Creating a custom Sionna RT scene (XML) without Blender

[natanzi]

Hello,
I am having a fundamental issue setting up a static scene in Sionna RT within a Docker environment. My goal is to load a static scene (walls, obstacles) from a Mitsuba XML file and then programmatically add Transmitters and Receivers.

Environment:

• Base Docker Image: nvcr.io/nvidia/tensorflow:25.02-tf2-py3
• Sionna Version: Installed via pip install sionna[rt] inside the container.
• Mitsuba Variant: mi.set_variant("llvm_ad_mono_polarized")

The Problem:
According to the documentation and API design, the correct workflow seems to be defining geometry in an XML file with a placeholder, and then linking a Python-defined RadioMaterial to it.

However, when I try to load a minimal scene with this placeholder, it fails.

My scene.xml file:

XML

<scene version="3.0.0">
    <integrator type="path"/>
    <bsdf type="holder_material" id="placeholder"/>
    <shape type="cube" id="my_object">
        <ref id="placeholder"/>
    </shape>
</scene>

My Python code:

import mitsuba as mi
import sionna.rt as rt
mi.set_variant("llvm_ad_mono_polarized")
scene = rt.load_scene("scene.xml")

The Final Error:
This code fails with the following error:

RuntimeError: ​[xml.cpp:1078] Error while loading "<string>": could not instantiate bsdf plugin of type "holder_material": 
[xml.cpp:1078]   [PluginManager] Plugin "holder_material" not found!

This error indicates that the Sionna holder_material plugin is not registered with the Mitsuba XML parser at the time of loading. All my previous attempts at purely programmatic scene creation (without XML) also failed due to persistent and unresolvable TypeError issues with mi.Transform4f.

My Question:
Is this a known issue with this version setup, or am I missing a crucial step for registering Sionna's custom plugins with the Mitsuba parser before loading an XML file? What is the definitive, robust method for creating a scene with custom static geometry in this environment?

Thank you for any help.
Milad

[merlinND]

Hello @natanzi,

The first issue is that the plugin name is holder-material, not holder_material:
https://github.com/NVlabs/sionna-rt/blob/38ff0af9ffcade89ab06391b2ebb39293ae484b2/src/sionna/rt/radio_materials/holder_material.py#L137

However, after fixing the name, you would run into another issue. In fact, the holder material is added automatically by the process_xml() function. You do not have to specify holder materials yourself.
https://github.com/NVlabs/sionna-rt/blob/38ff0af9ffcade89ab06391b2ebb39293ae484b2/src/sionna/rt/scene.py#L1060-L1080

Instead, you should include an actual radio material in your scene, for example:

<scene version="3.0.0">
    <integrator type="path"/>
    <bsdf type="itu-radio-material" id="placeholder">
        <string name="type" value="concrete"/>
    </bsdf>
    <shape type="cube" id="my_object">
        <ref id="placeholder"/>
    </shape>
</scene>

Note that this is unrelated to Docker.