Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Toward Artificial Palpation: Representation Learning of Touch on Soft Bodies
Authors: Zohar Rimon, Elisei Shafer, Tal Tepper, Efrat Shimron, Aviv Tamar
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We investigate a proof of concept for an artificial palpation method based on self-supervised learning. To validate our approach, we both develop a simulation environment and collect a real-world dataset of soft objects and corresponding ground truth images obtained by magnetic resonance imaging (MRI). We collect palpation sequences using a robot equipped with a tactile sensor, and train a model that predicts sensory readings at different positions on the object. We investigate the representation learned in this process, and demonstrate its use in imaging and change detection. |
| Researcher Affiliation | Academia | Zohar Rimon Elisei Shafer Tal Tepper Efrat Shimron Aviv Tamar Technion Israel Institute of Technology Correspondence to EMAIL. |
| Pseudocode | No | The paper describes the algorithms and architectures (e.g., encoder-decoder framework, GRU, MLP, flow matching) textually and with diagrams (Figure 2), but no explicit pseudocode or algorithm blocks are provided. |
| Open Source Code | Yes | Correspondence to EMAIL. Code and data are available at github.com/zoharri/Artificial Palpation |
| Open Datasets | Yes | Correspondence to EMAIL. Code and data are available at github.com/zoharri/Artificial Palpation |
| Dataset Splits | Yes | For the tactile imaging problem, we randomly split the trials to train and test (with the same split for the representation learning and the imaging). As for the change detection, we split per insert configuration, where all trials from the same configuration are either in train or test (this is done to have negative examples in the change detection task). |
| Hardware Specification | Yes | All of the training procedures, including the self-supervised phase and the image reconstruction training ran on a cluster of 12 A4000 GPUs, although each single run needed only a single GPU to run. |
| Software Dependencies | No | The paper mentions 'panda-py [Elsner, 2023]' as a tool used, but does not provide a specific version number for the software itself. It also mentions 'Adam [Kingma and Ba, 2014]' and 'Mu Jo Co [Todorov et al., 2012]' and other frameworks without specifying their versions. |
| Experiment Setup | Yes | Training Objective: In our implementation we set K = K = 64. Self-supervised Training Technical Details: The FLE is simply adding a linear projection for the forces and a basic sinusoidal Positional Encoding (PE) for the 6-dim locations (both the projection and the PE are of size 256). The sequence encoder is a one layer RNN with a GRU and with a 1024 hidden size. The force decoder first linearly projects the input representation and uses PE on the desired reconstruction to size 1020, and adds both together. Next, a three-layer MLP is used (with a 2048, 1024 hidden sizes) to predict the forces at the desired location. Adam optimizer hyperparameters (Table 3b): β1 0.2, β2 0.999, lr 0.001 |