Object contact_test() Segmentation Fault when colliding with non-floor object in Replica

[denfed]

Hello!

I am having an issue with the replica dataset and some custom code I have written to insert human objects into the scene. Piecing together some of the code from the ECCV 2020 tutorial, I got to the point where I have an agent at some random point in the scene, and I place a human object in front of the agent. I am currently trying to make a robust way of placing that object on the floor, or on top of whatever object it is over currently (so there isn't any clipping). Looking at tutorials, I can use an Object's contact_test() function, and it works well when that inserted object is clipping through the floor (i.e. contact_test returns true when it is clipping, and False when it is not. I can use this to iteratively lower an object until it starts clipping, which would result in the object being on the floor).

However, I noticed the simulator running into a segmentation fault when running the contact_test when the object is clipping with a non-floor object, such as a table. Here is a picture of an inserted object, that is clipping, and contact_test is returned true:

And here is an example where contact_test() results in a segmentation fault with the simulator:

I can provide the code that I created to do all this if needed, but maybe someone has a fast answer as to why this is happening.

Any help with this would be great!

[denfed]

Doing some digging on my own, it seems to be correlated to some of the other issues on habitat-sim in relation to issues loading the Replica dataset with bullet physics. I'm thinking that the complexity of the meshes within tables and the non-floor objects are too complex and cause the crash when trying to detect collisions. I tested this same idea within a matterport3d scene, and everything seems to work fine. I might instead try working with the HM3D dataset, instead of Replica.

I would still like to hear others thoughts though!

[aclegg3]

Hey @denfed,

This is a known issue with Replica and Bullet physics integration. You are correct that the geometry density is the root cause here.

I believe the best immediate solution is to use decimated meshes for collision instead of the render mesh. For example, you could use the MeshOptimizer project to generate a reduced-geometry mesh. The resulting mesh can be configured for use in Habitat by referencing it in the stage_config.json file as the "collision_asset".

Regarding your application of finding the correct placement for the object once collisions are working, you should check out the snap_down utility in Habitat-lab. Essentially this uses the object's bounding box and a raycast against the scene geometry to determine the best placement height for an object to rest on the support surface below it.

[denfed]

Oh wow, that snap_down function is perfect! I wrote some code to iteratively lower an object until it collides, but this seems to be a better way of doing it. Thank you!

[aclegg3]

Glad that helped. Let me know if you have success with the decimation. I don't think we'll have cycles for this in the near term. Note: related to #613