Bug Description

The kernel handle_contact_pairs [src] identifies and stores contact points between shapes of rigid bodies being in collision. One of the the kernel outputs is the array contact_tids. This kernel is necessary to identify which element in the other output arrays (contact_point0 etc) was written in which thread and thus, corresponds to which contact point of the corresponding shape.

For instance, if I want to figure out which mesh vertex is represented by contact_point0[x], I am able to figure it out using where(contact_tids) == x.

The kernel starts writing to the output tensors at index 0 (even though the variable name next_count here suggests something else, wp.atomic_add returns the old value). At the same time, contact_tids is initialized as a zero array. This makes it impossible to find the tid at which the very first element of the arrays has been written, because where(contact_tids) == x evaluates to true for this tid, but also all where no contact was detected.

We can see the effect, e.g. by doing:

tids = torch.tensor(model.rigid_contact_tids)
count = torch.tensor(model.rigid_contact_count)
print(f"Count: {count.item()}, Nonzero entries in tid: {(tids != 0).sum().item()}")

==> Count: 102, Nonzero entries in tid: 101

As far as I see, an easy fix would be initializing contact_tids with -1, as done for the contact shapes.

System Information

No response