Robots Performance Monitoring in Autonomous Manufacturing

Operations Using Machine Learning and Big Data

Ahmed Bendaouia

1,3 a

, Salma Messaoudi

, El Hassan Abdelwahed

2 b

and Jianzhi Li

1,3

Institute for Advanced Manufacturing (IAM), University of Texas Rio Grande Valley, U.S.A.

Cadi Ayyad University, UCA, Faculty of Sciences Semlalia, Computer Systems Engineering Laboratory (LISI),

Marrakech, Morocco

Manufacturing and Industrial Engineering Department, University of Texas Rio Grande Valley, U.S.A.

Keywords:

Additive Manufacturing, Autonomous Experimentation, Fault Detection, Machine Learning, Deep Learning,

Anomaly Detection, Predictive Maintenance.

Abstract:

Additive manufacturing has revolutionized industrial automation by enabling ﬂexible and precise production

processes. Ensuring the reliability of robotic systems remains a critical challenge. In this study, data-driven

approaches are employed to automatically detect faults in the UR5 robot with six joints using Artiﬁcial In-

telligence. By analyzing sensor data across different combinations of payload, speed, and temperature, this

work applies feature engineering and anomaly detection techniques to enhance fault prediction. New features

are generated, including binarized anomaly indicators using the interquartile range method and a difference-

based time feature to account for the sequential and irregular nature of robot time data. These engineered

features allow the use of neural networks (including LSTM), Random Forest, KNN, and GBM models to clas-

sify anomalies in position, velocity, and current. A key objective is to evaluate which anomaly type is the

most sensitive by analyzing error metrics such as MAE and RMSE, providing insights into the most critical

factors affecting robot performance. The experimental ﬁndings highlight the superiority of Gradient Boost-

ing Machine and Random Forest in balancing accuracy and computational efﬁciency, achieving over 99%

test accuracy while maintaining short training times. These two models outperform the others, which show

a noticeable gap either in training time or test accuracy, demonstrating their effectiveness in improving fault

detection and performance monitoring strategies in autonomous experimentation.

1 INTRODUCTION

Industrial robotics has become a cornerstone of mod-

ern manufacturing, enabling automation, precision,

and efﬁciency across various domains, including ad-

ditive manufacturing. The increasing complexity

of robotic systems demands robust fault detection

mechanisms to ensure reliability, reduce downtime,

and optimize predictive maintenance. Autonomous

robotic systems, such as the UR5 robot, operate un-

der dynamic conditions involving variations in pay-

load, speed, and temperature. These factors signif-

icantly inﬂuence system behavior, leading to unex-

pected anomalies that, if undetected, may compro-

mise production quality and system longevity.

Recent advancements in machine learning (ML)

and deep learning (DL) have demonstrated their ef-

https://orcid.org/0000-0003-0017-9285

https://orcid.org/0000-0002-2786-6707

fectiveness in real-time fault detection for industrial

robotics (Wescoat et al., 2021). Traditional rule-based

monitoring systems struggle to adapt to evolving pat-

terns of anomalies caused by environmental and op-

erational variations. Instead, ML techniques such as

Random Forest, Gradient Boosting Machine (GBM),

and deep neural networks have proven to be effective

in detecting irregularities and classifying fault types.

Deep learning models, particularly Long Short-Term

Memory (LSTM) networks, are well-suited for han-

dling sequential dependencies in robotic time-series

data, enabling the identiﬁcation of subtle deviations in

system behavior. Despite the growing application of

ML and DL in anomaly detection, several challenges

remain. Robotic fault detection involves multiple per-

formance metrics, such as position, velocity, and cur-

rent, each exhibiting varying levels of sensitivity to

operational disturbances. Identifying the most sensi-

tive metric is crucial for prioritizing predictive main-

Bendaouia, A., Messaoudi, S., Abdelwahed, E. H., Li and J.

Robots Performance Monitoring in Autonomous Manufacturing Operations Using Machine Learning and Big Data.

DOI: 10.5220/0013526800003967

In Proceedings of the 14th International Conference on Data Science, Technology and Applications (DATA 2025), pages 82-96

ISBN: 978-989-758-758-0; ISSN: 2184-285X

tenance efforts and minimizing false alarms. Com-

putational efﬁciency must also be considered, as real-

time applications require rapid inference with mini-

mal latency. Ensuring the robustness of ML-based

monitoring across different operating conditions is es-

sential to generalize the fault detection framework to

diverse industrial scenarios.

This study proposes a comprehensive machine

learning framework for fault detection and sensitiv-

ity analysis in industrial robotics. The research eval-

uates various ML and DL techniques, including Ran-

dom Forest, GBM, k-Nearest Neighbors, Neural Net-

works, and LSTM, for anomaly detection in the UR5

robot. Sensitivity analysis is conducted to deter-

mine which performance metric—position, velocity,

or current is most susceptible to anomalies by an-

alyzing error metrics such as Mean Absolute Error

(MAE) and Root Mean Squared Error (RMSE). The

computational efﬁciency of different models is as-

sessed in terms of training time and inference speed

to ensure suitability for real-time industrial applica-

tions. Additionally, a structured data pipeline is de-

veloped, integrating feature engineering, outlier de-

tection, and big data processing frameworks to en-

hance model reliability and scalability. The pro-

posed approach is validated using real-world data

from the National Institute of Standards and Technol-

ogy (NIST, 2017). The dataset includes controller-

level measurements such as joint positions, velocities,

currents, torques, and temperatures under varying op-

erational conditions(Qiao et al., 2017a; Qiao et al.,

2017b). Advanced feature engineering techniques,

including difference-based transformations and in-

terquartile range-based binarization, are applied to

improve anomaly detection capabilities. The models

are trained on six independent datasets corresponding

to different combinations of payload, speed, and tem-

perature, preserving temporal consistency and assess-

ing the generalizability of the model. The analysis

also reveals that the velocity anomalies ( y velocity)

are the most sensitive in most conditions, except in

high-stress environments where the current anomalies

( y current) dominate. These insights contribute to the

reﬁnement of predictive maintenance strategies and

the optimization of real-time fault detection frame-

works in autonomous additive manufacturing opera-

tions.

The remaining of this paper is structured as fol-

lows. Section 2 highlights the literature review

and similar applications of ML for fault detection

in additive manufacturing. Section 3 presents the

methodological framework, including data prepro-

cessing, feature engineering, and model selection.

Section 4 discusses the experimental results, high-

lighting model comparisons and sensitivity analy-

sis. Section 5 provides an in-depth discussion of the

ﬁndings, focusing on their implications for industrial

robotics. Finally, Section 6 concludes the study and

outlines potential future research directions.

2 LITERATURE REVIEW

Ensuring the reliability of robotic systems in au-

tonomous additive manufacturing operations requires

advanced machine learning-based fault detection, pre-

dictive maintenance, and human-robot collaboration.

This section reviews the current state of the art in

these domains, highlighting key methodologies and

technological advancements.

2.1 Machine Learning for Fault

Detection and Predictive

Maintenance

Machine learning plays a critical role in improv-

ing the reliability of robotic systems by enabling

fault detection, predictive maintenance, and real-time

anomaly detection. Several studies have demon-

strated the effectiveness of ML algorithms such as

Random Forest, Long-Short-Term Memory networks,

and Gradient Boosting Machines in diagnosing faults

in robotic manipulators (Wescoat et al., 2021; Van

and Ceglarek, 2021). Feature engineering techniques,

including binarized anomaly indicators and time-

sequential features, improve the precision of these

models in classifying anomalies in position, velocity,

and current (Scime and Beuth, 2018; Qi et al., 2019).

Research has also integrated deep learning methods

such as BN-LSTM for modeling energy consump-

tion, optimizing power efﬁciency while maintaining

performance (Zhang et al., 2024). In addition, predic-

tive maintenance strategies have been employed using

Digital Twin technology to synchronize real-time sen-

sor data with virtual models for improved fault predic-

tion and process control (Ko et al., 2019; Malik and

Brem, 2021).

2.2 Collaborative Robotics and

Human-Robot Interaction

The emergence of collaborative robots (cobots) in

industrial applications has redeﬁned traditional au-

tomation paradigms by facilitating human-robot inter-

action (HRI). Cobots integrate artiﬁcial intelligence

(AI) to enhance adaptability and task execution, re-

ducing programming complexity, and enabling intu-

Robots Performance Monitoring in Autonomous Manufacturing Operations Using Machine Learning and Big Data

Figure 1: Anomaly Detection Workﬂow for Robot Operations.

itive teaching methodologies such as lead-through-

learn programming (Canﬁeld et al., 2021; Zaatari

et al., 2019). Studies have explored the integra-

tion of zero moment control in collaborative weld-

ing robots to improve motion stability and process

monitoring (Antonelli and Astanin, 2016; Ferraguti

et al., 2023). In addition, the role of AI in cobot

applications spans various domains, including predic-

tive control, anomaly detection, and adaptation based

on reinforcement learning to dynamic environments

(Borboni et al., 2023; Rozo et al., 2016). Further-

more, research in Industry 4.0 has emphasized the in-

tegration of robotic wireless sensor networks for real-

time process planning and geospatial data manage-

ment, further enhancing smart manufacturing frame-

works (L

aroiu et al., 2022; Butt, 2020).

2.3 Autonomous Additive

Manufacturing and Digital Twin

Integration

Advancements in AM have increasingly leveraged

AI-driven process optimization, real-time monitor-

ing, and defect prediction techniques to enhance re-

liability and efﬁciency. Autonomous robotic AM

frameworks incorporating deep reinforcement learn-

ing have demonstrated improvements in adaptive tool

path planning and material deposition accuracy (Fel-

brich et al., 2022; Fan et al., 2025). Multi-robot sys-

tems for aerial AM have further expanded the po-

tential of decentralized robotic fabrication, enabling

autonomous 3D printing in hard-to-access locations

(Zhang et al., 2022; D

orﬂer and Gerke, 2022). The

integration of digital twin frameworks with machine

learning algorithms has facilitated continuous knowl-

edge engineering in AM, improving material prop-

erty predictions and defect classiﬁcation (Raza et al.,

2022; Djuric et al., 2016). Moreover, the incorpo-

ration of cyber-physical systems and real-time AI-

driven control has enhanced the scalability and efﬁ-

ciency of intelligent AM solutions (Sharma and Cu-

pek, 2023; Alghamdy et al., 2023).

The reviewed literature underscores the need to

integrate AI-driven anomaly detection, predictive

maintenance, and collaborative robotics to improve

the reliability of autonomous AM operations. Fu-

ture research should explore hybrid methodologies

that combine digital twin frameworks, reinforcement

learning, and real-time edge computing to further ad-

vance the robustness and adaptability of robotic AM

systems.

3 METHODOLOGY

The primary objective of this study is to establish

a framework to identify the most sensitive variable

among robot performance metrics (RPM), position,

velocity, and current in robot performance, using ma-

chine learning and deep learning techniques. The

framework allows the real-time monitoring of the

RPM to assess and alert when an anomaly is detected.

Sensitivity is assessed by evaluating the mean abso-

lute error and the root mean squared error, allowing

us to rank these three variables according to their sus-

DATA 2025 - 14th International Conference on Data Science, Technology and Applications

ceptibility to faults. Understanding which variable

is deviating is crucial for optimizing fault detection

strategies and improving predictive maintenance in

autonomous additive manufacturing operations.

To achieve this, various state of the art machine

learning methods, such as k-Nearest Neighbors, Gra-

dient Boosting Machine, and Random Forest, are

compared with deep learning models like neural net-

works and long-short-term memory. This compari-

son not only assesses their accuracy and precision in

fault detection, but also assesses their ability to clas-

sify anomalies into the three categories while con-

sidering computational efﬁciency in terms of train-

ing and inference time. By balancing these aspects,

this study aims to determine the most effective model

for anomaly detection while identifying which of the

three variables should be prioritized in the monitoring

and maintenance of industrial robotic systems.

The ﬁgure 1 summarizes the anomaly detection

workﬂow. Data are collected from robot sensors and

preprocessed separately for six distinct datasets, each

corresponding to a unique combination of payload,

speed, and temperature. This division is crucial be-

cause the robot-time column is sequential within each

dataset but would lose its temporal consistency if dif-

ferent combinations were merged, as each represents

an independent operating condition set at the begin-

ning of data collection. The ﬁrst input features ”X”

represent all the columns in the dataset except those

used to calculate ”y”, while ”y” denotes the target

variables used for anomaly detection. Four target

variables are used: the ﬁrst target detects anoma-

lies globally, serving as an initial detection mecha-

nism, while the other three (y pos, y velocity, and

y current) classify the speciﬁc type of anomaly (po-

sition, velocity, or current). Each of the six datasets

is used to train machine learning and deep learning

models independently. The models are then evalu-

ated for their fault detection and classiﬁcation perfor-

mance, ensuring that the best-performing approach is

selected for reliable and efﬁcient anomaly detection in

industrial robotic applications. The massive amount

of time-series data is processed through a big data

pipeline to guarantee proper streaming and storing of

data.

3.1 Data Collection and Preprocessing

3.1.1 Data Description

The data used for this study is a series of experi-

ments conducted on the Universal Robots UR5 robot

by the National Institute of Standards and Technology

(NIST, 2017). The main goal of these experiments

was to measure the degradation of the accuracy of the

position of the robot arm over time, which is essen-

tial to assess robot health. We have conducted the

preprocessing, cleaning, and preparation of the robot

fault detection data for model training and framework

validation. The dataset consists of six subsets, each

corresponding to different combinations of tempera-

ture (normal or cold start), payload (1.6 or 4.5), and

speed (half speed or full speed). Each dataset con-

tains 115 columns, ensuring a consistent feature struc-

ture across all datasets. However, the number of rows

varies for each dataset, as detailed in table 1, which

provides the exact number of observations along with

the corresponding temperature, payload and speed

conditions for each dataset. In total, there are 153,644

observations in all six datasets. These columns in-

clude engineered features, binary classiﬁcation labels,

target variables (y), and other relevant parameters that

are partially present in the datasets.

These data are essential for building robot health

monitoring algorithms and tools, thus supporting re-

search in robot health management and predictive

maintenance.The data include different compositions

of several factors inﬂuencing robot performance:

• Payload: Two different weight values (1.6 and

4.5) are used in the experiments to observe the

impact of weight on the degradation of robot per-

formance.

• Speed: Tests were performed at different speeds

(half speed and full speed), to simulate realistic

scenarios in which the robot could operate.

• Temperature: Temperature is a key factor inﬂu-

encing the accuracy of robots. Therefore, tests are

performed under two conditions: normal temper-

ature and cold start.

• Test: Each test is performed in a speciﬁc exper-

imental setting to observe the impact of previous

variables on robot performance.

Each joint of the UR5 robot is monitored using a

7-dimensional measurement system, capturing criti-

cal parameters that characterize its performance and

condition. These seven dimensions are summarized

in Table 2. Each of these features helps analyze the

behavior of the joint under different conditions such

as speed, payload, and temperature variations. Since

the UR5 robot has six joints and each joint is char-

acterized by the 7-dimensional measurement system,

the total number of measurable features amounts to

42 core features before considering actual and target

values where applicable. However, because position,

velocity, and current include both actual and target

values, the total number of features expands to 60

features, as detailed in Table 3. In addition to these

Robots Performance Monitoring in Autonomous Manufacturing Operations Using Machine Learning and Big Data

Figure 2: Visualization of TCP Articulation Forces for the First Combination of Payload, Temperature, and Speed: (1.6,

normal, half speed).

60 core features, the data set includes additional vari-

ables that provide further insight into the overall state

of the robot.

• ROBOT TIME: Time elapsed since the controller

was started.

• ROBOT CARTESIAN COORD TOOL: Carte-

sian coordinates (x, y, z) and rotation vector

(rx, ry, rz) representing the tool’s position and

orientation in space.

• ROBOT TCP FORCE: Generalized force mea-

surements at the Tool Center Point (TCP).

These additional features are not part of the 7D mea-

surement system since they describe the overall robot

state rather than individual joint behavior.

3.1.2 Cleaning and Pre-Treatment Process

Rigorous data cleaning and pre-processing were es-

sential to prepare the dataset for analysis and model

training within our proposed framework. This in-

volved merging measurement and test datasets based

on shared payload, temperature, and speed to create

six consistent and sequential datasets, preserving the

integrity of the ROBOT T IME column. Missing or

erroneous values were addressed using interpolation,

row deletion, or average substitution, depending on

context and extent.

The ﬁgure 2 illustrates the variations in the articu-

lation forces of the center point tool for the six joints

of the UR5 robot under speciﬁc conditions: a pay-

load of 1.6 kg, normal temperature, and a speed set

to half speed. The data represent the forces measured

in Newtons over time in seconds. Each subplot corre-

sponds to a speciﬁc joint (from 1 to 6), allowing a de-

tailed analysis of the force dynamics for each articula-

tion. The ﬂuctuations in the forces reﬂect the interac-

tions between the robot and its working environment.

Maximum forces and sudden peaks are evident, indi-

cating moments of high strain on certain joints. Each

joint exhibits distinct patterns of ﬂuctuations, likely

inﬂuenced by the dynamics of the payload and the

speciﬁc temperature and speed conditions. The av-

erage force levels vary across the joints, with some

maintaining relatively stable ranges. This visualiza-

tion is crucial for identifying potential anomalies or

unusual trends in the data, which could indicate per-

formance issues or the need for preventive mainte-

nance.

DATA 2025 - 14th International Conference on Data Science, Technology and Applications

Table 1: Overview of Datasets Based on Temperature, Payload, and Speed.

Dataset Temperature Payload Speed Number of

Observations

Data1 Normal 1.6 Half Speed 25,785

Data2 Normal 1.6 Full Speed 22,434

Data3 Normal 4.5 Half Speed 29,961

Data4 Normal 4.5 Full Speed 21,770

Data5 Cold Start 4.5 Half Speed 30,956

Data6 Cold Start 4.5 Full Speed 22,738

Table 2: The 7-dimensional measurement system per joint.

Measurement Dimension Type Description

Joint Position Actual & Target Measured and desired angular position of the joint (rad).

Joint Velocity Actual & Target Measured and desired angular velocity of the joint (rad/s).

Joint Current Actual & Target Measured and desired electrical current used by the joint motor (A).

Joint Acceleration Target Desired acceleration of the joint (rad/s²).

Joint Torque Target Desired torque applied by the joint (Nm).

Joint Temperature Actual Measured temperature of the joint (°C).

Joint Control Current Actual Measured control current applied to the joint (A).

Table 3: Breakdown of the 60 measurable features derived from the 6 joints & 7-dimensional measurement system.

Measurement Dimension Type Per Joint Total for 6 Joints

Joint Position Actual & Target 2 12

Joint Velocity Actual & Target 2 12

Joint Current Actual & Target 2 12

Joint Acceleration Target 1 6

Joint Torque Target 1 6

Joint Temperature Actual 1 6

Joint Control Current Actual 1 6

Total Features 10 per joint 60

3.2 Feature Engineering

Feature Engineering is a crucial step in the devel-

opment of machine learning models, as it allows to

transform raw data into more informative variables

suitable for prediction. In this study, several feature

engineering techniques were applied to improve the

performance of the robot fault detection model. Vari-

able transformations aim to make the data more suit-

able for machine learning, by highlighting relation-

ships or simplifying complex aspects of the system

studied.

3.2.1 Creating New Variables

One of the key steps in feature engineering is to create

new variables that better capture the relationship be-

tween different data features and robot performance

degradations. In this case, the creation of several

differences-based variables is particularly relevant. In

our framework, we have calculated the following fea-

tures:

• Diff Position: This variable represents the dif-

ference between the actual measured position of

the robot (measured TCP position) and the tar-

get position. It is crucial to evaluate the robot’s

accuracy over time and detect anomalies or per-

formance degradations. It explicitly quantiﬁes the

gap between the expected position and the actual

position reached by the robot. This variable is ob-

tained by subtracting the nominal position values

from the measured positions.

• Diff Velocity & Diff Current: Similarly, two ad-

ditional variables, diff velocity and diff current,

are created using the same approach. These repre-

sent the differences between the actual and target

values of velocity and current, respectively. These

variables help to assess deviations in motion and

electrical performance, contributing to better fault

Robots Performance Monitoring in Autonomous Manufacturing Operations Using Machine Learning and Big Data

detection.

• Diff Robot Time: Another important feature

introduced is diff robot time, which represents

the difference between sequential values in the

ROBOT TIME column. Since ROBOT TIME

is sequential but not evenly spaced, this vari-

able captures the irregularity in time intervals.

Adding this feature helps machine learning and

deep learning models, especially sequential mod-

els such as LSTMs, to understand the varying time

gaps between observations, improving the accu-

racy of temporal anomaly detection.

3.2.2 Transformation into Binary Variables

Based on IQR

Outliers can be indicators of malfunctions or abnor-

mal behaviors in mechanical systems. In this project,

the interquartile range is used to detect outliers and

transform them into binary variables. Table 4 shows

the anomalies detected by this method for position,

velocity and current in different data sets. The in-

terquartile range is a statistical method that captures

values that are outside the interquartile range of the

data. IQR is deﬁned as the difference between the

third quartile (Q3) and the ﬁrst quartile (Q1):

IQR = Q

− Q

The values considered outliers are those that fall out-

side the range deﬁned by: Suspicious value < Q

−

1.5 × IQR or Suspicious value > Q

+ 1.5 × IQR As

part of the study, some variables are transformed into

binary variables based on outlier detection. These

transformations are applied to several key variables

in the system, such as joint positions and velocity.

Figure 3: Anomaly Detection in Robot Joint Position

Across Multiple Test Intervals.

The ﬁgure 3 illustrates the anomaly detection re-

sults for the DIFF POSITION 6 variable, which was

introduced through feature engineering. This vari-

able represents the absolute difference between the

actual measured position and the target position of

the sixth robot joint. The data correspond to the

test conﬁguration where payload = 1.6, speed = half

speed and temperature = normal. To identify anoma-

lies, the Interquartile Range method was applied to

the DIFF POSITION 6 values. Points detected as

anomalies (shown in orange) represent signiﬁcant de-

viations from the expected position, indicating pos-

sible faults or inaccuracies in the robot’s movement.

Normal values are represented in blue. The x-axis

corresponds to ROBOT TIME, which represents se-

quential time steps. However, there are visible gaps

between the three intervals. This occurs because the

dataset consists of three separate test runs, where data

was collected during a certain time period, followed

by a pause in collection, and then resumed for the next

test. As a result, the ﬁgure shows three distinct time

segments in which the measurements were recorded

sequentially.

Figure 4: Distribution of DIFF POSITION 2 (Normal vs.

Anomalous Values).

This ﬁgure 4 represents the distribution of the

DIFF POSITION 2 variable, which corresponds to

the difference between the actual position and the tar-

get position of joint 2 for the speciﬁc combination of

normal temperature, payload of 1.6, and half speed.

In blue: Distribution of normal values (not consid-

ered anomalies). Most of these values are very close

to 0, indicating that the actual position of joint 2

generally matches the target position. This means

that there are no signiﬁcant positioning errors in most

cases.

In red: Distribution of anomalous values detected us-

ing the interquartile range method. These values are

concentrated mainly around 0.015, indicating a sig-

niﬁcant deviation between the target and actual po-

sitions of joint 2 in certain cases. These deviations

are considered anomalies, suggesting unusual vari-

ations in the robot’s movement. This visualization

highlights the distinction between normal values and

anomalies, clearly showing that most discrepancies

are minimal, but there are cases where the error be-

comes signiﬁcant.

DATA 2025 - 14th International Conference on Data Science, Technology and Applications

Table 4: Percentage of anomalies for Position, Velocity, and Current across datasets.

Dataset Position Anomalies (%) Velocity Anomalies (%) Current Anomalies (%)

Data1 5.81% 5.63% 1.41%

Data2 8.57% 7.74% 1.99%

Data3 4.66% 4.89% 1.86%

Data4 7.44% 8.09% 1.59%

Data5 5.25% 5.32% 4.82%

Data6 7.24% 7.90% 4.70%

Total 6.39% 6.37% 2.81%

3.2.3 Advantages of Binarizing Variables

Transforming some variables into binary values helps

simplify model interpretation, because binary classi-

ﬁcation models can often better handle variables that

clearly indicate the presence or absence of certain

behaviors such as an anomaly or failure. Binariza-

tion helps quickly identify anomalies in robot sys-

tems, facilitating early fault detection. Machine learn-

ing models, especially neural networks, can handle

binary variables more efﬁciently, reducing computa-

tional complexity, and improving model performance.

The use of binary variables is common in classiﬁca-

tion models, as it allows to predict the presence or

absence of a failure explicitly.

3.2.4 Feature Selection and Dimensionality

Reduction

Besides variable creation and transformation, it is

also crucial to select the most relevant features for

model training. Some variables may be redundant

or not very informative for failure prediction and can

therefore be eliminated to avoid making the model

unnecessarily complex. Techniques such as feature

importance-based, feature selection, or dimensional-

ity reduction with Principal Component Analysis can

be applied to optimize the model and reduce the risk

of overﬁtting. Feature engineering is a fundamental

step for the success of the study. The creation of new

variables, such as diff position, and the transforma-

tion of variables into binaries, based on interquartile

range and outliers, not only allow better capture ab-

normal behaviors, but also to increase the predictive

capacity of the model. These transformations help to

build a robust model capable of detecting failures in

varied conditions, based on the experimental data. By

strategically applying these techniques, we were able

to prepare the data for training the machine learning

model and improve the performance of the robot’s

predictive monitoring system.

3.3 Preparing Data for Training

Data preparation is a critical step in the process

of building a accurate AI models. Ensure that the

datasets are in the right format for training the model,

leading to more accurate and robust results. The num-

ber of input variables used for training the models cor-

responds to the number of features in X, which is 92

features. This is calculated as follows:

• 60 features from the 7D data.

• 13 features from non-7D data (force components

Fx, Fy, Fz, Tx, Ty, Tz; Cartesian tool coordinates

x, y, z, rx, ry, rz; and robot time).

• 19 features from feature engineering (6 joints with

3 features each, plus 1 feature for the difference in

robot time).

Thus, the total number of features in X is 92. Data

preparation includes several data preprocessing tech-

niques, such as normalization. In this study, nor-

malization was done using the Scikit-learn Standard-

Scaler method, which standardizes variables by trans-

forming them so that they have a mean of zero and a

standard deviation of 1. The formula used for normal-

ization is:

normalized

(x − µ)

Or :

• x is the original value of the variable.

• µ is the mean of the variable.

• σ is the standard deviation of the variable.

Thus, each variable is rescaled to have a normal dis-

tribution with a mean of 0 and a standard deviation of

1. Normalization is applied to all continuous numeric

variables to ensure consistency and reduce biases aris-

ing from differences in scale during model training.

Normalized variables include robot positions (x, y, z),

joint speeds, accelerations, and currents for each joint,

as well as temperature and torque. These quantitative

measures exhibit signiﬁcant variability, and normal-

ization ensures that all features contribute equally to

Robots Performance Monitoring in Autonomous Manufacturing Operations Using Machine Learning and Big Data

the learning process. These steps ensure that the data

is in an optimal format for training the model, facil-

itating more accurate prediction of robot failures and

degradations. The next step is to train the model us-

ing the prepared data to detect anomalies and improve

autonomous monitoring of the robotic system.

3.4 Model Architecture

The state of the art machine learning and deep learn-

ing algorithms compared in this study are k-Nearest

Neighbors, Gradient Boosting Machine, Random For-

est, Standard Neural Network, and Long Short-Term

Memory.

The models were trained using four distinct target

variables to address both the detection and classiﬁca-

tion of anomalies. The ﬁrst target, referred to as the

global target (CIBLE), was designed to detect the pres-

ence of anomalies without specifying their type. In

addition, three speciﬁc targets—y_pos, y_velocity,

and y_current—were used to identify anomalies re-

lated to position, velocity, and current data, respec-

tively. Each model was trained separately using the

global target to evaluate its performance in general

anomaly detection, as well as with the speciﬁc targets

to assess its ability to accurately identify the type of

anomaly.

3.5 Optimizing Data Management for

Sequential and Structured Data in

Robotics

HBase was chosen as the data storage tool due to

its suitability for our case study (see Table 5). Its

columnar storage model efﬁciently handles structured

sensor data (joint positions, velocities, currents, and

timestamps). Designed for sequential data process-

ing, HBase ensures optimal time-series management

while offering fast access and horizontal scalability

for large datasets. Its column-oriented architecture al-

lows quick retrieval of timestamped sensor readings,

and its seamless integration with Apache Spark en-

ables advanced data processing. These features make

HBase an ideal choice for storing and analyzing in-

dustrial robot data efﬁciently.

3.6 Selection of Data Processing

Frameworks for Autonomous

Experimentation

Choosing the right data processing framework is crit-

ical to ensuring the effectiveness of autonomous ex-

perimentation with the industrial robot. The table

below(Table 6) compares several major frameworks

based on key characteristics. Apache Storm is a

real-time streaming engine designed for low-latency

data processing, which can be useful for immediate

anomaly detection but does not support batch process-

ing or machine learning integration. Apache Spark

provides a hybrid approach that allows both batch

and streaming processing, making it particularly use-

ful for post-analysis of anomalies and training ma-

chine learning models, although it has higher latency

compared to Flink or Storm. Apache Flink is an

advanced real-time streaming framework that offers

stateful data processing with minimal latency, making

it an excellent choice for predicting anomalies before

they occur. Apache Kafka is a messaging system that

ensures fast and reliable data transmission from the

robot’s sensors to other processing frameworks, play-

ing a key role in maintaining a smooth data pipeline.

Dask is a framework optimized for parallel comput-

ing, making it suitable for handling large-scale experi-

mental data and optimizing machine learning models;

however, unlike Spark or Flink, it does not support

real-time streaming but is highly efﬁcient for batch

analytics. For this project, HBase was selected as

the primary storage solution due to its ability to han-

dle structured time-series data efﬁciently. In terms

of data processing, Apache Flink is the most suitable

framework because it enables real-time anomaly de-

tection with low latency and stateful stream process-

ing, which aligns with the need for continuous mon-

itoring of the UR5 robot’s operations. Additionally,

Apache Kafka will be used to facilitate data transmis-

sion between the robot’s sensors and the processing

framework, ensuring a reliable and scalable data ﬂow.

For batch analysis and model training, Apache Spark

will be employed as it offers efﬁcient large-scale data

processing and machine learning integration. This

combination of frameworks ensures a robust, scal-

able, and efﬁcient architecture for autonomous exper-

imentation and anomaly detection in the UR5 robot,

enabling real-time issue detection, performance op-

timization, and improved decision-making through

data-driven insights.

4 EXPERIMENTAL RESULTS

AND FINDINGS

All models are trained using the preprocessed

datasets, each corresponding to a different combina-

tion of temperature, speed, and payload. The met-

rics used for evaluation include training time for the

computational efﬁciency of each model, the precision

to measure the model’s ability to correctly identify

DATA 2025 - 14th International Conference on Data Science, Technology and Applications

Table 5: Comparison of Hadoop, Hive, MongoDB, HBase, and Apache Flink.

Feature Hadoop Hive MongoDB HBase Apache Flink

Data Model File System File System Document-Oriented Column-Based Stream-Based

Architecture Model Master-to-Slave Master-to-Slave Peer-to-Peer Master-to-Slave Distributed

Structured Data Yes Yes Yes Yes Yes

Sequential Data Handling Yes No No Yes Yes

Integration with Spark Yes Yes Yes Yes Yes

Fast Random Access Yes No No Yes No

Table 6: Comparison of Data Processing Frameworks for Autonomous Experimentation with the Industrial Robot.

Feature Apache

Storm

Apache

Spark

Apache Flink Apache

Kafka

Dask

Processing Type Streaming Hybrid (Batch

& Streaming)

Streaming Messaging Parallel Com-

puting

Real-time Processing Yes Partial Yes Yes No

Batch Processing No Yes No No Yes

Low Latency Yes No Yes Yes No

Fault Tolerance Yes Yes Yes Yes Yes

High Scalability Yes Yes Yes Yes Yes

ML Integration No Yes Yes No Yes

anomalies (Class 1) and normal states (Class 0) and

RMSE, MAE to assess the model’s prediction accu-

racy and error magnitude. The experimental results

demonstrate the performance of machine learning and

deep learning models in anomaly detection for the

UR5 robot. For the LSTM model, the architecture

consists of one hidden LSTM layer with 128 units,

followed by ReLU activation and a 20% Dropout reg-

ularization to prevent overﬁtting. The output layer

is a single neuron with a sigmoid activation function

for binary classiﬁcation. For the KNN model, the

number of neighbors is set to n neighbors=5, mean-

ing each prediction is based on the ﬁve closest neigh-

bors using the default Euclidean distance metric. For

the GBM model, the evaluation metric used is log

loss (eval metric=’logloss’), which measures the log-

arithmic loss for classiﬁcation. The random state=42

is set to ensure result reproducibility. For the RF

model, the number of trees in the forest is set to

n estimators=100, and random state=42 is used for

reproducibility. For the Neural Network (NN) model,

the architecture includes:

• Hidden layer 1: Dense layer with 128 neurons and

ReLU activation.

• Dropout layer (20%) to reduce overﬁtting.

• Hidden layer 2: Dense layer with 64 neurons and

ReLU activation.

• Dropout layer (20%).

• Output layer: Single neuron with sigmoid activa-

tion for binary classiﬁcation.

• Loss function: Binary cross-entropy, suitable for

binary classiﬁcation tasks.

• Optimizer: Adam, which dynamically adjusts the

learning rate.

• Evaluation metric: Accuracy.

The loss and accuracy curves show a stable con-

vergence of the LSTM model trained on the ﬁrst com-

bination (Temperature = normal, payload = 1.6, Speed

= half speed), with y pos as the target. The steady de-

crease in loss and increase in accuracy over epochs

demonstrate the effectiveness of the training process.

Figure 5: LSTM Accuracy in Fault Detection for First Com-

bination.

The ﬁgures 5 & 6 illustrate the accuracy and loss

curves, conﬁrming the model’s ability to progres-

sively improve its performance. These results indicate

that the LSTM effectively learns from the data while

avoiding overﬁtting because of the applied regulariza-

Robots Performance Monitoring in Autonomous Manufacturing Operations Using Machine Learning and Big Data

tion techniques.

The confusion matrix in ﬁgure 7 highlights the

performance of the model in correctly and incor-

rectly classiﬁed instances. This matrix demonstrates

the model’s capability to distinguish between anoma-

lous and non-anomalous cases based on y pos. Sim-

ilarly, the same analyses—accuracy and loss curves,

as well as confusion matrices—were conducted for

the other targets (y velocity and y current) to assess

the model’s ability to classify speciﬁc anomaly types.

These ﬁndings further validate the robustness and ver-

satility of the neural network in handling global and

type-speciﬁc anomaly detection tasks.

To analyze the performance of different models on

the ﬁrst combination of payload, speed, and temper-

ature, we compare their errors across the three tar-

get variables (Figure 8 & 9). y current demonstrates

the most stability with consistently lower MAE and

RMSE values across all models, whereas y velocity

appears to be the most sensitive target, showing

higher variations in error metrics, particularly in mod-

els like k-Nearest Neighbors and LSTM, which strug-

gle to maintain accuracy for this variable.

The combinations are based on variations in speed

(half/full), temperature (normal/cold start), and pay-

load (1.6/4.5). In all combinations except the last

one (cold start, full speed, payload 4.5), y velocity

is the most sensitive variable with the highest RMSE.

However, in the last combination, y current becomes

the most sensitive, indicating a signiﬁcant increase in

RMSE (Figure 10).

Figure 6: LSTM Loss in Fault Detection for First Combi-

nation.

The table 7 presents the performance of ﬁve ma-

chine learning models: gradient boosting machine,

k-nearest neighbors, random forest, neural network,

and LSTM—evaluated across four different targets

(y_pos, y_velocity, y_current and y). For each

model, the test accuracy, RMSE, MAE, and training

time are reported. The results highlight each model’s

ability to detect anomalies and classify their types

(position, velocity, current), as well as the computa-

tional efﬁciency of the training process.

Figure 7: LSTM confusion matrix in Fault Detection for

First Combination.

Figure 8: MAE Comparison Across Models.

Figure 9: RMSE Comparison Across Models.

DATA 2025 - 14th International Conference on Data Science, Technology and Applications

Figure 10: RMSE values for the LSTM model across dif-

ferent experimental conditions.

5 DISCUSSION

The experimental analysis demonstrates the effective-

ness of machine learning and deep learning tech-

niques for the real time monitoring of industrial

robots in autonomous operations, under the ﬁrst com-

bination of conditions: temperature (normal), speed

(half speed), and payload (1.6 kg). Five models, Ran-

dom Forest, Gradient Boosting Machine, k-Nearest

Neighbors, Deep Neural Networks, and Long Short-

Term Memory—were evaluated based on test accu-

racy, MAE, RMSE, and training time for real time

integration purposes. Among all tested models, Ran-

dom Forest emerged as the best performing model.

It achieved the highest test accuracy across all target

variables (y pos, y velocity, y current, and y), consis-

tently exceeding 99% accuracy for most cases. Addi-

tionally, it maintained relatively short training times,

ranging from 12 to 14 seconds, making it a highly

efﬁcient choice for real-time and industrial applica-

tions. The ensemble learning mechanism of Random

Forest, which aggregates multiple decision trees, en-

ables robust and reliable predictions while minimiz-

ing overﬁtting. This stability across different targets

makes it an ideal candidate for detecting anomalies

in the UR5 robot. Other models also exhibited com-

petitive performance. Gradient Boosting Machine

demonstrated excellent predictive capability, achiev-

ing test accuracies above 99%, with signiﬁcantly low

MAE and RMSE values. However, despite its strong

accuracy, GBM required slightly more training time

( 1.2–1.4 seconds per target) compared to Random

Forest. While GBM is a highly efﬁcient model, it

does not generalize as well across different targets as

Random Forest does. Deep Neural Networks showed

strong predictive capability, particularly for y pos,

y velocity, and y current, with accuracies exceeding

98%. However, its performance on the combined tar-

get (y) dropped to 92.75%, indicating potential limi-

tations in handling multi-dimensional fault detection

scenarios. Furthermore, DNN models required signif-

icantly longer training times (56–123 seconds per tar-

get), making them less practical for real-time deploy-

ment despite their high accuracy. Similarly, LSTM

performed well for sequential data, achieving over

97% accuracy for y pos, y velocity, and y current.

However, its accuracy dropped to 93.14% for y, and

it had the highest training time (225–306 seconds per

target), making it computationally expensive. While

LSTM is well-suited for capturing long-term depen-

dencies in time-series data, its efﬁciency trade-off

makes it less viable for real-time anomaly detection

in industrial settings. In contrast, KNN exhibited

the lowest performance among the models tested. Its

test accuracies ranged from 86.60% to 95.29%, with

signiﬁcantly higher MAE and RMSE values com-

pared to other models. However, it had the short-

est training times ( 0.01 to 0.02 seconds), making it

a viable option when computational speed is priori-

tized over predictive accuracy. Although KNN is not

the best choice for precision-driven applications, its

simplicity and speed could be useful in cases where

quick approximations are needed. Considering the

balance between accuracy, robustness, and compu-

tational efﬁciency, Random Forest is selected as the

best model for detecting and classifying anomalies in

the UR5 robot under these conditions. While GBM

offers competitive accuracy, Random Forest provides

greater stability across all targets and faster inference

times, making it ideal for industrial real-time appli-

cations. These ﬁndings highlight the importance of

leveraging ensemble learning methods for predictive

maintenance and fault detection in autonomous ad-

ditive manufacturing operations. By incorporating

sensor-based features (joint positions, velocities, cur-

rents, torques) and engineered variables (differences

between target and actual values), the models ef-

fectively distinguish between normal and anomalous

states, offering valuable insights into robotic system

performancemetrics.

5.1 Limitations and Future Work

Although the results are promising, several limita-

tions warrant discussion. First, current models were

trained and evaluated on a data set that, while repre-

sentative, may not capture the full range of variabil-

ity encountered in real-world scenarios. Factors such

as sensor noise, environmental conditions, and task-

speciﬁc variations could affect model performance.

The reliance on pre-engineered features, such as bi-

nary transformations based on the interquartile range,

while effective for anomaly detection, may lead to a

loss of nuanced information. Future work could ex-

Robots Performance Monitoring in Autonomous Manufacturing Operations Using Machine Learning and Big Data

Table 7: Comparison of the Performance of Different Models and Targets for the First Combination: Temperature (normal),

Speed (half speed), Payload (1.6).

Model Target (Y) Test Accuracy (%) MAE RMSE Training Time (s)

Neural Network

y pos 98.10 0.0382 0.1211 95.31

y velocity 98.10 0.0385 0.1201 121.25

y current 97.91 0.0351 0.1238 56.42

y 92.75 0.1105 0.2294 123.96

Random Forest

y pos 99.86 0.0014 0.0368 12.82

y velocity 99.92 0.0008 0.0279 12.34

y current 99.88 0.0012 0.0341 13.13

y 97.75 0.0225 0.1500 14.21

Gradient Boosting Machine

y pos 99.55 0.0045 0.0668 1.26

y velocity 99.81 0.0019 0.0440 1.05

y current 99.67 0.0033 0.0574 1.36

y 99.44 0.0056 0.0750 1.37

k-Nearest Neighbors

y pos 89.82 0.1018 0.3191 0.017

y velocity 86.60 0.1340 0.3661 0.015

y current 95.29 0.0471 0.2171 0.014

y 87.03 0.1297 0.3602 0.013

LSTM

y pos 97.50 0.0452 0.1374 232.58

y velocity 96.76 0.0531 0.1523 306.26

y current 97.38 0.0406 0.1358 225.78

y 93.14 0.1055 0.1284 265.25

plore the use of unsupervised learning techniques or

feature extraction methods to enhance data represen-

tation without predeﬁned thresholds. The models re-

quire further validation in live industrial environments

to assess their scalability, robustness, and adaptabil-

ity. Transfer learning techniques could be explored

to adapt DNNs to new robotic systems without re-

training from scratch. Similarly, hyperparameter op-

timization and the integration of domain knowledge

into machine learning models could further improve

their performance. Although recent trends highlight

the effectiveness of transformers for time series anal-

ysis, this study does not incorporate them. However,

future work could investigate their potential to en-

hance temporal dependency modeling and improve

anomaly detection in robotic systems. Additionally,

the methodology could be extended to other types of

robots, requiring adjustments based on their mechan-

ical characteristics and sensor conﬁgurations.

6 CONCLUSION

In this study, we explore a data-driven approach for

fault detection and anomaly classiﬁcation in indus-

trial robots, speciﬁcally the UR5 welding robot. Us-

ing sensor data and machine learning techniques,

we developed a robust system capable of identifying

anomalies under different operating conditions, con-

sidering variations in temperature, speed, and pay-

load. Our approach combines feature engineering,

statistical thresholding, and supervised learning, mak-

ing it an effective strategy for the monitoring of

robotic systems in autonomous additive operations.

The results demonstrate that Random Forest out-

performed other models, achieving the highest ac-

curacy while maintaining a relatively low computa-

tional cost. GBM also delivered strong performance,

particularly in terms of efﬁciency, making it a suit-

able alternative when fast predictions are required.

Deep learning models, including Neural Networks

and LSTM, showed promising results, particularly in

capturing complex temporal dependencies, but their

longer training times limit their practicality for real-

time fault detection. The KNN model, while compu-

tationally efﬁcient, exhibited lower accuracy, making

it less suitable for high-precision applications. Be-

yond model performance, this research aligns with the

principles of autonomous experimentation, where the

iterative cycle of data collection, model training, and

performance evaluation leads to continuous improve-

ment of the system. The integration of machine learn-

ing with robotic fault detection enables a data-driven

self-optimization framework, reducing the need for

DATA 2025 - 14th International Conference on Data Science, Technology and Applications

manual intervention to diagnose faults. This shift to-

wards automation enhances industrial efﬁciency by

enabling real-time anomaly detection and proactive

maintenance strategies, ultimately improving robot

reliability and operational lifespan. Future work will

focus on further reﬁning model performance, integrat-

ing real-time deployment strategies, and exploring re-

inforcement learning for adaptive fault detection. Ad-

ditionally, expanding the dataset with diverse oper-

ational scenarios and environmental factors will en-

hance the robustness of the models, ensuring their ap-

plicability across a wider range of industrial settings.

This study highlights the potential of AI-driven pre-

dictive maintenance, paving the way for smarter and

more autonomous robotic systems in manufacturing

and beyond.

ACKNOWLEDGEMENTS

This material is based upon work supported by the

U.S. Department of Defense under the Ofﬁce of

Local Defense Community Cooperation (OLDCC)

Award Number MCS2106-23-01. The views ex-

pressed herein do not necessarily represent the views

of the U.S. Department of Defense or the United

States Government. We would like to acknowledge

the National Institute of Standards and Technology

(NIST) for making their raw data available.

REFERENCES

Alghamdy, M., Almutairi, F. M., and Ahmad, R. (2023).

Job-scheduling model for an autonomous additive

manufacturing: a case of 3d food printing. Unknown

Journal, 217:1234–1242.

Antonelli, D. and Astanin, S. (2016). Qualiﬁcation of a col-

laborative human-robot welding cell. Procedia CIRP,

41:352–357.

Borboni, A., Reddy, K. V., Elamvazuthi, I., AL-Quraishi,

M. S., Natarajan, E., and Ali, S. S. A. (2023). The ex-

panding role of artiﬁcial intelligence in collaborative

robots for industrial applications: A systematic review

of recent works. Robotics and Computer-Integrated

Manufacturing, 11(1):111.

Butt, J. (2020). Exploring the interrelationship between ad-

ditive manufacturing and industry 4.0. Unknown Jour-

nal, 4(2):13.

Canﬁeld, S. L., Owens, J. S., and Zuccaro, S. G. (2021).

Zero moment control for lead-through teach program-

ming and process monitoring of a collaborative weld-

ing robot. Unknown Journal, 13(3):031016.

Djuric, A. M., Urbanic, R. J., and Rickli, J. L. (2016). A

framework for collaborative robot (cobot) integration

in advanced manufacturing systems. Unknown Jour-

nal, 9(2):457–464.

orﬂer, K. and Gerke, M. (2022). Additive manufactur-

ing using mobile robots: Opportunities and challenges

for building construction. Cement and Concrete Re-

search, 158:106772.

Fan, H., Liu, C., Bian, S., Ma, C., Huang, J., Liu, X., Doyle,

M., Lu, T., Chow, E., Chen, L., Fuh, J. Y. H., Lu,

W. F., and Li, B. (2025). New era towards autonomous

additive manufacturing: a review of recent trends and

future perspectives. International Journal of Extreme

Manufacturing, 7(3):032006.

Felbrich, B., Schork, T., and Menges, A. (2022). Au-

tonomous robotic additive manufacturing through dis-

tributed model-free deep reinforcement learning in

computational design environments. Construction

Robotics, 6(1):15–37.

Ferraguti, F., Villani, V., and Storchi, C. (2023). My-

welder: A collaborative system for intuitive robot-

assisted welding. Unknown Journal, 89:102920.

Ko, H., Witherell, P., Ndiaye, N. Y., and Lu, Y. (2019). Ma-

chine learning based continuous knowledge engineer-

ing for additive manufacturing. In 2019 IEEE 15th

International Conference on Automation Science and

Engineering (CASE). IEEE.

aroiu, G., Andronie, M., Iatagan, M., Geam

anu, M.,

tef

anescu, R., and Dijm

arescu, I. (2022). Deep

learning-assisted smart process planning, robotic

wireless sensor networks, and geospatial big data

management algorithms in the internet of manufactur-

ing things. Unknown Journal, 11(5):277.

Malik, A. A. and Brem, A. (2021). Digital twins for collab-

orative robots: A case study in human-robot interac-

tion. Unknown Journal, 68:102092.

NIST, U. (2017). Degradation measurement of robot arm

position accuracy. Accessed: 2025-02-06.

Qi, X., Chen, G., Li, Y., Cheng, X., and Li, C. (2019).

Applying neural-network-based machine learning to

additive manufacturing: Current applications, chal-

lenges, and future perspectives. Unknown Journal,

5(4):721–729.

Qiao, G., Schlenoff, C., and Weiss, B. A. (2017a). Accu-

racy degradation analysis for industrial robot systems.

Accessed: 2025-02-07.

Qiao, G., Schlenoff, C., and Weiss, B. A. (2017b). Quick

positional health assessment for industrial robot prog-

nostics and health management (phm). In 2017 IEEE

International Conference on Robotics and Automation

(ICRA), pages 1815–1820. IEEE.

Raza, A., Deen, K. M., Jaafreh, R., Hamad, K., Haider,

A., and Haider, W. (2022). Incorporation of machine

learning in additive manufacturing: a review. Interna-

tional Journal of Advanced Manufacturing Technol-

ogy, 122(3):1143–1166.

Rozo, L., Calinon, S., Caldwell, D. G., Jimenez, P., and

Torras, C. (2016). Learning physical collaborative

robot behaviors from human demonstrations. Un-

known Journal, 32(3):513–527.

Scime, L. and Beuth, J. (2018). A multi-scale convolu-

tional neural network for autonomous anomaly detec-

Robots Performance Monitoring in Autonomous Manufacturing Operations Using Machine Learning and Big Data

tion and classiﬁcation in a laser powder bed fusion

additive manufacturing process. Additive Manufac-

turing, 24:273–286.

Sharma, N. and Cupek, R. (2023). Real-time control and

optimization of internal logistics systems with collab-

orative robots. Unknown Journal, 225:248–258.

Van, M. and Ceglarek, D. (2021). Robust fault tolerant con-

trol of robot manipulators with global ﬁxed-time con-

vergence. Unknown Journal, 358:699–722.

Wescoat, E., Krugh, M., and Mears, L. (2021). Random for-

est regression for predicting an anomalous condition

on a ur10 cobot end-effector from purposeful failure

data. Procedia Manufacturing, 53:644–655.

Zaatari, S. E., Marei, M., Li, W., and Usman, Z. (2019).

Cobot programming for collaborative industrial tasks:

An overview. Unknown Journal, 116:162–180.

Zhang, K., Chermprayong, P., Xiao, F., Tzoumanikas, D.,

Dams, B., Kay, S., Kocer, B. B., Burns, A., Orr,

L., Alhinai, T., Choi, C., Darekar, D. D., Li, W.,

Hirschmann, S., Soana, V., Ngah, S. A., Grillot, C.,

Sareh, S., Choubey, A., Margheri, L., Pawar, V. M.,

Ball, R. J., Williams, C., Shepherd, P., Leutenegger,

S., Stuart-Smith, R., and Kovac, M. (2022). Aerial

additive manufacturing with multiple autonomous

robots. Nature, 609(7928):709–717.

Zhang, Z., George, A., Alam, M. F., Eubel, C., Vallabh,

C. P., Shtein, M., Barton, K., and Hoelzle, D. J.

(2024). An additive manufacturing testbed to evalu-

ate machine learning-based autonomous manufactur-

ing. Unknown Journal, 146(3):031008.

DATA 2025 - 14th International Conference on Data Science, Technology and Applications