onnx-model output does not equal Matlab DDPG-agent output

[vec123]

Dear Onnxruntime Team,

I have a Matlab defined RL-algorithm with an DDPG-agent which can chose values between 0 and 1.
This is defined in the Agent-definition through:

numAct = 1;
actionInfo = rlNumericSpec([numAct 1],'LowerLimit',0 ,'UpperLimit', 1);
actionInfo.Name = 'sine_amplitude';
 

During training only values between 0 and 1 are chosen.
However, when i export the Agent with

exportONNXNetwork(layerGraph(actorNetwork),'actorNet_test.onnx')
and load it into a python-script with
policy = rt.InferenceSession("actorNet_test.onnx")
I get a negative output for the input

input_data = [-0.1515581,1];
X = np.reshape(input_data,(1,1,2,1)).astype(np.float32);
output = policy.run(None, {input_name: X});

.

Also when I use the Agent to generate a Policy according to Matlab (see https://www.mathworks.com/help/reinforcement-learning/ref/rl.agent.rldqnagent.generatepolicyfunction.html)
and evaluate the function with the same input_data, I also recieve negative values, but not the same as returned by the onnx-model.

The evaluate_Policy function also expects an Input of shape [2 1 1], while the onnx-imported-model expects the shape[1,1,2,1].
I have used reshape-functions to reshape the data [-0.1515581,1] to the corresponding shape and then evaluate the two functions:

evaluatePolicy(reshape([-0.1515581,1],2,1,1))
which returns -1

and

     input_data = [-0.1515581,1]
    X = np.reshape(input_data,(1,1,2,1)).astype(np.float32)
    output = policy.run(None, {input_name: X})

which returns -0.0076.

My question is, why do the two outputs differ, eventhough they were generated from the same agent and also why do the two Policy-models expect different shapes?

Grateful for any tip. Currently I can deploy a trained Net on a raspberry, however i would like to have outputs which correspond (at the moment they don't).

BR,
V