xukechun
commented
2 years ago Hello Xiaowu:
(1) Yes, we checked our former repo and modified some mistakes due to our negligence when reorganizing the codes. (2) There is not fixed iteration number for training. It's the performance that matters.
- For the first step we trained the grasp net with an explore goal-agnostic grasp model to fasten the training process (we also encourage to do so in the readme), or you might spend more episodes for training the grasp net. The key for the grasp net training is to reach >=80% success rate, which can be supervised through tensorboard or saved logs. Note that the first stage is important for the latter stage training since it'll serve as a discriminator in the latter stages. In the first stage we trained around 3000 episodes.
- For the second stage, there is also one thing that must be paid attention to. In our paper, we emphasize that the Q value threshold depends on the first stage training. That is, this Q value should be chosen according to the Q curve of the first stage training, which should correspond to a high graspable probability(>=80%). In our training, we use 1.8 as our threshold (Note that it's an empirical hyper-parameter). And we trained around 2000 episodes for this stage.
- For the third stage, we aim to match the distribution via alternating training. Note that it can be regarded as an online training process. We trained grasping first for 1000 episodes and pushing for 500.
- Also you can train upon the models we provide in #4 .
Hope my reply will help you!
littlefiveRobot
Dear Xu,
I extremely appreciate your work. Thank you for your paper. Recently, I want to reproduce your paper ,but my trained model can't complete the test.And I found that the manipulator was always doing the same action in the test . In fact ,it always pushed in the same way,and it always pushed the air! So it could't finish the test.
(1) I saw the same question in the other issues that you had given the reason,So do you modify your code now? (2) I want to know how much iterations you trained your model in your three stages separately? I wonder if it's because I'm training the wrong way or too less iterations.
Hoping to hear from you! Xia Xiaowu