IEEE ARM 2026 (Supplementary material): Validation of autoencoders for collision detection of articulated robots using time series data of motor currents
Loading...
Date
2026-05-29
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Description
# Validation of autoencoders for collision detection of articulated robots using time series data of motor currents
The dataset contains code, autoencoder models and visualisations relating to the publication: "Validation of autoencoders for collision detection of articulated robots using time series data of motor currents" in the Proceedings of 2026 International Conference on Advanced Robotics and Mechatronics (ICARM), Paris, France.
## Abstract:
Articulated robots are nowadays an integral part of production lines. Due to their flexibility in application, accuracy and reliability, they are particularly well suited to today's automation requirements based on individual customer needs. To ensure the safety of machines and their environment, modern articulated robots are equipped with numerous sensors and undergo intensive monitoring to detect errors such as collisions. New artificial intelligence based approaches offer a alternative to expensive collision detectors, as they can efficiently process large amounts of data. This study examines the use of autoencoders, for collision detection in time series data of motor currents. The objective and novelty of this study is to ascertain whether autoencoders can help to detect collisions exclusively on the basis of time series data of the motor current in the joints of an articulated robots, without the necessity for additional domain knowledge. In order to evaluate the suitability of the autoencoder, a large-scale grid search has been performed on different autoencoder configurations. The objective of this study is twofold. Firstly, the objective is to evaluate the general suitability of autoencoders in reconstructing time series data of the motor current in the specified use case of articulated robots with regard to collision detection. Secondly, the study seeks to examine the impact of the hyperparameters of the configuration on the collision detection performance.
# The dataset contains:
## AE_Gridsearch_Tensorflow
- Contains the program for training autoencoders using a grid search with various parameter variations.
- The data from the “Trainingdata_without_Collisions” folder is used for training. The autoencoders are trained to map or reconstruct the data.
- The trained autoencoders and history of the training process are saved in this folder. The trained autoencoders are used for the validation process in the "Validation" folder.
## Trainingdata_without_Collisions
- Contains one time data series of the actual current from an elbow joint motor in a UR5e robotic arm during operation without collisions.
- The operation comprises 3,000 linear movements of the robot in random directions with a random acceleration between 0.1 and 1.5 m/s² and a random maximum speed of 0.1 to 3 m/s.
## RobotControl
- Contains the programs for controlling the robot arm and for reading the motor currents.
## Validation
- Contains all programs, models, and predictions required to verify the experiment.
- The main program uses the autoencoders and validates them on the dataset of motor currents involving collisions included in this folder.
- There are several collision data packages.
- xx_neg and xx_pos contain data of collisions with a negative or positive direction in x-Direction of the collision force, respectively.
- The Name of the subfolders indicate the strength of the collision with lvl form 1 to 7, "a" for the acceleration and "v" for the speed of the robot. The "nr" stands for the number of the collision scenario. For each collision and setting there are 10 scenario.
- Only data packages named with "short" were used in the end. The other data packages are also collision scenarios but where to weak for the autoencoder to detect. Short stands for the length of the collision leaver that was used. The shorter one had a stronger collision force.
- The folder contains also the prediction data, analyses and visualisations scirpts for the models, showing how well they performed.
## Load Cell
- Contains the programs for the load cell used to record the resistance force characteristics of the collisions. The code runs on the scale’s microcontroller.
## venv
- The virutal environment for the python code. It contains all necessary libraries and dependencies to run the code in the other folders.
Citation
Endorsement
Project(s)
Faculty
Collections
License
Except where otherwise noted, this license is described as CC BY-NC-ND 4.0 International