the joint position doesn't align with applied action #2810
Replies: 1 comment 4 replies
-
|
Hello! Have you tried increasing the stiffness of your actuation? Also, please share an XML of your model. |
Beta Was this translation helpful? Give feedback.
-
|
this is the joint setting I reduced the dampratio from 0.8 to 0.5 and kp from 1500 to 1000 that give me better performance. But again there comes a new problem robot angle: [ -1.14984286e+00 ] as you can notice, after some time, the robot angle exceeds the action and keep the gap. Actually, the real motor will just run after the action. After my test, the real motor will always smaller than action about 0.02 rad. For example, the robot angle in this situation should be -1.06 rather than -1.14. The simulation shows, the simulated robot angle after some time will exceed the action and keep the gap. what could be the reasons. thanks |
Beta Was this translation helpful? Give feedback.
-
|
Depending on the rest of the model this could come from a number of reasons. First of all check that there are no unexpected collisions in your model (e.g. arising from convex hull mesh collisions). It is important to keep in mind that with proportional actuators you will almost always have some errors - by definition they only apply forces if there is an error, so if there is an external force like gravity or a contact force you will have errors guaranteed. This is true for real systems too, but if you don't notice this there, there could be a couple reasons for that:
|
Beta Was this translation helpful? Give feedback.
-
|
Thanks for your reply, actually I'm using mjx with brax to train the reinforcement learning policy and there is fast no collision between model (the collision is only existed between base and ground). Currenlty, I'm using position control with global dampratio to excute joint, would it be better to use kv for each joint rather than dampratio? Also, the trained policy with domain randomizition is deployed on to real robot, which is also position controled, but the real robot doesnt work like simulation. Do you have some advises for that? thanks |
Beta Was this translation helpful? Give feedback.
-
|
thanks for your advise, I use gravcomp to solve this problem |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
Intro
Hi!
I am a researcher, I use MuJoCo and mjx for my project with brax reinforcement learn
My setup
mujoco 3.3.1 pypi_0 pypi
mujoco-mjx 3.3.1 pypi_0 pypi
mujoco-py 2.1.2.14 pypi_0 pypi
My question
as the pictured showed, the robot angle compare to actions has 0.2 rad gap at the beginning, whlie the real motor will accurately follow the action. which makes a sim2real gap. I would like to know how to solve this problem
Minimal model and/or code that explain my question
No response
Confirmations
Beta Was this translation helpful? Give feedback.
All reactions