This year, a Swedish startup, IntuiCell, released a video of a four-legged robot dog “Luna,” which learns to stand entirely on its own, and adapts through sensory feedback and real-world interactions, much like a newborn animal, with no pre-programmed intelligence or instructions.

It marks a significant shift from the notion of “pattern recognition at scale” in robotics to embodied, autonomous learning agents capable of improvising, adapting, and operating with genuine intelligence – and it’s just the beginning.

I spoke to CEO Viktor Luthman to learn more. 

IntuiCell aims to build AI that truly understands and learns, modelled on how brains work, not just mimicking the brain, but emulating its learning mechanisms.

Unlike most AI systems today — which depend on large static datasets, backpropagation, and a clear separation between training and inference — IntuiCell has developed a physical AI agent that learns continuously, mimicking the adaptive, real-time learning of biological nervous systems. This approach enables the system to operate effectively in dynamic environments where traditional AI often fails.

As CEO Viktor Luthman explains:

“They separate training and inference — we don’t. With us, learning never stops. It happens in real time. We’re building the brain for all non-biological intelligence.”  

In other words, a machine can learn directly from its surroundings—through real-world experience and interaction—without needing pre-training, massive datasets, or running endless simulations in the background.

A sci-fi vision too bold to ignore

According to Luthman, he’s spent his entire career building startups “within bleeding-edge science. I absolutely love working with top professors and research teams to commercialise their findings.” His last startup, Premune, was acquired in 2020. 

He came into contact with Intuicell through an old friend who was head of tech portfolio at Lund University’s holding company, and told him about a group of neurophysiologists in England with radical findings on how the brain predicts the world. 

“They had this sci-fi vision of building AI that works like the human mind. It sounded too crazy for me to ignore.”

Luthman visited the startup, fell in love with their contrarian mindset, and joined as CEO in January 2021. 

“I’m one of those people who think Europe needs more bold visions and deep breakthroughs. So I joined as the second employee. They already had a hacker genius translating the research into code.”

Turning neuroscience on its head

By translating decades of brain research into real-time learning systems, IntuiCell has carved out a unique space in AI.

While it’s easy to think of the tech as something new, it didn’t happen overnight. Rather, IntuiCell emerged from over 30 years of contrarian research at Lund University. 

Luthman contends that its researchers turned conventional neuroscience upside down:

“They didn’t win many popularity contests. Their work was hard to fund and didn’t get published in the most prestigious journals.

But five years ago, they found a way to communicate what their discoveries could mean for AI clearly. They weren’t AI researchers — that’s what I like about it. We see ourselves as the odd bird in the AI space. We don’t come from AI, we come from a deep understanding of how the brain works.

Over the past five years, we’ve translated and validated those findings in software. That’s what makes us unique.”

According to Luthman, Intuicell probably understands better than anyone how individual neurons can autonomously prioritise problems, make decisions, and solve local challenges.  Those mechanisms scale, from how an amoeba learns to avoid danger and find nutrients, all the way to how a 7-year-old learns to play football.

An abundance of usecases

To be clear, Intuicell is not selling a product or app — rather its building infrastructure — the brain for all non-biological intelligence.  According to Luthman, this can include both physical and digital agents, not just robots. Instead, the technology can be applied anywhere machines need to learn and adapt on the fly.

While IntuiCell started with robotics, for example, teaching a robot how to pick up garbage — and generalising that skill to any building or pavement — or learning how to clean a table, regardless of height or clutter,  the technology as the potential to power robots in space, underwater, disaster zones, last-mile delivery — anywhere requiring real-time adaptability.

The company did a feasibility study with ABB, through their SynerLeap program, which revealed that its system could perform anomaly detection in engine health monitoring, with no fine-tuning or pre-training.

Luthman detailed: 

“Take a service dog. You don’t preload it with everything it might encounter. You teach it. It interacts, learns from experience, understands intent, and refines its behaviour over time. We want to do that with machines. Create systems that can generalise—not just follow rigid instructions.”

He contends that if we want robots to go to Mars and build habitats, they need to learn and experiment on their own in unpredictable environments.

“But you don’t have to go to Mars to make it relevant. The real world is already the most dynamic system we know. Every millisecond is new.”

Further, IntuiCell is efficient. Luna runs on a few thousand neurons using off-the-shelf GPUs. There’s no massive cloud infrastructure, no country-sized data centres, but instead efficient, distributed learning. 

According to Luthman, “just a few hundred neurons were enough for our system to learn a normal engine state and detect new anomalies across different engines. No manual intervention, no costly deployment. That wasn’t about making money—it was about proving we can solve real problems.”

Challenging the AI status quo

I was interested in what Luthan says to sceptics. Luthman pushes back against the obsession with scale, arguing that real intelligence starts small:

“Some people scoff—”If an amoeba could do anomaly detection, is that really intelligence?” And I say: if you could replicate how an amoeba learns — which is fundamentally different from any existing tech — you’d be very close to advanced learning.

People are obsessed with bigger models and more data.

But we’re flipping that entirely. We’re solving learning from the smallest unit up. That’s how intelligence evolved on this planet, and it’s the only way to make scalable, efficient AI.”

In terms of commercialisation timelines, Inuticell’s go-to-market strategy is focused on the next couple of years, although the company is fortunate to have found aligned investors who aren’t pushing for premature monetisation. 

“We’ve been clear from the start: we needed to get the foundation right first. Neurons, synapses, sensors, learning algorithms—and our first problem-solving component, which we call the spinal cord. That’s what drives Luna,” shared Luthman.

The company plans to start with two or three high-value projects, once its scaled its tech and interfaces, it will open it up for broader applications.