PyTouch 0.4.2 and its current status
PyTouch is a tactile-processing library built on PyTorch. The project's official README calls version 0.4.2 a beta, while PyPI lists 0.4.2 as the latest release and dates it to November 13, 2021.
The official GitHub repository says it was archived on April 1, 2024 and is now read-only. This guide therefore documents the project's published interface; it does not imply active maintenance or compatibility with current Python and PyTorch releases.
Install PyTouch and initialize DIGIT
The project README gives this installation command:
pip install pytouch --upgrade
The 2021 paper initializes PyTouch for DIGIT touch and slip detection as follows:
import pytouch as pt
digit_pt = pt.init(
pt.sensors.DigitSensor,
tasks={pt.tasks.TouchDetect, pt.tasks.SlipDetect},
)
The paper says PyTouch downloads the corresponding pre-trained models when a supported sensor is selected. At publication time, its supported sensor list comprised DIGIT, OmniTact and GelSight.
Tasks, training and sensor support
For DIGIT, the paper's example loads touch-detection and slip-detection tasks. More broadly, the authors list contact classification, slip detection, contact-area estimation, and interfaces for training and transfer learning.
PyTouch also provides higher-level task abstractions for determining whether contact is occurring or an object is slipping. The paper says additional hardware can be configured with hardware-specific model settings, and that trained weights can be used with different input devices without changing application software.
For its DIGIT touch-detection experiment, the paper reports more than 2,278 samples collected across three DIGIT devices. The authors evaluated PyTouch with real-world tactile-sensor data on contact detection, slip detection and object-pose estimation.
What the paper establishes
The PyTouch paper was submitted in May 2021 and accepted at ICRA 2021. Its authors presented PyTouch as the first machine-learning library dedicated to touch-signal processing. That is the paper's attributed claim from 2021, not a statement about every tactile library available today.
The code remains publicly readable under an MIT license, but the archived repository and old package release are important constraints. Before using PyTouch in a new stack, test it in an isolated environment and verify its dependencies instead of assuming support for current runtimes.
Sources
This article was researched and fact-checked against the following sources:
- PyTouch | PyTouch (facebookresearch.github.io)
- [2105.12791] PyTouch: A Machine Learning Library for Touch Processing (arxiv.org)
- [2105.12791] PyTouch: A Machine Learning Library for Touch Processing (ar5iv.labs.arxiv.org)
- PyTouch: A Machine Learning Library for Touch Processing (arxiv.org)
- facebookresearch/PyTouch (github.com)
- pytouch · PyPI (pypi.org)