The Official PyTorch Implementation of "ACROSS: A Deformation-Based Cross-Modal Representation for Robotic Tactile Perception" (ICRA 2025)

Wadhah Zai El Amri   ·   Malte Kuhlmann   ·   Nicolás Navarro-Guerrero

Accepted to 2025 IEEE Conference on Robotics and Automation (ICRA 2025)

Paper page can be found here.

DIGIT BioTac Isaac Gym.

This package contains the code for the BioTac to Digit pipeline. It allows to convert BioTac Signal into Digit Images.

The Figure above describes the steps of the pipeline which is contained in this repository. The pipeline delivers results similar to the image below

To train the pipeline yourself the networks can be split in three phases:

Phase 1:

In this phase the network SVB, MVB and MVD are trained. The code for that can be found in BioTac_Signal_Reconstruction and Mesh_Reconstruction.

Phase 2:

In this phase the translation networks S2MPN and M2MPN can be trained as the build upon the networks of phase 1. The code for that can be found in across/BioTac_Signal_to_Deformation and across/BioTac_to_DIGIT_Deformation.

Phase 3:

This phase contains the Digit Image generation and does not need to be trained. The code for this can be found in DIGIT_Simulation.

Installation

Pipeline

To train and execute the BioTac to Digit pipeline follow these steps for installation:

1. Optional but Recommended Create a conda environment using conda create -n pipeline python=3.10.14
2. Install all required packages using pip install -e .[pipeline]

Citation

@InProceedings{ZaiElAmri2025ACROSS,
  author = {Zai El Amri, Wadhah and Kuhlmann, Malte and {Navarro-Guerrero}, Nicol{\'a}s},
  title = {{{ACROSS}}: {{A Deformation-Based Cross-Modal Representation}} for {{Robotic Tactile Perception}}},
  booktitle = {{{IEEE International Conference}} on {{Robotics}} and {{Automation}} ({{ICRA}})},
  year={2025},
}