Semi-mind controlled robots based on Reinforcement Learning for Indoor Application

Abstract

This paper presents a solution to this challenge by introducing interactive feedback derived from brain signals to train robots using deep reinforcement learning, particularly in the context of indoor maze navigation. Our objective is to enhance the learning process in a human-robot interaction scenario by incorporating human emotion or attention feedback. To accomplish this, we empowered the robot to learn new tasks through a dynamic policy network based on human feedback, and we augmented this input with other sensor data, including LIDAR. Various experiments are conducted to compare the eficacy of manual feedback, brain signal feedback, and no brain signal feedback, employing diverse Reinforcement Learning models. Additionally, we explore diferent models for emotion classification, employing Graph Neural Network models and traditional deep learning models, and subsequently compare the outcomes.

Keywords

• eol>EEG
• Deep Learning
• Graph Neural Networks
• Human-Robot Interaction

1. Introduction

IV dataset, and 88.51% and 85.62% for four classifications

and two dichotomies on the DEAP dataset, respectively.

Robotic applications have become indispensable across The key contributions of the paper encompass the introdiverse fields such as rescue operations, medical assis- duction of the CAN model, which combines deep-gated tance, and autonomous driving. Within the realm of recurrent neural networks [ 10 ] with canonical correlamobile robotics, autonomous navigation stands out as tion and an attention mechanism [ 11 ], and the exploa crucial research focus [ 1, 2 ], traditionally relying on ration of coordinated representation in tasks related to map-building techniques like SLAM [ 3 ]. Nevertheless, multimodal emotion recognition[ 12 ]. there is a growing interest in mapless navigation, which The literature reviewed diverse approaches establishes a direct link between sensory inputs and robot to autonomous robot navigation in unfamiliar actions. While deep reinforcement learning (DRL) has environments[ 13, 14, 15 ]. Simultaneous Localization made noteworthy advancements in autonomous naviga- and Mapping (SLAM) algorithms, such as the one tion, the challenges of real-world training persist [ 4, 5 ]. proposed in [ 16 ] and the lightweight OrthoSLAM

In the presented paper [ 6 ], the authors introduce a algorithm [ 17 ], emphasize the simultaneous mapping novel model named the Correlated Attention Network and localization process. In [ 18 ], an innovative method (CAN) for multimodal emotion recognition[ 7, 8 ]. This is introduced, combining Hector SLAM with an Artificial model extends the attention-based recurrent neural net- Potential Field (APF) controller, specifically designed work by integrating correlation calculations of difer- for indoor navigation in GPS-denied environments. ent gated recurrent units, specifically targeting the cor- The proposal in [ 19 ] advocates for a semantically rich relation between EEG and eye movement signals[ 9 ]. graph representation in the context of indoor navigation. Through the utilization of coordinated representation Additionally, research in [ 20, 21 ] explores end-to-end incorporating complementary features, the CAN model approaches utilizing Convolutional Neural Networks aims to achieve enhanced emotion classification accu- (CNNs) for autonomous robot navigation, particularly racy. The experimental outcomes on three real-world with RGB-D cameras[ 22 ]. datasets reveal that the proposed model outperforms ex- Methods based on Deep Reinforcement Learning (DRL) isting state-of-the-art methods, achieving mean accura- are also explored, such as the fusion of deep reincies of 94.03% on the SEED dataset, 87.71% on the SEED forcement learning with Recurrent Neural Networks (RNNs)[23, 24? ] for path selection discussed in [ 25 ]. In ICYRIME 2023: 8th International Conference of Yearly Reports on [ 26 ], an approach utilizing Deep Q-Network (DQN) for Informatics, Mathematics, and Engineering. Naples, July 28-31, 2023 autonomous robot navigation with visual observations is $ [e-mail] (I. Naidji); presented, albeit relying on RGB-D cameras. Addressing wahamlidegdu.teibttearlma@[e-mail]-(bWis.kGrau.deztt(aAla.);Tibermacine); collision avoidance and navigation, [ 27, 28 ] employ a [e-mail] (I. E. Tibermacine) Double Deep Q-Network (DDQN) approach. Lastly, [ 29 ] © 2023 Copyright for this paper by its authors. Use permitted under Creative Commons License introduces an Asynchronous Advantage Actor-Critic netAttribution 4.0 International (CC BY 4.0). The overarching design of this global system [Fig.1] adopts an AI perspective and consists of two principal components: the Brain Signal segment utilizing EEG and the DRL (Deep Reinforcement Learning) segment. The system architecture is defined by an environment featuring a maze with static obstacles representing maze walls and dynamic obstacles. A robot, guided by a DRL agent, endeavors to navigate through the maze with the goal of reaching a randomly generated red circular spot.

The DRL agent is reliant on two primary feedback signals for observations. The first is the EEG signal, portraying emotions like sadness, happiness, and neutrality.

The second feedback is derived from a 2D LIDAR sensor, work to enhance generalization and reduce learning time furnishing environmental information to the agent. Hufor indoor robot navigation. man interaction with the system is facilitated through

These studies ofer valuable insights into various tech- a Brain-Computer Interface (BCI) setup. Users observe niques and algorithms employed in autonomous robot the simulation and gather EEG recordings. The collected navigation, contributing to the comprehension of Simul- EEG data undergoes preprocessing, feature extraction, taneous Localization and Mapping (SLAM), end-to-end and model prediction to categorize emotional states into strategies, Deep Reinforcement Learning (DRL), and col- three classes: sad, happy, and neutral. lision avoidance methodologies. Concurrently, the 2D LIDAR sensor data is amalgamated with the EEG data and serves as observations for the DRL agent. The agent acquires the ability to make 2. Proposed model optimal decisions and execute actions by assimilating the combined feedback, facilitating efective navigation 2.1. Global System Design through the maze. Various iterations of the system have been created, incorporating versions that exclusively rely on the 2D LIDAR sensor or concentrate solely on emotion feedback. Additionally, diverse noise types and persistent feedback mechanisms have been investigated to improve the overall performance of the system.

In the upcoming sections, specific Deep Reinforcement Learning (DRL) algorithms and their intricate functionalities will be elucidated, ofering deeper insights into the operational and learning processes of the system.

3. EEG DATASET

3.1. Dataset Overview Concerning the brain signal feedback, it can be substituted with keyboard input involving humans as the The EEG dataset comprises recordings from 15 partici- source of feedback. In this scenario, users can input pants involved in the experiment. Each participant un- their emotional state, specifying whether they feel happy, derwent three sessions on distinct days, and each session sad, or neutral. Alternatively, the brain signal feedback encompassed 24 trials. The film clips displayed during system can be replaced with generated samples of emothese trials were meticulously selected to elicit emotions tional states adhering to a uniform distribution. In this such as happiness, sadness, fear, or a neutral emotional approach, three variables govern the probability of each state. Consequently, the dataset encompasses four emo- emotional event: happiness, sadness, and neutrality. tional classes: happiness, sadness, fear, and neutrality.

The SEED IV dataset [ 30 ] ofers an extensive compila- 3.2.4. Classification Models Used For Emotion tion of EEG signals, with a specific emphasis on emotion Dataset recognition utilizing both EEG and eye movement data.

This section provides an outline of the EEG segment We employed emotion classification as both a feedback within the SEED IV dataset utilized in this thesis. It en- system and an observation for deep reinforcement learncompasses details regarding the number of subjects, the ing. This feedback ofers insights into the individual’s iflm clips utilized, and the classes available. Additionally, emotional state within the simulation. To accomplish this, information about the preprocessing of the EEG signals we utilized diverse models for classifying brain signals and the process of feature extraction is presented. into four categories: angry, sad, happy, or neutral.

3.2.3. Replacmnet of Emotion feedback

3.2. Brain Signal Feedback Part

This is a sub-system design from the global system which has 2 phases:

3.2.1. Phase 1 During this stage, the model, specifically the CNN model, is trained utilizing the SEED-IV dataset. We implemented the preprocessing and feature extraction methods outlined in the dataset section. For our model, predictions were made for four distinct classes. However, within our overarching system design, the classes of sadness and fear were amalgamated into a single class. Subsequently, the batches were converted into 2D images to be inputted into the CNN model. Alongside the CNN model trained in the train data loader, we also conducted experiments with alternative models, including advanced CNN architecture, LSTM, and ViT. 3.2.2. Phase 2

Upon completing the initial phase, we employ the trained

weights and integrate them into our system as an ofline model. This ofline model encompasses a predetermined function that receives input from the BCI API, initially transforming the features to predict real-time human emotions. The predicted emotions are then stored in the EEG bufer. Subsequently, the EEG bufer is combined with the 2D lidar bufer, creating the sensor bufer and constituting the resultant batch.

3.2.5. Reward Function

The reward function comprises five versions denoted as

A, B, C, D, and E. In Version A, lidar serves as the observation, while Version B utilizes a uniform distribution of emotion. Version C combines the uniform distribution of emotion with lidar, Version D relies solely on emotion, and Version E combines emotion with lidar. Our focus will be on the two principal reward functions, with the other three representing variations of these main ones.

The first reward function is a hybrid model that incorporates lidar and emotion as observations. The second reward function relies on a uniform distribution of emotion based on a specified probability. Additionally, there is a variable named ’constant’ that is added to the reward when the robot successfully avoids collisions with obstacles. This constant value varies based on four types: Step-1, Step-2, Episode, and time, each defined in seconds. The constant is added according to one of these four types.

4. EXPERIMENTAL SETUP

4.1. Preprocessing and Feature Extraction

During the EEG preprocessing phase, several procedures

were implemented to ready the data for subsequent analysis. These steps encompass downsampling the raw EEG data to a sampling rate of 200 Hz and applying a bandpass filter ranging from 1 Hz to 75 Hz to eliminate noise. Subsequent to the preprocessing stage, EEG feature extraction was conducted to capture pertinent information from the data. Extracted features include Power Spectral Density (PSD) and Diferential Entropy (DE). PSD

6. Results

6.1. Simulation Challenges in Maze Environment quantifies the power distribution across five frequency bands, specifically Delta (1 Hz - 4 Hz), Theta (4 Hz - 8 Hz), Alpha (8 Hz - 14 Hz), Beta (14 Hz - 31 Hz), and Gamma (31 Hz - 50 Hz). Conversely, DE computes the diferential entropy within each segment and across the aforementioned frequency bands.

In this section, we delve into the diverse simulation chal

lenges encountered within the maze environment. These 4.1.1. Utilization of EEG Feature Data challenges are intentionally designed to assess the robot’s capabilities and performance across various scenarios.

The ℎ directory within the SEED IV An illustration showcasing the nine distinct challenges dataset served as the source for feature extraction, specif- can be found in Figure.2. ically employing the PSD and DE features. Data from each .mat file was loaded and transformed into a format 6.2. Challenge 1: Square Maze suitable for model training. To optimize the training process, samples from each .mat file were concatenated into The first challenge introduces a straightforward square batches of size 128, where each batch contained EEG maze with the robot positioned at the center. This chalsegments from diferent trials of the same subject. Fur- lenge serves as the baseline for assessing the robot’s thermore, one label from the four emotional classes was navigation capabilities. assigned to each batch. These batches were then used to construct a dataloader, which was subsequently divided 6.3. Challenge 2: Square Maze with into training, validation, and testing sets with ratios of 0.8, 0.1, and 0.1, respectively. To introduce randomness Dynamic Obstacles into the training process, the dataset was shufled.

By employing this methodology, our objective is to train a model capable of recognizing emotions based on EEG features extracted from the SEED IV dataset.

Challenge 2 expands on the preceding challenge by in

corporating four dynamic obstacles into the square maze. The introduction of these obstacles enhances the complexity of the robot’s path planning and obstacle avoidance strategies.

4.1.2. Simulation Feedback Data

The dataset utilized in this case study comprises environ

mental data acquired from a LIDAR sensor and feedback from an EEG (electroencephalogram) device. The LIDAR sensor supplies information about the surroundings, including distances to obstacles and the robot’s orientation. Concurrently, the EEG device gauges the brain activity of the operator, ofering feedback on their cognitive state.

5. Dynamical Graph Convolutional

Neural Networks (DGCNN)

6.4. Challenge 3: Varied Speed and

Movement of Dynamic Obstacles

In this particular challenge, the dynamic obstacles showcase varying speeds and movement patterns in contrast to Challenge 2. The robot is required to adjust its navigation strategy accordingly to evade collisions.

6.5. Challenge 4: Maze with Inner Walls and Two Dynamic Obstacles

Challenge 4 integrates inner walls into the maze structure

The DGCNN (Dynamical Graph Convolutional Neural and introduces two dynamic obstacles. The inclusion of Networks) model [? ] is designed for EEG emotion recog- inner walls adds an additional layer of complexity to the nition tasks. It leverages graph convolutional operations robot’s path planning process. to capture spatial dependencies between diferent electrodes in EEG signals. Graph Neural Networks, such 6.6. Challenge 5: Maze with Inner Walls as the DGCNN model, are essential for capturing and and Six Dynamic Obstacles modeling relationships between graph-structured data, where nodes represent entities and edges represent con- This particular challenge incorporates inner walls; hownections. In the con- text of EEG emotion recognition, ever, it introduces a higher number of dynamic obstacles, GNNs enable the DGCNN model to efectively learn and totaling six. This challenge is designed to assess the exploit the spatial dependencies between diferent elec- robot’s capability to navigate through an environment trodes. By leveraging graph convolutional operations, with an increased density of obstacles. the DGCNN model can capture complex patterns and representations from EEG signals, leading to improved emotion recognition performance. 6.7. Challenge 6: Maze with Dynamic

Obstacles (No Inner Walls) 6.9. Challenge 8: Larger Maze with Inner

Walls and Two Dynamic Obstacles Challenge 6 eliminates the inner walls present in Chal- Expanding on the foundation set by Challenge 7, Challenge 5 but retains the identical setup of six dynamic ob- lenge 8 introduces two dynamic obstacles to the maze. stacles. This challenge assesses the robot’s performance This challenge amalgamates the complexities of navigatin navigating through an open maze environment. ing in a larger maze with the additional challenge posed by moving obstacles. 6.8. Challenge 7: Larger Maze with More

Inner Walls Challenge 7 introduces a larger maze with a heightened count of inner walls. The robot is tasked with navigat- Finally, Challenge 9 preserves the larger maze and inner ing through narrower passages and surmounting more walls featured in Challenge 8 but augments the number intricate maze structures. of dynamic obstacles to six. This challenge poses an exceptionally demanding scenario, evaluating the robot’s capacity to navigate through intricate maze structures and contend with a high density of dynamic obstacles.

In summary, these nine challenges comprise a variety of maze configurations and obstacle arrangements, grad6.10. Challenge 9: Larger Maze with Inner

Walls and Six Dynamic Obstacles Model CNN LSTM GRU ViT DQCNN ually escalating in dificulty. Together, they facilitate a exhibits stable learning in the initial epochs. The converthorough assessment of the robot’s navigation and ob- gence of the plots becomes evident from epoch 80. stacle avoidance capabilities. Figure.2 provides a visual Additionally, the diminishing loss value as the model representation of the diferent challenges discussed. progresses through epochs indicates that the model is learning and refining its predictions. The training accu6.11. EEG Results racy shows a consistent increase with each epoch, signifying an improvement in the model’s ability to predict According to the provided table (Tab.6.11) displaying the the training data. It commences at 48.08% and gradually outcomes of models trained on the SEED IV dataset, each advances to 99.26%. model underwent training three times, and the average, Similarly, the validation accuracy follows a comparable as well as the variance of the scores, were computed. trajectory to the training accuracy, initiating at 46.58% The DGCNN (Dynamic Graph Convolutional Neural Net- and progressing to 98.67%. This progression implies that work) emerged as the top-performing model in terms of the model efectively generalizes and maintains consisaccuracy, delivering the most favorable results. tent performance on unseen data.

Specifically, DGCNN attained an average training ac- The final test accuracy also exhibits improvement over curacy of 99.52% and an average validation accuracy the epochs, commencing at 47.63% and culminating at of 99.03%. In comparison, ViT (Vision Transformer) 98.64%. This metric serves as a comprehensive gauge of achieved an average training accuracy of 99.67% and an the model’s performance on the test set. average validation accuracy of 98.60%. In summary, the training and validation accuracies

The outstanding performance of DGCNN can be of the DGCNN model exhibit a correlated movement, ascribed to its architectural features and capabilities. with the training accuracy marginally outperforming. DGCNN employs graph convolutional layers, which The model showcases stable learning, begins converadeptly capture and model intricate relationships and gence from epoch 80, and manifests enhancements in dependencies within the data. This renders it particu- loss, training accuracy, validation accuracy, and final test larly suitable for tasks involving graph-structured data, accuracy over the epochs. These outcomes signify that such as the SEED IV dataset. Furthermore, the high accu- the model learns efectively, generalizes proficiently, and racy demonstrated by DGCNN suggests its proficiency attains high accuracy on both the training and validation in generalizing efectively to unseen data. datasets.

In conclusion, the DGCNN model surpasses the other models in accuracy on the SEED IV dataset. The triumph 6.11.2. Analysis of Loss Curves and Convergence of this model can be attributed to its architectural design and its capability to efectively capture complex relationships and dependencies within the data.

Throughout the 100 epochs, both the loss curves of

DGCNN exhibit minimal variance observed across multiple trials. This suggests that the model consistently 6.11.1. Analysis of DGCNN Training and converges towards similar local minima, yielding highly Validation Results accurate predictions.

Furthermore, the loss curve illustrates a gradual deBased on the training accuracy and validation accuracy scent, initiating from an initial high value and steadily deplots of DGCNN [Fig.3], it is observed that the plots ex- creasing throughout the training process. This indicates hibit a synchronous pattern, with the training accuracy the model’s efective learning and adjustment of its pamarginally surpassing the validation accuracy. Never- rameters to better fit the training data. After 100 epochs, theless, the validation plot does not deviate significantly DGCNN attains an impressive loss value of 0.0005, infrom the training plot. This suggests that our model dicating a high level of accuracy and robust predictive learns efectively, demonstrates good generalization, and capability.

This consistency underscores the stability and relia- nizing happiness, sadness, and neutral emotions. It also bility of the solutions derived by DGCNN. In summary, performs commendably in detecting fear, with only a the analysis of the loss curves of DGCNN unveils its slight misclassification rate. remarkable capacity to minimize disparities between pre- In conclusion, the DGCNN model showcases rodictions and ground truth values, resulting in highly ac- bust classification capabilities across all emotion classes. curate models. DGCNN exhibits a robust ability to learn DGCNN delivers reliable predictions, establishing itself and converge towards dependable solutions, making it a as a valuable tool for emotion classification tasks. valuable tool for various applications.

7. Conclusion

6.11.3. Analysis of Confusion Matrix

The confusion matrix provides valuable insights into the

performance of classification models.

In this study, we conducted a thorough examination of the integration of deep reinforcement learning (DRL) with emotion feedback. Through a series of experiments, • Class 0 (Happiness):The DGCNN model achieves we identified optimal parameters to enhance learning perfect accuracy (1.00) in classifying happiness, outcomes by incorporating both lidar and emotion feedindicating no misclassifications in this category. back. Our investigations yielded improved overall perfor• Class 1 (Sadness): Similarly, the DGCNN model mance and a notable reduction in the number of episodes demonstrates flawless performance in identifying required for successful maze navigation. By incorposadness, with a precision of 1.00. rating natural sensor feedback, particularly human vi• Class 2 (Fear): The DGCNN model performs well sual perception and brain signals, we gained valuable in detecting fear, with an accuracy of 0.98. How- insights into the potential of utilizing such feedback in ever, it misclassifies a small portion (0.01) of the reinforcement learning tasks. Our findings highlight the samples in this class as other emotions. efectiveness of combining deep reinforcement learning • Class 3 (Neutral): The DGCNN model achieves with emotion feedback, opening up new possibilities for perfect accuracy (1.00) in recognizing neutral leveraging natural sensor feedback and presenting opemotions, implying no misclassifications in this portunities for future work in multi-robot scenarios and category. transfer learning using brain signals beyond emotions.

The DGCNN model demonstrates outstanding classification outcomes, achieving perfect accuracies in recog

Acknowledgments

This work has been developed at is.Lab() Intelligent Sys

tems Laboratory at the Department of Computer, Control, and Management Engineering, Sapienza University of Rome (https:// islab.diag.uniroma1.it). The work has also been partially supported from Italian Ministerial grant PRIN 2022 “ISIDE: Intelligent Systems for Infrastructural Diagnosis in smart-concretE”, n. 2022S88WAY - CUP B53D2301318, and by the Age-It: Ageing Well in an ageing society project, task 9.4.1 work package 4 spoke 9, within topic 8 extended partnership 8, under the National Recovery and Resilience Plan (PNRR), Mission 4 Component 2 Investment 1.3—Call for tender No. 1557 of 11/10/2022 of Italian Ministry of University and Research funded by the European Union—NextGenerationEU, CUP B53C22004090006. 51–59

References

[1] M. Khaled , B. Sirmaçek , S. Stramigioli , M. Poel , Towards continuous control for mobile robot navigation: A reinforcement learning and slam based approach , in: ISPRS Workshop Indoor 3D 2019 , 2019 .

[2] B. A. Nowak , R. K. Nowicki , M. Woźniak , C. Napoli, Multi-class nearest neighbour classifier for incomplete data handling , volume 9119 , 2015 , pp. 469 - 480 . doi: 10 .1007/978-3- 319 -19324-3\_ 42 .

[3] H. Durrant-Whyte , T. Bailey , Simultaneous localization and mapping: part i , IEEE robotics & automation magazine 13 ( 2006 ) 99 - 110 .

[4] L. Tai , G. Paolo, M. Liu, Virtual-to-real deep reinforcement learning: Continuous control of mobile robots for mapless navigation , in: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , IEEE, 2017 , pp. 31 - 36 .

[5] Y. Zhu , R. Mottaghi , E. Kolve , J. J. Lim , A. Gupta , L. Fei-Fei , A. Farhadi , Target-driven visual navigation in indoor scenes using deep reinforcement learning , in: 2017 IEEE international conference on robotics and automation (ICRA) , IEEE, 2017 , pp. 3357 - 3364 .

[6] J.-L. Qiu , X.-Y. Li , K. Hu , Correlated attention networks for multimodal emotion recognition, in: 2018 IEEE international conference on bioinformatics and biomedicine (BIBM) , IEEE, 2018 , pp. 2656 - 2660 .

[7] G. De Magistris , R. Caprari , G. Castro, S. Russo , L. Iocchi , D. Nardi , C. Napoli , Vision-based holistic scene understanding for context-aware humanrobot interaction , in: International Conference of the Italian Association for Artificial Intelligence , Springer, 2021 , pp. 310 - 325 .

[8] N. Brandizzi , S. Russo , G. Galati, C. Napoli , Addressing vehicle sharing through behavioral analysis: A solution to user clustering using recencyfrequency-monetary and vehicle relocation based on neighborhood splits , volume 13 , 2022 . doi: 10 . 3390/info13110511.

[9] E. Iacobelli , V. Ponzi , S. Russo , C. Napoli , Eyetracking system with low-end hardware: Development and evaluation , Information 14 ( 2023 ) 644 .

[10] K. Yao , T. Cohn , K. Vylomova , K. Duh , C. Dyer , Depth-gated recurrent neural networks , arXiv preprint arXiv:1508.03790 9 ( 2015 ) 98 .

[11] Y.-T. Lan , W. Liu, B.-L. Lu , Multimodal emotion recognition using deep generalized canonical correlation analysis with an attention mechanism , in: 2020 International Joint Conference on Neural Networks (IJCNN) , IEEE, 2020 , pp. 1 - 6 .

[12] I. E. Tibermacine , A. Tibermacine , W. Guettala , C. Napoli , S. Russo , Enhancing sentiment analysis on seed-iv dataset with vision transformers: A comparative study , in: Proceedings of the 2023 11th International Conference on Information Technology: IoT and Smart City , 2023 , pp. 238 - 246 .

[13] V. Ponzi , S. Russo , V. Bianco , C. Napoli , A. Wajda , Psychoeducative social robots for an healthier lifestyle using artificial intelligence: a case-study , volume 3118 , 2021 , pp. 26 - 33 .

[14] B. Nail , M. A. Atoussi , S. Saadi , I. E. Tibermacine , C. Napoli , Real-time synchronisation of multiple fractional-order chaotic systems: An application study in secure communication , Fractal and Fractional 8 ( 2024 ) 104 .

[15] S. Pepe , S. Tedeschi , N. Brandizzi , S. Russo , L. Iocchi , C. Napoli , Human attention assessment using a machine learning approach with gan-based data augmentation technique trained using a custom dataset , OBM Neurobiology 6 ( 2022 ). doi: 10 . 21926/obm.neurobiol. 2204139 .

[16] A. Garulli , A. Giannitrapani , A. Rossi , A. Vicino , Mobile robot slam for line-based environment representation , in: Proceedings of the 44th IEEE Conference on Decision and Control , IEEE, 2005 , pp. 2041 - 2046 .

[17] V. Nguyen , A. Harati , A. Martinelli , R. Siegwart , N. Tomatis , Orthogonal slam: a step toward lightweight indoor autonomous navigation , in: 2006 IEEE/RSJ International Conference on Intelligent Robots and Systems , IEEE, 2006 , pp. 5007 - 5012 .

[18] E. H. C. Harik , A. Korsaeth , Combining hector slam and artificial potential field for autonomous navigation inside a greenhouse , Robotics 7 ( 2018 ) 22 .

[19] G. Sepulveda , J. C. Niebles , A. Soto , A deep learning based behavioral approach to indoor autonomous navigation , in: 2018 IEEE international conference on robotics and automation (ICRA) , IEEE, 2018 , pp. 4646 - 4653 .

[20] Y.-H. Kim , J.-I. Jang , S. Yun , End-to-end deep learning for autonomous navigation of mobile robot , in: 2018 IEEE International Conference on Consumer Electronics (ICCE) , IEEE, 2018 , pp. 1 - 6 .

[21] J. Wang , X. Ding , H. Xia , Y. Wang , L. Tang , R. Xiong , A lidar based end to end controller for robot navigation using deep neural network , in: 2017 IEEE International Conference on Unmanned Systems (ICUS) , IEEE, 2017 , pp. 614 - 619 .

[22] B. Nail , B. Djaidir , I. E. Tibermacine , C. Napoli , N. Haidour , R. Abdelaziz , Gas turbine vibration monitoring based on real data and neuro-fuzzy system , Diagnostyka 25 ( 2024 ).

[23] S. Falciglia , F. Betello , S. Russo , C. Napoli , Learning visual stimulus-evoked eeg manifold for neural image classification , Neurocomputing ( 2024 ) 127654 .

[24] A. Alfarano , G. De Magistris, L. Mongelli , S. Russo , J. Starczewski , C. Napoli , A novel convmixer transformer based architecture for violent behavior detection 14126 LNAI ( 2023 ) 3 - 16 . doi: 10 .1007/ 978-3- 031 -42508-0\_1.

[25] H. Quan , Y. Li , Y. Zhang, A novel mobile robot navigation method based on deep reinforcement learning , International Journal of Advanced Robotic Systems 17 ( 2020 ) 1729881420921672 .

[26] P. Yue , J. Xin , H. Zhao , D. Liu , M. Shan , J. Zhang, Experimental research on deep reinforcement learning in autonomous navigation of mobile robot , in: 2019 14th IEEE Conference on Industrial Electronics and Applications (ICIEA) , IEEE, 2019 , pp. 1612 - 1616 .

[27] X. Xue , Z. Li , D. Zhang , Y. Yan , A deep reinforcement learning method for mobile robot collision avoidance based on double dqn , in: 2019 IEEE 28th International Symposium on Industrial Electronics (ISIE) , IEEE, 2019 , pp. 2131 - 2136 .

[28] X. Ruan , D. Ren , X. Zhu , J. Huang , Mobile robot navigation based on deep reinforcement learning, in: 2019 Chinese control and decision conference (CCDC) , IEEE, 2019 , pp. 6174 - 6178 .

[29] H. Surmann , C. Jestel , R. Marchel , F. Musberg , H. Elhadj , M. Ardani , Deep reinforcement learning for real autonomous mobile robot navigation in indoor environments , arXiv preprint arXiv: 2005 . 13857 ( 2020 ).

[30] P. Zhong , D. Wang , C. Miao , Eeg-based emotion recognition using regularized graph neural networks , IEEE Transactions on Afective Computing 13 ( 2020 ) 1290 - 1301 .