Any guides on general parameter tuning for articulated objects?

[StoneT2000]

Hi all,

I have been playing around with newton recently and was wondering about if there were any general robot/articulated object tuning guides. Looking at some of the franka examples there are a few parameter changes like

rigid gap: https://github.com/newton-physics/newton/blob/main/newton/examples/contacts/example_brick_stacking.py#L496

and

solimp https://github.com/newton-physics/newton/blob/main/newton/examples/contacts/example_brick_stacking.py#L537-L547

some (or all?) of which are derived from mujoco. Was curious if there was a general guide on how these parameters should be tuned with respect to sim accuracy and performance. E.g. I would imagine smaller rigid_gap usually means faster simulation (it seems similar to the contact offset parameter of physx?)

thanks!

[jcarius-nv]

Hi @StoneT2000 , I'm not aware we have more specific things than what you found already, but we're tracking this request in #3099 now. It is anyways on our roadmap to improve the documentation in this direction.

[StoneT2000]

perfect, thank you

[jcarius-nv]

A concrete note on rigid_gap: it is the default contact detection gap for shapes that do not set ShapeConfig.gap. It is additive per shape, expands the broad-phase AABB by margin + gap, and causes contacts to be generated earlier. Smaller values can reduce candidate contacts and sometimes improve performance, but they also reduce speculative contact detection and can make tunneling/missed contacts more likely. Larger values are usually more robust for fast/thin geometry but may increase contact count/solver work.

For solimp/solref, the MuJoCo-backed paths intentionally follow MuJoCo-style constraint parameters, so MuJoCo’s solver-parameter docs are currently the best reference until we add the Newton-specific tuning guide tracked in #3099.