Connect with us

NEWS

Visage Builds Adaptive Training Inside the EU Emotion Ban

Visage Technologies is using a €3.12 million EU-backed grant to read stress in ship simulators, inside the AI Act’s ban on classroom emotion inference.

Published

on

Visage Technologies has started an EU-backed project to build adaptive training technology that reads a trainee’s stress and mental load. The Croatian unit is leading work with a €3,117,060.40 budget, of which the EU is covering €1,904,735.93, from 1 April 2026 to 1 April 2029.

The first test bed is a ship-navigation simulator. The legal setting is harder: since 2 February 2025 the Union has banned AI that infers emotions in workplaces and education institutions, with a narrow medical and safety exception.

Brussels Is Paying for the Sensing It Restricted

The grant is an IRI S3 research-to-product award in Croatia, seven months into a three-year run. Visage AI Labs at Visage Technologies d.o.o. in Zagreb is the lead. Partners are the University of Zagreb Faculty of Humanities and Social Sciences and the University of Rijeka Faculty of Maritime Studies.

The company’s English project page sets the published funding figures for the project at those euro amounts, so the Union is paying 61% of the bill. Visage and its partners cover the remaining €1,212,324.47.

Hrvoje Šamija, chief business development officer, described the bet as two signals, stress and cognitive load, captured on whatever camera or wearable is in the room and processed on the device.

If someone is overloaded with stress, they will have a hard time learning anything. It can also work the other way. If a person is bored, we can make them more engaged in the task. We can integrate any camera and any hardware, and everything we do runs on the edge

Hrvoje Šamija, Chief Business Development Officer, Visage Technologies

Bruno Rosan, a senior R&D engineer at the firm, said the stack is being built in house, from data processing and machine learning models through to deployment on devices. The same page says the design already folds in the GDPR, the EU AI Act and the NIS2 cybersecurity rules, including limits on tools that monitor cognitive and emotional states.

One Medical and Safety Carve-Out

Article 5 of the EU AI Act has, since 2 February 2025, prohibited placing on the market, putting into service, or using AI systems to infer emotions of a person in workplaces and education institutions. The only written way out is a system intended for medical or safety reasons. Outside those rooms, emotion recognition is not banned outright. It is high-risk, and since 2 August 2026 deployers must tell the people exposed to it.

Recital 18 treats happiness, sadness, anger, surprise, disgust and several other moods as emotions. It leaves physical states such as pain or fatigue outside that definition. A drowsiness alert for a driver or a pilot does not need the safety exception, because it is not emotion recognition in the Act’s sense. A score of classroom engagement taken from a face or a voice sits inside the ban.

Visage’s current research targets four dimensions, and the team says the final set depends on what the models can estimate reliably after data collection.

FOUR TARGET DIMENSIONS

  • Cognitive load: Mental effort required to process information or perform a task, closer to fatigue than to mood.
  • Engagement: Involvement, attention and interest in the task, a measure workplaces are not free to harvest from faces.
  • Valence: Whether an experience registers as positive or negative, a classic emotion axis.
  • Arousal: How activated the person is, independent of whether the feeling is pleasant.

Cognitive load is the cleanest of the four under the Act. Valence and arousal are the ones that look like the banned practice once the software is plugged into a school platform or a staffed simulator. Ship training is how the project argues for the safety door.

Rijeka’s Simulators Already Have the Cameras

The University of Rijeka Faculty of Maritime Studies already runs a Navi-Trainer Professional 5000 full-mission setup for STCW watchkeeping and bridge-team courses. Three bridges sit in two rooms, with 49 mathematical ship models and 48 navigation areas. Each room already has three surveillance cameras, and the bridges can be joined into one shared scenario.

Assoc. Prof. Dario Ogrizović is the faculty lead, with Marko Strabić and Matej Šuper on the Rijeka faculty’s published project sheet. Visage and Zagreb’s humanities faculty have already walked the halls, looking at how the simulations run, which devices would capture the signals, where those devices could sit, and which of the four dimensions matter on a bridge.

That visit is the practical start of ACTrainer. The hardware problem is smaller than the model problem. The cameras are on the bulkhead. The open question is which states can be estimated well enough, in a fake storm at a dock in Rijeka, to stand in for a real night watch.

What Visage Already Ships for Cars

Visage already sells face analysis that estimates six emotions plus a neutral class, and in-cabin driver monitoring for anger and fatigue, including gaze, eye closure, head pose and signs of fear. The company was founded in 2002 in Linköping, Sweden, opened its Zagreb office in 2010, and says more than 300 clients use the face-tracking kit. A separate automotive unit works on forward-looking perception for driver-assistance systems at large carmakers. That unit is not this grant.

The training project is an extension of the cabin work, not a new company. Edge processing is already how visage|SDK is sold: analysis on the device, with no photos or names stored by default. Šamija’s line about any camera and on-device inference is the same architecture pointed at a simulator instead of a windscreen.

Most compulsory e-learning still grades the click. It knows whether the right box was ticked. It cannot tell whether the person was overloaded, bored, or reading something else. That gap is what ACTutor is for, and it is also where the education ban is tightest.

The SDK and the Two Plug-Ins

The grant funds one sensing core and two products that sit on it. ACInsight SDK is meant to fuse cameras, microphones, wearables and other sensors into an estimate of the person’s state. ACTutor is a plug-in for existing e-learning platforms. ACTrainer is an add-on for professional simulators, with ship navigation as the first case.

THREE MODULES ON ONE CORE

Product Job Where it sits
ACInsight SDK Combine sensor streams and estimate cognitive-affective state Core platform, any supported hardware
ACTutor Adapt curricula, materials and delivery from those estimates Existing e-learning platforms
ACTrainer Test whether a scenario loads a trainee like the real job Professional simulators, starting with ships

ACTrainer is the piece that can be framed as safety. If a simulated collision does not raise load the way a real one would, the session is a poor rehearsal, and the officer still meets the certificate. ACTutor is the piece that looks like classroom emotion AI: it changes the lesson because of how the learner feels. The project page is careful to say the four dimensions are preliminary, and that some states may be hard to elicit in data collection. That hedging is also a regulatory hedge.

Human Error Still Runs Through Most Casualty Files

Ship simulators exist because people, not hulls, still dominate the casualty record. The European Maritime Safety Agency’s 2025 annual overview of marine casualties, covering 2015 to 2024, put the human element in 78.8% of investigated casualties and incidents.

EMSA HUMAN ELEMENT, 2015 TO 2024

  • Human element share: 78.8% of investigated marine casualties and incidents involved the human element when action and behaviour were taken together.
  • Human action events: 64.5% of accident events were linked to human action, a different count from the behaviour factors.
  • Human behaviour factors: 50.5% of contributing factors were related to human behaviour.
  • 2024 volume: 2,659 marine casualties and incidents were reported in 2024.

Bridge-resource courses already drill fatigue, stress and communication. They still score the run by whether the trainee called the right heading, not by whether the scenario actually loaded working memory the way a night transit would. ACTrainer’s pitch is to measure that gap. If the models hold, instructors get a second instrument besides the answer key.

A Tighter Ban Around Classroom Software

ACTrainer on a maritime simulator can argue it is there for safety, in a trade where the human element still sits in most investigation files. Fatigue-style load is also the dimension recital 18 treats as a physical state rather than an emotion. That is the lawful path the grant is walking first.

ACTutor does not get that path for free. A plug-in that adapts school or workplace lessons from valence, arousal or engagement is the use Article 5 names, unless a medical or safety reason is real and necessary. High-risk duties still apply even when the ban does not. Fines for prohibited practices run to €35 million or 7% of worldwide turnover, whichever is higher.

The work ends on 1 April 2029. By then ACInsight has to estimate at least one state well enough to ship, ACTrainer has to show a simulator can be scored for realism rather than only for correct actions, and ACTutor has to live inside a ban that already treats classroom emotion inference as unacceptable risk.

Harry is the editor of IAQABA, an independent publication he owns and runs. A decade in journalism, beginning as a reporter and now as the editor of his own titles, has left him with a clear test for what deserves a story: it has to change what a reader knows or decides, and it has to rest on something he can point to. That rules out recycled press releases, forecasts with no data behind them and rumours that no document supports. It leaves room for a great deal, and the site covers news, business, science and technology alongside sports, entertainment and lifestyle, with travel, auto and gaming given the same standard rather than lighter treatment. Sources are primary wherever possible: the regulator's filing, the company's own statement, the transcript, the dataset, or the product on Harry's desk. Figures are checked before they are published and rechecked if a reader questions them. Mistakes are corrected under a published policy. Readers across the world can reach him directly at support@iaqaba.com.

Click to comment

Leave a Reply

Your email address will not be published. Required fields are marked *

Trending