Most cameras look at the world through a flat sensor, which is not how eyes work. Biological retinas are curved, and that curve lets a simple lens cover a wide view with little distortion. Researchers at the Hong Kong University of Science and Technology (HKUST) have now built a bionic eye that follows that design, with nearly 400,000 light-sensing pixels on a hemispherical surface.
The team’s results, published in Nature Materials and announced by HKUST on November 13, 2026, combine full-color imaging, a field of view beyond 160 degrees, and motion detection that happens inside the sensor itself.
Why curved sensors are hard to make
HKUST explains that conventional cameras need several optical elements and external processing to correct aberrations, which adds size, uses power, and narrows the field of view. Curved image sensors can support compact, wide-angle systems with better optical performance, but there is a catch: standard lithography was designed for flat surfaces, so packing dense pixel arrays onto a sharply curved one is difficult.
The group, led by Prof. Fan Zhiyong of HKUST’s Department of Electronic and Computer Engineering and Department of Chemical and Biological Engineering, developed a 3D laser fabrication approach for hemispherical substrates. The pixel count tells the story of the progress: the team’s first-generation eye in 2020 had only 100 pixels, while the new one has 367,500 at a density of 1,905 pixels per inch.
Three capabilities in one device
The new artificial retina is a tandem design with blue-, green-, and red-sensitive layers. According to the university, this gives:
- Full-color imaging across a spectral response of 300 to 800 nanometers.
- Wide-angle, aberration-corrected vision across a field of view exceeding 160 degrees.
- Event-driven motion detection inside the sensor, inspired by the trichromatic retina and motion sensitivity of cyprinid fish.
Putting all three on one platform is the point. Compact machine vision usually trades one of these for another.
Less data to move
Frame-based systems detect motion by processing a sequence of complete images. The HKUST sensor instead detects motion-related information within the sensor, so full image sequences do not need to be transmitted for the evaluated tasks. Under the study’s test conditions, HKUST reports that this cut bandwidth demand by more than 99.95% compared with frame-based imaging, while motion recognition accuracy reached 98.6%.
Those figures come from the researchers’ own test conditions, so real-world results will depend on the task and scene. Still, the direction is clear: do more at the sensor and send less downstream.
What it means
For robotics and autonomous systems, the appeal is a smaller, lighter vision module that sees widely and flags movement cheaply. Prof. Fan says the next-generation bionic eye could have more than one million sensing pixels, opening opportunities in robotics and autonomous driving. That is a stated ambition, not a product plan.
Several open questions remain that the announcement does not answer, such as cost, durability, and how the sensor would be manufactured at volume. The lead author, Dr. Long Zhenghao, now an assistant professor at Fudan University, and Prof. Zi Yunlong of HKUST (Guangzhou) collaborated on the work.
It is also worth noting what this is not: a medical implant. The release frames the device as a vision sensor for machines, and this article does the same.
How it compares with today’s machine vision
Robots and self-driving platforms today typically rely on flat sensors feeding a stream of full frames to a processor, which then looks for what changed. That pipeline is flexible but costly in bandwidth, power, and latency, especially when most pixels stay the same from one frame to the next.
Event-driven designs flip this: only changes are reported. HKUST’s contribution is building that idea into a curved, full-color retina so the sensor both gathers a wide view and reports motion without a separate processing stage for the tested tasks.
The curved geometry also simplifies optics. Because the sensor follows the shape of the focal surface, fewer corrective lenses are needed, which is why the HKUST team stresses compactness. For small robots, drones, or wearable vision modules, saving space and weight can matter as much as raw resolution.
The team’s 2020 Nature paper, which it cites as the starting point, showed the concept with only 100 sensing elements. Scaling to hundreds of thousands in six years shows how quickly fabrication methods can mature, though reaching a million pixels will bring new challenges in wiring and uniformity.
Bottom line
HKUST’s bionic eye jumps from 100 pixels in 2020 to 367,500 now, adds color and a 160-degree view, and handles motion at the sensor. If it can be scaled and ruggedized, it offers a leaner way for robots and vehicles to see.