Questions about applying the algorithm to other models

[TRYzero]

Hello,
I've been encountering some challenges while attempting to apply your algorithm to different bipedal robots. Specifically, I tested the algorithm on two robots with distinct configurations.

1. On a bipedal robot with six joints per leg, the robot can walk, but it continuously straddles.

1.30.2.mp4

2. On a bipedal robot with five joints per leg, the policy network still tries to straddle and fails to walk properly.

1.30.1.mp4

I'm wondering if these issues are related to how I'm adjusting the parameters. Could there be specific parameter settings or modifications needed for different robot structures? Any guidance or suggestions on how to optimize the algorithm for various joint configurations would be greatly appreciated.
Thank you for your time and assistance.

[Sujgaik2003]

How do you apply to other robot plz guide me

[TRYzero]

How do you apply to other robot plz guide me

I've made three main adjustments:

1. Switched to my own MJCF model.
2. Tuned the PD parameters.
3. Modified the mirror loss function.

If all these aspects are fine, the program should now be usable with your own robot model.

[Sujgaik2003]

Thanks for reply actually I want to apply it on nao robot so plz can u help me out . By adjusting the parameters for nao robot.

[Sujgaik2003]

@TRYzero can u provide me the above project link where can I see the code it will be great help 🙂.

[TRYzero]

@TRYzero can u provide me the above project link where can I see the code it will be great help 🙂.

Well, I'm not entirely sure which aspect you're confused about. Could you please describe your question in more detail? Are you unclear about the entire process?

1. If your question concerns how to convert URDF files to XML files, you might want to search for MuJoCo-related tutorials.
2. Once you have the XML file for the NAO robot, you can adjust your model using envs/jvrc/gen_xml.py. This script is designed to transform a complete robot model into one that is more suitable for training.
3. Next, you might need to refer to envs/jvrc/jvrc_walk.py to write your own nao_walk.py. In this file, you may need to adjust the PD gains and the mirror symmetry settings. Of course, you will need to replace the imported XML path with your own.
4. Another file you might need to adjust is envs/jvrc/robot.py, because the half_sitting_pose used for the JVRC may not be suitable for your robot.

[Sujgaik2003]

Hey I am done with step one by converting urdf to XML file not going ahead with 2,3,4 step as you mentioned can u guide me regarding the 4 step as u mentioned it could be a great help

[TRYzero]

Hey I am done with step one by converting urdf to XML file not going ahead with 2,3,4 step as you mentioned can u guide me regarding the 4 step as u mentioned it could be a great help

In the fourth step, I mentioned the "half_sitting_pose" as shown in the screenshot.

Well, this will not be the key factor.

[Sujgaik2003]

Okhy
So there is no need to change in robot.py ? It can also use with nao robot

[TRYzero]

Okhy So there is no need to change in robot.py ? It can also use with nao robot

yep

[Sujgaik2003]

(https://github.com/rohanpsingh/LearningHumanoidWalking/assets/118328948/d7f08070-c17e-453d-9000-8e4f272f722d)
WA0001.jpg…]()

Encounter with the error

Try to solve by below 👇🏼 links but not worked
https://github.com/openai/mujoco-py/issues/475

https://medium.com/@sayanmndl21/install-openai-gym-with-box2d-and-mujoco-in-windows-10-e25ee9b5c1d5

[formoree]

(https://github.com/rohanpsingh/LearningHumanoidWalking/assets/118328948/d7f08070-c17e-453d-9000-8e4f272f722d)（ https://github.com/rohanpsingh/LearningHumanoidWalking/assets/118328948/d7f08070-c17e-453d-9000-8e4f272f722d） WA0001.jpg…]() WA0001.jpg...]()

Encounter with the error遇到错误

Try to solve by below 👇🏼 links but not worked尝试通过以下👇🏼链接解决，但没有奏效 https://github.com/openai/mujoco-py/issues/475

https://medium.com/@sayanmndl21/install-openai-gym-with-box2d-and-mujoco-in-windows-10-e25ee9b5c1d5

Did you finally work it out?

[Sujgaik2003]

Still not yet !

[formoree]

Hello, I've been encountering some challenges while attempting to apply your algorithm to different bipedal robots. Specifically, I tested the algorithm on two robots with distinct configurations.

1. On a bipedal robot with six joints per leg, the robot can walk, but it continuously straddles.

1.30.2.mp4
2. On a bipedal robot with five joints per leg, the policy network still tries to straddle and fails to walk properly.

1.30.1.mp4
I'm wondering if these issues are related to how I'm adjusting the parameters. Could there be specific parameter settings or modifications needed for different robot structures? Any guidance or suggestions on how to optimize the algorithm for various joint configurations would be greatly appreciated. Thank you for your time and assistance.

Have you ever solve this problem? I think the key factors may be the setting of pdgains, the transformation of the .xml file and reward setting. I have change the robot to OP3, and it is still hard for it to walk steadily.

[formoree]

Still not yet !

I have read through your bugs, i think the problem is the mujoco library. Maybe you don't download the python packsites of mujoco. To be clear that the problem has nothing to do with your code.

[TRYzero]

Have you ever solve this problem? I think the key factors may be the setting of pdgains, the transformation of the .xml file and reward setting. I have change the robot to OP3, and it is still hard for it to walk steadily.

I solved this problem by introducing reference motions. Well, there shouldn't be any need for this. Did this issue also appear on your robot?