Where can I get the correct 3D models for running spatial_scene_grammars?

[siva-shiba]

Hello
I’m currently working with the Steerable Scene Generation system and would like to use spatial_scene_grammars to generate procedural datasets for training.
However, I’m not entirely sure how to correctly obtain and organize the 3D model assets required for the examples.

💭 Question

• Which 3D models are required to run the example grammars (especially dimsum_restaurant)?
• What’s the correct folder structure for those models?

🧩 Background

I'm trying to run the procedural scene generation examples in spatial_scene_grammars repository as part of the Steerable Scene Generation pipeline.
However, I'm not sure where to get all the required 3D model assets or how to organize them properly.

🔧 What I Did

Following the instructions in the Steerable Scene Generation README, I downloaded the following folders from the MIT SharePoint link:
🔗 MIT SharePoint Model Link
Then, I renamed and placed them under the root of spatial_scene_grammars repository as:

anzu/
greg_table/
gazebo/

⚠️ Issue

Even after placing these folders, several models still appear to be missing.
For example, I do not have the 3D model for a Coke can.

When I try to run the dimsum_restaurant (High-Clutter) grammar, I encounter the following error:

$ python spatial_scene_grammars_examples/dimsum_restaurant/generate_target_dataset_parallel.py ../dataset/a.pkl --points 1_000_000

error: URI 'package://anzu/models/home_kitchen/junk/coke.sdf' resolved to "*****" which does not exist.

🧠 Example

Currently, I placed the models as follows:

spatial_scene_grammars/
├── anzu/
│   └── package.xml
├── greg_table/
│   └── package.xml
├── gazebo/
│   └── package.xml

Any pointers or links to the correct 3D model set would be super helpful.
Thanks again!

[nepfaff]

Hi, thank you for your interest in our work!

You are on the right path and almost there. We had to rename some assets to be allowed to release them under TRI's licenses.

You already correctly renamed tri -> anzu and greg to greg_table. The code resolves the paths like os.path.abspath(os.path.expanduser("models/gazebo/package.xml")). Hence, they must be placed in the folder that you are running the script from. Your location is correct for it to work with the command that you shared.

We also renamed and retextured our can asset from coke to rlg_banana_concentrate. You can either rename it back to coke.sdf or change the path in the grammar.py file.
Also, carlisle_plate_mesh_collision was renamed to plate_v1_mesh_collision, ikea_dinera_plate_8in_mesh_collision to plate_v2_mesh_collision, and ikea_dinera_bowl to bowl.

As a heads up: The high-clutter restaurant is the example with the highest sample rejection rate. Hence, you might need to run that script for a while to get a sample, depending on the number of parallel processes. We documented our throughput for that scene type in the paper appendix. For testing, it might be best to start with a fast grammar, such as the living room shelf, or the low-clutter TRI table.
I also recommend debugging with our test_grammar.py scripts, as these will single-step visualize the scenes in the browser rather than writing them to a file. This makes it easy to check that they look as desired.

Let me know if that solves the issue or if you encounter any other ones.