How does A3C aggregate the model from different learner?

[dxu23nc]

For A3C, there are many learners which learning from the interactions generated in local actors. In some iteration, the learner has reach the condition that needs to update the model. The problem is: when the "model server" receives the model or gradients for local learners, how do the server aggregate the model sent by local learner with "the global model"?

[dxu23nc]

Thanks in advance.