Broadpahse not returning expected results

[rishavgmr]

Hello all,

I am just starting to learn about PhysX 5.0 and have been trying out some of the features that will be relevant to my application. I need to use PhysX for collision detection and distance measurement.

For testing collision detection, I loaded a octomap file (.bt) by reading in each leaf voxel as a box primitive into a scene and then running a overlap scene query against a test sphere. This test returned expected results, specifically that the sphere was in fact in collision with one of the voxels/box primitives. Here is the code I used for that:

// includes

int main() {
    // PhysX setup including, Foundation, Pvd and Physics

    // Setup scene
    physx::PxSceneDesc scene_desc(physics->getTolerancesScale());
    scene_desc.gravity = physx::PxVec3(0.0f, -9.81f, 0.0f);
    physx::PxCpuDispatcher* gDispatcher = physx::PxDefaultCpuDispatcherCreate(2);
    scene_desc.cpuDispatcher = gDispatcher;
    scene_desc.filterShader = physx::PxDefaultSimulationFilterShader;
    physx::PxScene* scene = physics->createScene(scene_desc);

    std::string filename = "/workspaces/advanced-coliisions/data/octomap.bt";

    std::unique_ptr<octomap::OcTree> env_octree;
    env_octree = std::make_unique<octomap::OcTree>(filename);

    double box_side_length = env_octree->begin_leafs().getSize();
    physx::PxMaterial* material = physics->createMaterial(0.1f, 0.1f, 0.1f);
    physx::PxShape* voxel_shape =
        physics->createShape(physx::PxBoxGeometry(box_side_length / 2.0, box_side_length / 2.0, box_side_length / 2.0), *material);

    physx::PxAggregateFilterHint hint = physx::PxGetAggregateFilterHint(physx::PxAggregateType::eSTATIC, false);
    physx::PxAggregate* aggregate = physics->createAggregate(env_octree->getNumLeafNodes(), env_octree->getNumLeafNodes(), hint);

    double x, y, z;
    for (auto it = env_octree->begin_leafs(); it != env_octree->end_leafs(); ++it) {
        physx::PxRigidStatic* actor = physics->createRigidStatic(physx::PxTransform(physx::PxVec3(it.getX(), it.getY(), it.getZ())));
        x = it.getX();
        y = it.getY();
        z = it.getZ();

        actor->attachShape(*voxel_shape);
        aggregate->addActor(*actor);
    }
    scene->addAggregate(*aggregate);

    physx::PxOverlapBuffer hit;
    physx::PxSphereGeometry overlap_shape(0.15f);
    physx::PxQuat quat2(0.0f, 0.0f, 0.0f, 1.0f);
    physx::PxTransform transf2(-0.7, 0.400000, -0.005950, quat2);
    physx::PxTransform shape_pose(transf2);

    physx::PxQueryFilterData query_filter;
    query_filter.flags |= physx::PxQueryFlag::eANY_HIT; 

    bool overlap_result = scene->overlap(overlap_shape, shape_pose, hit, query_filter);
    std::cout << "Overlap result : " << hit.hasAnyHits() << std::endl;

    return 0;
}

Now I wanted to go to the next step which was distance computation for which I figured I would need broadphase as a first step to calculate which voxels are closest to the sphere. For that I wrote this:

// includes

int main() {
    // PhysX setup including, Foundation, Pvd and Physics

    // Setup scene
    std::string filename = "/workspaces/advanced-coliisions/data/octomap.bt";

    std::unique_ptr<octomap::OcTree> env_octree;
    env_octree = std::make_unique<octomap::OcTree>(filename);

    double box_side_length = env_octree->begin_leafs().getSize();
    physx::PxMaterial* material = physics->createMaterial(0.1f, 0.1f, 0.1f);
    physx::PxShape* voxel_shape =
        physics->createShape(physx::PxBoxGeometry(box_side_length / 2.0, box_side_length / 2.0, box_side_length / 2.0), *material);
    std::cout << "Box side length : " << box_side_length << std::endl;
    std::cin.get();

    physx::PxBroadPhaseDesc bp_desc(physx::PxBroadPhaseType::eABP);
    bp_desc.mDiscardStaticVsKinematic = false;
    std::cout << "Broadphase spec valid : " << bp_desc.isValid() << std::endl;
    physx::PxBroadPhase* broadphase = physx::PxCreateBroadPhase(bp_desc);
    physx::PxAABBManager* aabb_manager = physx::PxCreateAABBManager(*broadphase);

    int obj_idx = 0;
    for (auto it = env_octree->begin_leafs(); it != env_octree->end_leafs(); ++it) {
        physx::PxBounds3 bounds;
        physx::PxGeometryQuery::computeGeomBounds(bounds, physx::PxBoxGeometry(box_side_length / 2.0, box_side_length / 2.0, box_side_length / 2.0),
                                                  physx::PxTransform(physx::PxVec3(it.getX(), it.getY(), it.getZ())));
        aabb_manager->addObject(obj_idx, bounds, physx::PxGetBroadPhaseStaticFilterGroup());
        obj_idx++;
    }

    physx::PxSphereGeometry overlap_shape(0.15f);
    physx::PxQuat quat2(0.0f, 0.0f, 0.0f, 1.0f);
    physx::PxTransform transf2(-0.7, 0.400000, -0.005950, quat2);
    physx::PxTransform shape_pose(transf2);

    physx::PxBounds3 bounds;
    physx::PxGeometryQuery::computeGeomBounds(bounds, overlap_shape, shape_pose);
    aabb_manager->addObject(env_octree->getNumLeafNodes(), bounds, physx::PxGetBroadPhaseDynamicFilterGroup(env_octree->getNumLeafNodes()));

    physx::PxBroadPhaseResults results;
    aabb_manager->update(results);
    aabb_manager->fetchResults(results);

    std::cout << "Num pairs " << results.mNbCreatedPairs << " and " << results.mNbDeletedPairs << std::endl;

    for (int i = 0; i < results.mNbCreatedPairs; i++) {
        std::cout << "Possible collision between " << results.mCreatedPairs[i].mID0 << " and " << results.mCreatedPairs[i].mID1 << std::endl;
    }

    return 0;
}

As you can see the sphere is placed at the same location as the last one and has the same size. However, I always see the following output:

Num pairs 0 and 0

i.e. The number of created pairs and deleted pairs are both 0. Is there something wrong I am doing in the setup?

[PierreTerdiman]

Isn't it because of the duplicate updates?

    aabb_manager->update(results);
    aabb_manager->update(results);

The results are incremental so if nothing changes between the two updates, the number of "created pairs" will be zero indeed (it will be non-zero after the first update, not after the second one).

[rishavgmr]

Ok so I looked through the source file you attached. The only difference I saw was that I was not using the centerExtents function to create a "PxBounds3" object. In the original, I would simply use computeGeomBounds and add the resultant PxBounds3 object to the aabb_manager using addObject. But now I tried by creating a new PxBounds3 object using centerExtents and then adding to the aabb_manager and it seems to suddenly work! I am also seeing it overlap with 2 voxels.

Earlier:

physx::PxBounds3 bounds;
        physx::PxGeometryQuery::computeGeomBounds(bounds, physx::PxBoxGeometry(box_side_length / 2.0, box_side_length / 2.0, box_side_length / 2.0),
                                                  physx::PxTransform(physx::PxVec3(it.getX(), it.getY(), it.getZ())));
        aabb_manager->addObject(obj_idx, bounds, physx::PxGetBroadPhaseStaticFilterGroup());
        obj_idx++;

Now:

physx::PxBounds3 bounds;
        physx::PxGeometryQuery::computeGeomBounds(bounds, physx::PxBoxGeometry(box_side_length / 2.0, box_side_length / 2.0, box_side_length / 2.0),
                                                  physx::PxTransform(physx::PxVec3(it.getX(), it.getY(), it.getZ())));
        const physx::PxVec3 center(bounds.getCenter().x, bounds.getCenter().y, bounds.getCenter().z);
        const physx::PxVec3 extents(0.5 * bounds.getDimensions().x, 0.5 * bounds.getDimensions().y, 0.5 * bounds.getDimensions().z);
        physx::PxBounds3 bounds_new = physx::PxBounds3::centerExtents(center, extents);
        aabb_manager->addObject(obj_idx, bounds_new, physx::PxGetBroadPhaseStaticFilterGroup());
        obj_idx++;

Is this expected behavior? What is the reason that one works but the other doesn't?

[PierreTerdiman]

Sorry, no idea, that makes no sense at all :)

This:

const physx::PxVec3 center(bounds.getCenter().x, bounds.getCenter().y, bounds.getCenter().z);
const physx::PxVec3 extents(0.5 * bounds.getDimensions().x, 0.5 * bounds.getDimensions().y, 0.5 * bounds.getDimensions().z);

Should be the same as:

const physx::PxVec3 center = bounds.getCenter();
const physx::PxVec3 extents = bounds.getExtents();

And thus bounds should be equal to bounds_new. And the bounds you computed with "computeGeomBounds" are the ones you gave me in that Excel file, right? So they're the ones I used as well, and they should be correct.

Maybe you can assert that bounds.minimum == new_bounds.minimum, same for maximum, and see if anything gets triggered. But otherwise I can't explain this.

[PierreTerdiman]

Could it be some compiler bug maybe? Are you compiling in Release or Debug? (Try both).

If it's a compiler bug, maybe adding any kind of code between computeGeomBounds and addObject is enough to avoid it. In that case, try adding something else instead of the centerExtents call (maybe just a printf).

Sorry, running out of ideas. Maybe there is something obvious I don't see at this time (past midnight here). Will have another look tomorrow.

[rishavgmr]

Thank you so much for all the help. So I found the issue was that I was calling aabb_manager->fetchResults(results) after aabb_manager->update(results). I guess that replaces the results that the update call was providing as no voxel had moved, thus no new pairs were created. However, calling aabb_manager->fetchResults(results) after aabb_manager->update() [NOTE the empty function argument], gives the expected output.
This solves the problem.

Tldr: Don't call fetchResults after update(results). Only call fetchResults after update().

[PierreTerdiman]

Oh wow. I should have seen this. I think it shows that this API could be clearer, and perhaps the single-function-helper should have been named something more explicit like "updateAndFetchResults" instead of just "update". Taking a note of that.

[rishavgmr]

That will make the purpose of the function clearer.