Summary: Can a pea-sized piece of lab-grown brain tissue learn to play a video game? According to researchers, yes. For the first time, scientists have rigorously demonstrated goal-directed learning in brain organoids—tiny clusters of neurons cultivated in vitro.
Using a closed-loop bioelectrical interface, the team coached organoids to solve the classic “cart-pole” balancing problem, a standard robotics and control benchmark where a system must keep an upright pole balanced on a moving cart. Although these organoids lack a body, dopamine signaling, or sensory experience, the study shows that the capacity for adaptive computation is intrinsic to cortical tissue itself.
Key Facts
- Cart-Pole Benchmark: Organoids were trained to balance a virtual pole, a common test for adaptive information processing in robotics, control theory, and artificial intelligence.
- Closed-Loop Interface: An electrophysiology system both recorded neural spikes to drive the cart and delivered electrical stimuli that encoded the pole’s angle back into the organoid.
- AI Coaching: A reinforcement learning algorithm acted as an artificial coach, selecting which neurons received training signals when the organoid failed to improve.
- Performance Gain: Adaptive coaching raised the organoids’ success rate from 4.5% with random stimulation to 46% using targeted reinforcement-learning-guided stimulation.
- Intrinsic Plasticity: The findings suggest neural circuits can be tuned via electrical feedback to solve problems without the usual bodily and hormonal scaffolding.
Source: UC Santa Cruz
Imagine balancing a ruler vertically in your palm: you must constantly monitor its angle and make tiny adjustments to keep it upright. That skill improves with practice.
Engineers call this the inverted pendulum or cart-pole problem. It tests whether a control system can adaptively process input and produce corrective output. The task is relevant across robotics, control theory, and artificial intelligence—and it mirrors basic sensorimotor challenges human infants solve on the path to walking.
Researchers at the University of California, Santa Cruz trained cortical organoids—small pieces of brain tissue grown from stem cells—to tackle this benchmark. By sending and receiving electrical signals to and from the organoids, their closed-loop software provided coaching that measurably improved the organoids’ ability to balance the virtual pole.
The team, led by Ph.D. student Ash Robbins, Professor Mircea Teodorescu, and Professor David Haussler, reports the results in Cell Reports. Their work explores how information is transmitted and adapted through neuronal spiking and has potential implications for basic neuroscience and biomedical research.
Understanding how neural circuits adapt could yield new tools for studying neurological disorders—such as Alzheimer’s disease, dementia, stroke, concussion, autism, schizophrenia, Parkinson’s disease, dyslexia, and ADHD—by revealing how disease alters the brain’s learning mechanisms.
“We’re trying to understand the fundamentals of how neurons can be adaptively tuned to solve problems,” Robbins said. “If we can figure out what drives that in a dish, it gives us new ways to study how neurological disease can affect the brain’s ability to learn.”
This work is the first rigorous academic demonstration of goal-directed learning in cortical organoids and establishes a platform for adaptive organoid computation: testing whether lab-grown neural tissue can be driven to learn and solve tasks through targeted electrical feedback.
“These are incredibly minimal neural circuits. There’s no dopamine, no sensory experience, no body to sustain, no goals to pursue. Yet when given targeted electrical feedback, this tissue is plastic and structured enough to be pushed toward solving a real control problem,” said Keith Hengen, an associate professor of biology not involved in the study. “That tells us the capacity for adaptive computation is intrinsic to cortical tissue itself.”
Organoid coaching
Organoids—miniature heart, liver, lung, brain, and other tissues grown from stem cells—have been used in biomedical research for about 15 years. Only recently have researchers begun exploring their potential to reveal how brains learn. Brain organoids model early brain development, structure, and function. Although smaller than a peppercorn, some contain networks of millions of neurons that generate electrical spikes.
Placing organoids on specialized chips allows researchers to record firing neurons and stimulate selected neurons. From an engineering standpoint, that combination of measurement, stimulation, and adaptation in one system makes it possible to study learning as a physical process. The closed-loop interface ensures the tissue’s outputs shape the next inputs, which is critical to studying plasticity directly.
The UC Santa Cruz Braingeneers group used organoids derived from mouse pluripotent stem cells and an electrophysiology platform to send and receive electrical information. By varying stimulus strength, they encoded the virtual pole’s angle to the organoid; the organoid’s spike patterns were decoded to choose forces applied to the virtual cart.
Each trial runs until the pole falls—an episode—after which the pole is reset. The organoid’s progress was evaluated in blocks of five episodes: if the recent five-episode average exceeded the previous 20-episode average, no training was given; if not, the organoid received a high-frequency training signal. Training occurs at episode end rather than during balancing, and a reinforcement learning algorithm chooses which neurons receive stimuli—functioning like an artificial coach that suggests targeted, incremental adjustments.
Observing improvement
The researchers established a strict success framework to distinguish genuine learning from chance. Using that framework, they found organoids guided by reinforcement-learning-selected stimuli improved substantially, with a winning rate rising from 4.5% under random stimulation to 46% under adaptive coaching.
The improvements represent short-term learning: organoids consistently shifted from one functional state to a more successful state when coached. However, gains largely decayed after rest periods. Following 15 minutes of training and a 45-minute pause, organoid performance typically returned to baseline, indicating limited retention.
Researchers believe longer-term memory may require more complex organoids that incorporate multiple interacting brain regions. They are also investigating which neuron targets and stimulus patterns are most effective and how sustained plasticity might emerge.
To accelerate this line of work, the team released an open-source software tool called BrainDance. It enables labs with tissue-culture expertise to run closed-loop neural learning experiments without building custom hardware or game environments, lowering the barrier to entry and broadening participation in organoid research.
“This software makes running complicated experiments much easier,” Robbins said. “Rather than spending years developing bespoke systems, researchers can begin experiments in minutes.”
The investigators emphasize their goal is to advance brain research and treatment of neurological disease—not to replace conventional computers or controllers with biological tissue. They also note ethical concerns that would arise if similar experiments used human brain organoids.
Key Questions Answered:
A: No. They do not possess consciousness, goals, or subjective experience. They perform adaptive computation—processing input and altering outputs to achieve a researcher-defined objective.
A: The experiment probes the basic physics of learning. By observing how healthy neural tissue adapts in a controlled setting, researchers can study mechanisms that break down in neurological disease.
A: Not yet. Performance decayed after a 45-minute rest. Researchers expect more elaborate organoids integrating multiple brain regions may be necessary for durable memory.
Editorial Notes:
- This article was edited by a Neuroscience News editor.
- The full journal paper was reviewed.
- Additional context was added by editorial staff.
About this neuroscience research news
Author: Emily Cerf
Source: UC Santa Cruz
Contact: Emily Cerf – UC Santa Cruz
Image: Image credit to Neuroscience News
Original Research: Open access. “Goal-Directed Learning in Cortical Organoids” by Ash Robbins et al., published in Cell Reports. DOI: 10.1016/j.celrep.2026.116984
Abstract
Goal-Directed Learning in Cortical Organoids
Advances in high-density electrophysiology and targeted electrical stimulation now allow single-cell-resolution recording and stimulation. Cortical organoids derived from pluripotent stem cells are promising in vitro models of brain development, function, and disease. In this study, researchers demonstrate goal-directed learning in brain organoids via feedback-driven neural plasticity. Using a closed-loop electrophysiology framework, mouse cortical organoids were embodied into a pole-balancing task (cart-pole), and performance improvements were evaluated when high-frequency training signals were delivered.
For most organoids, training signals selected by artificial reinforcement learning produced better performance than randomly chosen signals or no signal, though improvements did not persist after a 45-minute rest. The study also shows that training-induced plasticity depends on intact glutamatergic transmission: blocking AMPA and NMDA receptors abolished performance gains. This systematic in vitro approach to goal-directed neural plasticity opens new possibilities for neural rehabilitation research and the study of biological computation.