Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

weird robot rotating behavior #48

Open
Zhefan-Xu opened this issue Oct 30, 2024 · 8 comments
Open

weird robot rotating behavior #48

Zhefan-Xu opened this issue Oct 30, 2024 · 8 comments

Comments

@Zhefan-Xu
Copy link
Contributor

Hi @abizovnuralem, thanks for the great work! I have a general question regarding the behavior when we send the yaw angle speed command for robot rotation. From observation, during rotation, the robot's movement appears somewhat unnatural: one front leg remains still for a while until the robot has completed a certain amount of rotation, which looks unstable. Could this be a concern during sim-to-real transfer?

@abizovnuralem
Copy link
Owner

Hello, @Zhefan-Xu . You are right, the rotation is not natural for the real dog and it will not work with sim2real. To make it work, you need to do additional work. First, You need to modify the observation input layer and split it into observation for critic and observation for actor, and for the actor you need to delete lidar input because real RL balancing for unitree doesn't use lidar. For citric you should keep it all the same because we need some good feedback for the training process, then train RL again, second, you need to make omniverse simulation the same as real env (some work in actuators type and sync in omniverse, time step and etc) After that it will be ok I think, but this is only my theory.

@Zhefan-Xu
Copy link
Contributor Author

Thank you very much for your insights and that is very helpful @abizovnuralem . Because I saw the real unitree go2 has a pretty good control (with their own controller), do you know what controller it is using if it is open-sourced?

@abizovnuralem
Copy link
Owner

As far as I know, they use a basic controller with some modifications from MIT-cheetah open-source project + Other familiar GitHub open-source projects. Example: https://github.com/chvmp/champ
The latest update has RL balancing that was trained in Mujoco I guess. It is not open-sourced.

@Zhefan-Xu
Copy link
Contributor Author

Thank you for the insights. These are very helpful.

Another question: are you able to run the repo in real-time? it seems the simulator is slower than real time for me but in your demo video the robot walks at a fast speed which seems prettry real-time.

@Zhefan-Xu
Copy link
Contributor Author

NVM I saw your post #19 (comment).

I am currently running on isaac sim 4.2 which seems very slow in my case. Is there any update on the issue?

@abizovnuralem
Copy link
Owner

Hello, currently no updates, I hope I will find some free time and work on it

@abizovnuralem
Copy link
Owner

Hello @Zhefan-Xu , I just saw your repo, you have fixed the speed issue in new isaaclab versions, can you please share your experience here please ? BTW, great work !

@Zhefan-Xu
Copy link
Contributor Author

Hello @Zhefan-Xu , I just saw your repo, you have fixed the speed issue in new isaaclab versions, can you please share your experience here please ? BTW, great work !

Hey @abizovnuralem, apologize for the late reply as I got sick for a while. Here is my solution and I have created a PR: #67

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants