A startup spun out of UCL research into how the brain works is building a new kind of AI model — a much better one, it claims — for hardware to help it move more intelligently through the physical world, and it has now raised $8 million in funding to help get its business off the ground.
Stanhope AI, as the company is called, is certainly not the first company to build AI systems to power autonomous machines. But its approach is different. Based on the neuroscience concept of “active inference,” Stanhope’s AI learns as a human brain is thought to learn: in real time and continuously correcting what it does as a result. Stanhope believes its approach produces systems that are not only easier (and cheaper) to start using in active environments, but also more responsive to spontaneous situations.
If Stanhope’s tech works as it hopes, it could be a major step forward for autonomous systems. Introducing more autonomy has been a major priority in military, industrial and other verticals adopting newer generations of equipment and see AI as the future of their industries.
But the reality of adopting it has been far from perfect in physical scenarios — as events involving two of the highest-profile AI drone startups, Stark and Helsing, have shown. The standard, costly approach to date has been to use huge datasets to “train” an AI. Stanhope thinks its own paradigm for how an AI can “learn” can improve outcomes drastically.
“This is very distinct from deep learning,” Stanhope AI co-founder Professor Rosalyn Moran said in an interview with Resilience Media. “This is actually built by us [to be like] memory in a human brain. You get all the good efficiencies and reasoning, but you also get that fundamental thing: agency. It’s a real generative model.”
Indeed, Moran believes that the current version of “generative AI” is not really generative at all. “They stole that term,” she said curtly of the hyperscalers who use it liberally these days.
GenAI startups have taken more than just a term. Hundreds of billions of dollars have been thrown at the training challenge by companies like OpenAI, Anthropic and self-driving car companies building their models, and they’re not done.
Stanhope’s technology is already being tested with (unnamed) robotics and drone companies, said Moran. You can see some of the tests here, which show how drones powered by Stanhope “agents” are introduced to new environments with obstacles and figure out how to navigate around them to get to their destinations. The startup will be using the seed money to continue expanding its research as well as that business funnel.
Frontline Ventures is leading the seed. New backers Paladin Capital Group and Auxxo Female Catalyst Fund; and previous investors UCL Technology Fund and MMC Ventures are also participating.
A heady time at UCL
Stanhope was the brainchild of Moran and another professor, Karl Friston, who had been working together in the neuroscience department at UCL. Moran had come to studying and working in the field initially from the world of electrical engineering.
Friston’s research into understanding how the brain works was being crystallised into the “free energy” theory, which in a very basic nutshell postulates that self-organising biological creatures, in order to survive, continually learn from the outside environment by matching it against what they already know. Given the brain works by way of a series of electrical impulses, Moran joined the team to grow that angle of his research.
It was a heady time at UCL (pun intended), with Demis Hassabis in the same hallway doing his own neuroscience research that led to him eventually founding DeepMind. So unsurprisingly, perhaps learning from their own environment, the pair saw an opportunity to spin out their ideas as an alternative for physical world AI that better reflected how they saw the brain to work.
One of the unique aspects of how Stanhope’s tech is implemented is that the processing for the systems happen not in the cloud — as is the case with autonomous system training now — but at the “edge,” on the device itself.
That creates a massive reduction in cost and energy usage in building and operating systems. Potentially, it provides a solution to how, say, a military drone work even in electronic warfare environments where communications with operators are cut off.
That also makes the technology much more applicable across a wider range of use cases, something that attracted investors to the problem.
“We’ve looked at so many new hardware vendors in defence tech,” said Nazo Moosa, a partner at Paladin. “What we were really excited about with Stanhope is that this goes beyond drones. Drones are important test bed but we looked at it from a different angle. If you look at the physical or digital world, it’s hard to get data hasn’t already been eaten up, but in the physical world data is the crunch point.” She noted that Waymo had to launch in the market to start to collect physical world training data, and while warehouses are easier environments, war zones are not.
“In a war zone, you get surprises,” she said. “You don’t have the capacity and flexibility to train millions of hours. That is what Roslyn and her team are fixing.”
