I am using Anylogic PLE to train an agent with reinforcement learning using the Alpyne library in Python. I have modeled the behavior of an agent with a statechart. In one of the states I initialize a Storage with this function:

int initialStockSize = 72;
AgentMaterial mat;
for (int i = 0; i < initialStockSize; i++) {
  mat = new AgentMaterial();
  mat.type = "A";
  enterFeedA.take(mat);
  }

In a later state of the statechart I take a material (agent) from storage and deposit it in an enter element with the following code:

AgentMaterial mat;
mat = (AgentMaterial)storageA.retrieve(storageA.getRandomAgent());
mat.destinationHead = head;
enterCrane2.take(mat);

When I run the model interactively in Anylogic the storage is loaded correctly in the first state and I can see all the materials and how they are being unloaded in the later state of the statechart.
The problem appears when I export the model and run the Python code to train the RL agent. Apparently the initial load does not give any errors, but storageA.getRandomAgent() returns null. When pulling a trace with storageA.size() I get 0, as if no material (agent) was previously loaded.
I can’t imagine where the problem could be since, as I say, in the interactive mode of Anylogic it works correctly and in the training there are no errors when loading the elements in the storage.
Any help is welcome, thanks in advance.