Cubert’s cameras and developer tools are real; the podcast’s broader promise that AI makes hyperspectral inspection fast, expert-free and self-paying remains a vendor proposition without public outcome data.
By Land Offset PAI
Sept. 11, 2026
ULM, Germany — Cubert CEO and co-founder Dr. René Heine used a Sept. 9 inVISION Podcast episode to present artificial intelligence as the layer that can make the German company’s hyperspectral cameras practical for industrial vision. The public record supports a narrower conclusion: Cubert sells and documents snapshot hyperspectral cameras, and it maintains an open-source AI-development toolkit. It does not yet publicly establish the accuracy, false-alarm rate, throughput, customer adoption or return on investment needed to treat the episode’s more ambitious claims as proven production outcomes. 1 2 3
Hyperspectral imaging records many narrow wavelength bands for each pixel rather than the three broad red, green and blue channels of a conventional color camera. The additional spectral information can distinguish materials that look similar in RGB imagery, but it also creates a larger calibration, data-management and model-validation problem. A review of hyperspectral classification literature identifies calibration noise, spectral variability and limited or unbalanced labeled samples as persistent constraints; that is general field context, not a finding about Cubert’s systems. 8
Heine’s appearance was episode 29 of TeDo Verlag’s inVISION Podcast, published Sept. 9 and listed at 14 minutes. In the source video, he said Cubert’s AI layer helps users build and retrain models without an expert, and described anomaly detection in which a user trains on the acceptable product and the algorithm identifies what is different. Those are speaker claims about intended workflow, not released accuracy results. The episode names no model, data set, performance threshold, factory, customer or independently measured comparison with RGB inspection. 1 2
What the public record establishes — and what it does not
| Record or statement | Evidence classification | What it supports | What it does not support |
|---|---|---|---|
| Cubert’s ULTRIS XMR is specified as a 1,000-by-1,000-pixel, 61-band, 430–910 nm light-field snapshot camera, with maximum stated rates of 17 Hz at 8-bit and 12 Hz at 12-bit. An independent reseller lists the same core specifications. | Documented product capability | A commercial camera with defined spectral, spatial and acquisition specifications exists. | Classification accuracy, end-to-end decision latency, or performance in a customer’s environment. 4 6 |
The public cuvis.ai repository is Apache-2.0 licensed and describes a graph-based toolkit with supervised and unsupervised nodes plus preprocessing and postprocessing. Its README also says the project is still maturing and expects occasional breaking changes. |
Documented software availability and stated maturity | Cubert offers code intended to help build hyperspectral AI pipelines. | That a user can obtain a validated production model without domain expertise, labeled data, validation or integration work. 5 |
| Heine said Cubert was showing a thin-film wafer setup and food-sorting anomaly detection developed with two customers. | Planned or trade-show demonstration, as described by the speaker | Cubert said it had two application demonstrations. | A named customer deployment, acceptance test, production throughput, defect-detection rate, or commercial outcome. 1 |
| Cubert’s Sept. 4 wafer note reports a proof of concept on nominal 100, 300, 500 and 1,000 nm thermal-oxide wafers. It reports no more than 4% deviation from nominal in its demonstrated 300 nm to 1 µm range and says ellipsometry remains the reference for highest absolute precision on ultrathin films. | Vendor-authored proof of concept with an explicit limit | A bounded, disclosed application and the company’s own measured result. | Independent validation, a universal thin-film claim, or replacement of ellipsometry. 9 |
| Heine illustrated a business case in which a hypothetical RGB system at 95% specificity becomes 98% with hyperspectral imaging and pays for the added hardware through less waste or higher sorting quality. | Illustrative ROI claim | Cubert’s proposed value logic. | A measured Cubert result, a quoted price, a customer savings figure, or a contractual performance obligation. 1 |
Real hardware does not equal a validated AI outcome
Cubert says it was founded in 2012 by Heine, Rainer Graser and András Jung. Its current product page calls the XMR a one-megapixel, 61-band snapshot device, and Axiom Optics, a distributor, independently repeats the 430–910 nm range, 61 bands and 17-Hz maximum rate at 8-bit. Those matching technical descriptions substantiate the physical camera, though the distributor’s page is commercial product coverage rather than an independent performance trial. 3 4 6
The episode’s historical statement that Cubert introduced a first megapixel-resolution hyperspectral video camera in 2019 should remain attributed to Heine. Cubert’s own timeline says it introduced the ULTRIS 20 in 2019 as its first hyperspectral light-field camera; it does not independently substantiate the wider “first megapixel-resolution hyperspectral video camera” superlative used in the interview. 1 3
The distinction matters because a camera frame rate is not automatically an AI decision rate. Cubert’s XMR specification lists maximum acquisition rates, while a full inspection system still needs illumination, calibration, data transfer, model inference, rejection logic and validation for the relevant material mix. MITRE’s assessment of commercial snapshot spectral imaging similarly cautions that snapshot approaches involve spatial-versus-spectral tradeoffs and that deployable instruments must be matched to the mission rather than assumed to be plug-and-play. That assessment is not a test of Cubert, but it is a useful ceiling on claims that snapshot capture alone settles the operational problem. 4 7
The video also needs a timing qualification. Heine said the wafer demonstration measured a whole wafer “in a few milliseconds.” Cubert’s detailed wafer note describes a single exposure of “a few tens of milliseconds,” and its XMR page lists the maximum 17-Hz and 12-Hz rates noted above. The sources may refer to different portions of the process, but they do not publish a reconciliation or an end-to-end production-cycle time. The defensible statement is therefore single-exposure full-field measurement in a vendor proof of concept, not a demonstrated few-millisecond factory inspection cycle. 1 4 9
An AI toolkit is not an expert-free guarantee
Cubert’s code base is meaningful evidence that the company has moved beyond a slide-only AI message. The repository says the toolkit can assemble graphs from existing supervised and unsupervised nodes and provides preprocessing and postprocessing. It also states that working directly with Cubert session files requires installation of Cubert’s C SDK, and that the project is still maturing. Those published conditions are inconsistent with treating the episode’s “without experts needed” language as a general guarantee. 5
This is not an argument that the proposed workflow cannot work. It is an evidence boundary. The episode provides no confusion matrix, sensitivity and specificity definitions, false-positive burden, dataset split, model-drift result, retraining protocol, hardware configuration or independent benchmark. The academic literature explains why those omissions matter: hyperspectral data create high-dimensional processing demands, and models must contend with calibration effects, limited labels, overfitting and generalization. A production claim should therefore be tested against the actual materials, illumination, line speed and defect classes before it is treated as a result. 1 8
Demonstrations, not disclosed operational use
The two applications described at the end of the interview are best treated as demonstrations. The wafer work is the more concrete of the two because Cubert has since published its setup, sample thicknesses, wavelength window and stated limit near the lower end of its spectral range. The page is still written by Cubert, does not identify an outside laboratory or customer, and explicitly frames the work as a proof of concept. 1 9
For food sorting, the public episode describes a modern anomaly-detection demonstration but gives no performance data or customer identity. The 95%-to-98% specificity example is conditional language — “if” an RGB system achieves one figure, Cubert wants the higher one — rather than a published test result or contract term. It should not be reported as a three-percentage-point improvement, customer saving or deployment claim. 1
The source episode contains no claim of military, intelligence or combat deployment. Cubert’s separate defense marketing page promotes potential uses including reconnaissance, hazard detection and mine or improvised-explosive-device detection, but it identifies no government customer, procurement award, independent test or fielded operation. That page is evidence of a company marketing position, not evidence of operational military use. 10
Cubert’s strongest evidence is its cataloged camera hardware, a public AI toolkit and a technically bounded wafer proof of concept. Its unverified claims are the broadest ones: that users need no experts, that development falls from months to days, that anomaly detection generalizes from “good product” training, and that superior sorting economics will pay for the hardware. Those may be reasonable hypotheses for specific applications. The available public record does not turn them into demonstrated, general or operational facts.
References
- Source video: Episode 29 | Hyperspectral Imaging & AI — René Heine (Cubert). 1
- Apple Podcasts episode record: Episode 29 | Hyperspectral Imaging & AI — René Heine (Cubert). 2
- Cubert: About Cubert — Leaders in Hyperspectral Imaging. 3
- Cubert: ULTRIS XMR. 4
- Cubert Hyperspectral GitHub: cuvis.ai. 5
- Axiom Optics: ULTRIS XMR | Advanced Hyperspectral Snapshot Camera. 6
- MITRE: Commercial Snapshot Spectral Imaging: The Art of the Possible. 7
- Datta et al.: Hyperspectral Image Classification: Potentials, Challenges, and Future Directions. 8
- Cubert: SiO₂ Thickness Detection on Silicon Wafers. 9
- Cubert: Defense. 10