How to get mjcb_control to point to object member function?

Hi,

I wrote (in MuJoCo Pro, C++, VS2013) a sample program without classes, in which the model was loaded in the main() function (like your simulator example). After loading the model from xml I set: mjcb_control = myController;

Where I had:
void myController(const mjModel* mArg, mjData* dArg) {
//Some useful code here
dArg->act[someindex] = some actuator force/torque;
}
this worked perfectly fine.

Now I want to encapsulate everything in a class, e.g.
class Robot {
public:
Robot();
~Robot();
private:
mjModel* m = 0;
mjData* d = 0;
void myController(const mjModel*, mjData*);
double someUsefulState = 0.0;
};
To create multiple robots (separate simulations to step through) in parallel.

In the class constructor I want to attach/install the callback to the private myController(...) of this specific instance of the class, and maybe I want for this object instance to point it to a different controller function (i.e. it cannot be a static member function, also due to using instance specific data since it uses someUsefulState) that might be given as a constructor argument (not in the MWE below).

I tried something like this:
Robot::Robot()
{
//Something else happens first
//...

//Finally
void(*cbFunPtr)(const mjModel* m, mjData* d) = this->myController;
mjcb_control = cbFunPtr;
}

But installing the function pointer that way does not work. I don't know how to do the syntax for this properly. It mostly complains that whatever I do cannot be assigned to an entity of type mjfGeneric, or missing argument lists.

Thanks!

 

I was attempting this to allow for multi-threaded parallel calculation.

But even though, if mjcb_control is only 'global', is there a way to pass an extra argument to the callback function (apart from mjModel and the mjData of the thread) so that in my control callback function I knows which mjData it is operating on?

My current plan is to put some identifier, or even controller state data, in mjData->userdata.

 

Well, actually my controller is a neuromuscular controller with multiple muscles that have their own state that have to be integrated over time, but the joint actuator torques are a non-linear function of that state (i.e. I cannot use 'dyntype integrator' or 'mjcb_act_dyn'). The other mjData models should of course not use another model's muscle state.

The controller dynamics are of the form:
act = ctrl = f(s,s_dot,t)

with f some nonlinear function of some state s, its derivative s_dot (these are not actuator or states of the rigid bodies in the model) and possibly mjData->time. State s is the time integrated from s_dot. It is not of the form act_dot = g(act,ctrl,t), with some nonlinear function g.

As far as I understood, I cannot do this with 'dyntype integrator' or by using 'mjcb_act_dyn', since I cannot have MuJoCo integrate s_dot to s (I could start making 'useless' actuators that have 'integrator dyntype' to have s_dot integrate over time and use in function f to calculate ctrl/act?). It feels like I have to integrate s_dot myself with SIEuler or RK4 myself 'in parallel' to MuJoCo within the controller callback, since if I would store state s in mjData->userdata every time (sub)step, it is not integrated over time.
As Emo Todorov says: "Assuming each copy of mjData contains different state information", how can I get mjData to contain state information and integrate it by itself (so not userdata), without it being an act_dot to act state or a rigid body state.

Hopefully I completely missed the/some point and what I want actually can be done.
So actually my problem has shifted from what I mentioned in the OP. Since if this integration can be done, a single global controller callback function would be sufficient, as Emo Todorov suggested.

 

Ok thank you for the responses . Now I know the proper way(s) to solve the problem.

Will it be implemented in the future that one could have 'arbitrary' states be integrated the MuJoCo integrator? Like appending the q_pos/q_vel vectors?

 

Hello,

Let me refresh this thread! Have you implemented this feature yet? Your reply to my previous thread seems implying you haven't, but just in order to make sure that I did not misunderstand your answer.

http://www.mujoco.org/forum/index.php?threads/arbitrary-order-of-actuator-dynamics.3586/


Thank you!