Research / News
Why robots need to learn on the job: lessons from an artificial-life experiment published on CERN's Zenodo
A robot in a warehouse, a hotel or a home rarely works in the world it was tested in. Shelves move, floors get wet, people do things nobody planned for. So a fair question for any buyer is how much a robot can adjust on its own and how much it simply repeats what it was given. minimir, an artificial-life experiment designed and run by RoboHub founder Maksim Ilin, looks at a very simple version of that question.
At a glance · reported claims
0 of 40
Worlds that survived without lifetime learning
8 of 40
Worlds that survived with lifetime learning
128 × 128
Grid the creatures live on
DOI
Report on Zenodo, the open repository run by CERN
What minimir is
minimir (mini-world) is a small artificial world. Simple creatures live on a 128 by 128 grid, look for food, spend energy, reproduce and die. Each one carries a small neural network that passes to its offspring with mutations. Nobody gives them goals or rewards: whatever they end up doing, evolution has to find on its own. The project has its own site at minimir.ilinmaks.com.
The rules of this world change unpredictably. At random moments food that was safe to eat turns into poison. The question behind the experiment is an old one in biology: when does it pay for a creature to learn during its own life instead of relying on what it was born with? Learning costs something, so in a stable world fixed behaviour should win. In a world that keeps changing, learning might be the only way through.
How it was run and what it found
The experiment is preregistered: the hypotheses, conditions and success and failure criteria were written down and frozen before the first run, so the goalposts could not move once the data came in. The report on protocol B is published on Zenodo, the open research repository run by CERN, with the DOI 10.5281/zenodo.23713552.
The key result is a contrast. In worlds where lifetime learning was switched off, all 40 died out. With learning switched on, 8 of 40 survived. The report is also open about what did not work: the main preregistered hypothesis did not pass this round, because most worlds with learning also died at the first change of the food rule.
The survivors are the most interesting part. According to the write-up on Maksim's site, each of the 8 surviving worlds traces back to a few individuals that already had high learning rates when the rule changed. Biologists call this evolutionary rescue: a population survives because the right variety was already present when the environment turned against it.
What it does not show
minimir is an artificial-life model, not a robot study. Its creatures are tiny networks on a grid, not machines with arms, cameras and safety systems, and nothing in it was tested on a real robot. It does not tell you which robot to buy, and it does not prove that learning robots are better in general.
What it offers is a clean way to see one trade-off. Learning has a cost, and it pays off only when the world changes in ways that fixed behaviour cannot handle. Even then it is no guarantee: in minimir most learning worlds still died.
How to use the idea when judging robots
Start with how much your environment changes. On a fixed production line with the same parts every day, a robot with a fixed program is cheaper, more predictable and easier to certify. In a shop, a hotel, a home or a warehouse where layouts, items and people change all the time, a robot that can adjust or be retrained quickly has a real edge.
Then ask concrete questions. What happens when the robot meets a new item or a new layout? Does it adapt on site, or does it wait for an update from the vendor, and how long does that take? Who checks what it has learned before it goes back to work? If the answer is a remote operator, ask who that person is and what they see.
Live tests make this visible. In November 2026 six AgiBot G2 robots ran a Shanghai convenience store for 24 hours, and the cleaning robot failed when a full bin blocked its route. That is exactly the kind of moment where fixed behaviour runs out and something has to adapt, whether it is the robot or the people around it.
Open code and where to read more
Everything in minimir is open: the frozen preregistration, the code and the raw data are on GitHub, and the report is on Zenodo. The full story of protocol B and the plan for the next round, B2, is in the write-up on ilinmaks.com. Maksim's research profile is on ORCID, and there is more about him on the About page.
Robots in this story
Wheeled humanoid
Bipedal humanoid (home)
Bipedal humanoid (warehouse)
ChinaDelivery robot
Frequently asked questions
+What is minimir?
A preregistered artificial-life experiment by Maksim Ilin. Simple creatures with small inherited neural networks live on a 128 by 128 grid whose rules change unpredictably, and the experiment tests when learning during a lifetime pays off in evolution.
+Is this a study of robots?
No. It is an artificial-life model, and nothing in it was tested on a physical robot. It is useful as a way to think about when adaptation is worth its cost, not as evidence about any specific robot.
+What was the key result?
Without lifetime learning all 40 worlds died out; with it, 8 of 40 survived. The main preregistered hypothesis did not pass this round, because most learning worlds also died at the first rule change.
+Where is the report published?
On Zenodo, the open research repository run by CERN, under DOI 10.5281/zenodo.23713552. The code and data are on GitHub at akela1308/minimir, and the project site is minimir.ilinmaks.com.
+Should I choose a robot that learns on the job?
It depends on how much your environment changes. For a stable line a fixed program is usually cheaper and more predictable; for shops, hotels, homes and mixed warehouses, ask how the robot handles new items and layouts, how fast it can be retrained and who checks what it learned.
Sources & editorial note
- Zenodo (CERN) · minimir protocol B report · DOI 10.5281/zenodo.23713552
- ilinmaks.com · minimir on Zenodo: a preregistered artificial life experiment with a DOI, and what protocol B2 changes · 1 November 2026
- GitHub · akela1308/minimir (code, preregistrations, data)
- minimir.ilinmaks.com · project site
- ORCID · Maksim Ilin 0009-0009-4880-2153
Written by Maksim Ilin, founder of RoboHub, about his own research. minimir is an artificial-life model, not a robot study; the robot section is the author's interpretation for RoboHub readers.