Cortical Labs Uses Lab-Grown Neurons to Steer a Robot Dog Toward a Target

A robotic dog in Australia has been steered toward a target on a simple map by a small cluster of neurons grown in a lab. The cells sit on a chip, and they are doing the directing.

The work comes from Cortical Labs. The team grew the neurons directly on the chip and fitted it with electrodes that carry electrical signals in both directions. Signals go into the neurons, and the neurons' own activity comes back out. That returning activity is translated into movement for the attached robot.

Brett Kagan, co-founder and CEO of Cortical Labs, called the robotic dog a simple demonstration. He said it makes the prospect of applying biological systems in the real world tangible. Over the longer term, he said, the approach could support robots and autonomous systems that must make decisions in unpredictable environments.

His argument starts from a gap in robotics. Machines have become remarkably capable at physical control, he said, yet they still struggle with the higher-order control needed to learn quickly or adapt when conditions change. Walking, running and acrobatics are impressive, but deciding where and when to act remains hard with traditional methods.

The underlying appeal is the brain's economy. Living neurons as the basis of computer chips promise advantages in energy efficiency and adaptability. However sophisticated AI has become, the human brain runs on far less power and still handles the unfamiliar better.

The robot dog follows earlier Cortical Labs demonstrations in which clusters of neurons completed computer tasks, including playing Pong and Doom. Here the idea moves out of the screen and into a physical machine.

A related demonstration was run with Honda Research Institute Europe, entirely inside a virtual system. The aim was the same: guide an agent around a basic grid in search of a reward. The setup also matched, with electrical signals stimulating neurons on a chip and the chip sending back messages of its own. By varying the characteristics of the signal and giving feedback on the agent's moves, the researchers could train the living chip and interpret its movement. Over more than 4,000 hours of interaction between chip and agent, the team tested almost 1,300 different signal combinations.

Daniel Tanneberg, a Senior Scientist at Honda Research Institute Europe, said the collaboration lets researchers investigate fundamentally different approaches to intelligent agents. Using living cells as the core computational substrate brings many challenges, he said, but also potential in energy efficiency, robustness and adaptation that can now be explored and evaluated.

The technology is early. It is not yet particularly fast or efficient, and the Honda work involved a fair amount of trial and error. Next steps are scaling up the chips to handle more complex tasks in less time, since real-world robots will need to do much more than cross a simple grid. Researchers also want to learn how these chips retain information and what kind of memory they might have.

Johnson Zhou, a Research Software Engineer at Cortical Labs, framed the central difficulty: the challenge is not simply growing neurons on silicon but finding reliable ways to communicate with them, train them, and identify which approaches produce consistent results. The research, he said, provides a framework for testing those questions at scale. Possible future uses include neuron-controlled robots working in remote or hazardous conditions, or machines that give personalized responses that improve over time.