suddhu / gpis-touch-public

(In progress: see roadmap) Gaussian process implicit surface generation from manipulator contact measurements, for object modeling

GP implicit surfaces for shape estimation

 

      

This library generates an implicit surface representation from sparse tactile information, used for the paper Tactile SLAM: Real-time inference of shape and pose from planar pushing

Folders

• cpp/: The C++ .cpp and .h files
• data/: Has contact and shape data
  • python get_contact.py --shape rect1 lets you select contact points and normals to make your own dataset. First click sets the contact point, second click draws the normal. Normal must be away from the object.
  • shapes/: the ground truth shape .mat files and scripts
  • contacts/: The generated contact files are stored here
• matlab/: GPshape.m is the launch file and viz_shape.m has the visualization code
  • standalone/: old MATLAB code which implements a basic version of GPIS
• mex/: Contains the makefile, mexfile, and test_gp.cpp

CMakeLists Notes

• Modify the EIGEN3_INCLUDE_DIR in mex/CMakeLists.txt and cpp/CMakeLists.txt
• Modify EIGEN_PATH in mex/make_GPShape.m
• Modify BOOST_ROOT in cpp/CMakeLists.txt

Mex executable

Compile and run from MATLAB

Compile

cd mex/
make_GPShape

Run

cd matlab/
GPshape

C++ executable

Compile

cd mex/
mkdir build
cd build/
cmake ..
make -j

Run

./test_gp

Reference and Acknowledgements

• The mex functions have been adapted from Lee, Bhoram, et al. "Online continuous mapping using gaussian process implicit surfaces." 2019 International Conference on Robotics and Automation (ICRA). IEEE, 2019 (github)
• Another useful open-source reference was "Dexterous grasping under shape uncertainty", Miao Li, Kaiyu Hang, Danica Kragic and Aude Billard, Robots and Autonomous Systems, 2015 (github)
• The contouring functions are from conrec.
• C++ library to generate mesh grid: meshgen
• Shape models taken from the MIT push dataset
• cpp plotting uses matplotlib-cpp
• kdtree_eigen.h is developed by Fabian Meyer based on "Analysis of Approximate Nearest Neighbor Searching with Clustered Point Sets" by Songrit Maneewongvatana and David M. Mount

Citation

Feel free to use the library as you please. If you find it helpful, please consider referencing:

@article{suresh2020tactile,
  title={Tactile SLAM: Real-time inference of shape and pose from planar pushing},
  author={Suresh, Sudharshan and Bauza, Maria and Yu, Kuan-Ting and Mangelson, Joshua G and Rodriguez, Alberto and Kaess, Michael},
  journal={arXiv preprint arXiv:2011.07044},
  year={2020}
}

Roadmap

• 3D object reconstruction
• More kernels