What is a good way to know that the traning is enough?

[carlosedubarreto]

I'm testing some trainning, but I'm getting confused of how to know when to stop? Should I detect that the trainning is enought when on the visualization the models are doing the correct movement? or a number of iteration, or something else?

Thanks for the awesome work

[xbpeng]

Visualizing the models is probably the best way. But you can also look at the "test_return" in the learning curves and see if the progress has plateaued. That could be a good indication of whether the model is done training.

[carlosedubarreto]

Awesome, thanks.. that for sure helps a lot.
I ran a traning using the spinkinck that was already set up, and watching the models, when most of them were not falling worked well..

But, when testing my own pkl, the training was going well, maybe I stopped earlier, but when doing a test with the generated model, the test runs very quick. Does that happen because the model was going to fall on the floor and the simulation stopped automatically?

Probably the right question is, dos the simulation stops automatically when runnign a tes, if the model stats falling? I'm using the same settings from the humanoid ppo that was used on the spinkick.