[Question] An unrealistic physical phenomenon occurs during RL in a scene containing particles

[NAMHAUK]

Question

I'm trying to do reinforcement learning in a scene with particles, articulation characters, and rigid objects. However, I wonder why rigid objects and articulation characters create unrealistic physical phenomena when reinforcement learning in the scene which exist particles.

I know that I have to run on SimulationCfg with device = cpu to render and simulate the current particle moving.

Also, I know that when I run on SimulationCfg with device = cuda:0 (gpu), particle rendering seems to have stopped on the screen at first, but it actually moves and does physical motion. (Like the video attached below. A light rigid box bounces in the pouring water.)
https://github.com/user-attachments/assets/69fb598e-60f4-497c-b773-81b798cdb9a1

In SimulationCfg, device = cuda:0 (gpu) works without any physical problems when all these objects are brought to the scene without reinforcement learning, as shown in the video below. (The scene where the part is filled in the tank made of rigid objects, the light rigid object is floating on the water, and the character is in the water)
https://github.com/user-attachments/assets/5b54157e-5d7d-4fa8-a609-49cab329ee40

However, when I run train.py and play.py in device = cuda:0 (gpu) in Simulation Cfg and provided by the isaac lab, a strange physical phenomenon occurs, as shown in the video below. (rigid object, an articulation character goes down through the floor and bounces up)
https://github.com/user-attachments/assets/ce1b5cf9-4637-4f72-8ec3-fe95f7fc3823

If I run train.py or play.py from SimulationCfg to device = cpu, it works as well as the video below.
https://github.com/user-attachments/assets/f4426903-625b-4629-aeba-0acf5ea13615

in conclusion,

• When train.py and play.py are cpu,gpu in the absence of particles, all physical calculations run well.
• It just runs well when it's brought to scene, not train.py or play.py .
• simulationCfg print results when just fetching or RL, all simulationCfg states are the same
• Just in case, I upgraded to the latest version, but the same problem occurred.

What I don't understand is that when I just import and run a scene with particles, it works well without any problems, but why do strange physical collisions only happen when I run it with train.py or play.py.

Thank you for reviewing this.

[RandomOakForest]

This is a great post for our Discussions section. Will move to that section for the team to follow up. Thanks.