Question about the control frequency and joint states

[xyzacademic]

Hello, there.

Your work looks amazing, and I want to try it on my project. After several days strugglling, I cannot get any exciting result for a pick and place task. The robot arm I used is different (Mycobot 280m5) and I have serval questions.

1. Does the control frequency affect the algorithm a lot? Since my robot arm is much cheaper, looks like it can only support 5Hz during collecting episodes. Get the joint coordinates and execute takes long time.
2. What is the joint states for your robot? For my robot I can get the joint coordinates (x, y, z, rx, ry, rz), is it the same as yours? In inference, when the robot execute the predicted coordinates, it will go to a weired position or in a weired pose that did not show in training dataset.
3. My experiment environment is not clean. The lighting and other stuffs are in camera view. The object need to pick is small. So I think resnet18 cannot extract good features for locate the object. Then I replace it by a yolov8 backbone, which pretrained on my object's detection task. But I still cannot get good results.

Any suggestion is appreciated.

Thank you so much

[mamimaka]

for question 2, the joint states is qpos in the code, which means rotation angle.

[barsm42]

Hello, there.

Your work looks amazing, and I want to try it on my project. After several days strugglling, I cannot get any exciting result for a pick and place task. The robot arm I used is different (Mycobot 280m5) and I have serval questions.

1. Does the control frequency affect the algorithm a lot? Since my robot arm is much cheaper, looks like it can only support 5Hz during collecting episodes. Get the joint coordinates and execute takes long time.

2. What is the joint states for your robot? For my robot I can get the joint coordinates (x, y, z, rx, ry, rz), is it the same as yours? In inference, when the robot execute the predicted coordinates, it will go to a weired position or in a weired pose that did not show in training dataset.

3. My experiment environment is not clean. The lighting and other stuffs are in camera view. The object need to pick is small. So I think resnet18 cannot extract good features for locate the object. Then I replace it by a yolov8 backbone, which pretrained on my object's detection task. But I still cannot get good results.

Any suggestion is appreciated.

Thank you so much

Hello @xyzacademic,

Could you please explain me how did you change to yolov8 backbone? I want to learn it.