[ODE] Contact point merging strategy

[Jon Watte (ODE)]

Mark Williams wrote:
> Jon Watte (ODE) wrote:
>   
>> Yes, it's that "simple".
>>
>> You may actually want to ignore the similarity of normals, or allow 
>> for a big slop, depending on your intended result; 
>>     
>
> Under what circumstances might you want to ignore the similarity of the 
> normals?
>
>   

For example, if a ball hits an edge of a trimesh box, you will get two 
normals: one from the edge normal to one face, and one from the edge 
normal to the other edge. Averaging these two normals into a single 
normal will usually give you good results, and put less strain on the 
constraint solver (because the two normals may generate very different 
constraints).

>> driving over a "flat" trimesh with a ball can be pretty bouncy unless 
>> you filter pretty aggressively based on contact proximity. You do want 
>> to average all the normals you filter out per geometric point, though.
>>     
>
> Is it good enough to do a component-wise arithmetic average of the 
> normals followed by a renormalisation in your experience?
>
>   

In my experience, yes.