Yongtian Wang

Light field near-eye displays are promising for alleviating the vergence-accommodation conflict (VAC), yet their implementation in optical see-through augmented reality (AR) systems remains challenging. Microstructure-array-based approaches suffer from a spatial-angular resolution trade-off, while conventional multi-layer architectures place stringent demands on image quality and computational accuracy. Here, we present an optical see-through light field near-eye display based on multi-layer display panels and off-axis eyepieces. The proposed architecture avoids backlight occlusion while preserving the native spatial resolution of the display panels. To support this design, we develop a light field decomposition method for generating optimized multilayer modulation images and introduce a high-order polynomial precorrection method to compensate for irregular off-axis distortion. A binocular prototype using three liquid crystal display (LCD) panels is built and tested. The results demonstrate that the proposed system achieves optical see-through three-dimensional (3D) imaging with clear monocular focus cues, correct binocular parallax, and a geometrically corrected image field after distortion correction.

Quasi-bound state in the continuum (quasi-BIC) metasurfaces provide a promising platform for optical sensing due to their low radiative loss and narrow resonances. However, achieving stable sensing performance across different environments remains challenging. Here, we demonstrate a dual-resonance quasi-BIC metasurface based on asymmetric amorphous silicon nanopillar dimers, enabling sensing performance evaluation in both dry and liquid environments in the near-infrared region. The device exhibits two magnetic toroidal dipole (MTD)-dominated quasi-BIC resonances with refractive-index sensitivities of 162 nm/RIU and 609 nm/RIU. Both simulations and experiments demonstrate stable refractive-index sensing performance in dry and liquid environments, validating the applicability of the proposed metasurface under different operating conditions. The two resonances exhibit distinct sensing characteristics in different media, providing complementary responses for optical sensing. Owing to its simple structure, CMOS-compatible fabrication, and compatibility with microfluidic integration, the proposed metasurface provides a promising platform for optical sensing and biomolecular detection in different environments.

Beam steering devices are essential components in several applications, such as detection, LiDAR, and optical communication. However, traditional beam steering devices tend to be bulky or exhibit a limited steering angle. Cascaded metasurfaces exhibit strong capabilities for powerful dynamic light modulation. In this work, we propose and demonstrate a highly integrated beam steering device utilizing two cascaded, independently rotating metasurfaces. Through simple relative rotation, the output beam can be continuously and dynamically steered, generating pre‐designed trajectories that densely cover a wide angular space. We have numerically simulated and experimentally verified continuous beam steering of petal‐shaped scanning trajectory with discrete scanning technique. Furthermore, by integrating traditional gradient phase design with the Dammann grating profiles, we achieved a wider scanning range and a greater number of addressable beam positions. This technology is compact, and versatile to utilize in various fields, including but not limited to optical sensing, light detection and ranging (LiDAR), and free‐space optical communication.

Training generalist vision-language-action (VLA) policies requires large, diverse and precisely action-labelled data, yet real-robot teleoperation is costly and conventional simulation lacks photorealism. We present Gaussian2Sim, a real-to-sim pipeline built on 3D Gaussian Splatting (3DGS) that decouples reconstructed scenes into physically interactive, object-level assets with collision geometry and physical properties, populates them with tasks and language generated by a large language model, and re-renders simulated states into photorealistic RGB-D observations. The pipeline runs on AMD RDNA4 hardware through ROCm and Vulkan, without CUDA.Because every later stage consumes the decoupled assets, we make the decoupling stage measurable: evaluation scenes are authored in a physics simulator whose renderer emits exact per-pixel instance identity, giving labels over Gaussian primitives that are correct by construction rather than by annotation. Over 8 scenes, 87 object instances and 84 views per scene, lifting Segment Anything (SAM) masks by multi-view agreement raises 3D instance mean intersection-over-union (mIoU) from 67.3% for the single most informative view to 95.6%.The same benchmark shows that whether agreement can be trusted depends on the clustering threshold: at the cut whose cluster count matches the object count, only 2.0% of high-agreement primitives are misassigned, whereas at a strict cut 37.9% are, and they carry 60% of all decoupling error. Agreement should therefore be reported together with the threshold at which it was measured.A policy trained on Gaussian2Sim data grasps a target cube autonomously on a physical Franka Emika arm; we report this as a feasibility demonstration, without a success rate.

A compact dual-focal virtual reality display is presented to provide discrete accommodation cues for mitigating the vergence-accommodation conflict. The architecture combines a folded-pancake optical path with an electrically switchable liquid-crystal polarization modulator and a YVO ₄ birefringent doublet, enabling two steady-state focal states within a 15-mm total track length. A monocular prototype was fabricated and experimentally characterized, achieving diagonal fields of view of 43.61° and 46.03° at nominal −4 D and −0.2 D, respectively. Through-focus measurements yielded mean focal positions of −3.99 D and −0.19 D over three repeated scans. Peak spatial ghost-image intensity ratios were 2.2% and 2.3%, respectively. An idealized binary-weighted model further examines multi-focal scaling and associated practical constraints.

Holography enables modulation of light waves to create three-dimensional (3D) displays with natural depth cues. Previous research on holographic displays primarily focused on synthesizing real images in free space while neglecting the physical display optics. Here, we propose an end-to-end, system-level design framework that accounts for the complete path from the illumination source to the retina by integrating optimized eyepiece optics. Joint optimization of the hologram, eyepiece optics, and component positions across polychromatic wavelengths and multiple depth channels corrects chromatic dispersion, optical aberrations, and geometric distortions while incorporating eyebox deployment, depth representation, and hologram bandwidth, enabling apochromatic and high-performance holographic displays. Two prototypes, a full-color, large depth-of-field (DOF), large field-of-view (FOV) holographic near-eye display and a full-color, large FOV, multi-focus 3D holographic freeform augmented-reality (AR) near-eye display, validate the effectiveness of the proposed framework, providing valuable insights for advancing ergonomic holographic AR/virtual reality (VR) systems with realistic visual experiences.

Commercial off-the-shelf lenses can shorten optical prototyping but create a highly discrete design space. We present an automated stock-lens optical system generation framework that combines a transformer initial design generator with aberration-guided stock-lens replacement. Optical systems are encoded as singlet, cemented doublet, air-gap thickness, and aperture stop tokens. Principal component analysis based constraints are introduced during training to guide the generated structural and material parameters toward stock-lens distributions and improve replaceability. A database of 51,123 lenses is searched by Seidel Aberration Distance Index coarse ranking, followed by real ray tracing and damped least squares optimization. Two demonstrations were conducted to validate the proposed framework: stock-lens replacement of a whole-field eye measurement system and construction of a 50 mm F/2.8 imaging lens. In both cases, the resulting systems were composed entirely of stock lenses and achieved optical performance comparable to that of the initial designs; the fabricated imaging lens further demonstrates the practical value of the proposed method.

Selecting suitable starting points for optical systems often requires extensive manual screening and adaptation of patent or historical designs. To reduce this dependence on designer experience, we propose a physics-constrained neural method that generates diverse optical starting points conditioned on lens sequences and target specifications. A conditional variational autoencoder predicts surface curvature, thickness, refractive index, and Abbe number. Structural encoding, material constraints, total track length / back focal length conditioning, diversity regularization, and ray-tracing-based evaluation are incorporated to improve physical feasibility and generation diversity. An LLM-agent interface is further introduced to translate natural-language requirements into generation conditions and organize optional subsequent optimization and reporting. Experiments demonstrate that the proposed method can generate diverse, ray-traceable, and optimization-ready starting points. By integrating conditional generation and LLM-agent-assisted requirement parsing, the framework significantly improves design efficiency, lowers the difficulty of optical system design, and dramatically reduces the dependence on designer experience, existing starting points, and manual intervention.

This paper proposes a freeform imaging system generation framework based on a performance-aware deep neural network. Existing optical system generation methods usually focus on directly predicting system structures from system specifications, while the generated results often lack direct evaluation and constraints on performance targets. In this paper, a freeform imaging system design method using a performance-aware deep neural network is proposed to predict both system structures and image performance. A training dataset is built using an automated dataset generation method, and the network is trained with a multi-objective joint data-driven strategy. After training, the network can be regarded as a freeform system generation model with performance awareness. For a specific design task, once the design requirements are provided, the network can immediately generate multiple candidate freeform imaging systems and simultaneously predict the image performance that each system may achieve. By comparing the predicted image performance with the predefined performance requirements, candidate structures with the potential to satisfy the target performance requirements can be identified before complete optimization, thereby allowing invalid candidates to be screened out in advance. The effectiveness and feasibility of the proposed framework are verified through design examples of freeform off-axis three-mirror imaging systems with multiple folding geometries.

Compact and lightweight varifocal imaging systems are increasingly important for emerging applications such as portable sensing, wearable optics, and aerospace instrumentation. However, conventional varifocal architectures rely on mechanically complex axial or lateral translations of optical components, limiting their miniaturization and integration. Here, we present a hybrid refractive-metalens varifocal system that enables focal tuning through rotational modulation of cascaded freeform metalenses rather than component displacement. The proposed architecture combines the high imaging capability of refractive optics with the wavefront-engineering flexibility of metalenses, enabling continuous varifocal operation within a significantly reduced system footprint. To enable the practical design of such hybrid systems, a real-ray-tracing-based co-design strategy is developed, together with a multi-configuration optimization framework that suppresses parasitic zeroth-order diffraction while maintaining imaging performance across the full varifocal range. A fabricated prototype demonstrates continuous 2× zoom operation with stable aberration control and low distortion, showing good agreement between experimental measurements and theoretical predictions. These results validate both the proposed hybrid refractive-metalens architecture for compact translation-free varifocal imaging and the accompanying system-level co-design methodology, providing a practical framework for future hybrid refractive-metasurface optical systems in infrared sensing, AR/VR, aerospace instrumentation, compact cameras, and portable intelligent devices.