[ODE] discontinuities/accuracy of integration

[Jon Watte (ODE)]

Could you describe what you mean by "discontinuities" ?

ODE is time stepped. This means that it's discontinuous at every time 
step; it's a stairstep function, not a smooth matematical function.

Regarding the solver, if you're using dWorldStep, the solver uses a 
traditional full matrix solver; you could probably take out the part 
where it starts solving the matrix (after creating it), and put in 
whatever you want. There are already three solvers to choose from: 
dWorldStep, dWorldStepFast, and dWorldQuickStep, so there should be 
ample samples for a fourth implementation :-)

Cheers,

			/ h+


Marc Toussaint wrote:
> Hi there!
> 
> I'm using ODE in a scientific context, simulating robots to test 
> machine  learning techniques. Actually the accuracy of the physical 
> integration is  not critical for this application -- as long as the data 
> collected from  the simulation is smooth.
> 
> However, ODE quite often produces discontinuities in joint velocities,  
> especially if there are substantial torques applying on the joints.  
> Consequently, the accelerations that I monitor for the system become  
> unrealistic.
> 
> I figured that the origin of such discontinuities might be how ODE  
> realizes joint error reduction. (There are NO collisions/contacts in my  
> scenario.) I tried playing around with ERP and CFM setting for 
> different  joints, but it seemed to me that (1) also for an optimal 
> setting I  couldn't get rid of the discontinuities completely, (2) 
> finding an optimal  setting (which is different for each joint, 
> depending on the attached  loads) is itself a hard problem.
> 
> Q1: Can you confirm that the joint error correction mechanisms is most  
> likely the origin of such discontinuities?
> 
> Q2: Is there a way to access the ``internal error-reducing forces'' (or  
> whathever other mechanisms there is) in an accurate quantitive way?
> 
> Q3: Is there a principled way to get rid of the discontinuities (other  
> than playing around with ERP and CFM parameters)?
> 
> Q4: Is there a chance that the core of ODE (the physical integration  
> engine) can be replaced/modified such that hard constraints are 
> fulfilled  exactly? E.g., could one easily replace the engine by a DAS 
> solver like  DASPK?
> 
> Thanks!
> Marc.
>