Can't solve task for fetchPickAndPlace-v2
Closed this issue · 4 comments
Hi, thanks for providing this code. I failed to reproduce the result you have achieved with the updated environment from gymnasium Robotics and I observed the success rate remains around 0. I have modified the code a little bit so it matches the updated environment, and I tested the code with fetchReach-v2 which works well. I'm wondering what aspects I should investigate? or Perhaps you've encountered a similar challenge in the past and could offer some pointers. Thanks for your help!!!!!
Hi Qiyang!
Thank you for your comment!
To be honest, I can't give any comments on this matter as I wrote the code close to 3.5 years ago where the gym api was the common way to have our rl algorithms and I know that gymnasium has changed a lot of things that mandates changes to my code.
So, I didn't encounter that issues except that I had at least 8 workers for my experiments and that's the very minimum to get results. I suggest to make sure you have at least that amount of compute.
Also, if you could make the code work with the updated apis, then a pull request from you would be highly appreciated on my side! :)
Hi! Thanks for your quick reply! It really helps! Really Appreciate! I tried the algorithm with longer epochs and a single worker and the success rate started increasing around 50 epochs. I think the reason why it wasn't showing anything is because of the number of workers. But the speed of training doesn't seem to improve when I change the number of workers by setting these parameters below as my laptop has an M1 Pro chip with 8 cores (6 performance and 2 efficiency), I'm wondering if I'm doing it in the wrong way?
os.environ['OMP_NUM_THREADS'] = '6'
os.environ['MKL_NUM_THREADS'] = '6'
os.environ['IN_MPI'] = '1'
Oh, that's not the right way, I guess.
Please take a look at:
#1 (comment)
and, the readme file using MPI
:
https://github.com/alirezakazemipour/DDPG-HER#readme
I see! It works now, thanks for your help!!