Multi-target positioning and motion tracking enabled by a compound meta-eye system Authors: Li Zhang, Lingyun Zhang, Xiaoguang Zhao, Jin Li, Hongjing Cao, Jianing Yang, Chi Zhang, Zilun Wang, Yuhang Li, Haiyan Shi, Zhe Zhang, Junbin Li, Hanwen Guo, Fei Xing, Zichen Zhang Published: 2026 ABSTRACT AbstractAccurate perception of spatial position is essential for both biological vision and intelligent unmanned systems. Existing radio-based positioning approaches are susceptible to interference and require bulky infrastructures, while optical systems often trade accuracy for compactness. Here, we present a compound meta-eye system (CMES) that integrates an array of metalens sub-eyes to capture angular parallax from multiple targets simultaneously. Each sub-eye focuses light onto a detector to form arrayed images, from which a global-ratio algorithm reconstructs spatial coordinates with high precision. The CMES enables multi-target positioning and motion tracking within a meter-scale range, achieving a relative depth error below 2% and trajectory-fitting deviations under 0.5 mm. The metalens design provides diffraction-limited focusing and wide angular tolerance, combining biological compound eye compactness with flat meta-optics. This compact optical-perception platform offers an efficient solution for real-time multi-target spatial perception, with potential applications in formation control, visual navigation, and environmental perception for unmanned aerial vehicles and embodied intelligent agents. Subject terms: Optical sensors, Applied optics Introduction Perception of spatial position is a fundamental function of biological vision1, enabling organisms to navigate, hunt and avoid predators. Nowadays, unmanned systems and embodied intelligent agents likewise require the ability to perceive and determine the locations of targets2–4. Typical examples include the formation control and relative navigation of low-altitude drones5,6, as well as the perception of surrounding obstacles by autonomous vehicles and service robots. Existing spatial-positioning techniques can be broadly categorized into radio-based and optical-based approaches. Radio-based systems such as the Global Navigation Satellite System (GNSS), including China’s Beidou satellite positioning system and the US Global Positioning System, determine target positions via the virtual intersection of propagation spheres between multiple satellites7,8. Although GNSS achieves global coverage, it suffers from multipath interference in urban environments and significant signal attenuation indoors, leading to degraded accuracy9. On smaller scales, ultra-wideband (UWB) positioning determines distances from the time-of-flight of narrow pulses between base stations and receivers, achieving centimeter-level precision but requiring a dense infrastructure layout10,11. Optical and vision-based positioning methods offer higher spatial resolution but face fundamental geometric constraints. A conventional single-aperture camera loses depth information during three-dimensional (3D) to two-dimensional (2D) projection, and depth recovery based on prior features such as intensity or shape typically requires complex computation12,13. Binocular stereo vision estimates target depth from parallax, analogous to how mammals and birds perceive spatial position14. However, its accuracy strongly depends on the baseline length—high precision demands a large baseline, inevitably increasing the system volume15. The compound eyes of insects are composed of numerous ommatidia, each observing from a distinct angle, naturally collecting parallax information within a miniature structure16,17. This biological architecture has inspired the development of highly compact target-positioning systems. In addition to systems that mimic binocular vision using a pair of artificial eyes18,19, several approaches based on the intersection of beam vectors from multiple sub-eyes have also been proposed20–23, where target coordinates are typically estimated via least-squares fitting. Other methods exploit the image morphology variations24–29 or optical flow techniques30–32. However, these systems share a common limitation: their short focal length and small baseline constrain spatial perception to short distances, preventing accurate positioning of distant targets. Our previous work demonstrated a planar compound eye optical system employing a ratio-based positioning method33. By averaging the imaging results from multiple sub-eyes, random errors were reduced, and the measurable range of a 5 mm baseline system was extended to approximately 10 m. Nevertheless, the sub-eyes in that design were merely micron-scale transmitting pinholes, limiting operation to a single bright target. In this paper, we develop a compound meta-eye system (CMES) that replaces each ommatidium of a biological compound eye with a metalens sub-eye. The CMES focuses light from distant targets to form an array image and reads out parallax information for multi-target positioning. Experimental results verify that the CMES simultaneously performs spatial positioning and trajectory tracking of multiple non-overlapping targets within a 1 m range, achieving a relative depth-error below 2% and a tracking standard deviation under 0.5 mm. This compact and precise optical-perception platform shows strong potential for applications in formation control, visual navigation, and obstacle avoidance of unmanned aerial vehicles (UAVs), intelligent robots, etc. Results Working principles of the CMES The concept of the proposed CMES is illustrated in Fig. 1. The CMES can be mounted on a UAV or other mobile platforms to perceive the spatial positions of cooperative targets (Fig. 1a). Each target carries a light source or reflector that emits near-infrared radiation toward the CMES. The n × n metalens sub-eyes collect light from slightly different viewing angles, and their corresponding images are formed on the detector as multiple distinct spots for one target (Fig. 1b). The parallax among these spots encodes the target’s spatial location relative to the CMES, enabling accurate spatial positioning and motion tracking within a compact optical structure (Fig. 1d). Working principles of the compound meta-eye system (CMES). a Conceptual illustration of the CMES mounted on unmanned aerial vehicles (UAVs) for cooperative target perception. b Schematic of parallax formation: one target produces multiple focal spots through the metalens sub-eyes. c Local magnification of a sub-eye, showing nanoscale amorphous-silicon pillars for phase modulation. d Spatial positioning and motion-tracking process reconstructed from the parallax data Taking the center of the detector as the origin, the row and column directions are defined as the y and x directions, respectively, while the z axis is perpendicular to the detector plane. The vertical separation between the sub-eye plane and detector is fixed to the sub-eye focal length f. For a single target with spatial coordinates P = (xp, yp, zp), any pair of sub-eyes together with their corresponding imaging spots form similar triangles. The difference in the incident angles of light rays from these sub-eyes represents the parallax. The ratio R, defined as the distance between each pair of imaging spot centroids divided by the spacing between the corresponding sub-eye centers33, quantitatively expresses this parallax. The spatial coordinates of the target can therefore be determined using Eq. (1): where N (= n × n) is the number of imaging spots collected in a single measurement, xsi and ysi (i∈1, …, N) are the centroid coordinates of the collected spots, xai and yai (i∈1, …, N) are the coordinates of the corresponding sub-eye centers. To make full use of the available parallax information and to suppress random measurement noise, a global-ratio algorithm is introduced (see Fig. S1 for comparison with the previous methods). Specifically, a temporary ratio R′ is calculated for each possible pair of imaging spots and the corresponding sub-eye spacing. The global ratio R is obtained by averaging all R′ values across the collected dataset. When a target forms N imaging spots after passing through N sub-eyes, the global ratio R can be calculated by Eq. (2). Substituting the calculated R into Eq. (1) yields the refined spatial coordinates of the target. When multiple targets (m in total) are simultaneously present within the field of view (FOV), the detector records m × N imaging spots. Provided that these spots do not overlap, the N sub-eye images associated with each target can be separated according to the known geometric arrangement of the sub-eyes. Starting from the imaging spot at the top-left corner of the detector, the algorithm sequentially searches for neighboring spots following the sub-eye spacing pattern until the complete array corresponding to one target is identified. This process is repeated iteratively to extract the imaging arrays of all targets. The target 3D positioning algorithm based on the global-ratio processes one image each time, and its computational complexity is O(N), where N is the number of photosensitive pixels of the detector. Compared with the traditional binocular vision depth estimation algorithm, the order of magnitude of the computational complexity is comparable. Moreover, the number of images collected and the number of feature points processed in each measurement are less than those of the binocular vision depth estimation algorithm, so theoretically the processing speed is faster. Small differences in imaging spacing caused by parallax variations may influence the grouping and identification process; this effect and its mitigation are discussed in Supplementary Note 1. Once the spot arrays corresponding to each target are correctly identified, their 3D coordinates can be independently computed based on the multi-sub-eye parallax model, thereby enabling simultaneous spatial positioning and dynamic motion tracking of multiple targets within a compact and efficient optical framework. Design of the metalens sub-eyes The implementation of the CMES primarily relies on the optical design and fabrication of the compound meta-eye (Fig. 2), in which each metalens sub-eye is responsible for converging light emitted or reflected from a target onto the detector plane. By modulating the incident wavefront, each sub-eye encodes the target’s spatial position information into its focused image. To accommodate different incident angles, the phase profile34 of each metalens sub-eye follows Eq. (3): where k0 is the free-space wave number (k0 = 2π/λ), and r denotes the radial distance from any point on the sub-eye surface to its center. In this design, the focal length f is set to 7 mm to achieve a high angular measurement resolution and enhance parallax sensitivity. Design and fabrication of the metalens sub-eyes. a Simulated amplitude transmission of nanopillars and a magnified view at the design wavelength of 940 nm. b Simulated phase-modulation capability of nanopillars. The red points in (a) and (b) are the selected D-size libraries. c Photograph of the fabricated 3 × 3 metalens-array CMES. d Scanning-electron-microscope (SEM) image of the fabricated sub-eyes The designed phase profile is implemented by subwavelength amorphous silicon nanopillars standing on the fused silica substrate (Fig. 1c). The symmetry of cylinder unit cells guarantees polarization insensitivity. For the working wavelength of 940 nm, the nanopillars are arranged in a regular hexagonal lattice with a period of 450 nm and a height (H) of 600 nm. By varying the diameters (D) of the nanopillars, the local phase shift of the transmitted light can be precisely controlled while the amplitude transmission keeps close to unity, demonstrated by finite-different time-domain simulation (Fig. 2a, b). As a proof-of-concept, we fabricated 3 × 3 sub-eyes with a square shape of 1.6 × 1.6 mm for each and a center-to-center separation of 1.8 mm (Fig. 2c, d). In order to reduce the influence of stray light in the environment, a chromium aperture is deposited on the other side of the substrate to suppress environmental stray light, ensuring that only light within the designated FOV reaches the corresponding sub-eye while blocking unwanted light from adjacent FOVs. The fabrication of the metalens sub-eyes is detailed in the “Materials and methods” section, and the design of the aperture size and position is illustrated in Supplementary Note 2. A collimated light beam at a wavelength of 940 nm was directed vertically onto each sub-eye to evaluate its focusing performance. The imaging performance testing system is described in the “Materials and methods”. The focal spot was observed through a magnifying imaging system, as illustrated in Fig. S6, and the measured intensity profiles are shown in Fig. S7. The experimentally obtained focal size approaches the diffraction limit of approximately 10 μm. As the incident angle varied within ±15°, the sub-eyes maintained clear and well-defined focal spots, confirming high-quality imaging across the designed FOV and minimal parallax distortion. These results demonstrate that the metalens sub-eyes achieve robust focusing and angular tolerance, thereby providing a reliable optical foundation for the subsequent multi-target positioning and tracking experiments. Performance characterization of the CMES Combining the metalens sub-eyes array with a flat detector forms an optical sensing microsystem, namely CMES. Compared with traditional binocular vision devices, the size of CMES detection end is extremely compact and tiny. To evaluate the positioning and motion-tracking capability of the proposed CMES, an experimental platform was constructed as shown in Fig. 3. The CMES sample was mounted vertically on an optical plate holder, with its optical axis perpendicular to the table surface. A near-infrared (NIR) detector (ME2P-2622-15U3M, Daheng Imaging) featuring a 5120 × 5120 pixel array and a pixel pitch of 2.5 μm was employed for image acquisition. The detector operates in global-shutter mode, thereby avoiding motion-induced distortions and parallax inconsistencies that commonly occur in rolling-shutter sensors. The detector was coupled to a 3D motorized translation stage (GCM-901602M, Daheng Optics) to allow precise alignment of its center with the CMES optical axis and fine adjustment of the working distance between the metalens array and the image plane. Single-target depth-perception experiment. a Experimental setup for depth measurement using a 940 nm LED mounted on a translation guide rail. b Arrayed imaging of the single target at different depths. c Calculated spatial coordinates for all measurement points. d Reconstructed depth values for multiple detector frames and magnified view at z = 800 mm, showing <2% relative error across 1 m range Single-target depth perception experiment A 940 nm LED bead was used as a cooperative target to simulate the NIR beacon mounted on a UAV. The LED was fixed to a linear translation stage to vary its position along the optical axis, with the rail’s displacement serving as the ground-truth reference (Fig. 3a). A 940 ± 12.5 nm narrow-bandpass filter was placed in front of the CMES to suppress ambient stray light (see Supplementary Note 4 for the analysis of system robustness). The array images of the single target captured by the CMES are displayed in Fig. 3b. The spatial position of the target was reconstructed using Eqs. (1) and (2), and the results were compared with reference displacements to evaluate depth-perception accuracy. The target was sequentially positioned at 200, 400, 600, 800, and 1000 mm, with 100 frames recorded at each position. The reconstructed spatial coordinates and the corresponding depth estimations are summarized in Fig. 3c, d. At the reference distance of 200 mm, the retrieved depth averaged 201.22 mm with a standard deviation (σ) of 0.06 mm. At 400 mm, the mean was 399.42 mm (σ = 0.20 mm); at 600 mm, 589.72 mm (σ = 0.54 mm); at 800 mm, 796.13 mm (σ = 1.43 mm); and at 1000 mm, 1020.30 mm (σ = 2.42 mm). The deviation and variance increase with distance, primarily due to the reduced parallax sensitivity at longer ranges. Within the measured 1 m working range, the maximum relative depth error was below 2%, validating the system’s high-precision spatial positioning performance. Although the experimental setup limited measurements to 1 m, the CMES conceptually supports longer working distances with a trade-off between precision and range (See Supplementary Note 6 for the measurement and analysis of longer distances). Multi-target motion tracking experiment The CMES also exhibits moderate sensitivity in the visible band, enabling the use of a programmable display as a multi-target simulator to emulate UAV patterns and dynamic trajectories. Multiple luminous targets on the screen can be independently controlled to move along arbitrary paths, as illustrated in Fig. 4a. The corresponding array images acquired by the CMES are shown in Fig. 4b. The display was positioned 1000 mm from the CMES, with its screen plane parallel to the metalens array, serving as the reference depth plane. Calibration and multi-target tracking setup. a Experimental configuration for dynamic trajectory tracking using a programmable display as a multi-target simulator. b Example of four-target array images captured by the CMES. c Displayed calibration pattern for spatial correction. d Corresponding array images recorded on the detector. e Calibration matrix of global-ratio correction ΔR(x, y) used to compensate alignment errors In practical implementation, small mechanical misalignments between the CMES and the detector, such as lateral offsets, tilt, or rotation—introduce systematic errors, especially for off-axis targets. These distortions are negligible in the near-axis single-target tests but become non-trivial for wide-FOV tracking. To compensate for these effects, a spatial calibration procedure was developed using a 5 × 5 array of calibration points displayed on the target simulator (Fig. 4c). Each calibration point produced a distinct 3 × 3 array of spots on the detector (Fig. 4d). Using Eqs. (1) and (2), the initial global ratio R and corresponding spatial coordinates were computed for all calibration points. The deviations between the measured and reference R values were then mapped into a correction matrix (Fig. 4e), expressed as a function ΔR(x, y) of the calculated target coordinates. For new measurements, R was corrected via interpolation of ΔR(x, y), significantly improving positioning accuracy across the FOV. After calibration, dynamic multi-target tracking tests were performed, as shown in Fig. 5. For each captured frame, the spatial positions of all visible targets were computed, and their temporal associations were established based on motion continuity between consecutive frames. The system has the ability of real-time processing and tracking. The analysis of its processing time is shown in Supplementary Note 5. Experimental results of multi-target motion tracking. a–c Lateral motion of four targets: schematic, reconstructed 3D trajectories, and Oxy-plane projection. d–f Longitudinal motion of three targets: schematic, reconstructed 3D trajectories, and Oxy-plane projection. g–i Rotational motion of four targets along circular paths: schematic, reconstructed 3D trajectories, and Oxy-plane projection. j–l Arbitrary trajectories shaped as letters “BH”: schematic, reconstructed 3D trajectories, and Oxy-plane projection. All reconstructed trajectories exhibit <2% relative deviation and <0.5 mm standard deviation of fitting residuals Four targets were programmed to move laterally along straight trajectories—two along the +x direction and two along the –x direction over a 100 mm displacement (Fig. 5a). The reconstructed trajectories (Fig. 5b, c) yielded path lengths of 99.96, 99.77, 99.21, and 99.77 mm, with a maximum deviation of 0.79 mm from the reference true value. Linear fitting of the motion paths produced residual standard deviations of 0.17, 0.20, 0.20, and 0.18 mm, respectively. In a second experiment, three targets moved longitudinally along the ±y axis by 100 mm (Fig. 5d). The retrieved path lengths were 98.07, 99.36, and 99.26 mm, with a maximum deviation of 1.93 mm, and residual standard deviations of 0.36, 0.33, and 0.29 mm (Fig. 5e, f). To verify rotational tracking, four targets were programmed to move clockwise along circular trajectories with a real-world diameter of 100 mm (Fig. 5g). The fitted diameters of the reconstructed paths were 101.82, 101.68, 101.76, and 101.81 mm, corresponding to a maximum deviation of 1.82 mm. The root-mean-square (RMS) radial errors relative to the fitted circles were 0.41, 0.38, 0.41, and 0.40 mm, respectively (Fig. 5h, i). Finally, two targets were driven to trace arbitrary alphanumeric trajectories (“BH”) on the display (Fig. 5j). The reconstructed 3D and projected 2D trajectories (Fig. 5k, l) matched the designed paths closely, demonstrating the CMES’s capacity for simultaneous multi-target trajectory reconstruction with high spatial fidelity. Across all dynamic experiments, the maximum relative deviation between reconstructed and true trajectories remained below 2%, while the standard deviation of trajectory-fitting errors was consistently under 0.5 mm. These results confirm the CMES as a compact, high-precision optical perception system capable of real-time multi-target positioning and tracking within a meter-scale workspace. Detailed video demonstrations of the tracking experiments are provided in Supplementary Movie 1. Discussion In summary, we have developed a compound meta-eye system (CMES) capable of achieving simultaneous multi-target spatial positioning and motion tracking within a compact optical architecture. Each metalens sub-eye focuses incident light from the target to the detector plane, forming a 3 × 3 array of imaging spots for a single target. For multiple targets within the FOV, distinct spot arrays are produced and separated according to the sub-eye arrangement pattern. Based on multi-sub-eye parallax computation, the spatial coordinates of each target can be precisely reconstructed. By exploiting the temporal continuity between consecutive frames, the motion trajectories of multiple targets are identified and tracked in real time. A global-ratio algorithm was introduced to fully exploit the parallax information between all detected sub-eye images. Compared with previous ratio-based methods, this approach significantly enhances depth accuracy and robustness against random noise. The introduction of metalens sub-eyes, replacing the simple aperture structure used in our previous planar compound eye design, results in markedly improved imaging performance. The sub-eyes enable light convergence rather than transmission-only sampling, leading to smaller and brighter focal spots, higher signal-to-noise ratio, and enhanced centroid positioning precision. Consequently, the CMES can detect and position weak targets that were previously unresolvable. The enlarged sub-eye apertures also improve the diffraction-limited resolution, allowing fine spatial features of the targets to be discerned. This capability suggests that the CMES could not only determine spatial coordinates but also provide information on target morphology or identity. Although the proof-of-concept experiments demonstrated positioning of up to four targets, the system architecture imposes no intrinsic limit on target number, as long as their spot arrays do not overlap. Thus, the number of simultaneously trackable targets can be further increased by optimizing the sub-eye arrangement and imaging separation strategy (see Supplementary Note 3). At present, calibration was conducted at a fixed depth of 1000 mm. For other working distances, recalibration is required to generate new correction matrices for the global-ratio R. For full 3D operation, a volumetric calibration matrix can be constructed as a superposition of multiple 2D calibration layers, following the same procedure proposed here. This method remains conceptually simple but introduces additional workload, which could be alleviated through automated calibration or machine-learning-based correction. Furthermore, expanding the FOV by enlarging the sub-eye array is feasible. In that case, the phase profiles of individual sub-eyes must be re-optimized to ensure minimal aberration over larger angular ranges. In the current study, trajectory recognition relied on temporal continuity between consecutive frames. Future work will explore intelligent recognition algorithms based on neural networks to directly identify and distinguish multiple targets with different motion patterns. Overall, the CMES represents a versatile optical perception platform that combines the compactness and wide angular coverage of biological compound eyes with the wavefront-shaping precision of metasurfaces. It provides an efficient solution for multi-target spatial positioning, motion tracking, and potentially morphological recognition. Such capabilities are promising for applications in formation control, visual navigation, and environmental perception of UAVs and embodied intelligent agents operating in complex low-altitude or indoor environments. Materials and methods Fabrication of the metalens sub-eyes The CMES sample was fabricated using standard semiconductor microfabrication techniques, as illustrated in Fig. S5. A 0.5-mm-thick fused silica substrate with a side length of 40 mm was first cleaned ultrasonically in acetone, ethanol, and deionized water. A 600 nm amorphous silicon (a-Si) layer was deposited on one side of the substrate by plasma-enhanced chemical vapor deposition (PECVD). The thickness and dispersion of the a-Si film were characterized by spectroscopic ellipsometry to ensure design accuracy. A 100 nm chromium layer was then deposited on the opposite side of the substrate via electron-beam evaporation to form a light-shielding layer. Apertures defining the individual sub-eye FOVs were patterned by ultraviolet lithography followed by ion-beam etching. These apertures effectively block environmental stray light and confine illumination to the designed optical channels of each sub-eye. Subsequently, the substrate was flipped, and a positive electron-beam resist (ZEP 520 A) was spin-coated onto the a-Si surface. The metalens pattern was written using electron-beam lithography (EBL) and developed to form a resist template. A 50-nm-thick aluminum oxide (Al₂O₃) layer was deposited by electron-beam evaporation and lifted off to produce a hard mask. The underlying a-Si was then etched through the mask using inductively coupled plasma (ICP) etching to generate the designed nanopillar structures. Finally, the Al₂O₃ mask was removed, revealing the completed metalens nanostructures. Imaging performance test of sub-eye Figure S6 shows the experimental setup for measuring the diffraction limit of a sub-eyes. Here, an LED emits light at 940 nm, which passes through Aperture 2 and a collimator (with a focal length of 200 mm and an aperture of 40 mm) to emerge as parallel light. After passing through Aperture 1, the light irradiates a certain sub-eye. The focal plane of the sub-eye coincides with that of the objective lens (Olympus_LCPLN50XIR, 50x, WD = 4.5 mm). The converged light spot is magnified 50 times by the objective lens and the tube lens (Olympus_U-TLUIR, focal length 180 mm) before reaching the detector (Xenics_Bobcat320). Supplementary information Supplementary Informations-clear Supplementray Movie Acknowledgements This work was financially supported by the National Key Research and Development Program of China under Grant Number 2023YFB3906300, the Postdoctoral Fellowship Program of CPSF under Grant Number GZC20252756, the China Postdoctoral Science Foundation under Grant Number 2025M784275, the Beijing Natural Science Foundation-Non-Consensus Innovation Project under Grant Number F251046, the National Natural Science Foundation of China under Grant Number 62304249 and the Fundamental Research Funds for the Central Universities. Author contributions L.Z. was responsible for instrument design, experiment, data analysis and paper writing. L.Y.Z. was responsible for metalens design and paper writing. X.Z., J.L., H.C., J.Y., and C.Z. were responsible for metalens design. L.W., Y.L., H.S., Z.Z., J.B.L., and H.G. were responsible for the experiment. Z.C.Z. and F.X. were responsible for the guidance of the whole process. All authors read and approved the final manuscript. Data availability The datasets generated during the current study are available from the corresponding author upon reasonable request. Conflict of interest The authors declare no competing interests. Contributor Information Lingyun Zhang, Email: zhangly_imet@hust.edu.cn. Fei Xing, Email: xingfei@mail.tsinghua.edu.cn. Zichen Zhang, Email: zz241@buaa.edu.cn. Supplementary information The online version contains supplementary material available at 10.1038/s41378-026-01319-9. References Associated Data Supplementary Materials Supplementary Informations-clear Supplementray Movie Data Availability Statement The datasets generated during the current study are available from the corresponding author upon reasonable request.