Confused: Dynamics Non-Linear Effects

[kklankers]

Hey there,

i am starting to use RBDL. I am quite familiar with robot dynamics but I was confused by the documentation. I am implementing different controllers in pybullet using the python bindings and urdf addon.

Basically I want to have: tau = M*ddq + C(q,dq) + G(q).

I used the CompositeRigidBodyAlgorithm to get M, and I wanted to get the Coriolis term via NonlinearEffects. I thought I still need the to compute the gravity term ( I did with InverseDynamics and putting dq and ddq to zero) , because the documentation says about the "NonlinearEffects" function:

/** \brief Computes the coriolis forces
*

• This function computes the generalized forces from given generalized
• states, velocities, and accelerations:
• \f$ \tau = N(q, \dot{q}, f_\textit{ext}) \f$
• \param model rigid body model
• \param Q state vector of the internal joints
• \param QDot velocity vector of the internal joints
• \param Tau actuations of the internal joints (output)
• \param f_ext External forces acting on the body in base coordinates (optional, defaults to NULL)
  */

Am I missing something? Because I observe an error, when putting C(q,dq) and G(q) on the robot.

I found in the issue sections, that NonlinearEffects calculates both gravity and coriolis combined. Just putting C(q,dq) in my control law fixes the problems i faced and everything works like expeceted.

Thanks for your effort and this amazing library !

[mjhmilla]

Dear @kklankers,

Just to confirm your suspicion, the gravity term is indeed included in the NonlinearEffects function. You can see this by opening Dynamics.cc and scrolling to the NonlinearEffects function: the first lines of code add gravity to the system and then perform the calculation to evaluate the nonlinear effects. I imagine that @martinfelis included the gravity terms in the NonlinearEffects function because its such a trivial amount of code. However, your experience shows that this is a bit confusing.

@martinfelis, we could update the NonlinearEffects function so that it really does match Featherstone's equations one-to-one. Or we could improve the documentation. Given the spirit of RBDL is to be as fast as possible, I'm leaning towards improving the documentation. What are your thoughts?

Matt

[kklankers]

Dear @mjhmilla ,

thanks for your response and considerations.

Maybe I can add another thought? Some robots control interfaces i worked with already compensate for gravity. In this case, it would be benefical to just have access to the coriolis forces. I faced situations were I needed the gravity term as well.

Maybe you can consider a function call that is flexible to these situations, e.g. a boolean flag to also consider the gravity, because I also see your point of making it as fast as possible. Anyway, thanks for your time !

Kilian

[martinfelis]

Hi @kklankers (and @mjhmilla!),

the function NonlinearEffects() was derived from InverseDynamics and therefore includes the gravity term. By doing so it very efficiently propagates gravity throughout the tree.

@kklankers easiest solution to support robots that already compensate the gravity term would be to set Model::gravity to zero as that is used for the virtual body at index 0.

Best regards,
Martin

[mjhmilla]

Good point, @martinfelis.