Home | Tutorials | Wiki | Issues
Ask Your Question

Is it possible to reduce inertial sensor noise by adjusting gazebo parameters?

asked 2013-01-14 16:48:58 -0600

asomerville gravatar image

updated 2013-01-14 16:52:04 -0600

We've noticed that the IMU and force-torque sensors are extremely noisy (well beyond expectation for real sensors). I imagine this is likely a result of the normal imperfections of simulating physics in discrete time steps.

Are there parameters that we can adjust to significantly improve the fidelity of the simulation (perhaps at cost of a much slower simulation)?

Related, but not quite the same question:
* http://answers.gazebosim.org/question/530/gazebo-imu-noise-possible-imu-bug-gazebo-131

edit retag flag offensive close merge delete


An issue was created for the linked post.https://bitbucket.org/osrf/drcsim/issue/85/imu-too-noisy

isura gravatar imageisura ( 2013-01-14 20:27:23 -0600 )edit

1 Answer

Sort by ยป oldest newest most voted

answered 2013-02-26 16:28:32 -0600

hsu gravatar image

updated 2013-02-26 17:05:20 -0600

It's an error inherent in the existing LCP solver given performance requirements. You can run simulation slower to get better results. For example, figure below from running:

roslaunch atlas_utils atlas.launch
rxplot /atlas/imu/linear_acceleration/z

image description

If you increase physics->solver->iterations from default of 40 to 80, the oscillation magnitude decreases (the spike is just an artifact of parameter update):

image description

Setting iterations to 500 reduces oscillation magnitudes even further:

image description

Alternatively, you can low pass filter the data. As an extremely crude example, exponential filter where

accel(new) = (1 - e) * accel(new) + e * accel(old)

where e = 0.9 leads to:

image description

edit flag offensive delete link more


I also noticed that the noise increased as the joint PID gains of Atlas was increased. This is a problem since the low gain joint controllers (specifically the ones in drcsim whole body yaml) cannot follow trajectories very well.

isura gravatar imageisura ( 2013-03-07 10:22:54 -0600 )edit

Specifically, is this a numerical problem that will be solved at some point or should we plan on using low gains, filter and/or increase the iterations?

isura gravatar imageisura ( 2013-03-07 10:27:05 -0600 )edit

Not sure if it's correct to ask here, but along the lines of this question: Will time step be configurable? Do you know if such adjustments to time-step or iterations will be allowed for the DRC?

asomerville gravatar imageasomerville ( 2013-03-18 12:36:37 -0600 )edit

Contestants in the VRC will not be allowed to control properties of the physics engine.

nkoenig gravatar imagenkoenig ( 2013-03-18 13:00:54 -0600 )edit
Login/Signup to Answer

Question Tools


Asked: 2013-01-14 16:48:58 -0600

Seen: 779 times

Last updated: Feb 26 '13