This Documentation is functionally complete as of 12/1/2024. More minor details will be added from time to time.
🆕 [2024-12-1] Tutorial on Module 2 Active Touch Selection and Module 3 Local Geometric Optimization Updated.
🆕 [2024-11-15] Installation for VLM Reasoning & Active Touch Selection Updated.
🆕 [2024-10-17] Installation for Hardware Integration/3D Printing Updated.
🆕 [2024-10-15] Installation for Robotics Software Updated.
🆕 [2024-10-11] Made Public
This is the official implementation of FusionSense: Bridging Common Sense, Vision, and Touch for Robust Sparse-View Reconstruction
Irving Fang, Kairui Shi, Xujin He, Siqi Tan, Yifan Wang, Hanwen Zhao, Hung-Jui Huang, Wenzhen Yuan, Chen Feng, Jing Zhang
FusionSense is a novel 3D reconstruction framework that enables robots to fuse priors from foundation models with highly sparse observations from vision and tactile sensors. It enables visually and geometrically accurate scene and object reconstruction, even for conventionally challenging objects.
This repo has been tested on Ubuntu 20.04 and 22.04. The real-world experiment is conducted on 22.04 as ROS2 Humble requires it.
We used a depth camera mounted on a robot arm powered by ROS2 to acquire pictures with accurate pose information. We also used a tactile sensor for Active Touch Selection.
If you have no need for this part, feel free to jump into Step 1 for the 3D Gaussian pipeline of Robust Global Shape Representation and Local Geometric Optimization.
- For installing robotics software, please see Robotics Software Installation.
- For hardware integration, please see 3D Printing Instructions.
Note: ROS2 doesn't play well with Conda in general. See official doc and this issue in the ROS2 repo. As a result, in this project, ROS2 uses the minimal system Python environment and have limited direct interaction with the Python perception modules.
We will need two independent virtual environments due to some compatibility issue.
Please see DN-Splatter and Metric3D Installation
Please see Grounded-SAM-2 Installation
Please see Active Touch Selection Installation
You can see here for an example dataset structure.
Note that a lot of the folders are generated during the pipeline. The data needed to start this projects are: images, realsense_depth, and transforms.json.
The ROS2 packages I shared can be used to acquire the aforementioned data. Or you can manually format your own dataset this way.
The project assume that all the folders in the HuggingFace repo are put under FusionSense/datasets/.
If you want to let VLM classify the object, click this line.
If you want to manually specify the name, please read ahead.
Inside our main conda env
conda activate fusionsenseSet your OpenAI API key as an environment variable. Please do not directly use your API key in your code, as you may accidentally commit it to an online public repo.
export OPENAI_API_KEY='Your Key'Run this script.
python scripts/VLM.py --mode partname --data_name {DATASET_NAME}mode: Operation mode of our VLM pipeline. In this case we want it to give us partname.data_name: Name of the specific dataset folder. Example: transparent_bunny
Whether you got the name from VLM or your own brain, we can proceed with that.
Switch your conda env first
conda activate G-SAM-2Inside the submodule of our Grounded-SAM2
cd Grounded-SAM2-for-maskingRun the script to extract masks by setting your dataset path and object name prompt text. The prompt text ends with an '.' at the end.
You can use something you came up with, or one proposed by the VLM. In our experience, both works fine.
eg. --path /home/irving/FusionSense/dataset/transparent_bunny --text 'transparent bunny statue.'
python grounded_sam2_hf_model_imgs_MaskExtract.py --path {ABSOLUTE_PATH} --text {TEXT_PROMPT_FOR_TARGET_OBJ}You will see mask_imgs in the newly created /masks folder, and you can check /annotated folder to see the results more directly.
set train.txt with images id.
You can pick images that have better masking for better final result. Although in our experiment we didn't cherrypick which images to use except that we want images to be relatively evenly spread out.
This pipeline is mostly run in Nerfstudio.
You can change configs at configs/config.py
First go back to our main conda environment and main folder
conda activate fusionsensecd ..Then we run
python scripts/train.py --data_name {DATASET_NAME} --model_name {MODEL_NAME} --load_touches False --configs {CONFIG_PATH} --verbose {True, False} --vram_size {"large", "small"}data_name: Name of the dataset foldermodel_name: Name of the model you train. It will impact the output and eval folder name. You can technically name this whatever you want.load_touches: Whether to load tactile data. Default=Falseconfigs: Path to the Nerfstudio config fileverbose: False: Only show important logs. True: Show all logs. Default=Falsevram_size: "large" or "small". Decides the foundation models variants used in the pipeline. Default="large"
An example using the provided data would be:
python scripts/train.py --data_name transparent_bunny --model_name 9view --configs configs/config.py --vram_size largeThe previous step should produce output in the outputs/{DATASET_NAME}/{MODEL_NAME} folder.
In the same fusionsense conda environment, we run
python scripts/VLM.py --mode touch --data_name [DATASET_NAME] --model_name [MODEL_NAME] --mesh_name [MESH_NAME]data_name: Name of the dataset folder. Should be the same as the one in the last module.model_name: Name of the model you just trained.mesh_name: Name of the mesh you want to use. We recommandpoisson_mesh_surface_level_0.3_closest_gaussianas it strikes a good balance between details and smoothness. But you can use whatever you want in theoutputsfolder.
Optionally, you can use the following parameters
object_name: The name of the object you come up with. Should be a string. Default:None, a VLM will be called to classify the object.part_name: The name of the parts of this object. Should be a list of string. Example:--part_name ear head body base. Default:None, a VLM will be called to come up with the part names.--vlm_name: Name of the specific VLM model. Currently, only OpenAI API is supported. Default:gpt-4o
An example using the provided data would be:
python scripts/VLM.py --mode touch --data_name transparent_bunny --model_name 9view --mesh_name poisson_mesh_surface_level_0.3_closest_gaussianWe will have a few proposed points for touching. After collecting tactile reading on these points with teleoperating or other methods provided by my ROS2 packages, we can proceed to the next module.
After acquiring the tactile data, we should make sure it's put in datasets/{DATASET_NAME}/tactile, along with a datasets/{DATASET_NAME}/gelsight_transform.json.
Then, in the fusionsense conda environment, we run
python scripts/train.py --data_name {DATASET_NAME} --model_name {MODEL_NAME} --load_touches True --configs {CONFIG_PATH} --verbose {True, False} --vram_size {"large", "small"}Note that, this time we want to modify the MODEL_NAME so that we do not overwrite the results from the first module.
An example using the provided data would be:
python scripts/train.py --data_name transparent_bunny --model_name 9view_touch --load_touches True --configs configs/config.py --vram_size largeCongrats! You have run through the entire pipeline!
For render jpeg or mp4 outputs using nerfstudio, we recommend install ffmpeg in the conda environment:
conda install -c conda-forge x264=='1!161.3030' ffmpeg=4.3.2To render outputs of pretrained models:
python scripts/render_video.py camera-path --load_config your-model-config --camera_path_filename camera_path.json --rendered_output_names rgb depth normalmore details in nerfstudio ns-render.