Should episode_length be increased to 500?
#1
by
IliaLarchenko
- opened
In the original dataset, each episode is 10 seconds (500 frames) but for validation, you use 8s (400 frames) which is the default for the cube transfer problem. I think the fair validation for this test should use episode_length=500.
Just add the following parameter to train and eval commands to fix it. In this case, the training and validation are more aligned.
--env.episode_length=500
It improves the success rate to 23.2%.