How to obtain training data for diffusion planner

[mcudsm]

Hi, thanks for the great work!

I am currently trying to train the diffusion planner , but I am a bit confused about how to obtain or prepare the training data.

Could you please clarify:

Which dataset is used for training (nuScenes? internal dataset?)
Whether there is a preprocessing pipeline available

Thanks a lot!

[SakodaShintaro]

We used the internal data to train the published weights.

The converter from rosbags to dataset is implemented in the following directory:

https://github.com/tier4/Diffusion-Planner/blob/tier4-main/cpp_tools/src/autoware_diffusion_planner_tools/src/data_converter.cpp

You can build it by

cd cpp_tools
./prepare_repos.sh
./build.sh

Note: Rosbags need the following topics:
https://github.com/tier4/Diffusion-Planner/blob/1ec7495620c82090c03c926021e6bdfacd1a55d4/cpp_tools/src/autoware_diffusion_planner_tools/src/data_converter.cpp#L477-L483

[mcudsm]

Thanks !

I have another question about incorporating traffic participants. Is it possible to collect training data by integrating with CARLA for co-simulation?

Or do you have any other recommended approaches for adding traffic participants?

Thanks in advance!

[SakodaShintaro]

For the first trial, you can use simple_planning_simulator in autoware.

It is a difficult question how to execute a more realistic traffic simulation. We are also searching for a way.

At least, CARLA will be used as a hub to connect autoware and traffic simulators, I think.

As far as I know, there are some candidates:

• Vissim
• SUMO
• TeraSim

They are not totally integrated with autoware, so we have to make some efforts.

[salighan]

Thank you
Under autoware_diffusion_planner, there are some model weights released as ONNX.
Do you provide model checkpoints for each model weights released for fine-tuning purposes?