Gymnasium: [Question] Problems with Collisions in Pusher-v4
Question
I’m testing out the RL training with cleanRL, but I noticed in the video provided below that the robotic arm goes through both the table and the object it is supposed to be pushing.
The code below also has a similar behaviour:
import gymnasium as gym
env = gym.make("Pusher-v4", render_mode="human")
observation, info = env.reset(seed=39)
for _ in range(1000):
action = env.action_space.sample() # this is where you would insert your policy
observation, reward, terminated, truncated, info = env.step(action)
if terminated or truncated:
observation, info = env.reset()
env.close()
Any idea what is going on or what might be missing?
System info
Installed via pip version 0.29.1
Python 3.8.12
Gymnasium 0.29.1
Mujoco 3.1.2
https://github.com/Farama-Foundation/Gymnasium/assets/86664284/0e28d81f-8633-4a78-a716-4f244dd1ff32
About this issue
- Original URL
- State: open
- Created 4 months ago
- Comments: 39 (4 by maintainers)
@HernandezEduin I’m guessing you using the pretrained model to test with this. Once we have the final xml model complete, it would be helpful to retrain the agent for scratch on the new model and check that the agent will learn the correct model
I agree with the suggested change for version 4.
Normally, i would not agree with you, about changing the version 5 pusher, but because the bug is so critical and it was released in an alpha version of gymnasium we should fix it.
Okay, so I’ve tested it out with various versions of Mujoco and it the results are as follows:
The collisions stopped working with Ver. 3.0.0. I’ll report it soon to the Mujoco team.
@HernandezEduin also, if you have some example actions list for a fixed seed for the MuJoCo team to use, it would be helpful