Visual perception technologies for micro- and nanorobots: Principles, methods, and applications (<italic>invited</italic>)

Significance Micro- and nanorobots have emerged as promising tools for precise manipulation, biomedical intervention, and micro/nano manufacturing in complex confined environments. Their capability to access narrow spaces and perform delicate operations offers new technological paradigms for targeted drug delivery, minimally invasive diagnosis and therapy, single-cell manipulation, and high-precision microassembly. However, the practical deployment of such robots relies fundamentally on reliable perception. Unlike macroscopic robotic systems, micro- and nanorobots operate in low-Reynolds-number environments and are highly susceptible to Brownian motion, background flow disturbances, imaging noise, and environmental uncertainty. Without real-time feedback on robot position, orientation, and surrounding context, open-loop control cannot support accurate navigation, stable manipulation, or autonomous decision-making. Therefore, visual perception has become a key enabling component for closed-loop control and intelligent operation of micro- and nanorobotic systems. Yet this capability remains difficult to realize because in vivo applications are constrained by tissue opacity, scattering, attenuation, and biosafety requirements, whereas in vitro operations are limited by optical diffraction, shallow depth of field, low contrast, and the demand for high-speed sensing and computation.Progress Recent advances in visual perception for micro- and nanorobots can be summarized in two representative scenarios, namely in vivo navigation and in vitro micromanipulation, along three major trends: the evolution of imaging modalities, the deepening of perception dimensions, and the shift from model-driven to data-driven algorithms (Fig.1). For in vivo navigation, imaging methods have progressed from radiation-based X-ray imaging to safer multimodal strategies such as ultrasound, photoacoustic imaging, and magnetic resonance imaging (MRI) (Fig.2). These developments have improved the balance among penetration depth, spatiotemporal resolution, and biosafety, while multimodal fusion has further enabled joint perception of microrobots and anatomical environments (Fig.3). In parallel, perception capability has advanced from coarse localization to six-degree-of-freedom (6-DoF) pose estimation, environmental context fusion, and simultaneous localization and mapping in complex endoscopic scenes. Correspondingly, algorithms have evolved from geometric and physics-based modeling to machine learning and deep learning, which offer stronger robustness against noise, artifacts, and nonlinear imaging distortions (Fig.4). For in vitro micromanipulation, research has focused on overcoming depth ambiguity, diffraction blur, shallow depth of field, and the transparency of micro-objects. Perception tasks have expanded from two-dimensional tracking to three-dimensional positioning, 6-DoF pose estimation, and dense morphology reconstruction, supported by stereo microscopy, depth-from-defocus analysis, and learning-based depth inference (Fig.5). The perception targets have also extended from single rigid robots to multiple robots, biological samples, and microbial swarms, driving the development of deep detection and multi-object tracking methods (Fig.6). More recently, simulation-to-real transfer, physics-informed learning, and benchmark datasets have promoted system-level integration, real-time closed-loop control, and standardized evaluation (Figs.7-9).Conclusions and Prospects Visual perception for micro- and nanorobots is undergoing a clear transformation from single-modality, low-dimensional, and model-dominated methods to multimodal, high-dimensional, and intelligence-driven systems. Significant progress has been achieved in biosafe imaging, deep-tissue tracking, pose estimation, morphology reconstruction, multi-target perception, and closed-loop system integration. Nevertheless, several challenges remain unresolved, including the scarcity of cross-scale annotated data, the instability caused by severe imaging artifacts in extreme environments, the difficulty of real-time fusion across multiple physical fields, and the lack of universally accepted benchmarks for fair comparison. Future development is expected to focus on multimodal intelligent fusion, foundation-model-enabled perception, high-throughput collaborative sensing for multi-robot and biohybrid systems, and tighter integration of perception, decision-making, and control. These advances are essential for translating micro- and nanorobotic technologies from laboratory demonstrations to practical deployment in clinical medicine and advanced manufacturing.