Robotics vendor OpenMind has unveiled Iris, a humanoid robot that uses lab-grown human neurons to help shape its decision-making.
The robot uses neurons derived from stem cells to help it interpret its surroundings and social cues and choose appropriate responses. The neurons work alongside conventional computing, including data fusion, perception and language models, rather than replacing them.
According to Jan Liphardt, a bioengineer at Stanford University and founder of OpenMind, the neurons also give Iris basic memory and set the robot’s mood — making them “less like a brain and more like a reflex.”
How it works
Liphardt and his team — in a demonstration disclosed on Nov. 13 — remotely connected Iris to two commercial neural platforms, both made of hundreds of thousands of human cerebral cortex neurons: 2D cultures from an Australian company and 3D spheroids from a Swiss company.
Each was grown on an electrode array that sends electrical impulses to and from the neurons.
An API linked Iris to the cells, using a camera to capture the outside world and feed it to the neurons for analysis. That neural activity then fed back into Iris’s behavior, enabling the robot to interpret basic social cues, such as a person engaging with it, and to adjust its responses accordingly.
While in its nascent stages, the research shows that neural cultures can operate within a robot’s control loop, with the potential to ultimately support learning, adaptation and decision-making.
Across nearly 500 stimulations, fresh 2D cultures decoded signals with up to 94% accuracy within 150 to 240 milliseconds. Spheroids reached up to 91% accuracy, with delays of up to two seconds.
Lian Jye Su, robotics analyst at Omdia, said the work is “meaningful but limited.”
“It is an important proof of concept that demonstrates the combination of living human cells with artificial control systems, as the biological system was applied to higher-level social and decision-making responses,” he said. “But its capabilities remain narrow, and the systems still rely heavily on conventional electronics for sensing, actuation and interfacing.”
Why OpenMind used neurons
Asked what prompted the project, Liphardt told TechTarget: “curiosity.”
“It’s critical to explore alternative compute architectures,” he said. “Right now, robots think with silicon, and animals think with wet electrochemistry. I’m a bioengineer who runs a robotics company, so at some point I was always going to ask what happens if you start to mix the two.”
“We are not building RoboCop,” he added, referring to the 1987 American science fiction action film. “We wanted to know whether living human neurons could do useful things for real robots. The answer is yes, but only a tiny little bit. But it’s a start.”
Su said biological neural networks could be much more energy-efficient than conventional designs, learn from limited data, and also excel at storing large amounts of information. They could make “flexible, context-sensitive decisions in unstructured environments,” he said.
Another potential benefit to using human neurons is security — an increasingly pertinent factor given ongoing concerns over AI governance and digital safety.
“A robot that is partly alive can’t be easily copied,” Liphardt said. “The information inside the neurons cannot be easily read out, which matters if it’s been living in your home.”
Biological neural networks in robotics
While Iris remains, as Liphardt says, a “basic” research project rather than a commercial product, possible applications are wide-ranging, particularly as developers look to create robots capable of operating alongside humans.
“Some people may prefer a companion or caregiver that’s a little bit alive, the way people prefer a biological plant to a plastic one,” he said. “And any work that helps us understand how neurons and electrodes get along is groundwork for brain-machine interfaces.”
Yet a number of challenges remain before the approach can be expected to reach commercial viability, the primary one being the maintenance of living neural culture. Others highlighted by Su include scalability, technological maturity, integration with existing systems, reliability and reproducibility, and integration between bio and mechanical interfaces.
OpenMind is looking to address some problems, with next steps for the team including improving the network’s stability, ensuring longer culture life, and creating more electrodes and better spatial architectures and cell compositions.
“Once reliability and culture lifetime are under control, we can start to look at learning and memory,” Liphardt said.
Su was more skeptical about the future of such an approach, though he said he does not disregard biological integrations into robots.
“I don't see a realistic route at this stage,” he said. “The most optimistic scenario points to a hybrid system. I would like to think that we can achieve biomimicry without adopting a fully biological system, and biomimicry is not necessarily the only way to achieve human-level intelligence.”
Scarlett Evans is a freelance writer with a focus on robotics and emerging technologies. Previously, she was assistant editor at IoT World Today, where she specialized in robotics and smart city technologies. Scarlett also has a background in the mining and resources sector, with experience at Mine Australia, Mine Technology and Power Technology.