Question: Clarification on ONNX model and missing DPM-Solver in inference code

[FleverX]

Hi, thanks for your excellent work on this project!

I have a question about the inference code of the diffusion planner — I noticed that it directly loads an ONNX model. Could you please clarify where this ONNX model comes from?

Also, is the model output directly the predicted trajectory? In the original Diffusion Planner paper, the model’s output represents the denoised direction rather than the full trajectory. So I’m wondering why the inference code here doesn’t use a DPM-Solver (or any diffusion sampling process)?

[SakodaShintaro]

Thank you for your interest.

Could you please clarify where this ONNX model comes from?

We are developing in the following repository. Please note that it is very unstable.

https://github.com/tier4/Diffusion-Planner

create_onnx.sh is here.

https://github.com/tier4/Diffusion-Planner/blob/tier4-main/ros_scripts/create_onnx.sh

So I’m wondering why the inference code here doesn’t use a DPM-Solver (or any diffusion sampling process)?

Currently, the number of denoising steps is fixed, and the step width and other parameters are statically determined, so I understand that the entire denoising step becomes a single deterministic feed-forward process. Therefore, ONNX contains the whole process, and the trajectory can be output without reimplementing the solver in C++. (We don't use the classifier-free guidance.)

[FleverX]

Thank you very much for your kind and detailed response！