Shortlist der besten Einreichungen

Company: Lidwave
Author: Eyal Assa

Every conventional vision system, cameras and pulsed LiDAR alike, captures geometry. To understand whether something is moving, the system must compare frames over time and infer motion through computation. That inference is slow, fragile under poor lighting, and breaks down in fog, dust, spray, and direct sunlight.

Odem removes the inference step. Built on Lidwave's patented Finite Coherent Ranging (FCR(TM)) technology, Odem is a 4D coherent vision sensor that measures range, reflectivity, and instantaneous velocity (Doppler) for every single pixel, in the same optical instant. Velocity is no longer estimated, it is measured natively, as a fourth data dimension delivered directly from the sensor. The entire optical engine: laser, amplifiers, and detectors, is integrated onto a single silicon chip in a true monostatic architecture, where each channel both transmits and receives. The result is a calibration-free, vibration and temperature-resilient sensor that streams millions of depth-and-velocity points per second over hundreds of meters, and that can be manufactured at wafer scale. Odem is the first vision sensor to measure true instantaneous velocity at every pixel - optically and in real time - on a single, calibration-free silicon chip. It turns motion from something machine vision computes into something the sensor sees, delivering range, reflectivity, and Doppler velocity in one robust device that is manufacturable at a scale conventional LiDAR cannot match.

Technical details and advantages of the innovation:

1. Coherent FMCW sensing (based on our unique FCR(TM)): Each pixel emits an optical signal with unique phase and frequency manipulations, and mixes the return against a local reference arm. The beat frequency yields range and Doppler velocity simultaneously - no frame stitching, no motion estimation, no post-processing.

2. True single-chip monostatic design: Laser, amplification, and detection are integrated on one photonic chip, with transmit and receive co-located at the every channel. This eliminates the multi-component optical alignment that limits conventional LiDAR.

3. Calibration-free operation: Because there is no separate transmit/receive optical path to align, the sensor requires no field or factory calibration - enabling streamlined, large-scale production and drastically lower unit cost at volume.

4. Coherent rejection of interference: As a coherent receiver, Odem is intrinsically immune to ambient light, sunlight saturation, and cross-talk from other sensors, and it maintains clean depth and velocity data through rain, fog, dust, and snow.

5. Mechanical robustness: The monolithic design delivers high resilience to vibration and temperature swings without performance loss - critical for moving platforms and industrial environments.

6. Rich native data stream: Range + reflectivity + per-pixel instantaneous velocity, at millions of points per second, over hundreds of meters.

Relevance and application possibilities of the described innovation for the machine vision industry:

Per-pixel velocity reshapes the hardest problems in machine vision. Real-time segmentation becomes trivial when moving and static objects are separated by their measured Doppler signature rather than by frame-to-frame computation - slashing the perception compute load that today's AI-driven vision systems struggle with.

For robotics and embedded vision (AMRs, AGVs, drones, humanoids), Odem gives the platform an immediate, lighting-independent understanding of what is moving and how fast, enabling safe high-speed operation where camera-only stacks fail.

In industrial automation and in-line inspection, the same sensor delivers high-resolution 3D geometry and instantaneous motion in one device, supporting bin-picking, conveyor and high-speed process monitoring, and dimensional measurement of objects in motion. Its robustness to dust, mist, and glare opens harsh-environment applications - logistics yards, ports, agriculture, construction, and outdoor robotics - that conventional machine vision cannot reliably serve. And because it produces clean raw data without fusing multiple sensors, Odem simplifies system architecture and reduces the integration burden across the entire vision stack.

This is the data layer that physical AI has been missing: not just where things are, but how the scene is moving - measured, not inferred.

Unique Selling Point (USP):

Odem is the first vision sensor to measure true instantaneous velocity at every pixel - optically and in real time - on a single, calibration-free silicon chip. It turns motion from something machine vision computes into something the sensor sees, delivering range, reflectivity, and Doppler velocity in one robust device that is manufacturable at a scale conventional LiDAR cannot match.

When will the innovation be available?

Odem 2.5 engineering samples are shipping to evaluation partners today, production grade Odem 3 is scheduled for 2027.

Company: Medabsy GmbH
Author: Trang Nguyen

Medabsy is a CAD platform for designing machine vision systems -- the cameras and lighting that automatically inspect products on factory lines. Today they are largely designed by trial and error; Medabsy replaces guesswork with simulation. In our editor, the user imports a 3D model of their part and defines its material, texture, and the target surface defects, all derived from physical parameters. The inspection setup can be planned manually or with our optimization algorithm, using a database of market-available components. We then run physically accurate light simulation to generate synthetic datasets -- with pixel-perfect defect masks, full control over content, and rare edge cases on demand. This matters because real defect samples are scarce and hard to collect, yet reliable inspection AI depends on large, controlled, well-labeled training data.

Technical details and advantages of the innovation:

Currently, experts perform the inspection design manually in a lab by rearranging imaging hardware and assessing performance which can take weeks.
With Medabsy, the customer can design the digital twin of the inspection system on their own, by loading a 3D model of the inspected product, choosing components from hardware library and simulating the results.
The digital twin is further used for bill of materials, system specification and synthetic data Generation.

AI needs 1000s of labelled training images, which are not available for inspection - developers are lucky if they get 300 labelled images! And the dataset they get usually lacks the examples of the hardest defects.
Currently, the few available images are multiplied, by introducing rotation, perspective and color transformations - however, the defect still stays the same, no new information is obtained. With Medabsy, the customer can generate as many training images as they need! With each new image providing new appearance information. On their own, user can control surface appearance, dirt, defect characteristics and simulate any common or edge-case scenario they can think of.

Many are trying to use generative AI for synthetic data generation - however, pixel precise labelling is very difficult, and it is impossible to guarantee for the dataset content due to network hallucination.
Medabsy's synthetic images are generated using Monte Carlo simulation and stochastic geometry modelling for texture and defects. This means that the user has complete control over the results, knowing exactly which content is going to be generated for a given set of Parameters.

Added Value Proposition
- Sales efficiency increase
- Early risk assessment and cost planning
- Avoid penalties - stress test before deployment
- Train AI before having real data
- Train for edge-case scenarios

Relevance and application possibilities of the described innovation for the machine vision industry:

Using Medabsy software platform, it is possible to

* speed up inspection system design by 40% by designing and simulating inspection results virtually
* speed up AI project by 94% by using synthetic datasets
* simulate critical defects and edge cases which rarely occur in production
* simulate hardware performance before ordering hardware components
* quickly estimate inspection feasibility
* simulate a wide variety of defects using physical parameters such as type, width, depth, location (e.g. 0.2 mm wide crack, 2mm deep with branching)
* control texture parameters created by production process (milling, turning, casting...)

It is already tested by customers.

Unique Selling Point (USP):
Our software lets users virtually design machine-vision systems using real market-ready hardware, then generate photorealistic synthetic training data with pixel-perfect annotations. Unlike generative-AI tools, we create datasets via physical light-transport simulation and mathematical texture/defect models--no generative AI--ensuring physical accuracy, full controllability, and systematic coverage.

Company: PhotonicSENS
Author: Sergio Mantecón, Chief Commercial Officer, photonicSENS S.L.

Plenoptic imaging has been understood theoretically since the early 20th century. The physics are well established. What has been missing, until now, is a manufacturable, calibrated, industrially deployable implementation at a price point that makes it viable outside a research budget.
photonicSENS was founded in Valencia in 2015 with the specific aim of closing that gap. The company has developed a proprietary plenoptic (light field) camera platform that captures a full 2D image and a high-resolution depth map simultaneously, from a single sensor, in a single shot, in real time. There is no projected pattern, no second exposure, no external projector, and no on-site calibration procedure. The camera ships calibrated and outputs metric depth immediately.

The optical principle is straightforward to describe, though demanding to manufacture. A Micro-Lens Array (MLA) is positioned between the main imaging lens and the image sensor. The MLA enables to reconstruct multiple slightly different angular perspectives of the scene, encoding directional as well as spatial information into a single raw frame. This four-dimensional light field is then processed by a proprietary epipolar analysis algorithm that extracts per-pixel depth from the angular shift between adjacent sub-aperture views, without iterative correspondence search, and without the failure modes that affect stereo matching on repetitive or textureless surfaces.
The depth values that come out are measured, not inferred. Every reported point in the depth map corresponds to a genuine angular parallax measurement. The 2D image and depth map are inherently co-registered at pixel level, not because they have been aligned in post-processing, but because they come from the same optical path at the same instant. One of photonicSENS's key strengths is the engineering effort invested in an ultra-efficient implementation of the depth-estimation algorithms. Years of optimisation have brought the processing pipeline to the point where it delivers real-time point clouds  on a compact, self-contained system, without the powerful and expensive workstations or GPU clusters that comparable 3D approaches typically demand. This efficiency is what makes the embedded SMART line possible and what keeps total cost of ownership within reach for industrial deployment.
Making this work reliably at industrial scale required solving a manufacturing problem that has no off-the-shelf solution. Assembly of a light field camera demands sub-micrometre alignment between the MLA and the image sensor. photonicSENS developed a fully automated, proprietary assembly process using six-axis active alignment with real-time image quality feedback, implemented in a dedicated cleanroom facility in Paterna, Valencia (Spain). Each camera is individually calibrated against traceable reference targets to establish the pixel-to-depth relationship: a one-time factory process that holds for the operational life of the camera without further adjustment on site. 

Most machine vision sensors require in-situ calibration as part of every system installation. photonicSENS cameras do not.
The resulting product portfolio spans eight camera families (MICRO through VIVID+), covering depth precision from 0.6 µm to the millimetre range, with 1.4 and 2.1 megapixel depth maps and fields of view from under 2 mm to over 100 mm diagonal. The most recent addition to the range is the apiCAM SMART series, which integrates the full acquisition and depth-processing pipeline into an embedded industrial camera body. All computation runs onboard. The output is a calibrated 2D image and metric depth map, delivered over Ethernet with GenICam compatibility, external trigger support, and optional IP65 protection. No external PC, no GPU cluster, no specialist infrastructure.
A portfolio of granted patents covers the MLA optical design, the depth estimation method, and the manufacturing assembly process. A superresolution algorithm developed in-house reduces the spatial resolution penalty inherent to light field imaging by a factor of N rather than N², partially recovering the resolution trade-off that is the technology's principal limitation.

Technical details and advantages of the innovation:

The dominant 3D sensing methods in industrial machine vision, structured light projection, laser line triangulation, stereo vision, and time-of-flight, each carry constraints that are well understood by anyone who has tried to apply them at the boundary of what they can do.
Structured light systems need multiple exposures and a projector in the optical path. They work well in controlled conditions, but single-shot capture of moving targets is not possible, and specular or transparent surfaces cause the projected pattern to break down. Laser line triangulators require a scan axis, which limits throughput and rules them out wherever a static single-shot acquisition is needed. Time-of-flight sensors are fast and cover long ranges, but close-range precision is limited and multipath interference on reflective surfaces introduces systematic errors. Stereo vision requires two mechanically stable, thermally calibrated optical paths and correspondence algorithms that fail predictably on
textureless areas and repetitive patterns.
What none of these methods delivers is a true per-pixel depth map from a single shot, from a single compact sensor, with the depth and 2D data inherently aligned. Light field imaging delivers exactly that, which is why it has attracted sustained research interest for twenty years. The reason it has not previously appeared in production environments is the manufacturing difficulty described above, combined with the computational cost of earlier processing approaches. 

photonicSENS has addressed both. The manufacturing process exists, is automated, and is running. The processing pipeline is efficient enough for real-time operation on standard GPU hardware, and has been further reduced to embedded implementation in the SMART product line. On surfaces where established 3D methods struggle most, the advantages are concrete. Highly polished metallic components, machined screws, connectors, milled fasteners, produce specular reflection artefacts in laser profilometers: data voids, ghost points, degraded accuracy. The passive light field approach, combined with appropriate illumination, produces reliable depth data on these surfaces, enabling geometrical model fitting to detect dimensional deviations below 10 µm. On transparent substrates such as cover glass or optical windows, depth-based classification separates genuine surface defects from dust and foreign objects, something 2D imaging cannot address and most 3D methods handle poorly. In microelectronics assembly, the combination of 2.1 megapixel depth resolution at 2.4 µm precision and single-shot capture supports inline 3D metrology of wire bonds and solder joints at production cycle rates.
Light field imaging does have genuine limitations: Performance is strongest at short working distances and compact fields of view: the physics of the disparity-depth relationship mean that precision degrades with distance, and field of view and depth range trade off against each other within a given optical configuration. For applications requiring large fields of view, long working distances, or sub-micrometre precision at scale, other methods are often more appropriate. The photonicSENS contribution is not to claim the technology works everywhere. It is to make it work reliably where it is the right choice, and to provide sufficient portfolio range to address the majority of precision inspection scenarios that fall within its envelope.

Relevance and application possibilities of the described innovation for the machine vision industry:

The machine vision industry is well served by mature, proven 3D sensing technologies. What it has lacked is a commercially available light field option: a sensor that delivers true single-shot depth without a projector, without a scan axis, and without post-hoc registration between the 2D and 3D data streams. photonicSENS fills that gap.
For system integrators, a calibrated photonicSENS camera arrives ready to output metric depth. The SDK supports Python, MATLAB, LabVIEW, and Halcon, with full GenICam compliance in the SMART line. There is no alignment step between the 2D image and depth map because there is no alignment to do.
Applications that previously required a structured light projector and a 2D camera, two components to mount, calibrate, and maintain, can now be addressed with a single sensor. For OEMs and integrators working in space-constrained environments, this is a real simplification. The SILICON project, co-funded by the European Union under the ERDF Comunitat Valenciana 2021-2027 programme, provided independent third-party validation of the photonicSENS approach against established 3D solutions across agricultural sorting and automotive inspection use cases. The project confirmed performance and delivered hardware improvements including a reduction in energy consumption exceeding 50% and significant cost reduction relative to prior configurations.
Application areas include wire bonding inspection in semiconductor packaging and PCB solder joint volume measurement. In wire bonding, bonds as fine as 20 µm are reconstructed cleanly, allowing verification that each bond is correctly attached to the wedge and that the height and shape of the wire matches specification.  Glue bead continuity and geometry verification, thread pitch measurement on fasteners, and surface and subsurface defect classification on transparent components are further demonstrated use cases. In pharmaceutical manufacturing, a live production demo with Edmund Optics at MedtecLIVE Stuttgart in 2026 detected engraved tablet markings at under 100 µm depth, outputting simultaneous 2D and 3D data in a compact, line-ready form factor. The compact form factor and low weight also make the cameras well suited to mounting on a robotic arm, opening up emerging applications under active development: robotic bin-picking of metal components, inspection of additively manufactured medical implants, and large-area inspection through X-Y-Z stitching.
The 2026 product roadmap includes global shutter sensor integration, removing the current constraint on fast-moving targets, extended depth range configurations, and full-resolution output modes. A single camera, a single shot, a calibrated depth map, ready to use. Light field imaging has been discussed at VISION for years. photonicSENS is the company that has made it a product.

Unique Selling Point (USP):

The only camera that captures a full 2D image and a metric, per-pixel depth map in a single shot, from a single sensor, with zero projector, zero scan axis, and zero on-site calibration. Every depth point is a real angular measurement, not an inference. Factory-calibrated for life, running in real time on embedded hardware.

When will the innovation be available?

Available now. The core apiCAM platform (eight camera families) is in production and shipping. The SMART line is available in select models, with the global shutter variant launching Q3 2026.