Collision detection without stepping

[RodsCoimbra]

Hi,
I'm creating a new task in IsaacLab where a mobile manipulator must navigate to a goal while avoiding collisions with the environment. The goal for the base is defined by an x, y position and orientation, while the goal for the arm is a set of seven joint positions (not the final position of the end-effector).

The problem is that there will be obstacles, so I wanted to spawn a robot at the goal and check whether placing the arm in the target configuration would result in any collisions, to check if the goal is feasible. However, when using the overlap_mesh function, I noticed that even after moving the rigid objects to the desired pose, collisions are only detected after a simulation step.
The issue is that during training, I can't step a single environment without stepping all of them. So I'm wondering if there's a way to perform collision detection without actually advancing the simulation—something similar to PyBullet's performCollisionDetection() function.

Here’s the simple example I'm currently using to check this:

from omni.physx import get_physx_scene_query_interface
from pxr import Sdf, PhysicsSchemaTools, UsdGeom

def report_hit(collision):
    global counter
    counter+=1
    print("Collision with: ", collision.rigid_body)
    return False #stops search

def check_overlapping(path):
	global counter
	counter = 0
	path_tuple = PhysicsSchemaTools.encodeSdfPath(path)
	_ = get_physx_scene_query_interface().overlap_mesh(path_tuple[0], path_tuple[1],report_hit, False)
	print(path, " -> ", counter)
 
def get_all_descendants(prim):
	children = prim.GetChildren()
	for child in children:
		if UsdGeom.Mesh(child):
			check_overlapping(child.GetPath())
		get_all_descendants(child)

[RandomOakForest]

Thank you for posting this. Have you considered using a ray caster to detect collisions?

[RodsCoimbra]

Could you please clarify how that would work? I thought ray casters were mainly used to measure distances. How can I check if, for example, the arm of my mobile manipulator is inside another mesh? Also, don’t ray casters require a simulation step after moving the articulation (mobile manipulator) to the desired position in order to update the distances?
Also, the arm and gripper together are made up of about 15 meshes. Would I need to place a sensor on each of these meshes in every environment? I'd really appreciate some clarification on this. Thank you!