robot cycling

[BowenY]

Hi there,

I am not sure if I should put this question here. Please move it if it's not appropriate.

When the robot is being trained on stage_1 and stage_2, should it get to the infinite cycling after training finished? I noticed the robot just keep cycling after couple of episodes and the reward recached to over 1500.

Thanks

[kijongGil]

Hi @BowenY :) The agent has to subscribe to the scan topic from gazebo, but the learning is sometimes stopped because gazebo does not publish scan topic. It is a bug. You can see that the learning has stopped, the action graph is not moving. If learning stops, try again from the last saved episode.

Thanks, Gilbert.

[BowenY]

Hi @kijongGil ,

Thanks for replying. That makes sense to me cause I didn't find any code mentioning cycling around in the code base. As to the train from the last saved episode, I just would to double check to make sure I found the place currently: set the "self.load_model = true" and "self.load_episode = last saved episode number which can be found in the save_model folder"

Is that correct?

Thanks!

[kijongGil]

Yes, right. If you have a 'stage_1_100.h5' file, you can load save model.

self.load_model = true
self.load_episode = 100

[BowenY]

@kijongGil Wonderful. Thanks. It's working now. The gazebo stops sending "/scan" so often though.

[kijongGil]

I'm going to close this issue. You can reopen this issue to show this issue to the users whenever.

[EdgarJP006]

Hi @kijongGil ,

Thanks for replying. That makes sense to me cause I didn't find any code mentioning cycling around in the code base. As to the train from the last saved episode, I just would to double check to make sure I found the place currently: set the "self.load_model = true" and "self.load_episode = last saved episode number which can be found in the save_model folder"

Is that correct?

Thanks!

your recommendation worked