Real2code: Reconstruct Articulated Objects via Code Generation

Mandi Zhao, Yijia Weng, Dominik Bauer, Shuran Song

Arxiv | Website

Installation

Use conda environment with Python 3.9, and install packages from the provided .yaml file

conda create -n real2code python=3.9
conda activate real2code
conda env update --file environment.yml --prune

Code Overview

Data Generation & Processing

1. Download the pre-processed version of PartNet-Mobility data from UMPNet and from the Sapien official website (this website has web-based interactive visualization for the objects). Note that the objects were manually inspected and de-duplicated: for Eyeglasses category, 11 objects from the original site was removed, and results in 54 objects in total, from which we selected 5 test objects.

2. Use blender_render.py to process and render RGBD images from PartNet-Mobility data.

• If you see error xcb_connection_has_error() returned true, try unsetting DISPLAY variable (e.g. export DISPLAY=).

3. Use preprocess_data.py to generate OBB-relative MJCF code data from the raw URDFs for LLM fine-tuning.

See data_utils/ for detailed implementations of the helper functions.

Kinematics-Aware SAM Fine-tuning

See image_seg/ Example commands to start fine-tuning:

cd image_seg 
DATADIR=xxx # your data path
python tune_sam.py --blender --run_name sam_v2 --wandb --data_dir $DATADIR --points --prompts_per_mask 16 --lr 1e-3 --wandb --fc_weight 1

Shape Completion

See shape_complete/, we use Blender-rendered RGBD images to generate partially-observable point clouds inputs; kaolin for processing ground-truth mesh to generate occupancy label grids.

LLM fine-tuning

We use a custom fork of Open-Flamingo: https://github.com/mlfoundations/open_flamingo. More details avaliable soon.

Real World Evaluation

See real_obj/. We use DUSt3R to achieve reconstruction from multi-view pose-free RGB images, the DUSt3R-generated 3D pointmaps are provided in the real world dataset below.

Dataset

Synthetic Data

Our dataset is built on top of PartNet-Mobilty assets, and the same set of objects are used for training and testing throughout our SAM fine-tuning, shape completion model training, and LLM fine-tuning modules. The full dataset will be released here: https://drive.google.com/drive/folders/1rkUP7NBRQX5h6ixJr9SvX0Vh3fhj1YqO?usp=drive_link

Real-world Objects

We have released the real objects data used for evaluating Real2Code. These are objects found in the common lab/household settings around Stanford campus. Raw data is captured using a LiDAR-equipped iPhone camera and the 3dScanner App

• Download: Google Drive Link
• Structure: each object folder is structured as follows:

  ls obj_id/
  - raw/
  - sam/
  - a list of (id.jpg, id_mask.png, id_scene.npz),
  

  Each id corresponds to one 512x512 RGB image selected from the raw dataset, e.g. 00000.jpg; id_mask.png is the foreground object mask obtained from prompting the SAM model with randomly sampled query points in the image margin area; id_scene.npz is the globally-aligned 3D point-cloud obtained from DUSt3R.