THE ESSENTIALS IN 30 SECONDS
- Camera input from Iris is sent over an API to remote cultures of human iPSC-derived cortical neurons, and electrode spikes are decoded into outputs that return to the robot.
- Across nearly 500 simulations, 2D cultures reached up to 94% decoding accuracy in 150–240 ms; 3D spheroids reached up to 91% but sometimes took up to two seconds.
- Flat cultures stayed viable for more than 100 days, spheroids only a few days; an outside researcher called the work a very humble first step.
OpenMind, the robotics software company founded by Stanford bioengineer Jan Liphardt, has described a humanoid robot called Iris whose control loop runs partly through living human neurons rather than only through silicon. The work was reported this weekend by The Scientist and The Gadgeteer. Neither article links to a paper or preprint, and no OpenMind announcement detailing the results was found, so the figures below come from the team as relayed by those outlets.
How the setup works
Iris is meant to pick up basic social cues, such as whether a person seems engaged, and adjust its behaviour. Images from its cameras are sent through an API to neuron cultures hosted by two commercial providers, which the team rented instead of building a wet lab. The cells are cerebral cortex neurons derived from human induced pluripotent stem cells. One provider, in Australia, grows them as flat 2D sheets on electrode arrays; The Scientist indicates this is Cortical Labs' platform. The other, in Switzerland, grows them as 3D spheroids. Spikes recorded by the electrodes are converted into outputs that are sent back to the robot.
Reported results
According to The Scientist, the team ran nearly 500 simulations. Fresh 2D cultures decoded signals with up to 94% accuracy and responded in 150 to 240 milliseconds. Spheroids reached up to 91% accuracy, but some responses took as long as two seconds. The 2D cultures survived for more than 100 days, while the spheroids lasted only a few days. As the cultures aged, lag increased and behaviour drifted, so the robot can respond differently from one week to the next. Liphardt presents that drift as a possible source of individual personality, and argues that information held in living cells is harder to copy or steal.
Limits and open questions
The coverage does not make clear how much of the closed-loop testing involved the physical robot rather than simulation. Steve Potter of Georgia Tech, who was not involved, told The Scientist the project is "a very humble first step", noting that human hand and mouth control works in loops of roughly 10 to 100 milliseconds. Cortical Labs chief executive Brett Kagan said the hard part is making such systems reliable at scale in real settings. On ethics, Liphardt says these simple cultures raise no sentience concerns, though more complex structures or links to animals or people would change that; donor consent for the cells was obtained. The team says its next goal is keeping the neurons alive longer.
