search

IanYangChina / DRL_Implementation

A DRL implementation repo
MIT License
19 stars 6 forks source link

Reproducing pick and place result #3

Closed seann999 closed 2 years ago

seann999 commented 2 years ago

Hi, I was trying to reproduce this result from the paper image So I replaced the task argument in https://github.com/IanYangChina/DRL_Implementation/blob/master/drl_implementation/examples/KukaPushPHER.py to pick_and_place, but this is the result I got for one seed (test success rate): image If I am reading the y-axis in the paper correctly, I was expecting almost perfect success rate by epoch 30, but my result seems to settle at 60%. Is there something that is different that I need to modify?

seann999 commented 2 years ago

Here is a video during testing. The agent learned to push quite well but seems to have had a hard time learning mid-air goals. Peek 2021-10-01 18-03

IanYangChina commented 2 years ago

Hi, thanks for the report. I have noticed this problem with other tasks as well. It comes from the environment side, not the algorithm. I'm working on it at the moment, let's hope that I can get it done ASAP :) You are very welcome to take a look into the environment package as well, if you'd like to. Thanks!

IanYangChina commented 2 years ago

Hi, I was trying to reproduce this result from the paper image So I replaced the task argument in https://github.com/IanYangChina/DRL_Implementation/blob/master/drl_implementation/examples/KukaPushPHER.py to pick_and_place, but this is the result I got for one seed (test success rate): image If I am reading the y-axis in the paper correctly, I was expecting almost perfect success rate by epoch 30, but my result seems to settle at 60%. Is there something that is different that I need to modify?

Hi I have fixed the issue. It turns out that the problem was with the joint forces I setup to control the kuka arm. As expected, it's not with the algorithms. Please install the lastest version of the environment package using:

git clone https://github.com/IanYangChina/pybullet_multigoal_gym.git
cd pybullet_multigoal_gym
conda activate YourEnv (optional)
python -m pip install .

Then return to the example script and have another try. Thanks

IanYangChina commented 2 years ago

Issue closed as solved.