Fix instability in model #2387
Replies: 2 comments 5 replies
-
|
Is it colliding with your coordinate system feature? (the red, green and blue triad of cylinders) |
Beta Was this translation helpful? Give feedback.
-
|
No, It is very far away. |
Beta Was this translation helpful? Give feedback.
-
|
Are there any discontinuities in your reference motion? Does the error happen before the reference motion ends, or when it ends? |
Beta Was this translation helpful? Give feedback.
-
|
I'll say that there is not discontinuities, as you can see in the green character (the one that fully follow the reference movement) is smooth. I'm also able to see the smooth bhv motion using Blender, so no discontinuities, at least as I understand it. I'm additionally checked the values in the reference motion matrix and in the reference velocity matrix they are also smooth. The reference motion is never applied to the right wrist, the wrist is "free" during all the simulation that you can see in the video. Not sure if this answer your question, if not please let me know |
Beta Was this translation helpful? Give feedback.
-
|
Since the wrist is not constrained, then roughness in the motion can be translated as sudden impulses onto the unconstrained dofs, which can cause instabilities. I do see that you have applied damping and armature on the joint already which I would have recommended. In general, since the dynamics of the kinematic tree are coupled, simply setting the parent joint configuration will produce unrealistic movement in the unconstrained child joint. Applying forces on the joints of the parent would have some effect on the child joints too, which is skipped/undefined with your current approach. Can you tell a bit more about what are you investigating by letting the wrist be free? |
Beta Was this translation helpful? Give feedback.
-
|
Of course. The final goal is to create a dataset to train a world model. In this dataset the X's are the current state of the character (qpos for all the joints) and the actions (torque applied via actuator to each joint and calculated using a PD controller) and the Y's or labels will be the next state of the character (i.e. qpos in the next frame) in this way I can train a model where given the current state and the action it can predict the next state. To build this dataset I have as starting point a BVH file which contains just the position for each joint at each frame. So I need to calculate the torques to apply to each joint to take the character from one state (qpos in a frame) to the following one (qpos in the next frame). I need to carefully adjust the gains (kp and kd) for the PD controller to apply to each motor, so I thought that it was a good idea to isolate a leaf joint, in this case the right wrist to check if my PD was working correctly. When I applied the PD I found that the movement was broken at certain frame due to a unexpected and huge increase in the rotation velocities of the wrist. Now trying to isolate the error I am not applying any force to the wrist and just trying to found the cause of this unexpected increase in the velocity. Do this answer your question? Thanks for your help, time and interest in the issue! |
Beta Was this translation helpful? Give feedback.
-
|
I've been doing some additional tests, the "spike" in the velocity seems to be related to the mass of the "hand" (the blue ball). The larger the mass of the geometry representing the hand, the greater the increase in speed. If I remove the hand geometry from the model, the speed peak still occurs but is not visually noticeable, although it does appear in the logs. So I would say that the issue is not caused by the hand, but worsened (at least visually) by the mass of the geom attached. Here I attach 3 videos with different hand's mass value: 1, 0.1 and 0.001 ScreenRecording2025-01-27at14.14.58-ezgif.com-video-cutter.movScreenRecording2025-01-27at14.15.36-ezgif.com-video-cutter.movScreenRecording2025-01-27at14.24.24-ezgif.com-video-cutter.1.mov |
Beta Was this translation helpful? Give feedback.
-
Intro
Hi!
I am a Master student, I use MuJoCo for my research on human motion generation.
My setup
MuJoCo Python
My question
I'm trying to replicate the human motion from animation in a BVH file. The idea is achieve the same movements but applying torques to joints. The problem is that my model works correctly until some joint velocity explode by an unknown reason. You can see this in the attached video, I do not why this is happening I'm not applying any force, I apply qpos and qvel to all the body except for the right wright. (The green is the full directed character and the orange is the one that have the wright "free"). Here is the part of the script that I use for this approach:
Any idea about why this instability appear? In addition the logs do not show the message that usually says that the model is unstable.
As you can see in the model I'm using implicit integrator with small timestep 0.001.
Not sure if this information is enough to explain the issue, if not please let em know.
Thanks a lot for all the help that I'm getting from this forum.
ScreenRecording2025-01-27at00.39.57-ezgif.com-video-cutter.mov
Minimal model and/or code that explain my question
Confirmations
Beta Was this translation helpful? Give feedback.
All reactions