This Bionic Eye Sees in Full Colour — and Detects Motion Without Filming Every Frame

Most cameras see the world in a surprisingly inefficient way. They capture complete frames again and again, even when almost nothing in the scene has changed. They also rely on flat image sensors behind multiple optical elements that correct distortion and focus light across the frame.

A bio-inspired vision system reported in Nature Materials on 17 August 2026 takes a very different route. Researchers built a hemispherical tandem artificial retina that combines full-colour imaging, a field of view greater than 160 degrees and event-driven motion detection in the sensor itself.

The result is not a medical implant that restores human sight. It is an artificial vision sensor aimed at machine vision, robotics and compact intelligent devices. That distinction matters, because the phrase “bionic eye” can easily suggest something the research did not claim.

Why ordinary cameras are flat

Modern cameras are built around planar semiconductor sensors because flat chips are straightforward to manufacture at enormous scale. The problem is optical geometry. A lens naturally forms a curved focal surface, while the detector is flat. Camera makers compensate with carefully designed stacks of lenses and computational correction.

Biological eyes approach the problem differently. The retina curves around the inside of the eyeball, allowing light from a wide range of angles to land on a naturally curved sensing surface.

Engineers have wanted to reproduce that geometry for years. Curved artificial retinas can simplify optics and expand field of view, but manufacturing high-density pixels on a strongly curved surface has been difficult. Earlier systems generally traded curvature for practical resolution.

The new work attacks that bottleneck directly.

A hemispherical retina with 367,500 pixels

The researchers report a high-curvature image sensor with 367,500 total pixels and a density of 1,905 pixels per inch. The artificial retina covers a hemispherical geometry rather than behaving like a conventional flat camera chip.

That matters because the shape is doing part of the optical work.

Instead of forcing a lens system to project a wide scene onto a flat plane, the curved detector more closely follows the focal geometry. According to the study, the system achieved aberration-corrected imaging across a field of view exceeding 160 degrees.

For compact machines, drones and robots, fewer corrective optical elements could eventually mean smaller camera modules and less complicated hardware.

It sees across a broad colour range

The system provides full-colour imaging over wavelengths from roughly 300 to 800 nanometres. That range extends from near-ultraviolet through the visible spectrum and toward the near-infrared edge.

Colour matters for machine perception for many of the same reasons it matters to humans. It helps distinguish objects with similar shapes, identify materials, interpret signs and signals, inspect surfaces and separate a target from its background.

The important point is not that the artificial eye “sees like a human.” It does not. The researchers have created a machine sensor whose curved architecture can capture useful colour information over a very wide viewing angle.

The clever part: motion without conventional video

The second half of the design may be even more important than its shape.

A normal video camera records frames at fixed intervals: perhaps 30, 60 or hundreds of times every second. Every pixel is repeatedly read whether its part of the scene changed or not. That produces enormous amounts of redundant data.

Event cameras work differently. Instead of continuously reporting the entire scene, pixels respond primarily when brightness changes. Movement therefore generates a stream of events while static areas generate little information.

The tandem artificial retina incorporates event-driven motion sensing directly into the device.

In the reported experiments, this reduced bandwidth demand by more than 99.95 percent compared with frame-based imaging while achieving 98.6 percent motion-recognition accuracy.

Those are laboratory results for the demonstrated system, not a guarantee that every future robot camera will achieve the same numbers. But they illustrate why neuromorphic and event-based vision is attracting so much attention.

Why bandwidth matters to robots

A robot does not merely need to capture an image. It must move that information through memory, process it and make a decision quickly enough to act.

Imagine a mobile robot moving through a warehouse. Most of the walls, floor and shelving may remain visually unchanged from one instant to the next. A conventional camera still sends complete frames containing all of those static pixels.

An event-driven sensor can concentrate data on what changed: a worker stepping into the aisle, a box beginning to fall or another robot crossing the path.

Less redundant data can mean lower communication load, less processing and potentially lower energy consumption. It also fits the broader movement toward giving machines more intelligence at the sensor itself rather than sending every raw measurement to a central processor.

That same philosophy appears in other emerging hardware. Our article on electronic skin for humanoid robots describes tactile sensors that process contact and slip information close to where physical interaction happens.

Why a 160-degree field of view is useful

Wide-angle vision is particularly valuable for autonomous systems.

A robot navigating a room needs peripheral awareness. A drone benefits from seeing obstacles that are not directly in front of it. A compact inspection camera may need to observe a large area while fitting through a small opening.

Traditional wide-angle cameras can achieve enormous fields of view, of course, but often with distortion that must be corrected optically or computationally.

A curved sensor gives engineers another option: change the geometry of the detector itself.

The Nature Materials system demonstrated an aberration-corrected field of view beyond 160 degrees. That does not make conventional lenses obsolete, but it shows that sensor shape can become part of the optical design rather than being treated as a fixed flat surface.

The artificial retina is becoming a computer

The phrase “image sensor” increasingly undersells what experimental vision hardware can do.

Traditional architecture separates sensing and computation. A sensor measures light, converts it into digital values and sends those values somewhere else for interpretation.

Bio-inspired hardware is beginning to blur that boundary.

Event detection can happen in the sensor. Noise reduction can happen close to the pixels. Future devices may perform feature extraction, motion estimation or simple recognition before a conventional processor ever receives the data.

This approach is often described as in-sensor or neuromorphic computing because it borrows a principle from biological nervous systems: do useful information processing near the point of sensing.

The benefit is not biological imitation for its own sake. It is efficiency.

Could this become a robot eye?

That is one of the obvious possibilities, but substantial engineering remains.

Robots need cameras that survive vibration, dust, temperature changes, impacts and years of operation. They need standardized electronics, calibration procedures and manufacturing processes that can produce sensors cheaply and consistently.

A laboratory prototype can demonstrate a powerful architecture without yet satisfying those requirements.

If curved event-driven sensors become manufacturable at scale, however, they are a natural fit for humanoids and autonomous machines. Wide peripheral vision, compact optics and low-bandwidth motion detection are all valuable when computing power and battery capacity are limited.

What about prosthetic vision for humans?

This research should not be confused with a retinal implant or a device ready to restore sight.

The “bionic eye” described in the paper is a bio-inspired machine vision system. Its artificial retina converts light into electronic information for imaging and motion detection. Connecting an artificial sensor to the human visual nervous system is an entirely different challenge involving biocompatibility, neural interfaces, surgical safety and the brain’s interpretation of encoded signals.

Research in prosthetic vision exists, but this particular result is better understood as advanced camera technology inspired by the architecture of the eye.

That distinction is important both scientifically and ethically. Interesting technology does not need an exaggerated medical claim to be significant.

Another bionic-eye result shows how fast the field is moving

The Nature Materials work is not happening in isolation.

A separate paper published in Nature Sensors on 31 July 2026 described a neuromorphic spherical eye with an ultradense flexible artificial retina. That system reported 7,000 pixels per inch, more than 17 million pixels in its demonstrated retina, depth-tunable three-dimensional imaging, a 110-degree field of view and motion tracking.

The two systems use different architectures and should not be treated as a head-to-head product comparison. Together, however, they show a clear research direction: artificial vision is moving away from simply copying flat phone-camera sensors and toward curved, computational and task-specific retinas.

What still has to be solved

Manufacturing

Curved semiconductor systems are much harder to mass-produce than planar chips. Yield, repeatability and assembly cost will determine whether the concept escapes the laboratory.

Resolution

1,905 ppi is a major step for a highly curved sensor, but conventional imaging technology has decades of manufacturing optimization behind it. Pixel count alone also does not determine image quality.

Packaging

A practical device needs durable encapsulation, electrical connections and optics without destroying the advantages of the curved geometry.

Calibration

Every unusual optical architecture requires reliable calibration across temperature, ageing and manufacturing variation.

Software

Event streams are fundamentally different from normal video. Algorithms and AI models must be designed to exploit sparse asynchronous data rather than assuming a sequence of ordinary frames.

Why this development matters

The most interesting part of the research is not that engineers built something shaped like an eye. Nature-inspired shapes have appeared in laboratories for decades.

The important development is convergence.

High pixel density, wide-angle curved imaging, colour sensing and event-driven motion detection are beginning to coexist in a single architecture.

That could change how engineers design cameras for machines. Instead of building a general-purpose sensor that records everything and asking software to discard most of the data later, future vision systems may physically capture information in a form already optimized for the task.

A warehouse robot might prioritize motion. A drone might prioritize extreme peripheral awareness. An industrial inspection system might combine colour with wavelengths outside normal human vision.

The camera becomes less like a passive photographic device and more like a specialized sensory organ.

The bottom line

Researchers have demonstrated a hemispherical tandem artificial retina with 367,500 pixels, 1,905 ppi density, full-colour imaging across 300–800 nm and an aberration-corrected field of view greater than 160 degrees. Its event-driven mode reduced bandwidth demand by more than 99.95 percent in the reported comparison and achieved 98.6 percent motion-recognition accuracy.

Those numbers make this an impressive machine-vision prototype, not a finished commercial eye and not a human retinal implant.

But the direction is significant. Cameras have spent decades becoming better flat cameras. The next generation of machine vision may stop being flat at all.