Tech • AI • Robotics • Game
New AI Robot Is Run By Living Human Brain Cells (Absolutely Terrifying)
5/10TL;DR
Indiana University researchers reported that human cortical organoids linked to a small humanoid robot improved a simple decision task from 72% to 94% in about two days, an early result that highlights both the promise and ethical complexity of biological computing.
KEY POINTS
Human cells guided a robot
The system, called Brainoware Bot or Brainobot, uses living human neurons grown from embryonic stem cells into a cortical organoid. The tissue sits on a chip with 26,400 electrodes inside an incubator at 98.6°F, while a separate educational robot receives movement commands over Bluetooth. The organoid never enters the robot; it functions as a remote biological controller.
How the setup works
A camera on the robot sends images to a conventional PC, which compresses them into simple patterns, sometimes a grid of just 36 black-and-white squares. Those patterns are converted into electrical pulses delivered through only six electrodes to the organoid. The chip records the neurons’ responses, and software translates the activity into one of three actions, such as moving left, right, or forward.
Tasks stayed simple but measurable
The robot first learned basic left-right choices, then a block task in which it responded differently to the letters I and U. It also chased a laser dot across the floor, making one decision every 5 seconds and in one test reaching the target in about 35 seconds. The most complex demonstration had it assembling blocks into the Indiana University logo.
A toxin test suggested the neurons mattered
To test whether the organoid was actually contributing, researchers applied tetrodotoxin, the nerve poison associated with puffer fish. Performance on the block task fell from roughly 90% to about 30%, near random guessing for a three-choice system. The camera, computer and robot still worked, but the decision output collapsed, indicating that active neural signaling was central to the result.
The tissue appeared to learn
Across four training rounds over about two days, one organoid’s score on the block task rose from 72% to 94%. Researchers also reported visible changes in coordinated firing patterns, from sparse activity to thicker bundles of linked responses. When they added a drug that prevents neurons from strengthening connections, performance stalled around 62% to 64% and those network changes largely disappeared.
Training transferred between tasks
An organoid trained only on laser chasing later scored 86% when switched to the block task for the first time. A fresh organoid with no prior training scored about 70%. That suggests some generalization, though still within a narrow set of highly simplified tasks.
The comparison with AI has limits
Against a small artificial neural network trained for the same four rounds, the organoid averaged in the high 80s, while the software model stayed in the low 60s. But the comparison was limited: given 100 rounds, the artificial network reportedly reached nearly 100%. The main claim is not superiority to modern AI, but faster learning from very little data.
Important caveats remain
The headline 94% was the strongest single run, while the average came from only five organoids. The robot’s walking and grasping motions were pre-programmed, the PC handled most of the vision processing, and the readout software was retrained after each round, meaning some gains came from ordinary code rather than the tissue alone. The broader system also remained energy-intensive despite the neurons themselves using little power.
A wider field is emerging, with ethical scrutiny
Similar work includes Cortical Labs in Melbourne, which has developed systems using up to 800,000 human neurons and sells a platform priced around $35,000. Other groups have built machines powered by living muscle or covered with living skin, pointing to a broader push toward hybrid biological machines. Bioethicists, including Insoo Hyun and Paola Arlotta, have called for global oversight, warning that larger arrays of organoids could eventually raise questions about memory, continuity of experience, and whether any features of consciousness might emerge.
CONCLUSION
The Indiana results suggest living human neurons can contribute to fast, low-data learning in a robotic system, but only in tightly constrained tasks supported by substantial conventional computing. As organoid-based machines grow more capable, the scientific challenge will increasingly be matched by an ethical one.
Ask a question