Joint Velocities Incorrect?

[davesarmoury]

I am running a simulation task using OmniIsaacGymEnvs. The simulation has reached steady-state and the robot is stationary (no joints are visibly moving).

Run the Anymal task from this branch in test mode: j_vel

Calling get_joint_velocities() on the ArticulationView returns non-zero values (+/- 0.3). Why would this be? Are the joint velocities relative to the joint itself, or relative to some other frame? Are they in radians per second, or degrees for some reason? The current targets are being set using set_joint_position_targets()

I'm using 2022.2.1

[kellyguo11]

Hi there, thanks for the report. We will look into this.

[AndrePatri]

Hi @davesarmoury, @kellyguo11 I am having the same issue on IsaacSim2023.1.0 (and I remember having it also on the 2022 version). It might be some IsaacSim bug though, since I am not using OmniIsaacGymEnvs at all and I am getting nonzero velocity anyway when the robot I am using is stationary. This happens for all kinds of velocity (e.g. linear and joints). For example: and These are graphs of the quantities I am getting from IsaacSim, linear velocity of my root link and joint velocities of the links. As you see, even if the positions are clearly stationary, velocities are non-null

[AndrePatri]

Also, going through IsaacSim discussions I spotted this and this, where they say this is an expected behaviour. It is not clear to me however how this data can then be used for control purposes.

[kellyguo11]

Hi there, there are a few parameters that you can try tuning to achieve more accurate velocities. Decreasing the simulation dt can help with numerical issues when the velocities converge, which is likely the main issue in this case. Other parameters include increasing the position iteration count, or trying with a significantly larger stiffness value to avoid oscillations.

[AndrePatri]

Hi @kellyguo11, thanks for the feedback. I already tried to decrease the integration_dt (I was using 0.01 and I lowered it to 0.001). However, from what I saw, this didn't make much difference (just a slight improvement). From the documentation I can see methods for setting the iteration count on Articulations/ArticulationViews on both positions and velocities. Should I try to play also with the velocity one? What is their meaning? P.s. For control applications, like in my case, where I aim for an almost pure torque controlled robot (i.e. very low or 0 joint impedance), increasing stiffness and damping of the articulation controller is not an option.

[kellyguo11]

You can try setting velocity iteration to 0, since the velocity iterations will tamper with the velocities, but those updates won't be reflected in the positions. Our physics team is also working on a set of recommendations for different applications for working with the velocities, which hopefully we'll be able to share soon.

[AndrePatri]

Thanks @kellyguo11. I think for robotics applications it's crucial for all the quantities you get from the simulation. Having for example this issue with the velocity without realizing it exists can completely alter the personal of any controller, including RL ones. At least a warning on this in the documentation of the getter methods is imperative, in my opinion. Btw, I tried to increase both the position and velocity iteration count, but the behavior is the same