rule. Within the studies reviewed, geometric navigation and instrument-level feedback represent the most mature autonomy-supporting capabilities. Tissue-state-driven adaptation with validated intervention limits, human takeover and safe termination remains undemonstrated.

Cross-cutting clinical translation, governance, and accessibility

Translation also depends on requirements shared across the four technical groups. Review-level Internet-of-Robotic-Things and digital-twin frameworks provide conceptual context for bidirectional data integration across surgical systems, but do not demonstrate nano-enabled online control or autonomous surgery [58]. The operating room is a high-risk, multidisciplinary environment in which data movement, equipment arrangement, communication and team preparation affect safe performance. A review emphasises integrated data flow, ergonomic workflow design, simulation and team training rather than adoption of an isolated device [59]. Accordingly, integration requires clear definitions of hardware placement, information display, calibration responsibility, interruption routes and failure communication. The exact arrangement will remain platform specific, but operator roles and rehearsable workflows require definition before clinical testing.

The clinical-translation and readiness reviews make these dependencies explicit at whole-system level [16, 18]. Development should begin from an unmet clinical need and evaluate the robot body together with reproducible manufacturing, quality-management practices, sterilisation, human-scale actuation, imaging compatibility, deployment, retrieval and workflow integration. Ceylan et al. further frame control authority as risk dependent. The interventionalist retains live monitoring and immediate override, while failure management addresses visibility loss, uncontrolled migration, inadequate actuation margin and recovery pathways [18]. Clinical value therefore depends on whether the integrated system improves a defined task relative to current care under traceable supervision and contingency rules. Patient-centred clinical outcomes are also necessary when judging clinical value. In a Chinese cohort of 347 men, open, laparoscopic and robot-assisted radical prostatectomy showed broadly similar 12-month urinary-continence and erectile-function outcomes [60]. This comparison does not establish equivalence across all outcomes or superiority of any approach. It illustrates why emerging robotic capabilities require prespecified functional and patient-reported endpoints, rather than an assumption that technological complexity confers clinical benefit.

Manufacturing quality and reproducibility, batch consistency, scale-up, sterilisation, storage, maintenance, long-term reliability and clinical evidence depend on intended use. A review highlights standardised methods, independent reproducibility, material compatibility and retrieval or fate data for micro/nanorobotic drug-delivery systems [57]. A China-specific regulatory-science analysis adds lifecycle responsibility, traceability and evidence-based evaluation [61].

Regulatory, ethical and legal requirements also depend on the level of robotic autonomy and the clinical context in which the system is used. Yang et al. proposed that increasing autonomy changes the allocation of human oversight, responsibility and risk [56]. Their framework therefore requires regulatory assessment to consider not only technical performance but also the intended task, operating environment and retained human authority. For nano-enabled robotic systems, these considerations are further complicated by material safety, software validation, retrieval or degradation, data governance and failure management. A single universal regulatory pathway therefore cannot be inferred from the current evidence.

Human oversight and responsibility should be defined before control authority is increased. A review argues that sensitive robotic tasks should remain supervised and that responsibility cannot be displaced onto an autonomous system [57]. Infrastructure and accessibility are best evaluated within a defined use case. For colonoscopy microdevices, readiness included safety, biopsy capability, usability, training, staffing, facility needs and cost [62]. This example is specific to colonoscopy and does not establish field-wide access effects. It nevertheless shows that technical maturity alone cannot determine clinical value or deployment feasibility.

An evidence-based roadmap towards clinically relevant autonomy

The near-term stage comprises bounded sensing, localisation, sampling, monitoring and intervention modules within supervised workflows. Current examples include image-guided sampling with post-retrieval interpretation, external imaging cues, operator-directed nanomanipulation and geometric feedback navigation [21, 27–30, 37]. A separate study adds a benchtop closed-loop navigation architecture and a separate in vivo imaging-guided motion demonstration [33]. These capabilities should remain distinct until validated within the same task and model. Progress at this stage is best judged by repeatability, safety, workflow compatibility and recovery from failure. Combining functions is informative only when the links between them are also demonstrated.

At an intermediate stage, a validated task-relevant state could inform a human decision or a narrowly bounded robotic adaptation. Advancement depends on independent or prospective validation, uncertainty reporting, latency characterisation and prospective safety testing. The decisive evidence would show that the state improves a prespecified task outcome without allowing erroneous or missing inputs to produce uncontrolled action. This stage remains decision-supporting or adaptive rather than autonomous unless control authority and task boundaries are explicitly defined.

These capability stages should not be conflated with adjacent roadmap frameworks. The broad technology roadmap by Ju et al. organises advances in propulsion, materials, intelligence, applications and scale-up [17]. Bozuyuk et al. emphasise the platform barriers to mobile-microrobot clinical translation [16]. The mTRL framework by Ceylan et al. stages whole-system readiness and clinical value [18]. The present information-flow pathway instead asks whether each signal-to-state and state-to-action link is demonstrated. Technology breadth, translational readiness and information-flow completeness are complementary but non-equivalent dimensions.

The long-term stage is bounded task autonomy in which a validated state influences action under explicit supervision, intervention limits, fallback, safe-stop and governance provisions. The clinical nanorobot remains a conceptual goal rather than a demonstrated surgical system [63]. None of the studies reviewed establishes the complete pathway from nano-enabled acquisition through validated state estimation to autonomous surgical action under explicit safety constraints. Progression between these capability stages is neither automatic nor inevitable. It depends on evidence for each information-flow transition, not on technological complexity or the accumulation of isolated functions. Figure 2 summarises the platform-specific translational barriers and the evidence milestones required to progress from supervised functional modules towards bounded task autonomy.

Conclusion

This review used information flow, rather than device class alone, to assess how nano-enabled technologies may contribute to robot-assisted intervention. Localised signal acquisition is comparatively well represented across imaging, spectral, flexible-sensor and micro/nanorobotic platforms. Selected processing and computational interpretation have converted some signals into loading classes, force estimates, tissue classifications or geometric states. Geometric states derived from localisation and motion have also entered feedback controllers in selected phantom, surrogate-device and instrument-level systems. These demonstrations remain heterogeneous in experimental model and evidence stage. Task-relevant tissue states are less consistently validated, and biochemical, spectral, tactile and physiological estimates rarely modify an online surgical controller. The studies reviewed therefore do not establish a continuous pathway from nano-enabled sensing to clinically relevant surgical autonomy.

Geometric feedback, automated motion and navigation autonomy should remain distinct from tissue-state-driven surgical control. A system may correct position or follow a planned route without recognising tissue condition, selecting a therapeutic response or completing a bounded surgical task. Near-term translation is more plausibly based on nano-enabled sensing, localisation or intervention modules embedded within supervised workflows. Advancement towards bounded task autonomy requires predefined task boundaries, uncertainty handling and traceable state-to-action rules. Human supervision should include explicit fallback, takeover and safe-stop provisions. Evidence should also address workflow compatibility, material safety, reproducible manufacturing and prospective validation under anatomically realistic conditions. Patient-centred outcomes and prespecified task benefits should guide clinical evaluation. These requirements provide a more appropriate basis for judging autonomy-supporting capability than technological complexity or isolated component performance.

Author contributions

J.Z., S.T., and Z.Z. contributed equally to this work. J.Z., S.T., and Z.Z. conducted the literature review, screened the relevant studies, organised the manuscript framework, drafted the manuscript, and contributed to figure preparation. R.-C.W., D.-X.L., and X.J. supervised the project, provided conceptual guidance, and critically revised the manuscript for important intellectual content. All authors reviewed and approved the final manuscript.

Funding

This work was supported by the Zhejiang Provincial Traditional Chinese Medicine Service Capacity Building Project (Grant No. 2A12510); the Health Innovation Talent Program / Modernization Capacity Enhancement for High-Quality Development of Public Hospitals (Grant No. 1S22513); the Clinical efficacy of a Chinese patent medicine combined with L-carnitine in the treatment of prostatitis complicated by asthenospermia (Grant No. 2B12442); and the Clinical Research Program of Traditional Chinese Medicine of the Zhejiang Provincial Administration of Traditional Chinese Medicine (Grant No. 2025ZL262). The funders had no role in the study design, literature collection, analysis, interpretation, manuscript preparation, or decision to submit the article for publication.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.