Implementation progress?

[yoann-fleytoux]

Hi I see you are trying to implement rl100 :) I was wondering if you had any success?

[Starsshine21]

Hi I see you are trying to implement rl100 :) I was wondering if you had any success?

I implemented it based on dp3. Right now I've basically completed the IL stage (mainly reuse dp3) and the offline RL stage. I'll probably get the online RL running soon too, but the original paper misses a lot of details, so I'm not sure if my reproduction is really accurate.

[Starsshine21]

Hi I see you are trying to implement rl100 :) I was wondering if you had any success?

now you can give it a try with train_v2.py. Parameters are still being debugged

[itsmathtime]

Hi. I'm also interested in implementing the RL100 paper, I would love to contribute to this effort.

I've read your code. I have set up a working environment for running it. I am trying to get the basic pieces of RL100 working. And I am starting with Phase 1. IL policy trained on expert demonstrations.

• Is it correct that the env and task you have choosen to test the algorithm on is MetaWorld with the metaworld_dial-turn
• I have found the collect_dail_turn.py script. How many episodes do you collect for the initial expert demonstration dataset provided to the IL training?
• What success rate do you get for the initial IL policy trained only on successfull trajectories from the initial expert demonstration dataset? I am getting 0% so I must be doing something wrong. I am training on a dataset of 100 episodes for 1000epochs.
• Maybe you can share how you have trained the IL part, and how you generated the data you trained on?

[Starsshine21]

Hi. I'm also interested in implementing the RL100 paper, I would love to contribute to this effort.

I've read your code. I have set up a working environment for running it. I am trying to get the basic pieces of RL100 working. And I am starting with Phase 1. IL policy trained on expert demonstrations.

• Is it correct that the env and task you have choosen to test the algorithm on is MetaWorld with the metaworld_dial-turn
• I have found the collect_dail_turn.py script. How many episodes do you collect for the initial expert demonstration dataset provided to the IL training?
• What success rate do you get for the initial IL policy trained only on successfull trajectories from the initial expert demonstration dataset? I am getting 0% so I must be doing something wrong. I am training on a dataset of 100 episodes for 1000epochs.
• Maybe you can share how you have trained the IL part, and how you generated the data you trained on?

Hello, I'm curious about your current progress. I believe there should be no issues with the current version of the IL stage. I collected 100 pieces of demonstration data, and the success rate on the metaworld dial-turn task reached around 60%. Even with just 10 pieces of data, the success rate was similar.

[Starsshine21]

Hi. I'm also interested in implementing the RL100 paper, I would love to contribute to this effort.

I've read your code. I have set up a working environment for running it. I am trying to get the basic pieces of RL100 working. And I am starting with Phase 1. IL policy trained on expert demonstrations.

• Is it correct that the env and task you have choosen to test the algorithm on is MetaWorld with the metaworld_dial-turn
• I have found the collect_dail_turn.py script. How many episodes do you collect for the initial expert demonstration dataset provided to the IL training?
• What success rate do you get for the initial IL policy trained only on successfull trajectories from the initial expert demonstration dataset? I am getting 0% so I must be doing something wrong. I am training on a dataset of 100 episodes for 1000epochs.
• Maybe you can share how you have trained the IL part, and how you generated the data you trained on?

Sorry for taking so long to reply to you. I'm a student, and I didn't work during the holidays. I just recently started to resume working on RL100