git clone --recurse-submodules https://github.com/notvenky/eFlesh.git
cd eFlesh
conda env create -f env.yml
conda activate eflesh

System pre-requisites

sudo apt-get update && sudo apt-get install -y build-essential cmake libgmp-dev libmpfr-dev libcgal-dev libeigen3-dev libsuitesparse-dev libboost-all-dev

cd microstructure/microstructure_inflators && chmod +x build.sh && ./build.sh

You're now all set to use regular.ipynb and cut-cell.ipynbto make your own eFlesh sensors, ensure to provide the correct paths against all marked palceholders - like path to your OBJ/STL fle.

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.

Please upload the arduino code located in arduino/5X_eflesh_stream/5X_eflesh_stream.ino to the qtPy. 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:

1. Cut-Cell Microstructures for Two-scale Structural Optimization
2. Learning Precise, Contact-Rich Manipulation through Uncalibrated Tactile Skins
3. AnySkin: Plug-and-play Skin Sensing for Robotic Touch
4. 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}
}