Number of contacts between boxes and spheres in MJX

[S-A-M-J]

Hi im using MJX for reinforcement learning through brax. I have a robot with 4 legs and feet (spheres) which interacts with a voxelized environment represented by boxes with sizes that vary but are bigger in every axis than the spheres diameter. Logically for very hard material this should reduce the max contact points between 1 sphere and the boxes to 3. Now even with not infinite material i would assume the number of contact points should be more but not exessively more since. But with these values for both spheres and boxes: solref="0.001 1" solimp="0.99 0.99 0.001" i get multiple hundred contact points (by checking ncon). Usually about 300 to 500 to be exact. Therefore i was wondering if this is expected behaviour and how i might reduce them for training performance without having the feet clip through the box geoms

[btaba]

Hi @S-A-M-J , some light reading https://mujoco.readthedocs.io/en/stable/mjx.html#mjx-the-sharp-bits, notably the "contact" bits. We have some contact culling parameters you can use, but we don't expect an environment like this to have decent behavior in the current version of MJX (sphere-filtering between boxes and spheres, has pathologically failing cases if the boxes are oblong). The good news, is that https://github.com/google-deepmind/mujoco_warp doesn't have the same failing pattern, and we are integrating this into MJX, so you'll be able to get decent performance for the scene you're describing soon. For now, I would read the docs and futz with MJX-specific contact culling parameters.

[btaba]

There should be one contact point between a single sphere and a box. The number of contact points is fixed by max_contact_points, not all of which are active contacts (dist < 0). Can you use hfield instead or no? You can optionally send an XML or screenshot so we can better understand what you're trying to do?

[S-A-M-J]

Unfortunately i cant use hfield because i need geometry above the robot since it should be able to hang from the ceiling or walls. The image kind of shows the environment i am working with. The robot consists of extendable boom arms that are supposed to attach to walls etc.

If there should be only one contact point between a sphere and a box in theory i should be able to reduce the max_contact_points to 12 (4 spheres touching max. 3 boxes at once) + the box to box contacts of the main body without clipping right? At least if contacts are handled by highest penetration to lowest?

Is there any way to filter active contacts in mjx to only see the ones with dist < 0. Or even access the dist value? I would need that to implement the sticking to walls and ceiling feature.

[btaba]

Here is an example futzing with the contact data

The boxes look not too oblong, I think you may be able to narrow to 4 potential boxes per sphere with the sphere filter, then the box-box might have at least 4 pairs as well. Set max_contact_points high and see how low max_geom_pairs can go without some gnarly behavior. With that being said, this is where the Warp stuff should come to save the day. I may OSS the MJX-Warp integration before the official supported release if there's sufficient interest. It's usually good to have more eyes on a piece of code!

[S-A-M-J]

Can you quantify what high means in this context?
That would be great!

[btaba]

You can omit max_contact_points altogether and see what ncon is.

[S-A-M-J]

Alright so i completetly omitted the max_contact_points and set the max_geom_pairs and yet i still quite often get clipping where by robot foot will go straight through the cave wall voxels. Is there anything else i could do beside turning that further up which does not really seem to help and im running out of memory, or perhaps decreasing the sim timestep from 0.004.
Could it help to change the solref/solimp values of the feet/walls to be even more stiff? Current values are: solref="0.001 1" solimp="0.99 0.99 0.001"