-
|
Hello, I am a beginner in reinforcement learning. I am currently trying to integrate my own model into the playground reinforcement learning environment. I designed a task where a 7-degree-of-freedom robotic arm reaches a target position, and I set the training parameters by following the examples in the notebook. During training, I encountered a GPU memory overflow. To fix this, I modified the STL model of the robotic arm (reducing the number of triangular faces in the collision detection model). Although the memory overflow issue was resolved, my training time exceeded 8 hours, which is much longer than the 8-10 minutes shown in the notebook demo. I have the following questions: Why is the training speed of the model in the example so fast? Are there any tricks to improve the training speed of my model? |
Beta Was this translation helpful? Give feedback.
Replies: 2 comments 2 replies
-
|
Hi @ppap36, looks like I'm encountering similar problem like you had. Have you solved the issue? |
Beta Was this translation helpful? Give feedback.
-
|
I find that using STL model for collision may make training quite slow, if you try to use mujoco elements, like box/cylinder/sphere, training goes much faster. |
Beta Was this translation helpful? Give feedback.
-
|
even if there are only 6 faces in a STL file, it is slower than using box |
Beta Was this translation helpful? Give feedback.
-
|
Got it, thank you so much. |
Beta Was this translation helpful? Give feedback.
I find that using STL model for collision may make training quite slow, if you try to use mujoco elements, like box/cylinder/sphere, training goes much faster.