Event camera

An event camera (also called a dynamic vision sensor, DVS, or neuromorphic camera) is an image sensor whose pixels work independently and asynchronously. Instead of every pixel exposing together on a fixed clock to build a frame, each pixel watches its own brightness and fires a small message, called an event, the instant that brightness changes by more than a set threshold.^[1] An event encodes just a few values: the pixel's x and y address, a timestamp accurate to roughly a microsecond, and a polarity bit marking whether the brightness rose or fell.^[1][2] Because pixels report only what changes and stay silent otherwise, event cameras combine dynamic range beyond 120 decibels, latency in the microsecond range, low power draw, and freedom from motion blur, at the cost of never reporting an absolute brightness value and requiring algorithms built around a sparse, asynchronous stream of events rather than a sequence of images.^[1][2] The sensor traces its ancestry to 1980s analog silicon-retina research and has since become a specialized but growing option in high-speed robotics, automotive perception, and experimental humanoid robot hands.

In brief: a conventional camera works like a flipbook, redrawing the whole page 30 to 240 times a second whether anything moved or not. An event camera is more like a room full of people who each call out only when the light on their own seat changes. Most of a static scene stays quiet, so the messages that do arrive are sparse, extremely fast, and point directly at motion: exactly the kind of signal a robot needs to catch a slipping grip or a falling object before a frame-based camera would even finish exposing its next image.

How it works

Each pixel in an event camera contains a small analog circuit instead of a simple charge well. A photodiode generates a current proportional to incident light, and a logarithmic amplifier converts that current into a voltage proportional to the logarithm of intensity. Working in this log domain, rather than on raw linear brightness, is what lets the pixel respond to relative contrast instead of absolute light level, and it is the main reason the sensor achieves such wide dynamic range.^[1] The pixel stores this log-intensity value as a reference and continuously compares the incoming signal against it. When the difference exceeds a positive threshold, the pixel fires an ON event; when it drops past a negative threshold, it fires an OFF event; either way, the reference resets to the new level and the comparison starts again.^[1][2] Because this comparison runs inside each pixel in continuous time rather than on a shared clock, two neighboring pixels can fire microseconds apart, and a pixel that is not changing produces nothing at all. The threshold, often called the contrast sensitivity, is configurable and typically set so a relative brightness change of somewhere between about 10 and 50 percent triggers an event, trading sensitivity to subtle motion against susceptibility to noise.^[2]

The sensor's output is not an image but a stream of tuples: pixel coordinates, a timestamp accurate to roughly a microsecond, and a polarity bit.^[1] Software can pack a slice of that stream into a pseudo-frame for display, but the sensor itself never produces one. A completely static scene under constant lighting produces no output at all, while a fast-moving edge produces a dense trail of events tracing its path, similar in effect to a very high frame rate camera but only where something is actually happening.^[1][2] This also means a single event carries no information about how bright a surface is, only that it changed by roughly a fixed ratio; recovering anything resembling a conventional image takes either a companion frame sensor or an algorithm that integrates many events over time, both covered later in this article.

History

Event cameras descend from neuromorphic computing, a field that builds analog circuits as literal models of biological neurons rather than simulating them in software.^[2] Working at Caltech in the late 1980s, computer scientist Carver Mead and his graduate student Misha Mahowald built the first "silicon retina," an analog VLSI chip that mimicked the layered, adaptive response of biological photoreceptor and ganglion cells, describing it in the 1988 paper "A Silicon Model of Early Visual Processing."^[3] Mead later credited the idea to Mahowald, who earned a PhD at Caltech in 1992 for further silicon-retina and neural-modeling work and died in 1996; neuromorphic engineering's leading award was created in her name and later renamed the Mahowald-Mead Prize once Mead agreed to share the honor.^[4][5] Mead's Caltech program, the first dedicated to computation and neural systems, later trained Tobi Delbruck, who earned his own PhD there in 1993 studying under Mead and others before moving to the Institute of Neuroinformatics, a joint institute of the University of Zurich and ETH Zurich, where he would go on to build the sensor described in this article.^[6]

In 2008, Delbruck, Patrick Lichtsteiner, and Christoph Posch published the dynamic vision sensor, a 128 by 128 pixel chip with 120 decibels of dynamic range and 15 microsecond latency that generated an asynchronous stream of per-pixel brightness-change events, the same architecture described above.^[7] Posch subsequently developed a variant called the Asynchronous Time-based Image Sensor, published in 2011, which added a second circuit that measures an approximate absolute gray level only at pixels where a change was just detected, reaching a dynamic range above 143 decibels at the cost of a lower fill factor and slower per-pixel readout.^[8] In 2014, Delbruck's group closed a different gap with the Dynamic and Active-pixel Vision Sensor, which shares a single photodiode between a change-detection circuit and a conventional global-shutter frame readout, so one chip outputs both an asynchronous event stream and synchronous intensity frames.^[9]

These research chips seeded a small industry. iniVation, founded in Zurich in 2015 as a spinoff of the Institute of Neuroinformatics with Delbruck among its founders, commercialized DVS and DAVIS-style sensors for research and industrial customers; in February 2024 the neuromorphic-processor maker SynSense acquired iniVation to form a combined vision-and-processing group.^[10][11] In Paris, Christoph Posch co-founded a company originally named Chronocam in 2014 alongside Ryad Benosman, Bernard Gilly, and Luca Verre; it was renamed Prophesee around 2017 and has since become the most prominent independent event-sensor supplier, with Sony among its investors.^[12][13] Large image-sensor manufacturers followed. Samsung Electronics published a series of research chips, including a roughly 1.3-megapixel, million-events-per-second design shown in 2020, though these have mostly stayed at the research and sampling stage rather than shipping as a catalog product.^[14] A Shanghai-based developer, CelePixel, founded in 2017, built its own CeleX line of event sensors before Will Semiconductor, which also owns the image-sensor maker OmniVision, acquired it in 2020.^[15] Sony took two paths into the technology: it acquired the Swiss startup Insightness, another Institute of Neuroinformatics spinoff, in late 2019, and separately became Prophesee's sensor-fabrication and co-design partner, jointly releasing the stacked IMX636 and IMX646 sensors in the early 2020s using Sony's copper-to-copper pixel-stacking process.^[16][17]

Types and variants

Commercial and research event sensors fall into a few architectural families, distinguished mainly by whether and how each recovers absolute brightness alongside the event stream.^[1]

Pure DVS sensors output only change events and no brightness information at all. This is the simplest, lowest-power, highest-speed design, and it underlies most current commercial parts, including Prophesee's Metavision line.^[1][7]

Time-based hybrid sensors, following the ATIS pattern, add a second per-pixel circuit that measures an approximate gray level using the timing of a ramping signal, but only for pixels that just reported a change, keeping most of the power and bandwidth savings of a pure DVS design while recovering some usable brightness information.^[8]

Frame-event hybrid sensors, following the DAVIS pattern, dedicate part of each pixel, or a companion sensor plane, to a conventional shutter-based frame capture running alongside the event stream, giving software both a sparse, fast event feed and an occasional full image to anchor it to.^[9] Sony and Prophesee's IMX636 and IMX646 sensors take a related but distinct approach: a stacked design puts the event-detection pixels on one silicon layer and the readout and digital logic on another, connected by Sony's copper-to-copper bonding process, which is how the pair reached a 4.86 micrometer pixel pitch, among the smallest reported for this sensor class.^[17]

Since the original 128 by 128 pixel DVS, resolution and event throughput have both climbed, though a high-resolution event sensor is harder to build than a high-resolution frame sensor: more pixels means more independent comparator circuits competing for the same output bus bandwidth the instant a scene changes everywhere at once.^[1] The table below traces that progression through sensors documented in the academic literature and in vendor specifications.

Scroll sideways for more →

Event cameras versus frame cameras

The table below summarizes how event and conventional frame-based cameras differ on the properties that matter most for robotics and machine vision.^[1][2]

Scroll sideways for more →

Tradeoffs and key evaluation criteria

Temporal resolution and latency. Because each pixel is its own independent sensor, an event camera can report a change within microseconds of it happening. A fast-moving object, a sudden slip, or an unexpected collision shows up in the event stream while a conventional camera running at 30 to 60 frames per second would still be several milliseconds from its next exposure, which is the main reason robotics researchers reach for the sensor.^[1][2]

Dynamic range. Because the log-domain pixel circuit responds to relative rather than absolute brightness, an event camera can watch a scene spanning deep shadow and direct sunlight in the same field of view without the sensor saturating in bright areas or going dark in shadow, a common failure mode for a frame sensor that must pick one exposure setting for an entire image.^[1]

Power and bandwidth. Because pixels only report change, a mostly static scene produces almost no data, so both the sensor's power draw and the data volume a downstream processor has to handle scale with how much is moving rather than with resolution and frame rate alone.^[1][19] This is why the lowest-power event sensors target always-on, battery-constrained uses such as edge AI devices and augmented and virtual reality headsets, where a conventional camera streaming full frames continuously would drain a battery much faster.^[19]

No absolute intensity. The same change-only design that gives event cameras their dynamic range and low bandwidth also means a single event says nothing about how bright a surface actually is, only that it changed by roughly a set ratio.^[1][2] Recovering anything resembling a normal image, useful for a human operator or a model trained on ordinary photographs, requires either a companion frame sensor, as in DAVIS-style hybrid designs, or a reconstruction algorithm that integrates a window of events into an image, which adds back some of the latency and computation the sensor was chosen to avoid.^[1][9]

Unconventional data, new algorithms. Decades of computer vision and deep learning tooling, including most of the datasets, model architectures, and pretrained weights behind modern vision language action models, assume input arrives as a sequence of dense images.^[1] An event stream is sparse, asynchronous, and arrives at a rate that depends on scene motion rather than a fixed clock, so it cannot simply be dropped into that pipeline. Using it well generally means either converting events into an image-like representation first, at some cost in latency and information, or adopting genuinely event-native algorithms, an active research area covered next.^[1]

Event-based vision algorithms

Because an event stream has no fixed frame rate and no absolute brightness, most classical computer vision algorithms, built around comparing one full image to the next, do not apply directly, and a distinct body of event-native methods has grown up around the sensor since the late 2000s.^[1]

Event-to-video reconstruction treats the event stream as a signal to invert back into ordinary images. E2VID, introduced by Henri Rebecq, Rene Ranftl, Vladlen Koltun, and Davide Scaramuzza in 2019, uses a recurrent convolutional network trained on simulated event data to reconstruct high dynamic range video from an event stream alone, letting existing frame-based computer vision tools run on event-camera footage after the fact.^[20]

Motion estimation by contrast maximization avoids reconstructing images at all. Proposed by Guillermo Gallego, Rebecq, and Scaramuzza in 2018, the method searches for the motion trajectory that, when used to warp and stack a window of events onto a single image plane, produces the sharpest possible edge image; the best-fitting motion becomes the estimate of optical flow, depth, or ego-motion, without ever needing to know the scene's actual brightness.^[21]

Event-based SLAM and visual odometry, estimating a camera's own trajectory and building a map from event data, follows several of the same strategies as optical flow: feature-based methods track corners or edges detected directly in the event stream, direct methods fit motion and scene structure straight to raw events without an intermediate feature step, and a growing share of recent systems fuse events with frames or an inertial measurement unit to combine the event stream's speed with a conventional sensor's absolute scale and appearance information.^[1][22] A 2026 survey in the International Journal of Computer Vision groups the resulting literature into feature-based, direct, motion-compensation, and deep-learning categories, reflecting how much the field has diversified since the DVS's introduction.^[22]

Spiking neural networks pair naturally with event data, since an event is functionally similar to a spike in a biological neuron.^[1] Neuromorphic processors such as Intel's Loihi run spiking networks directly on incoming events at very low power, rather than on a conventional GPU pipeline, and researchers have used Loihi paired with DVS cameras for tasks including gesture and lane recognition, keeping data in event form from sensor to decision instead of converting it to frames along the way.^[23] Because well-labeled real event data is scarce, much of this research also relies on event-camera simulators that generate synthetic event streams from ordinary video or rendered 3D scenes, a workflow that echoes the broader sim-to-real transfer approach used to train robot control policies before deploying them on physical hardware.^[1]

Use in humanoid robots and robotics

High-speed perception and obstacle avoidance

The clearest demonstrated advantage of event cameras in robotics is reacting to fast motion that a frame camera would miss or blur. Researchers at the University of Zurich's Robotics and Perception Group showed a quadrotor drone using a single event camera to detect and dodge objects thrown directly at it, at relative speeds up to about 10 meters per second, with an end-to-end detection-to-avoidance latency as low as 3.5 milliseconds, far faster than the 20 to 40 millisecond reaction time typical of a frame-based pipeline.^[24] The same property, catching a fast-changing edge in microseconds rather than waiting for the next frame, is what motivates event cameras for robot manipulation tasks where a slipping or falling object must be caught quickly, and for legged and humanoid robots that need to notice a foot slip or an unexpected collision as early as possible.^[1]

Slip and contact detection in robot hands

A research application closer to the humanoid-hardware bill of materials uses an event camera as the sensing element inside a tactile sensing module. In one design, an event camera looks through a soft, transparent silicone pad from behind; as a grasped object begins to slip, the resulting tiny, fast deformation of the pad shows up as a burst of events well before the object visibly moves, which a controller can use to tighten its grip. A 2018 study by Amin Rigi, Fariborz Baghaei Naeini, Dimitrios Makris, and Yahya Zweiri detected this kind of incipient slip with about 44 milliseconds of average latency across a range of grasped objects, using a DAVIS sensor mounted behind a silicone fingertip.^[25] A newer design called Evetac, built by researchers at TU Darmstadt, places printed markers on a similar silicone gel and uses an event camera to track how the markers shift under contact and shear: an event-driven cousin of the broader family of vision based tactile sensors that infer touch from a camera watching a deformable surface, with the added benefit of the event camera's microsecond timing for catching the earliest moments of slip.^[26] This approach sits alongside, rather than replaces, other ways of sensing contact at a robot's end effector, including dedicated force torque sensors and electronic skin arrays: an event-camera fingertip adds very high temporal resolution at the cost of needing a transparent or semi-transparent medium and enough internal volume for a small camera and lens, a real constraint inside a dexterous hand sized to match a human hand.^[25][26]

SLAM, optical flow, and ego-motion

A walking or manipulating humanoid robot moves its cameras quickly and sometimes unpredictably, and frame-based visual tracking can blur or lose lock during fast rotations, exactly when an accurate self-motion estimate matters most. The event-based SLAM and visual-inertial odometry work described above targets this gap directly, and several recent systems fuse event data with an inertial measurement unit for that reason.^[1][22] As of the mid-2020s, this remains largely a research-stage capability rather than a standard feature of production humanoid robots: no major humanoid platform has publicly confirmed an event camera in its bill of materials, and most current humanoid vision stacks instead rely on conventional RGB and depth cameras alongside lidar, processed on platforms such as NVIDIA's Jetson Thor.^[1] Event sensors remain an active line of academic and industrial research aimed at the fast, contact-rich moments, such as a grasp, an impact, or a slip, where a humanoid's onboard vision otherwise has the least time to react.

Automotive and industrial adoption

Outside robotics, the automotive industry is the largest declared source of demand for event-camera development. Prophesee has described work with Renault Group, one of its strategic investors, evaluating event-based sensing for driver assistance functions, and has also worked with Mercedes-Benz and the Swedish company Terranet on related automotive perception projects, positioning event cameras as a complement to conventional cameras, radar, and lidar rather than a replacement for them.^[27] Prophesee also cites industrial machine vision, including high-speed counting, vibration monitoring, and quality inspection, as a nearer-term commercial market than automotive, and expanded an industrial-vision partnership with camera maker IDS in 2026.^[28] On the consumer side, Prophesee and Qualcomm have collaborated since 2023 to integrate event sensors with Snapdragon mobile platforms for computational-photography features such as reduced motion blur, reaching production readiness in 2024.^[29]

Suppliers and landscape

The event-sensor market remains small and concentrated compared with conventional image sensors, split between a handful of specialists and larger sensor manufacturers that have added event sensing to a much bigger conventional-camera business.^[1]

Prophesee, headquartered in Paris and founded in 2014 under the name Chronocam, is the most prominent independent supplier, selling evaluation cameras and its Metavision software stack and co-designing the IMX636 and IMX646 sensors with Sony; its investors have included Sony, Bosch, Intel Capital, and Renault Group.^[12][17][27] iniVation, the University of Zurich and ETH Zurich spinoff that commercialized the original DVS and DAVIS designs, sold research-grade event cameras until SynSense acquired it in 2024 to pair event sensing with SynSense's low-power neuromorphic processors.^[10][11] Sony Semiconductor Solutions participates from two directions: as Prophesee's stacking and fabrication partner on the IMX636 and IMX646, and independently through its 2019 acquisition of Insightness, a second Institute of Neuroinformatics spinoff.^[16][17] Samsung Electronics has published a series of high-resolution DVS research chips through its semiconductor division but, unlike Prophesee and Sony, had not brought a standalone event sensor to market as a catalog product as of the mid-2020s.^[14] In China, CelePixel developed the CeleX line of event sensors before Will Semiconductor, which also owns OmniVision, acquired it in 2020, folding event-sensing work into a much larger conventional image-sensor business.^[15]

Scroll sideways for more →

See also

• Neuromorphic computing
• Computer vision
• Tactile sensing
• Vision based tactile sensor
• Dexterous hand
• Humanoid robot hands
• Robot manipulation
• Lidar
• Inertial measurement unit
• Edge AI
• Force torque sensor

References

1. Event-based Vision: A Survey, Gallego et al., IEEE Transactions on Pattern Analysis and Machine Intelligence (2022; arXiv 2019) ↩
2. Event camera, Wikipedia ↩
3. The First Analog Silicon Retina, History of Information ↩
4. Misha Mahowald, Wikipedia ↩
5. Home, Mahowald-Mead Prize for Neuromorphic Engineering ↩
6. Tobias Delbruck, Wikipedia ↩
7. A 128x128 120 dB 15 us Latency Asynchronous Temporal Contrast Vision Sensor, Lichtsteiner, Posch, and Delbruck, IEEE Journal of Solid-State Circuits (2008) ↩
8. Event-Based Tone Mapping for Asynchronous Time-Based Image Sensor, Frontiers in Neuroscience (describing Posch et al.'s 2008/2011 ATIS sensor) ↩
9. A 240x180 130 dB 3 us Latency Global Shutter Spatiotemporal Vision Sensor, Brandli, Berner, Yang, Liu, and Delbruck, IEEE Journal of Solid-State Circuits (2014) ↩
10. iniVation incorporates Dynamic Vision Sensor business from iniLabs, iniVation ↩
11. SynSense and iniVation join forces to form leading neuromorphic technology provider, SynSense, February 2024 ↩
12. Company presentation, Prophesee ↩
13. Prophesee (formerly Chronocam) announces initial closing of $19M funding round, GlobeNewswire, February 21, 2018 ↩
14. Samsung Presents 1.3MP Event-Driven Sensor with 4.95um Pixels, Image Sensors World, October 2020 ↩
15. DVS Company Celepixel Acquired by Will Semiconductor, Image Sensors World, September 2020 ↩
16. Sony Quietly Acquires Insightness, Image Sensors World, December 2019 ↩
17. Sony to Release Two Types of Stacked Event-Based Vision Sensors with the Industry's Smallest 4.86 um Pixel Size, Sony Semiconductor Solutions, September 9, 2021 ↩
18. IMX636 HD Sensor, Prophesee Metavision SDK documentation ↩
19. Event-Based Sensor GenX320, Prophesee ↩
20. High Speed and High Dynamic Range Video with an Event Camera, Rebecq, Ranftl, Koltun, and Scaramuzza, IEEE Transactions on Pattern Analysis and Machine Intelligence (2019) ↩
21. A Unifying Contrast Maximization Framework for Event Cameras, with Applications to Motion, Depth, and Optical Flow Estimation, Gallego, Rebecq, and Scaramuzza, CVPR (2018) ↩
22. Event-based Simultaneous Localization and Mapping: A Comprehensive Survey, International Journal of Computer Vision (2026) ↩
23. Neuromorphic Computing and Engineering with AI, Intel ↩
24. Dynamic obstacle avoidance for quadrotors with event cameras, Falanga, Kleber, and Scaramuzza, Science Robotics (2020) ↩
25. A Novel Event-Based Incipient Slip Detection Using Dynamic Active-Pixel Vision Sensor (DAVIS), Rigi, Baghaei Naeini, Makris, and Zweiri, Sensors (2018) ↩
26. Evetac: An Event-based Optical Tactile Sensor for Robotic Manipulation, Funk et al., arXiv (2023) ↩
27. Event-based Vision for Automotive Applications, Prophesee ↩
28. IDS and Prophesee to Advance Event-Based Industrial Vision, Prophesee, March 11, 2026 ↩
29. Qualcomm partners with image sensor pioneer Prophesee, eeNews Europe ↩