Barcelona
18/06/2026
18/06/2026
🦊
Gorgs
IA Studio
IA Studio
Gorgs IA Studio · Catalonia
12 → 17 June 2026 · intensive week
Four sessions, from what AI is to a robot that works.
One intensive week, fifteen teams, a laptop and a robotics kit each. They sorted real AI from mere automation, trained their own vision models and broke them on purpose, then built and tested a robot for a Sustainable Development Goal they chose.
4Sessions
12hin the lab
15Teams
The week, session by session
Four sessions at Gorgs IA Studio12 June
AI Detectives.
15 June
Reactivating the machine.
16 June
Designing an AI for a robot.
17 June
From prototyping to testing.
01
12 June
Done
Done
AI Detectives · how machines actually learn.
▸
What we ran
The week opened with a brainstorm on where AI already sits in daily life. Each team then sorted scenario cards (voice assistants, automatic doors, navigation apps, spam filters, calculators, traffic lights) into three piles: definitely AI, sometimes AI, definitely not AI, and wrote its own working definition. The long middle activity set two ways of classifying against each other: first sorting shape cards by explicit rules given by the facilitators, then sorting a second set where the criterion stayed hidden and had to be inferred from examples. Test cards and deliberately ambiguous ones showed what a forced decision looks like. The session closed with Teachable Machine: each team trained a live vision model on concave and convex shapes, tested it under different lighting, distances and orientations, then filled a synthesis sheet: one idea, one piece of evidence, one question.
Who
15 teams
Mood
Sceptical, then hooked. Sorting the cards turned into an argument about the automatic doors, and nobody wanted to give up their pile.
Skill tree
AI literacy
supervised learning
training data
forced decision
Best practice
Letting the youth train a model themselves in the last half hour, right after they have argued about what counts as AI, is what moves the machine from magical to explainable. They stop describing AI as infallible and start describing it as pattern recognition learnt from data they chose.
02
15 June
Done
Done
Reactivating the machine · breaking your own model.
▸
What we ran
The teams trained a binary image classifier, then spent the hour trying to break it: changing the lighting, the angle, the distance, the background, adding accessories the model had never seen, and logging every failure in order to deduce the implicit rules it had learnt. A short discussion moved from technical failure to social consequence, using two real cases, blink detection in cameras and automated CV screening, to separate a bias that is a bug from a bias that lands on people. The longest activity was unplugged: a perceptron built from plastic cups, water and measuring beakers, where input weights, the activation threshold and the hidden layers became something you could pour. The session ended on a brainstorm of social uses tied to the Sustainable Development Goals, each idea written with the problem it targets, the pattern the AI would have to recognise, and the risk of misuse.
Who
15 teams
Mood
Competitive and slightly gleeful. Finding the exact accessory that fooled the model became the game of the morning.
Skill tree
model testing
edge cases
algorithmic bias
perceptron
Best practice
Asking the youth to break the model they have just trained works far better than showing them a broken one. Once they have found the accessory that fools it, the question of who a real system fails lands on its own, without any lecturing.
03
16 June
Done
Done
Designing an AI for a robot · starting from a need.
▸
What we ran
The session turned from technique to design. Each team picked a Sustainable Development Goal challenge and had to argue that it was actually feasible, not merely desirable. They then filled an empathy map centred on one precise place or user (a classroom, a garden, a recycling area) and condensed what they found into a single problem statement, reframed as a "how might we" question. Ideas were filtered through three criteria: useful, possible, and trainable or programmable. The rest of the morning went into a first layout of the whole chain, from input to AI classification to robot action, as a storyboard, paper mockup, early training run or first piece of robot code.
Who
15 teams
Mood
Slower and more argumentative. The feasibility filter killed several favourite ideas, and the teams had to defend what was left.
Skill tree
empathy mapping
problem framing
SDG design
system architecture
Best practice
Making the youth argue for the feasibility of their challenge before they design anything is what keeps a week like this honest. It is also the moment the project stops being a technical exercise and becomes theirs, because the need they picked is one they can point at in their own building.
04
17 June
Done
Done
From prototyping to testing · and defending it out loud.
▸
What we ran
The last session opened with each team naming what was still missing, splitting the roles and setting a minimum viable product for the day. They then finished the functional core: training or retraining the model, wiring the classification output to a robot action, and firming up the physical structure. Cross-testing followed: each team demonstrated to its peers with expected cases, edge cases and cases nobody had planned for, with presenters and observers both logging what happened. The week closed on presentations of three to four minutes covering the project title, the Sustainable Development Goal it targets, the user context, the classification logic, the robot action, the tests that passed, the errors found, and what the team would change next.
Activity
Del prototip a la prova
Who
15 teams
Mood
Deadline energy, then relief. The unexpected test cases given by the other teams were the part everyone remembered.
Skill tree
prototyping
test logs
error analysis
public presentation
Best practice
Having the youth test each other's prototypes, rather than only their own, surfaces the cases the builders could not imagine. They end the week presenting their errors alongside their successes, and that is the habit worth keeping.
What the week shipped
Fifteen teams, one full chain eachTrained
A vision model per team.
Trained on the images the team chose, then stress-tested under changed lighting, angles and backgrounds until it failed.
Built
A robot prototype per team.
The classification output wired to a physical robot action, aimed at a Sustainable Development Goal the team had argued for itself.
Defended
A test log and a pitch.
Expected, edge and unexpected cases written down as they happened, then a three-minute defence in front of the other teams.