Ask Your Question
0

Real vs simu params and gains (behavior difference)

asked 2014-08-05 05:00:41 -0600

arennuit gravatar image

updated 2014-08-06 07:49:06 -0600

Dear all,

I am wondering up to what extent the parameters and gains of a real robot are re-usable on a simulation (and vice versa: up to what extent the parameters and gains of a simulated robot are re-usable on a real robot).

Let me give some more details. Imagine I have identified a real robot and I have a good estimate of its geometric and dynamic parameters, I have also set the low-level PID gains in such a way that my real robot is well parametrized. Would I be able to reuse the same geometric and dynamic parameters and gains in the simulation and get a stable simulation with a behaviour in simulation comparable to the real life behaviour?

If the behaviours are different are they totally different and unrelated, or are the gains / parameters from the real world a good start to tweak the gains / parameters of the simulated world?

I know that gazebo uses ODE as its default engine, would the answer be the same with bullet or another engine?

Thanks,

Antoine.

PS: I am currently using the UR5 simulation from ROS-industrial and I am wondering up to what extent the simulation results could be transferred to the real robot. And how the simulation could be made as close as possible to the real life behavior.

EDIT

There is a discussion on a related matter in this link.

edit retag flag offensive close merge delete

2 Answers

Sort by » oldest newest most voted
1

answered 2014-08-06 07:06:05 -0600

jalfonso gravatar image

updated 2014-08-06 07:06:59 -0600

Hello arennuit.

Question 1: Would I be able to reuse the same geometric and dynamic parameters and gains in the simulation and get a stable simulation with a behaviour in simulation comparable to the real life behaviour?

In the Community Forum 2014 ROS-Industrial, people COMAU (manufacturer of robots) showed GAZEBO model of one of his robot. The model runs same algorithms and parameters that control real robot, so you can reproduce the behavior of the real robot in the gazebo.

From my experience developing controllers and simulation models for the automotive sector, if you have a good model, you can use the same PID in the real world as a starting point. This is the base of Hardware In the Loop (HIL) testing.

Question 2: If the behaviours are different are they totally different and unrelated, or are the gains / parameters from the real world a good start to tweak the gains / parameters of the simulated world?

If behaviours are different, this point the model is not correlated with the real plant. So the values of gains will not correlate. This could apply to UR5 because we do not have a model of the controlers of the robot and we use standar PID from ROS. For sure, the controller of the real UR5 is better than PID.

If you find information on the control algorithms for robot motors, I offer to colaborate with you to model them in GAZEBO.

Question 3: I know that gazebo uses ODE as its default engine, would the answer be the same with bullet or another engine? I don´t know.

edit flag offensive delete link more

Comments

Hi Alfonso, and thanks for you answer, it gives an interesting insight on the situation. FYI I opened a discussion on a related matter [here](https://github.com/ros-industrial/universal_robot/issues/87).

arennuit gravatar imagearennuit ( 2014-08-06 07:44:11 -0600 )edit

Indeed the UR5 low-level controllers are certainly not simple PIDs... Is their design proprietary? They do not use some kind of well known maxon controller or so? Anyone knowing anything on this matter?

arennuit gravatar imagearennuit ( 2014-08-06 07:59:05 -0600 )edit

I guess this is the moment a bit of help from universal robots corp. would help... Though I guess they will either do a package for ROS all by themselves or do nothing... rather than participating to an open effort they do not have full control on... Any feedback from someone at UR?

arennuit gravatar imagearennuit ( 2014-08-07 02:39:32 -0600 )edit
1

answered 2014-08-07 11:44:31 -0600

Adolfo Rodriguez T gravatar image

I don't expect the gains to work for both the real hardware and its simulation because I doubt that the dynamics are the same. They might be qualitatively similar, but things like system friction and damping are hard to get right in simulation (if they are simulated at all). In my experience, gains can and will be considerably different.

Also, some hardware platforms implement controllers on the actuators, not the joints. Most Gazebo robot simulations I know of have no notion of actuators, so at least there's a scaling factor to apply between joint and actuator controllers. This is assuming simple reducers, if you have more complex transmissions then things get more hairy.

In the Community Forum 2014 ROS-Industrial, people COMAU (manufacturer of robots) showed GAZEBO model of one of his robot. The model runs same algorithms and parameters that control real robot, so you can reproduce the behavior of the real robot in the gazebo.

@jalfonso Which parameters are you referring to here?. I don't think the statement applies to the controller gains.

edit flag offensive delete link more

Comments

Hi Adolfo and thanks for you answer. From my experience I also believe that the simulation has a different physical behaviour than real life. Now you stated "the simulation does not have the concept of actuators" but isn't that the reason of being of "transmissions" like the SimpleTransmission?

arennuit gravatar imagearennuit ( 2014-08-08 03:35:50 -0600 )edit

Your Answer

Please start posting anonymously - your entry will be published after you log in or create a new account.

Add Answer

Question Tools

2 followers

Stats

Asked: 2014-08-05 05:00:41 -0600

Seen: 432 times

Last updated: Aug 07 '14