Scientists Use Living Human Brain Cells To Decode Braille at 83% Accuracy and Even Guide a Robot Dog

A University of Bristol study by Tianyi Liu, Hemma Philamore and Benjamin Ward-Cherrier, published 16 October in Frontiers in Artificial Intelligence, used human forebrain organoids about 500 micrometres in diameter on microelectrode arrays to decode Braille letters, reaching 61 per cent average accuracy with a single organoid and 83 per cent when combining three, according to the paper. A support vector machine classifier working directly on the encoded stimulation data with no organoid reached 82 per cent under clean conditions, limiting the organoids' advantage. Separately, Melbourne-based Cortical Labs and Honda Research Institute Europe tested about 1,300 parameter configurations across 26 neural cultures over more than 4,000 hours of real-time interaction, with biological agents outperforming optimised Deep Q-Network software agents on a robot-dog navigation task under an equivalent interaction budget.

Scientists Use Living Human Brain Cells To Decode Braille at 83% Accuracy and Even Guide a Robot Dog Researchers used lab-grown neural tissue to process tactile signals, while separate experiments explored biological networks controlling a robot dog Scientists have used living human brain cells to help a computer distinguish Braille letters with up to 83 per cent accuracy, according to new research that also describes how similar 'biohybrid' systems have been used to steer a robot dog in a navigation task. The study, published on 16 October in Frontiers in Artificial Intelligence , was written by Tianyi Liu, Hemma Philamore and Benjamin Ward-Cherrier at the University of Bristol. It tested whether human forebrain organoids could process tactile information https://www.ibtimes.co.uk/robotic-arm-emarm-touch-recognition-obstacle-detection-1826415 . Organoids are laboratory-grown, three-dimensional tissue models containing neural cells, not miniature brains. Turning Braille Into Brain Stimulation The organoids, about 500 micrometres in diameter, were cultured on microelectrode arrays that could both stimulate the tissue and record its electrical activity. The team converted Braille letters into patterns of electrical stimulation and recorded the organoids' responses. A machine-learning classifier https://www.ibtimes.co.uk/ai-researcher-challenges-black-box-view-language-model-audit-1826618 known as a support vector machine then learned to identify letters from those responses. The organoids themselves were not trained. The paper describes an open-loop task, meaning the tissue received no feedback. Accuracy, Limitations and Noise Tests Using responses from a single organoid, the classifier identified letters with an average accuracy of 61 per cent. Combining the responses of three organoids raised this to 83 per cent. The three-organoid set-up was also more robust to several types of artificial noise introduced by the researchers. However, a classifier working directly on the encoded stimulation data, with no organoid involved, reached 82 per cent under clean conditions. In that comparison, the advantage of including the organoids was limited. The study does not show that the organoids learned to read Braille as a person would. The paper states that the organoid material was obtained commercially from FinalSpark and that no new ethical approval was required for the experiments, as the material already existed. From Lab Dishes to Robot Dog Navigation Separately, Melbourne-based Cortical Labs and Honda Research Institute Europe have investigated how cultured neural networks interact with digital environments, including work in which a quadruped robot platform https://www.ibtimes.co.uk/china-humanoid-robots-everyday-work-challenge-1821099 was guided through a navigation task. Their research on embodied neurocomputation used Cortical Labs' CL1 biological computers, linked to a simulated navigation environment and to a physical robot dog. The team tested about 1,300 parameter configurations across 26 neural cultures, accumulating more than 4,000 hours of real-time interaction. Under the selected configurations, the biological agents achieved higher performance than the study's optimised Deep Q-Network software agents, when both were given an equivalent interaction budget. The findings do not establish that biological networks outperform conventional AI https://www.ibtimes.co.uk/indian-workers-record-movements-train-ai-robots-1826013 across tasks generally. The framework treats the interface between biological tissue and conventional computing as an optimisation problem. It examines how input signals, biological responses, decoding and feedback can be configured to improve consistency. It is a method for studying variability in biological responses, not a way of eliminating it. Unanswered Questions on Energy Use Researchers hope biological networks could eventually offer energy-efficient computing https://www.ibtimes.co.uk/nvidia-pledges-uk-ai-safety-research-computing-power-1825563 . The human brain performs complex tasks on roughly 20 watts of power, but the efficiency of a complete biohybrid system remains an open question. Any comparison with conventional AI would need to account for the culture environment, stimulation and monitoring equipment and supporting software, as well as the neural tissue itself. Both sets of experiments are early-stage demonstrations. The long-term aim is to explore whether living neural networks could complement silicon-based systems https://www.ibtimes.co.uk/ghost-core-screenless-ai-computer-launch-1824005 in specialised applications where adaptive processing may be valuable. The results do not show that the technology is ready for general-purpose or commercial use. © Copyright IBTimes 2026. All rights reserved.