You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Issues with Collisions
I've been exploring reinforcement learning for a project and decided to utilize the Pusher-v4 environment available through Mujoco and Gymnasium. My approach was informed by examples from CleanRL, aiming to gain a deeper understanding of the coding practices within this domain. However, I encountered a notable issue: the robot, which is expected to learn to push an object towards a goal, instead passes through the object without any collision.
Upon consultation with the community and specific members (@Kallinteris-Andreas), we've pinpointed that this issue arises exclusively with Mujoco versions post 2.3.7. For reference and further investigation, I've detailed this observation at the following Gymnasium GitHub issue.
I'm looking forward to any insights or solutions to address this collision detection problem.
Reproducibility
You can test this behaviour with the following code I modified from clearnrl and trained model: cleanrl.zip
Mujoco Ver. 2.3.7
mujoco.2.3.7.mp4
Mujoco Ver. 3.0.0
mujoco.3.0.0.mp4
Context
OS: Windows 10
Python: 3.8.18
Issues in: MuJoCo version 3.0.0 and up
Bindings:
pytorch
mujoco == 2.3.7 or 3.0.0,
gymnasium == 0.29.1,
gymnasium[mujoco],
The text was updated successfully, but these errors were encountered:
HernandezEduin
changed the title
ollision Detection Issue in Pusher-v4 Environment with Mujoco > 2.3.7
Collision Detection Issue in Pusher-v4 Environment with Mujoco > 2.3.7
Mar 11, 2024
This is not a collision detection problem, this is a modeling problem. Visualize contacts and you will see that they are fine.
The underlying problem is density="0.00001" which is 1e-5 the density of air, your object weighs 1e-8Kg. Why would you do that? The reason this affects contact forces is that you also have armature and damping set for all joints. Combined with the silly mass of your cylinder, it effectively locks it in place.
The reason for the discrepancy between the versions is bugfix 33 in 3.0.0. You have a positive margin set for all geoms, which is making your cylinder collide with the table. Again, just visualize contacts and you will see this...
Issues with Collisions
I've been exploring reinforcement learning for a project and decided to utilize the Pusher-v4 environment available through Mujoco and Gymnasium. My approach was informed by examples from CleanRL, aiming to gain a deeper understanding of the coding practices within this domain. However, I encountered a notable issue: the robot, which is expected to learn to push an object towards a goal, instead passes through the object without any collision.
Upon consultation with the community and specific members (@Kallinteris-Andreas), we've pinpointed that this issue arises exclusively with Mujoco versions post 2.3.7. For reference and further investigation, I've detailed this observation at the following Gymnasium GitHub issue.
I'm looking forward to any insights or solutions to address this collision detection problem.
Reproducibility
You can test this behaviour with the following code I modified from clearnrl and trained model:
cleanrl.zip
Mujoco Ver. 2.3.7
mujoco.2.3.7.mp4
Mujoco Ver. 3.0.0
mujoco.3.0.0.mp4
Context
pytorch
mujoco == 2.3.7 or 3.0.0
,gymnasium == 0.29.1
,gymnasium[mujoco]
,The text was updated successfully, but these errors were encountered: