Question about Sionna‑RT’s path‑tracing design

[LuHaofan]

Hi everyone – I’m trying to understand why Sionna‑RT’s ray‑tracing pipeline
differs from modern photorealistic renderers (e.g. Mitsuba).
I may be missing something obvious, so any clarifications would be welcome!

---

1. Forward vs. backward tracing

• Graphics engines usually trace backwards: rays start at the camera
  and are connected (via NEE, BDPT, etc.) to light sources.
• Sionna‑RT traces forward from every transmitter and later stitches
  mirror segments with the image method.

Typically backward tracing is more efficient than forward tracing. So why Sionna adopt to use forward tracing?

---

2. Image method vs. Monte‑Carlo for specular paths

Renderers sample even perfectly specular surfaces stochastically and rely on
variance reduction (MIS, MNEE, …).
Sionna‑RT instead:

• uses the image method for mirror / refraction chains, then
• Monte‑Carlo samples only the diffuse component.

The image method is great for zero variance but seems fragile when the CAD
geometry deviates from reality.
Couldn’t we widen the specular lobe a bit, sample it MC‑style, and avoid the
deterministic post‑pass?

---

TL;DR

What fundamental constraints of wireless‑channel simulation make the
graphics approaches (back‑tracing + MC everywhere) less attractive for
Sionna‑RT?

Thanks a lot for any insight!

[faycalaa]

Hi @LuHaofan ,

This is a very good question!

I assume that you are referring to the path solver, as the radio map solver does not use the image method.

The aim of the path solver is to find a set of paths that connect two points in space, a source (transmitter or transmit antenna) and a target (receiver or receive antenna). Specifically, the aim of the path solver is not to compute an integral over paths observed at the receiver, but to identify the individual paths that connect the transmitter to the receiver and for each path, compute the corresponding complex-valued coefficient, real-valued delay, angles of arrival and departure, etc.

This is quite different from rendering, where the aim is to integrate over a set of paths that impinge on a sensor. When aiming to connect two endpoints, there is no difference between forward or backward tracing as the two endpoints play a symmetric role.

The image method is used to identify specular chains (and specular suffixes) among a set of path candidates identified by shooting and bouncing rays. Using other techniques, such as relaxing the directions of specular reflections and refractions to lobes or using detection spheres, would result in approximate values for the path coefficients, delays, angles of arrival/departure, etc. Such approximations are not desired in the built-in path solver. Note that specular chains are important in wireless propagation because they typically carry a significant part of the transported energy.

Radio maps are much closer to rendering, as we aim to integrate the received power over measurement cells.

[LuHaofan]

Hi @faycalaa ,

Thank you so much for your clarification!