Reading time: about 12–14 minutes · 31 October 2026 · Cell Reports (2026)

Can a cortical organoid learn to keep a pole balanced—and what exactly is being trained?

Robbins and colleagues connect mouse cortical organoids to a virtual CartPole task through a high-density microelectrode array (HD-MEA). The pole angle is encoded as electrical stimulation, while recorded neural activity is decoded into a force that moves the cart. The experiment combines a fast sensorimotor loop with a slower training loop that selects additional stimulation patterns to influence the biological network.

The distinction between these two loops is the key to understanding the paper.

Figure 1A. Two nested loops: fast CartPole control and slower adaptive training.

Figure 1A. Two nested loops: fast CartPole control and slower adaptive training.

1. Two loops with different jobs

The fast control loop runs continuously:

pole angle → electrical input → organoid activity → decoded force → cart movement → new pole angle.

The slow training loop operates after an episode ends:

task performance → external learning algorithm → selection of a stimulation pattern → stimulation of the organoid → next episode.

The authors organize the experiment into three phases: recording spontaneous activity to characterize the network; targeted stimulation to map stimulus–response relationships; and closed-loop training in CartPole. This sequence helps identify useful input, output, and training electrodes before running the task.

The organoid receives pole-angle information through two input units. Activity from two output units is converted into leftward or rightward force. When the pole falls beyond ±16°, the episode ends.

2. Who chooses the training signal?

The paper compares four conditions:

  • Null: no additional training stimulation.
  • Random: stimulation locations are selected randomly.
  • Adaptive: an external reinforcement-learning algorithm selects stimulation patterns based on their estimated association with better task performance.
  • Continuous adaptive: the adaptive strategy is applied continuously rather than in separated training/rest cycles.

The adaptive algorithm uses an eligibility trace: recently selected stimulation patterns leave a decaying record, and the algorithm updates their estimated value based on subsequent task performance. This is a digital algorithm running outside the organoid. It chooses which stimulation to deliver; it is not evidence that the organoid itself runs a reinforcement-learning algorithm.

The typical cycled protocol used 15 minutes of training followed by 45 minutes of rest. Improvements were often visible during training but were not reliably retained across the rest period.

3. How large was the improvement?

The authors report the proportion of successful training cycles as follows:

Condition Successful cycles
Null 2.3%
Random 4.5%
Adaptive 23.1%
Continuous adaptive 46.4%

Figure 1B. Successful-cycle proportions reported for the four training conditions.

Figure 1B. Successful-cycle proportions reported for the four training conditions.

These percentages describe cycles, not the percentage of organoids that learned the task. The study included 19 organoids from eight independent batches overall, with 16 organoids from seven batches used for CartPole training. Across the training experiments, the authors report 432 cycles and more than 125 hours of cumulative recording.

The same organoid can contribute multiple cycles. Those cycles are repeated measurements, not independent biological replicates. The authors also report that stimulation-evoked causal connectivity between selected input and output units predicts task performance better than spontaneous functional connectivity measured using STTC. This suggests that directly probing how a stimulus drives downstream activity may be more informative for configuring the interface than relying only on correlations in spontaneous activity.

4. What does the pharmacology tell us?

The researchers applied NBQX and APV to block AMPA/kainate and NMDA receptor-mediated glutamatergic transmission. Under receptor blockade, the training-induced performance improvement was abolished; the reported effect supports a requirement for intact glutamatergic transmission in this training paradigm.

This experiment does not isolate one specific synaptic learning rule. Blocking these receptors can affect both the network’s basic ability to transmit signals and its capacity to adapt.

The study also reports that improvements generally did not persist through a 45-minute rest period. The result is best described as short-timescale, task-related adaptation rather than stable long-term memory.

5. What exactly happens inside one CartPole episode?

CartPole is an unstable control problem. The simulated environment is described by four state variables: cart position, cart velocity, pole angle, and pole angular velocity. At each time step, the controller chooses a horizontal force of either −10 N or +10 N. The episode ends when the pole angle exceeds ±16°, which the authors define as an unrecoverable state.

The organoid receives the pole-angle information through rate-coded stimulation of two input units. Neural activity from two output units is converted into a leftward or rightward force. The authors use short read phases of 200 or 300 ms to balance responsiveness against noisy spike counts. This is a real-time sensorimotor loop: the decoded force changes the cart state, and the new angle is encoded and sent back to the organoid.

The training signal is separate from this online control. It is delivered only after an episode ends. In the cycled protocol, a training pattern is sent when the five-episode mean performance falls below the twenty-episode mean. This design reduces the chance that the training pulse itself directly creates an apparent improvement in the same episode.

6. What does “successful training” mean statistically?

The authors use episode duration—the time the pole remains balanced—as the performance measure. Their primary cycle-level metric is the 90th percentile of episode duration within a training cycle. An organoid configuration is considered proficient when this value exceeds a threshold set three standard deviations above the best of five simulated random-firing baselines.

This definition is important because the headline percentages refer to training cycles, not independent organoids. The study includes 19 organoids from eight batches overall, with 16 organoids from seven batches used for CartPole training. Across the experiments, the authors report 432 cycles and more than 125 hours of recording. The same organoid can contribute many cycles, so cycle-level points are repeated observations.

The authors also compare connectivity measures. In proficient experiments, stimulus-evoked causal connectivity predicts performance more strongly (R² = 0.59, p < 0.01, n = 12) than spontaneous functional connectivity measured using STTC (R² = 0.20, p = 0.14, n = 12). In practical terms, a direct perturbation test may reveal which input–output pathways can support control better than passive correlation alone.

7. What kind of reinforcement learning is this?

The reinforcement-learning algorithm is external. It maintains value estimates for candidate pairs of training electrodes and uses eligibility traces to remember which patterns were recently selected. The reported learning rate and trace-decay factor are both 0.3. Patterns associated with improved episode duration become more likely to be selected again.

The biological network is changed by the electrical training patterns; the digital algorithm decides which patterns to try. The organoid is not running the reinforcement-learning code, and the experiment does not use a biological dopamine signal as the reward. It is a hybrid training design in which an external optimizer searches for stimulation patterns that alter the organoid’s input–output dynamics.

8. The limitations are part of the result

Only a minority of adaptive cycles reached the authors’ proficiency threshold: 23.1% for cycled adaptive training and 46.4% for continuous adaptive training. Success depends on the chosen input, output, and training units. The authors manually select neural configurations using connectivity analysis, and they acknowledge that better configurations may have been missed.

The HD-MEA records activity from the organoid surface in contact with the array. A single electrode can contain multi-unit activity, which complicates single-neuron interpretation. The improvements also fail to persist through the 45-minute rest period, indicating short-timescale adaptation rather than durable memory.

The pharmacological result is strong evidence that glutamatergic transmission is necessary for the observed training effect, but it does not identify one precise synaptic learning rule. Blocking AMPA and NMDA receptors also disrupts ordinary signal transmission.

My perspective

The most useful contribution of this paper is its separation of control and training.

A real-time controller needs to react quickly to the current state. A training algorithm can operate more slowly, evaluate the previous episode, and change the stimulation strategy for the next one. Keeping these loops distinct makes the experiment easier to interpret: does the control pathway work, does a training intervention change it, and does that change survive a rest period?

For future closed-loop organoid experiments, I would log each sensory input, decoded action, task outcome, selected stimulation pattern, and subsequent change in neural response. This would help distinguish biological adaptation from changes introduced by the external optimizer.

The limitations are instructive: the reinforcement-learning algorithm is external; the benefit is short-lived; electrode selection depends on network characterization; and the same organoid contributes repeated cycles. A strong next step would test whether a stimulation strategy selected in one session improves a later session under a frozen decoder and a pre-registered evaluation protocol.

Takeaway

  • The system has a fast sensorimotor control loop and a slower training loop.
  • An external reinforcement-learning algorithm selects stimulation patterns; the organoid does not run the digital RL algorithm.
  • Successful-cycle proportions were 2.3%, 4.5%, 23.1%, and 46.4% across the four conditions.
  • NBQX + APV blockade abolishes training-induced improvement, supporting a role for glutamatergic transmission.
  • The observed adaptation generally does not persist across a 45-minute rest period.

References / 参考资料