Using sequence distribution on multiple USD assets, to spawn 1 different object each frame

Omniverse Synthetic Data Generation (SDG)
1

Randomizing appearance, placement and orientation of an existing 3D assets with a built-in writer — Omniverse Extensions latest documentation.
based on this example i am trying to use sequence distribution to spawn different variations of similar assets for every frame, just like spawning the worker object but a little bit different. For example, I have two similar USD files of truck models (target object) that I am trying to spawn in the scene at the same spot everytime. I have referenced the USD paths and put them in a list, but i can’t seem to use sequence distribution to spawn a different truck each frame. Is there guidance for this or if someone has tried this before?

2

I’m unsure if this is the correct or the most efficient way of doing this but using rep.modify.visibility() got working results in a similar case.

Basically, spawn all the models and just show one at a time.
Here’s a scalable version, of course with two models you can just alternate true and false.

paths = ['omniverse://localhost/assets/truck1',
         'omniverse://localhost/assets/truck2']

trucks = []
for x in paths:
    truck_usd = rep.create.from_usd(x, semantics=[('class', 'truck')])
    trucks.append(truck_usd)

viz_matrix = []
for x in range (len(trucks)):
    arr = [False] * len(trucks)
    arr[x] = True
    viz_matrix.append(arr)

with rep.trigger.on_frame(num_frames=10):
    for idx, truck in enumerate(trucks):
        with truck:
            rep.modify.visibility(rep.distribution.sequence(viz_matrix[idx]))

Context to above example, just dealing with tables not trucks: reference-SDGenerator/final_generator.py at main · jkuhno/reference-SDGenerator (github.com)

If I recall correct, there has been discussion of a similar case on this forum but I failed to find the post to include in this answer.

1 Like
3

I would use the randomizer.instantiate way of doing it.

1 Like
6

@j.kuhno - Your 2nd post is correct.

Loading multiple assets and hiding them is the more efficient solution.

Using instantiate or a similar method with create and destroy the objects each frame, which would likely be slower overall.

You could likely simplify it a little bit by not using a custom function and instead using 2 Replicator triggers.

See a code example of toggling visibility here: How to generate objects binary masks for synthetic data offline pose generation - Omniverse / Isaac Sim - NVIDIA Developer Forums

1 Like
7

If you have unlimited VRAM…