An Incremental Learning Approach to Detect Muscular Fatigue in Human– Robot Collaboration

Human–robot collaboration aims to join the distinctive strengths of humans and robots to compensate for the weaknesses associated with each party and, thus, to enable synergetic effects. Robots are characteristically considered fatigue-proof. Hence, they are utilized to assist human operators during heavy pushing and pulling activities. To detect physical fatigue or high payloads held by a human operator, wearable sensors, such as electromyographys (EMGs), are deployed. The EMG data are typically processed via machine learning, which includes training models offline before an application in an online system. However, these approaches often demonstrate varying performances between offline and online applications due to subject-specific characteristics within the data. An opportunity to tackle this challenge can be found in incremental learning, as these models purely learn online and constantly fine-tune the model's structure. In this article, a Mondrian Forest is applied to predict payloads and physical fatigue of human operators during an assistance scenario with a collaborative robot. An experiment was conducted with a total of 12 participants, where the payload was increased until participants initiated an assistance request from a Universal Robots model 10 cobot. This allowed for testing whether the Mondrian Forest can accurately predict the payload and fatigue levels from the acquired EMG signals. Overall, the approach demonstrates a promising potential toward higher awareness when an operator might require assistance from a robot and ultimately toward a more effective human–robot collaboration.


I. INTRODUCTION
H UMAN-ROBOT collaboration is considered a key paradigm for combining the best of both worlds: a robot's endurance, speed, and precision, with human dexterity, perception, and adaptability [1]. At the same time, this is also envisioned to compensate for weaknesses associated with each party [2]. Human weaknesses typically include being susceptible to fatigue and stress, which can apply both mentally and physically [3], [4]. Consequently, human-robot collaboration aims to assign physically demanding tasks to the robot or assist the humans in lifting heavy payloads [5]. In such physical collaborations, robots support human operators via force amplification to perform heavy pushing or pulling activities [6], [7]. Thus, it not only allows for creating a more ergonomic environment but also mitigates potential physical disadvantages caused by age, sex, or disability.
One of the challenges, however, is to establish a collaborative robot's awareness of its human partner's need for assistance [4]. For this purpose, wearable devices have faced increased popularity to enable sensing of the human operator's state [8], [9]. In the case of physical interactions, electromyographys (EMGs) are deployed, which allow for measuring human muscle activity [10], [11]. Since interpretation and integration of these signals are often challenging, machine learning is utilized to classify patterns within the data. Approaches are typically based on supervised offline training before the application of a trained model in an online system. These models often demonstrate considerably lower classification performance in an online system than during the training sessions prior [12]. This effect can be observed during different recording sessions, even for the same individual [13]. The issue of varying performance and high manual training and programming efforts could be tackled through advances in machine learning regarding incremental learning. Incremental learning algorithms purely learn online, which enables constant automatic fine-tuning and adaption of the models [14], [15]. Therefore, they could allow us to minimize the training and programming effort while delivering more persistent results.
In this work, an incremental learning approach based on a Mondrian Forest is utilized to predict payload and muscle fatigue from EMG data during a human-robot collaborative task. Overall, the approach offers a promising potential for machine learning models to learn "on the fly" and adapt to the uniqueness of human operators in terms of strength and endurance.

II. RELATED WORK
Human-robot collaboration is widely considered the highest form of human-robot interactions since it includes joint tasks, shared workspaces, a common aim, and permitted physical contact between the two parties [16]. To establish effective communication for physical interactions, two options are plausible: direct control [17] or deploying wearable sensors, such as EMGs. EMGs are utilized to detect muscle contraction forces of the human operator [13]. This is due to the contraction of muscles generating electrical activity, which can be measured on the surface [18]. An increase in mean amplitude and a decrease in frequency are indicators of intense muscle contraction [19]. The EMG signals are typically acquired from the upper limbs since they are mainly involved in the physical interactions with a robot [10]. Subsequently, the EMG data can provide insights on human intentions, such as applying forces in a certain direction [20]. It can also be utilized to detect human muscle fatigue due to high payloads or endurance stress [19], [21]. In both cases, a collaborative robot could assist its human partner in creating more ergonomic working conditions [10], [13]. This could prevent strain injuries, as well as long-term health issues related to muscular fatigue. Long-term damages are often referred to as musculoskeletal disorders in this context, which might not appear immediately [19]. Thus, overall, there is a strong motivation to detect and integrate these signals into the human-robot collaborative loop.
Measurement and integration of EMG signals into a humanrobot collaborative system, however, is often challenging. This is due to subject-specific characteristics within the EMG data. EMG data streams demonstrate unique features for each individual [13], [22]. A change in these features can be observed even during different recording sessions for the same person [13]. Consequently, manual fine-tuning and programming efforts are often required to adapt to these features. Hence, machine learning is utilized to identify relevant patterns and to perform continuous classification. For this purpose, different strategies can be applied.
The most commonly used processing and interpretation technique is based on Fourier transform in conjunction with a machine learning classifier, referred to as "Standard" in Fig. 1. A wave transform, such as Fourier transform, is required since most human physiological data follow sinusoidal patterns [18]. Moreover, raw EMG signals often contain high levels of noise. Thus, filters are applied, such as Butterworth filters, with a cutoff frequency of 2-20 Hz [13]. Afterward, relevant features (frequency and amplitude) can be extracted to train a machine learning classifier. This feature extraction is described to have a larger impact on the classification performance than the selection of the classifier itself [12], [13]. The classifiers are trained offline until they reach satisfactory performance levels before they are applied online. According to Hakonen et al. [12], support vector machines and linear discriminant analyses are commonly used to classify EMG signals. One of the advantages Fig. 1. Data processing and machine learning methodologies to process EMG signals (adapted) [25]. of this methodology is its high transparency. The results of each intermediate step, such as feature extraction, can be visualized. The disadvantages, however, are twofold: First, many systems obtain high classification accuracies during offline training. Yet, the online performances of such systems are often significantly lower. Second, programming and training the models require high manual efforts [12], [13]. Thus, other approaches have been established to minimize manual fine-tuning.
In recent years, artificial neural networks (ANNs), including various variations, have gained a high research interest. In the case of human sensor data, the main idea is to stream raw data into the classifier. Due to their advanced nature, the ANNs are expected to identify relevant patterns by themselves [20], [23]. In these approaches, time-series-based ANNs, such as recurrent neural networks (RNNs) and long short-term memory RNNs, delivered promising results. The advantages of this methodology were the lower variance in prediction accuracy between offline and online systems. Moreover, manual fine-tuning efforts are substantially lower than during the "Standard" methodology. Disadvantages, however, are the large quantities of training data required to train a model. This can also result in several hours of training time until a classifier can be used online.
An opportunity to cope with the challenges of excessive training times and varying accuracies between offline and online systems can be found in incremental learning. Incremental learning algorithms offer the ability to purely learn online or "on the fly" [15]. They typically offer the following characteristics [14], [15]: 1) ability of life-long learning; 2) ability to incrementally fine-tune the model's performance; and 3) no prior knowledge about the data or their properties are required. Thus, this would allow for the use of an incremental learner to continuously adapt to a human operator while improving its performance over time. However, there are two challenges, namely the plasticity-stability dilemma, which entails that the model must continuously learn new knowledge without forgetting previously obtained knowledge [15], and second, convergence time, Fig. 2. Overview of the approach. Signal acquisition via EMG sensors on different muscles, filtering noise, such as powerline interference at 50-60 Hz, performing an STFT to retrieve underlying frequencies and intensities, and performing the online/incremental learning via the Mondrian forest. The predicted payload and fatigue levels are then communicated to the collaborative robot, which in turn can assist the operator. which includes the time to perform a learning operation and classification [24]. Generally, the more complex a dataset and the subsequent model, the longer the convergence time. These factors require investigation during an application to predict muscular payloads and fatigue from EMG data in human-robot collaboration.

III. APPROACH
In order to establish an effective human-robot collaboration where both parties can compensate each other's weaknesses, a high mutual awareness is required. While there are comanipulation and load-sharing approaches in place, they do not necessarily qualify as human-robot collaboration since the robot is lacking the cognitive skills to determine when an operator requires assistance. Hence, in this article, wearable sensors are deployed to detect muscle activity, where the acquired signals are interpreted via incremental learning, which allows for learning and classifying "on the fly." Thus, this is envisioned to predict when an operator reaches high levels of payload leading to muscular fatigue and, consequently, requires assistance from the collaborative robot.
In a first step (see Fig. 2), wearable EMGs are deployed on an operator's upper limb muscles that are primarily involved in lifting or holding activities (biceps and forearm). EMGs are chosen as they are the most popular sensor for detecting muscle activity [13]. However, simply feeding the acquired signals into the incremental learner does not yield sufficient results. This is mainly due to the nature of the EMG signal, which follows sinusoidal wave patterns and also contains noise (such as powerline interference at 50-60 Hz in Europe). Therefore, a feature extraction step is added, which entails a low-bandpass filter to remove noise, followed by a short-time Fourier transform (STFT). The resulting spectrum analysis (frequency, amplitude, and phase of the spectrum) is then used in the incremental learner. The main advantages of the incremental learner in this context are twofold: First, it allows for continuous adaptation to the uniqueness of a person (e.g., individual muscle activity signals and training/fitness level); Second, it allows for distinguishing different levels of payload from EMG signals over time. For instance, if a nonincremental classifier, such as a support vector machine or a random forest, is trained offline with three different payloads (low, medium, and high), it can only distinguish these three classes in an online application. In contrast, an incremental learner can learn and distinguish new classes during an online application. In the example, this could include "very low," "low," "medium," "high," and "very high" payloads. Consequently, incremental learners offer the opportunity of lowered fine-tuning and programming efforts, alongside potentially higher prediction accuracies in an online application.
In the following, this methodology, also shown in Fig. 2, will be applied and further explained in Section IIIA-D.

A. Data Acquisition
Based on the application, EMG sensors can be placed on various muscles, which are required to perform an activity. For example, the human upper limbs, which are predominantly involved in performing tasks and collaborating with robots, consist of a wide variety of muscles. Thus, EMG signals could be obtained regarding movements and forces in the shoulder, arms, and fingers [13]. In the current context, signals will be acquired from the biceps brachii and brachioradialis (forearm) since these muscles are mainly involved in lifting or steady holding of a workpiece. Overall, correct placement and a careful selection (a) Decision tree and its key elements: split nodes (blue) containing the decision logic and leaf nodes (dark blue), which are the classes/labels that the model is predicting. The higher the depth (horizontally and vertically) of the tree, the higher the complexity. (b) Forest structure, in which several decision trees make a prediction that is then aggregated to a final prediction of the model. (c)-(e) Incremental learning operations that are continually performed "on-the-fly"/online. Note that the tree structure is not a representative example, and its simplified structure is intended to demonstrate the learning and classification principle. of the EMG device (channels and sampling rate) are essential since both factors have a significant impact on the signal quality. Even advanced classifiers, such as ANNs, achieve low prediction accuracies if the EMG signal quality is poor [25].

B. Low-Bandpass Filter
Although the general idea of incremental learning is to minimize manual programming and filtering efforts, this might not be fully applicable in the current context. As described in Section II, EMG signals are susceptible to contain noise, which could interfere with prediction accuracies. Noise sources include ambient noise, such as electromagnetic radiation from power sources, as well as the inherent noise in electronic equipment itself [13]. Therefore, an additional preprocessing step was added, which includes a low-pass filter with a cutoff frequency of 30 Hz. This is aligned with the literature, where most approaches consider the frequency band of 2-30 Hz to retrieve relevant muscle activity features while removing the majority of dominant external noise sources [13]. The filtered signal is used in an STFT.

C. Short-Time Fourier Transform
Similar to a guitar, where a cord consists of different notes, EMG sensors provide muscle activity data in a complex signal within a time domain. The signal consists of different frequencies and associated intensities. Consequently, it is necessary to perform a wave transform to retrieve relevant features (underlying frequencies and their intensity), which can be mapped to payloads and muscular fatigue. In this approach, an STFT is chosen, as it allows for a fast processing of the signal [26]. The STFT is performed on 0.75 s interval windows and an overlap of 0.05 s. The overlap of windows prevents the loss of features if they occur at the beginning or the end of a window. The resulting spectrum analysis is then used in the incremental learner. Rather than selecting relevant features (frequencies) manually, all resulting features of the STFT are used. Thus, the machine learning algorithm is able to identify correlating frequencies with payloads and muscular fatigue by itself.

D. Learning and Classification
Incremental and online learning offer a wide variety of classifiers [14]. One of the most prominent is an online random forest (ORF) due to its high accuracy, scalability, and robustness [24]. In this work, a subcategory of ORFs is chosen in Mondrian Forest, which is described to obtain comparable accuracies with its state-of-the-art batch/offline counterparts on the same datasets [24]. Mondrian Forests consist of several decision trees, which are generated on random subsamples of the data. The structure of the decision tree, shown in Fig. 3(a), is generated automatically. Each tree consists of split nodes, which contain the decision logic. For instance, if the signal intensity of a certain frequency is above or below a threshold, the split node determines which path the signal takes down the tree. The labels or classes are at the end of a "branch" and referred to as leaf nodes. In the current context, these leaf nodes contain payload and associated fatigue levels of operators ranging from "very low" to "very high." Both the split nodes and the leaf nodes would be unique for each operator depending on the signals and the variety of payload handled. The more complex the classification logic due to a combination of split nodes, the higher the tree complexity would be. Moreover, the model complexity also depends on the number of trees in a model. The number trees, however, can be manually selected as a hyperparameter. Typically, the performance increases with a higher number of trees until saturation is reached. The individual predictions of the trees are merged based on voting or averaging, as shown in Fig. 3(b). Afterward, a final prediction will be made.
To enable incremental learning, the model learns and updates itself purely online (on-the-fly). Thus, three types of learning operations can be performed: introducing a new split node [see Fig. 2(c)], updating the split condition in an existing node ([see Fig. 2(d)], and finally adding a new class via splitting a leaf node into two [see Fig. 2(e)]. In the current context, introducing a new split node can be based on a new feature (frequency and frequency intensity of one EMG sensor and the associated muscle). A split condition can be updated in an existing node, for example, a different threshold for frequency intensity is set. Finally, a new leaf node can be established when a new type of payload or fatigue level is identified, for example, from "medium payload" to "low-medium" and "higher medium" payloads. These learning and prediction operations are continuously performed, which enables constant fine-tuning of the model.
There are, however, two challenges when applying Mondrian Forests, namely maximizing the prediction performance of the model and minimizing the processing/convergence time. The convergence time depends on the time the model takes to make a prediction and perform a learning operation. Both the prediction performance and the convergence time can be controlled via the hyperparameter: number of trees in the model. A higher number of trees typically produces a higher prediction performance until saturation is reached. After this threshold, any additional tree in the model does not contribute to an increased performance anymore. At the same time, each additional tree in the model adds to the complexity, which in turn increases convergence time. This is mainly because each tree needs to be updated and maintained during the learning operation.
Thus, on one hand, the convergence time should be kept as low as possible. On the other hand, the prediction performance should be maximized. To obtain this optimum (number of trees) and to validate whether a Mondrian Forest can accurately predict different payloads and muscle fatigue from EMG data, an experiment was performed, which is introduced in the following.

IV. EXPERIMENT
The experimental design aims to generate EMG data samples while participants lift and hold different payloads ranging from low to subjectively high. The weights are increased until participants manually trigger an assistance request, as soon as they feel fatigued. The EMG data stream is collected and labeled based on the payload. Thus, this allows for applying the Mondrian Forest to validate whether it can correctly predict the varying levels of payload and subsequent muscular fatigue from EMG data. Moreover, instead of participants pressing the assistance request manually, the Mondrian Forest is envisioned to identify the fatigue threshold and trigger a robotic assistance request automatically afterward.
The experiment was conducted with 12 participants aged between 20 and 39 years old. The participants had various levels of prerequisites regarding exercise and strength. While some participants performed strength-/conditioning-related exercises three times per week and more, other participants stated they did not perform any type of exercise at all. Thus, varying levels of performance were expected during the experiment. Consequently, it can be validated whether the Mondrian Forest can adapt to the uniqueness of participants in terms of EMG signal (variety in strength and muscle size/cross section), as well as the different number of classes to predict (varying levels of payload per participant).
In the first step, EMG data need to be acquired. For this purpose, four Myoware muscle activity sensors were placed on the participant's biceps and forearms, as shown in Fig. 4(a). These sensors provide two output modes: raw data and already-filtered data streams. For the following data processing, the filtered data streams will be utilized, which includes the low-bandpass filter (see Section III-B). Each Myoware device is connected to a Raspberry Pi 3 within the blue box shown in Fig. 4(b). The blue box contains analogue-to-digital converters, a battery pack, and a WiFi interface. This interface is used for streaming the collected EMG data to the main workstation (ROS master). Further details on the hardware and ROS setup can be found in [9]. Throughout the experiment, participants were tasked to stand on the markers and hold the engine cover, as shown in Fig. 4(a) and (c). The cover itself has a weight of 8 kg, which is not considered particularly heavy. However, one of the specific features of this container is its uneven weight distribution, as it would be faced in an industrial scenario. Subsequently, different levels of muscle activity are expected for the left-arm and right-arm muscles. In order to enable the classification of different payloads, additional 2.5 kg weights were added at 20 s intervals. The initial EMG data stream, while holding the 8 kg cover, was labeled "0," and after each additional 2.5 kg weight, the data label integer was increased by one, e.g., "0"-8 kg, "1"-10.5 kg, "2"-13 kg, and so on. The addition of weights not only increased the payload but also changed the weight balance of the engine cover. Thus, it became increasingly difficult to hold. Although the main goal was not to test the physical capabilities of participants, they were briefed to press the assistance request button on the engine cover, shown in Fig. 3(c), as soon as they reached an uncomfortable level. The assistance request button is connected to the ROS master and would trigger a Universal Robots model 10 (UR10) to run a predefined assistance program. In the current setup, the UR10 is equipped with a ROBOTIQ 2F-140 gripper and an integrated force/torque sensor. The gripper is holding a thin metal plate, which was fitted with a Styrofoam edging to avoid scratching participants. As soon as the assistance button is pressed, the UR10 would slowly move upward in a straight line, until it touched the engine cover. Afterward, it would move an additional 7 cm upward at the 10 kg payload setting to assist the human operator in lifting the weight. During this assistance operation, the force/torque readings of the gripper would allow for quantifying the amount of weight that the UR10 is holding.
One of the main drawbacks of the current setup is the UR10's limited capability to cope with high payloads. As the name suggests, it has a maximum payload of 10 kg. However, most participants are expected to hold higher payloads before pressing Authorized licensed use limited to the terms of the applicable license agreement with IEEE. Restrictions apply. the assistance request button. Therefore, the robot could not fully take the burden off the operator and rather initiate load sharing. Nevertheless, this would still relieve stress from muscles and joints.
Overall, different performances of participants are expected regarding holding time and weight, due to varying levels of strength, body composition, and stress tolerance. Hence, this would allow the Mondrian Forest to adapt to the uniqueness of one individual. The EMG data results, as well as the Mondrian Forest learning and prediction performance, are presented and discussed in the following section.

V. RESULTS AND DISCUSSION
In this section, the results are presented and discussed in two sections. Section V-A shows the participant's individual performance regarding payload, muscle fatigue, and associated EMG signals. In Section V-B, the results include the payload/fatigue prediction performance as well as the optimization of performance and convergence time.

A. Experimental Results
At first, the experiment described in Section IV was conducted. Overall, different performances regarding payload and holding time before pressing an assistance request were expected, due to the variety of strengths and body composition among participants. This expectation was confirmed, as shown in Table I. Participants 3 and 5 pressed the assistance button after one additional 2.5 kg weight was added, which equals 30-40 s of holding the container. In contrast, participant 2 managed to hold the engine cover until six additional 2.5 kg weights were added, totaling 130-140 s of holding time. Most participants pressed the assistance request button after the third increase in weight This can be achieved by establishing individual, data-driven models, such as those based on EMGs. Therefore, EMG data were collected from both left and right arms throughout the experiment, where variations in EMG signals could be observed among different participants. Moreover, variations could also be observed between the left-arm and the right-arm signals, due to the uneven load balancing of the container. One low-bandpass filtered sample data stream of participant 6 is shown in Fig. 5. For simplification purposes, only data of the left bicep are illustrated. As can be observed, the signal is increasing with each additional weight. Finally, after the fourth addition of 2.5 kg, the participant pressed the assistance request button. Following the assistance request, the robot is slowly moving upward (peaks on the UR10 F/T z-axis) before it touches the container. Immediately, the forces on the x-, y-, and z-axes increase, whereas the EMG signal decreases. In fact, the EMG signal power decreases to the initial levels when the participant was holding the engine cover without any additional weights. Thus, evidently, the effect of robot's assistance can be observed, where the burden of the load is shared between the operator and the robot.
Yet, in a collaborative scenario, the system is envisaged to detect an operator needing assistance by itself (from EMG sensor data, for instance), rather than an operator manually initiating a request by pressing a button. Thus, providing a collaborative robot with cognitive skills regarding its human partner will improve teamwork by compensating each other's weaknesses.
Since the collected EMG signal shows an obvious change in pattern regarding different payloads, high classification performances are expected via machine learning. In the following, the Mondrian Forest will be applied and optimized to predict the different payloads and the fatigue threshold from the collected EMG data.

B. Mondrian Forest Results
Throughout the experiment, EMG data were collected from four channels, namely the biceps brachii and the brachioradialis (forearm) for both the left and right arms. Ideally, the classifier could predict the different payloads and the fatigue limit. This would allow the classifier to trigger an assistance request rather than a participant manually pushing a button. For this purpose, the bandpass-filtered and STFT-transformed EMG data were utilized to train and test a Mondrian Forest. Since the Mondrian Forest learns purely online, the collected data were streamed into the classifier at the same sampling rate as they were acquired (70 Hz). Thus, this allowed for investigating the online learning capabilities regarding prediction performance and convergence time. One of the main challenges was to find an optimal number of trees, which delivered sufficient prediction accuracy, while minimizing convergence time. For evaluating the prediction performance, the metric root mean square error (RMSE) was chosen, since it provides an indication of how close the predictions were to the actual classification. For example, four different payloads can be distinguished for participant 6 (see Fig. 5), ranging from class 0-"low payload" to class 3-"fatigued." If the classifier correctly predicted a window from class 3, the RMSE would be 0%. If, however, it predicted a window from class 3 as class 2, the RMSE would be 33%. Similarly, it would be 67% if it predicted class 1, 100% if it predicted class 0, and so on.
As for the convergence time, the maximum time limit would be one STFT window size, which is currently 750 ms. A longer period for classifying and learning would result in the classifier "lagging behind," which means that it would not be able to classify new data fast enough, resulting in a continuously increasing queue.
In order to find the aforementioned optimum, a hyperparameter optimization was performed based on a loop starting from 1 to 20 trees for each participant individually. However, in contrast to expected high prediction performances, the models did not deliver these results. In fact, whenever a new class was introduced, the model's performance went below 10% prediction accuracy for 1 or 2 s, until it correctly identified the new class. Consequently, due to the short observation window (2-3 min  TABLE II  MODEL TRAINED ON PARTICIPANT 6 AND TESTED ON ALL REMAINING  PARTICIPANTS max) of EMG signals, the relatively poor performance rating would be misleading. Instead, a so-called "warm start" was attempted, rather than training models from scratch. This included pretraining the model on one participant (also online) before applying it to other participants, as it would be faced in an authentic application. Sample results are shown in Table II, which presents the optimization scores of a model trained on participant 6 and then tested on all other participants. Participant 6 was chosen for the warm start due to the average performance and subsequent number of classes (four different payloads, as shown in Fig. 5). As highlighted in Table II, the prediction accuracy reaches saturation at 17 trees. After that, the prediction performance does not improve anymore. Thus, the optimum would be considered at an RMSE of 25% and a convergence time of 528 ms. Generally, this accuracy is considered acceptable due to overlay in frequency/power of EMG signals during different predicted classes (classes 1 and 2 in Fig. 5). Regarding the convergence time, 528 ms is below the set threshold of 750 ms and is therefore considered satisfactory. Overall, two factors need to be considered concerning the convergence time: hardware and code optimization. In the current setup, the Mondrian Forest was trained on a Linux Ubuntu 18.04 machine with an Intel Core i5 processor, an NVIDIA NVS 5200M graphics card, 16-GB RAM. Potentially lower convergence times would be anticipated with a stronger CPU. For code optimization, the Python 3.7 code was compiled into C through Cython. This substantially decreased convergence times from seconds to milliseconds.
After the general optimization, the "warm-started" model, which was trained on participant 6, was applied and further trained on each of the remaining participants individually. In this context, the goal was to investigate the unique features among individuals and subsequent potential differences in tree complexity. As shown in Table III, the data complexity and associated number of trees revealed similar results for different participants. In general, the optimum was reached at either 16 or 17 trees, suggesting similar levels of data complexity. The highest prediction performance was achieved for participant 7 at an RMSE of 7%. The lowest performance was attained for participant 2 at an RMSE of 22%. This could be due to participant 2 having the longest recording and most different payloads (and subsequent classes), which would also result in the most diverse EMG signals. Nevertheless, the model could correctly predict the "fatigue limit" for participant 2 from the EMG data, as in when the participant pressed the assistance request button.
Overall, the Mondrian Forest achieved promising prediction accuracies while keeping the convergence time below the set threshold of 750 ms. As a side note, due to the Mondrian Forest generating the underlying trees randomly, it demonstrates a minor variation in terms of prediction performance and convergence time after each training.

VI. CONCLUSION AND FUTURE WORK
In this article, a novel application of an incremental learner (Mondrian Forest) was proposed to predict payloads/muscular fatigue from EMG signals in human-robot collaboration. This is envisioned to provide the collaborative system with awareness when an operator requires assistance from the robot. Since strength and endurance levels demonstrate a large variety among different individuals, muscle activity and fatigue are directly measured on the operator's muscles. For this purpose, wearable EMG devices were placed on the human operator's biceps brachii and brachioradialis (forearm) since they are mainly involved in lifting/holding activities. However, other placements of EMG sensors are plausible, such as shoulder muscles and triceps (for pushing activities). Thus, this would allow for a wider range of scenarios regarding the prediction of payloads and associated fatigue levels.
In the current context, an experiment was performed, in which the payload held by participants was incrementally increased while the EMG data stream was labeled until they reached an uncomfortable/fatigued level. Participants then pressed an assistance request button, which triggered the support of the collaborative robot. These data were used in the Mondrian Forest. At first, the acquired raw EMG data were streamed directly into the Mondrian Forest to minimize fine-tuning efforts. This, however, did not achieve sufficient results for all participants. Therefore, additional preprocessing steps were added in low-bandpass filters and short-term Fourier transform to extract relevant features (such as spectrum analysis). From these processed data, the Mondrian Forest could identify the payloads and the fatigue threshold more accurately. Afterward, a hyperparameter optimization of the Mondrian Forest could be performed regarding the number of trees in the model. Generally, the higher the number of trees, the higher the prediction performance, until saturation is reached. After that threshold (so-called elbow), the performance does not increase anymore. At the same time, a higher number of trees in the model results in a higher convergence time (time to perform a prediction and learning simultaneously). Thus, the optimization considered both prediction performance and convergence time. Overall, promising prediction results were achieved due to the adaptation of the models to the uniqueness of individuals. Moreover, since incremental learners learn purely online, this negates the effect of varying performances between offline and online applications. For instance, the Mondrian Forest correctly identified new classes (payloads) "on the fly" without having seen the data before.
Limitations of the current setup were the collaborative robot's restricted maximum payload capabilities, which are often between 10 and 15 kg. For instance, the UR10 used in the experiment has a maximum payload of 10 kg. Consequently, the robot is fairly limited in terms of assisting operators with high payloads. On one hand, this maximum payload restriction is intended to create a safe collaborative working environment by protecting the health and safety of human operators. On the other hand, the collaboration between humans and robots aims to compensate for each party's weakness, where human operators are considered susceptible to fatigue and robots are deemed fatigue-proof. Hence, current collaborative robots would be required to evolve in terms of maximum payload while maintaining safety for the human operator. Consequently, future research into safely coordinating higher levels of force/torque of cobots is considered necessary. Ultimately, this is envisioned to create a more ergonomic working environment and, thus, further synergetic effects between humans and robots.