Venkatesh Pattabiraman, Zizhou Huang, Daniele Panozzo, Denis Zorin, Lerrel Pinto and Raunaq Bhirangi
git clone --recurse-submodules https://github.com/notvenky/eFlesh.git
cd eFlesh
conda env create -f env.yml
To run the cut-cell microstructure optimizers and generate the lattice structures, there are some dependancies to be installed. Please use the following links provided and download oneTBB and BOOST from source.
cd eFlesh/microstructure/microstructure_inflators
mkdir build && cd build
Please replace the path placeholders below to the correct local paths, during the installation.
cmake -DCMAKE_BUILD_TYPE=release .. -DTBB_ROOT=</path/to/oneTBB/installation> -DBoost_NO_SYSTEM_PATHS=ON -DBOOST_ROOT=</path/to/boost_1_83_0>
make -j4 stitch_cells_cli
make -j4 cut_cells_cli
make -j4 stack_cells
In the conversion notebooks regular.ipynb and cut-cell.ipynb, ensure to provide the correct paths against all marked palceholders.
We slice the generated STL file with pouches, using OrcaSlicer or Bambu Studio and 3D print it with TPU 95A on a Bambu Lab X1 Carbon 3D printer.
We use N52 neodymium magnets of dimensions: 1/8" thickness, 3/8" diameter for the standard cuboidal instance and many of the medium-large form factor sensors. For the fingertips, we use N52 magnets of dimensions 1/16" thickness, 3/16" diameter. According to the user's requirements, the magnet pouches can be easily tweaked, and so magnets of any dimensions can be used.
We use the rigid magnetometer PCBs used in Reskin and AnySkin. Details can be found in the circuit section of Reskin's repository.
We characterize eFlesh's spatial resolution, normal force and shear force prediction accuracy through controlled experiments, The curated datasets can be found in characterization/datasets/. For training, we use a simple two layered MLP with 128 nodes (python train.py --mode <spatial/normal/shear> --folder /path/to/corresponding/dataset).
We grasp different objects using the Hello Stretch Robot equipped with eFlesh, and tug at it to collect our dataset. The dataset can be found in slip_detection/data, and the trained classifier is slip_detection/checkpoints/eflesh_linear.pkl.
We perform four precise manipulation tasks, using the Visuo-Skin framework, achieving an average success rate of >90%. Representative videos of trained policies can be found on our website.
eFlesh draws upon these prior works:
- Cut-Cell Microstructures for Two-scale Structural Optimization
- Learning Precise, Contact-Rich Manipulation through Uncalibrated Tactile Skins
- AnySkin: Plug-and-play Skin Sensing for Robotic Touch
- ReSkin: versatile, replaceable, lasting tactile skins
If you build on our work or find it useful, please cite it using the following bibtex
@article{pattabiraman2025eflesh,
title={eFlesh: Highly customizable Magnetic Touch Sensing using Cut-Cell Microstructures},
author={Pattabiraman, Venkatesh and Huang, Zizhou and Panozzo, Daniele and Zorin, Denis and Pinto, Lerrel and Bhirangi, Raunaq},
journal={arXiv preprint arXiv:2506.09994},
year={2025}
}