On November 13, 2026, the Hong Kong University of Science and Technology (HKUST) announced a hemispherical bionic eye with 367,500 pixels. The curved sensor mimics the biological retina and stacks light-sensing layers for red, green and blue, combining wide-angle color imaging with motion detection in a single device. The team succeeded in forming high-density pixels on a curved surface, which had been difficult until now, and showed that detecting motion inside the sensor could greatly reduce the amount of data sent to downstream processors. Applications in robotics and autonomous systems are anticipated, but technology for rapidly reading out signals from so many pixels remains a challenge for future development.
The research was led by Professor Zhiyong Fan of HKUST, with Zhenghao Long and colleagues handling development. The results were published in the peer-reviewed journal Nature Materials, and the paper, "An aberration-corrected bionic eye for full-colour vision and event-driven motion detection," went online on August 17, 2026. The university's announcement is dated November 13, which differs from the date the findings themselves were published.
Three light-sensing layers stacked on a hemisphere, integrating 367,500 pixels
At the core of the new bionic eye is a hemispherical, stacked artificial retina modeled on the retina inside a biological eyeball. Three layers, each sensitive to blue, green or red light, are stacked to capture color images. Each layer holds 122,500 pixels in a 350 × 350 array, for a combined total of 367,500 pixels across the three layers. According to the university, the pixel density is 1,905 ppi, the responsive wavelength range is 300 to 800 nanometers, and the device secures a field of view exceeding 160 degrees with aberration, the optical distortion of the image, corrected.
Conventional image sensors are flat, and securing a wide field of view while suppressing aberration can require multiple optical components and image processing. A curved sensor closer to a biological retina, by contrast, may allow a configuration that favors a smaller, wider-angle optical system. However, forming many fine pixels on a strongly curved surface such as a hemisphere is not easy. Lithography, commonly used in semiconductor manufacturing, assumes a flat substrate, so high-density patterning on surfaces with large curvature has been difficult.
To address this, the research team developed a three-dimensional processing technique that uses a laser to form fine electrode patterns on a hemispherical surface. According to the paper's supplementary materials, after bonding the hemispherical substrate, a gold film formed on its surface is laser-processed to create electrodes that separate the pixels. There are 350 connection pads on each side, and the electrode spacing is 40 micrometers when projected onto a plane. This processing technique made it possible to form a high-density pixel array on a curved surface, something previously hard to achieve.
The first-generation bionic eye the same team published in Nature in 2020 had only 100 pixels. The new device has increased the combined pixel count across the three layers to 367,500, a major step toward higher-resolution curved sensors. Note, however, that this figure is the sum of the pixels in the three light-sensing layers; it does not mean that 367,500 pixels are lined up in a single layer.
Motion detected inside the sensor, reducing the required data bandwidth
The new bionic eye not only captures color images but also has a built-in function for detecting motion inside the sensor. In typical frame-based imaging, a camera captures successive images, and an external processor compares them to detect motion. With this approach, pixel information from every frame must be processed even when most of the image has not changed. The "event-driven" approach adopted by the team, by contrast, uses the stacked light-sensing layers to extract motion information at the sensor and aims to limit the data sent downstream to what is needed.
The team evaluated how much this approach reduces the required data bandwidth. They reconstructed the same video segment with both the conventional and new methods and calculated the data volume based on the number of pixels required. The conventional method counts all 350 × 350 × 3 pixels in every frame, whereas the new method counts only the pixels corresponding to motion, those with brightness above 127. This comparison estimates how much the required bandwidth can be reduced by sending downstream only the pixels associated with detected motion, rather than all the pixels that make up the video.
According to data released with Figure 4 of the paper, assuming 60 frames per second (60 fps), the conventional method produced 21.02852 MB per second, while the new method produced 0.00984 MB per second. The reduction under these conditions exceeds 99.95%. However, this is a calculation based on the number of effective pixels obtained from the same video; it does not mean the prototype shot video in real time at 60 fps, nor that an equivalent bandwidth reduction was confirmed with actual communication equipment. While it demonstrates the effectiveness of a design that extracts necessary information inside the sensor, it must be evaluated separately from the actual processing speed of the device.
Motion recognition performance was also evaluated. The research team compared the case where one frame obtained from the bionic eye was used as input with the case where one to nine frames from a conventional camera were used, after five epochs of training, meaning five passes over the full training data. The recognition accuracy recorded for the bionic eye in the published data was 0.98567, equivalent to about 98.6%. However, the published abstract and supplementary materials alone do not make it possible to confirm the total number of videos used in the evaluation, how the data were split into training and validation sets, or the denominator used to calculate accuracy. This figure is therefore performance obtained on a recognition task set up by the research team, and does not show that equivalent accuracy would be obtained in general motion recognition.
High resolution achieved, but pixel readout speed remains a challenge
What matters in evaluating these results is that forming 367,500 pixels and reading out all of them at high speed are separate technical challenges. The paper's supplementary materials describe the specific procedure used to capture images with the prototype. The team used a computer-controlled projector to repeatedly project a video that alternated between two images. They then selected pixels one by one through a multiplexer and sampled each pixel's current 20 times over about 0.5 seconds.
With this method, the area that could be connected and measured at one time was limited to 122 × 122 pixels. Measuring a different region therefore required manually changing the connection of the flexible substrate. The three light-sensing layers were also measured in turn, and the resulting data were combined to reconstruct a 350 × 350 × 3-pixel image. In other words, the images were not acquired by reading out all pixels simultaneously but by repeatedly projecting the same scene and collecting measurements from each region.
Based on pages 30 to 32 of the supplementary materials and the Figure 4 data, the main results and how they were verified can be summarized as follows.
| Type of result | Target | Verification method and content |
|---|---|---|
| Measured | Fig. 4b, 4e | Images were projected repeatedly and each pixel measured in sequence to reconstruct color images and motion information |
| Simulation | Fig. 4f | Motion detection and object trajectory prediction were evaluated |
| Calculation | Fig. 4i | Required data bandwidth compared from the effective pixel counts of the same video, assuming 60 fps |
As shown, the results presented in the paper include actual sensor measurements, simulations and values calculated from pixel counts. In particular, the bandwidth reduction of more than 99.95% comes from a calculation based on the effective pixel counts of a video and does not directly indicate the prototype's real-time processing performance. The authors also explain that the current measurement system is limited in the number of multiplexer and instrument channels, and lacks the processing capability to read out tiny currents on the order of picoamperes. As a result, the effective readout speed is low, and the need to switch connections manually during measurement is a major constraint at this stage.
Next goals: parallel pixel readout and expansion to 1 million pixels
The improvements the research team cites for the future are a parallel readout scheme that acquires signals from multiple pixels simultaneously and the development of multichannel low-noise circuits. Even if the pixel count increases, a mechanism that reads out signals one at a time sets a limit on the speed at which moving targets can be continuously imaged and recognized. In applications such as robotics and autonomous driving, where the surroundings change constantly, what matters is not only image resolution but also whether the necessary information can be acquired with low latency. Future prototypes will need to be evaluated on a combination of factors, including the number of pixels that can be read out simultaneously, the frame rate during continuous operation and motion detection accuracy.
Some materials for checking the reproducibility of the research have also been released. The paper comes with supplementary materials and numerical data for the figures, allowing the experimental conditions and results to be checked. The code used for the simulations and classification, however, is made available by the corresponding author upon reasonable request. Whether independent research groups can reproduce the results with the same processing method and measurement conditions, and whether equivalent performance can be obtained in different imaging environments, will be important questions for future verification.
In the university's announcement, Professor Fan said that with further technological development, a next-generation bionic eye exceeding 1 million pixels could be built. This, however, is a goal premised on future improvements, and the current study has not demonstrated a 1-million-pixel device. By combining high-density pixel formation on a hemispherical surface with motion detection inside the sensor, the results open new possibilities for compact vision sensors for robots and autonomous systems. What will determine their practicality is not only whether the pixel count can be raised further, but also whether parallel readout circuits can capture high-resolution information fast enough and whether performance can be maintained in real operating environments.