Summary: In vitro biological neural networks (BNNs) embodied in robots demonstrate diverse intelligent behaviors, including supervised and unsupervised learning, memory formation, object tracking, obstacle avoidance, and even the ability to learn simple games.
Source: Beijing Institute of Technology Press Co., Ltd
A review by researchers at the Beijing Institute of Technology synthesizes recent advances and future prospects for using in vitro biological neural networks to realize biological intelligence, with a particular emphasis on applications in robot intelligence.
Published on Jan. 10 in the journal Cyborg and Bionic Systems, the review outlines: 1) the fundamental features of intelligence displayed by in vitro BNNs, including memory and learning; 2) how these networks are embodied in robots through bidirectional interfaces to create BNN-based neurorobotic systems; 3) the preliminary intelligent behaviors demonstrated by such systems; and 4) current directions and challenges facing this multidisciplinary field.
“The human brain is a vast biological neural network composed of billions of neurons that underlies consciousness and intelligence, but studying the whole brain remains extremely difficult because of its complexity,” explained Zhiqiang Yu, an assistant researcher at the Beijing Institute of Technology and one of the study’s authors.
By cultivating a subset of brain neurons in vitro, researchers can form simpler networks—often called mini-brains—that are easier to observe and manipulate. These simplified BNNs offer a practical window into how networks of neurons give rise to memory, learning, and other cognitive functions.
Yu noted that these in vitro networks mirror several foundational properties of in vivo neural circuits: neurons connect through synapses, networks exhibit short-term memory via fading traces and hidden-state dynamics, and they can encode and retain information. Researchers have trained BNNs with supervised protocols so they produce specific responses to given stimuli, and recent experiments have shown that in vitro networks can perform unsupervised tasks such as separating mixed signals.
“This capability may relate to the free energy principle: BNNs tend to reduce uncertainty about their environment,” Yu added.
However, a cultured mini-brain alone cannot realize full cognition. The brain depends on a body to sense, interpret, and act in the world; similarly, in vitro BNNs require embodiment to interact with their surroundings. Robots provide an ideal physical substrate for that embodiment, giving rise to a growing interdisciplinary area at the intersection of neuroscience and robotics: BNN-based neurorobotic systems.
A stable bidirectional connection between the biological network and the robot is essential. The review categorizes these connections by direction: from robots to BNNs and from BNNs to robots. Signals from sensors on the robot can stimulate BNNs using electrical, optical, or chemical approaches. Conversely, neural activity in BNNs can be recorded and decoded to generate control commands for the robot using extracellular, calcium-based, or intracellular recording techniques.
“When embodied in robots, in vitro BNNs can display a wide range of intelligent behaviors,” Yu said. “Examples include supervised and unsupervised learning, memory retention, mobile object tracking, active obstacle avoidance, and even learning to play simple games such as ‘Pong.’”
The review separates observed behaviors into two categories based on their underlying mechanisms: computing-capacity-dependent behaviors and network-plasticity-dependent behaviors. In the first category, the BNN functions as an information processor that produces specific neural responses to stimuli without requiring learning. In the second category, learning and synaptic plasticity are central: the BNN adapts as it is trained, and those adaptations enable the robot to perform tasks.
To help readers compare techniques and outcomes, the authors summarize representative studies, detailing stimulation and recording methods, encoding and decoding schemes, training protocols, and the robotic tasks demonstrated. The review also presents a chronological overview of notable experiments to trace the development of BNN-based neurorobotic systems over time.
The authors identify several major research trends and challenges. Chief among them is the need to develop three-dimensional BNNs that more closely resemble in vivo networks. Another pressing difficulty is establishing effective training methods for robot-embodied BNNs. Cultured networks lack the full complement of neuromodulatory systems present in intact animals, which complicates direct transfer of animal training paradigms to in vitro preparations.
BNNs also face inherent limits: tasks requiring complex, higher-level cognition remain far more difficult to achieve in simplified networks. While some animal behaviors—such as motor skills—can be approximated, sophisticated deliberative tasks remain out of reach.
“How consciousness and higher-order intelligence emerge from networks of cells in the brain is still an open question,” said Yu. “By continuing to embed in vitro BNNs in robotic bodies, we expect to observe increasingly complex behaviors and to gain deeper insights into the neural basis of cognition.”
About this robotics and biological neural network research news
Author: Ning Xu
Source: Beijing Institute of Technology Press Co., Ltd
Contact: Ning Xu – Beijing Institute of Technology Press Co., Ltd
Image: The image is in the public domain
Original Research: Open access.
“An Overview of In Vitro Biological Neural Networks for Robot Intelligence” by Zhiqiang Yu et al., Cyborg and Bionic Systems
Abstract
An Overview of In Vitro Biological Neural Networks for Robot Intelligence
In vitro biological neural networks (BNNs) interconnected with robots—BNN-based neurorobotic systems—can interact with the external world and exhibit preliminary intelligent behaviors, including learning, memory, and robot control. This review provides a comprehensive overview of the intelligent behaviors demonstrated by such systems, with a focus on robot intelligence applications.
The authors introduce the biological foundations necessary to understand two defining characteristics of BNNs: nonlinear computing capacity and network plasticity. They then describe typical neurorobotic architectures and outline mainstream techniques for implementing bidirectional interfaces, covering both sensory-to-BNN stimulation and BNN-to-robot decoding.
Next, the review classifies intelligent behaviors by whether they rely solely on computing capacity or also depend on network plasticity, and it details representative experiments in each category with emphasis on robot-relevant outcomes. Finally, the paper discusses development trends and scientific and technical challenges facing the field of BNN-based neurorobotic systems.