A Deep Learning-Based Autonomous RobotManipulator for Sorting Application
AA Deep Learning-Based Autonomous RobotManipulator for Sorting Application
Hoang-Dung Bui, Hai Nguyen, Hung Manh La, Shuai Li
Abstract —Robot manipulation and grasping mechanisms havereceived considerable attention in the recent past, leading todevelopment of wide-range of industrial applications. This paperproposes the development of an autonomous robotic graspingsystem for object sorting application. RGB-D data is used by therobot for performing object detection, pose estimation, trajectorygeneration and object sorting tasks. The proposed approachcan also handle grasping on certain objects chosen by users.Trained convolutional neural networks are used to perform objectdetection and determine the corresponding point cloud cluster ofthe object to be grasped. From the selected point cloud data, agrasp generator algorithm outputs potential grasps. A grasp filterthen scores these potential grasps, and the highest-scored graspwill be chosen to execute on a real robot. A motion plannerwill generate collision-free trajectories to execute the chosengrasp. The experiments on AUBO robotic manipulator show thepotentials of the proposed approach in the context of autonomousobject sorting with robust and fast sorting performance.
Index Terms —robot manipulation; grasping; deep learning;object sorting;
I. I
NTRODUCTION
Object sorting has numerous applications in a diverse rangeof environments and contexts, ranging from household andindustrial settings to agriculture and pharmaceutical industries.However, objects sorting tasks performed by human beings aretedious and error-prone in nature, especially over extendedperiods of time. In order to improve on the deficiencies ofhuman-based sorting applications, the use of autonomous orsemi-autonomous robots has been proposed in recent studies[1]–[4]. In [1], Gupta at al. proposed a framework, whichsorted the simple Duplo bricks by size and color by usingdepth data to determine the grasp pose for the bricks. However,its application was limited to very light objects with simple
This material is based upon work supported by the National Aeronauticsand Space Administration (NASA) Grant No. NNX15AI02H issued throughthe NVSGC-RI program under sub-awards No. 18-94 and 20-16. This workis also partially supported by the U.S. National Science Foundation (NSF)under grants NSF-CAREER: 1846513 and NSF-PFI-TT: 1919127. The views,opinions, findings and conclusions reflected in this publication are solely thoseof the authors and do not represent the official policy or position of the NASAand NSF.The first two authors have made equal contributions to this work.H.D. Bui, H. Nguyen and H. La are with the Advanced Robotics andAutomation (ARA) Laboratory, Department of Computer Science and En-gineering, University of Nevada, Reno, NV 89557, USA.S. Li is with the College of Engineering, Swansea University, Fabian Way,Swansea, SA1 8EN, Wales, UK.Corresponding author: Hung La ( e-mail: [email protected] geometries. In [2], Zeng at al. proposed an robotic pick-and-place system, which was able to grasp and recognizeobjects in cluttered environment. In this algorithm, the objectrecognizing happened after grasping, and it seemed suitablefor the cleaning task than object’s sorting. In [3], Guerin etal. applied an unsupervised deep neural network for a roboticmanipulator to classify the object based on feature extractionand standard clustering algorithm. It sorted well the objectswith similar geometries, however, it failed to classify the sametype objects with different dimensions.To work in complex environment [5], [6], an autonomoussorting machine should have the following capabilities: (i)detection and classification of objects with different shapes,sizes and physical properties, (ii) optimal object grasping, and(iii) trajectory generation and motion planning within the 3Denvironment. In this paper, we present our development ofan autonomous object sorting system using robotic graspingmechanism. By using deep neural networks, the system is ableto detect multiple types of objects, select an optimal graspingobject, and its optimal grasp pose, perform the grasp actionon a real robotic manipulator. Particularly, the contribution ofthis paper are: • development of a complete integration system of roboticsorting manipulator, • combination of two convolutional neural networks(CNNs) to be able to process RGB-D data to do bothduties of object detection and object grasping.Combining with Trajopt [7] motion planner, the experimentson AUBO robotic manipulator show the potentials of theproposed approach in the context of autonomous object sortingwith robust and fast sorting performance.This paper has been divided into five sections. In sec-tion II, state-of-the-art, related to object detection, graspingmechanisms and robotic motion planning for pick-and-placeoperations, will be presented. Section III will discuss thesalient features of the proposed method for development ofobject sorting system using autonomous robotic manipulatorand Deep CNNs. Section IV will discuss the different aspectsof the experimentation and associated results. Section V willconclude the research findings and provide recommendationsfor future research in relevant research area.II. R
ELATED W ORK
This section discusses the state-of-the-art technologies inObject Detection and Object Grasping.
Object Detection . The sorting system runs in real-time a r X i v : . [ c s . R O ] S e p ig. 1: Proposed approach.manners and performs both detection and grasping work.Thus, it requires that the used algorithms should meet thereal-time running, and provide the coordinates of detectedobjects. Applying deep neural network in object detection hasimproved in term of accuracy and real-time processing evenwith limited computational resources [8]. The CNNs in [9],[10] are region proposal based framework, which mappedstraightly from image pixels to bounding box coordinatesand class probabilities, thus reduce time expense for sharedconvolution parameters. Liu et al. [10] proposed a SingleShot MultiBox Detector (SSD), which takes advantage of aset of default anchor boxes with different aspect ratios andscales to discretize the output space of bounding boxes. Tohandle objects with various sizes, the CNN fuses predictionsfrom multiple feature maps with different resolutions. Giventhe VGG16 [11] backbone architecture, SSD adds severalfeature layers to the end of the network to predict the offsetsto default boxes with different scales and aspect ratios andtheir associated confidences. The network is trained with aweighted sum of localization loss and confidence loss. SSDruns at 59 frame per second (FPS) with 28.1 mean AveragePrecision (mAP), however, it does not handle well with smallobjects.Redmon et al. [9] proposed a novel framework calledYOLO to predict both confidences for multiple categoriesand bounding boxes. The YOLOv3 consists of 53 conv layersof which some conv layers construct ensembles of inceptionmodules with 1x1 reduction layers followed by 3 x 3 convlayers. It is able to process an image in 22 ms at 28.2 mAPand classify more than 80 object classes. With real-timeoperation capabilities, efficient performance and versatilityfor object detection, YOLOv3 was a good option for ourproposed system. Even though, we need to modify the outputof YOLO to get the centroid of each object selection. Grasp Detection . The use of point cloud data withCNN has been used to provide a reliable object grasp posewith varying pose and finger gripper configurations [6], [12].In [13], [14], the authors tried to find good grasp poses usingRGB-D image frames with single-state regression to obtainbounding boxes containing target objects. The algorithm in[13] reached great speed up to 3 FPS on a GPU and highaccuracy of 90.0 % on image-wise split. However, it onlyprovided the pose in 2D, and lacked of pose orientationin depth direction. The algorithms in [14] considered the depth data and outputted reliable grasp poses, which can begraspable for finger gripper configuration. The drawback isthat it is is not a real-time application due to the processingtime.In [15], [16], the authors used another neural network calledGrasp Pose Detection (GPD) to improve the quality of detectedgrasp poses. The input of the CNN was the object pointcloud data, which processes the local geometry and graspablesurfaces of the objects. To speed up the processing time, theauthors proposed two new representation of grasp candidatesand trained the CNN with large online depth datasets obtainedfrom idealized CAD model. Their approach, however, failed toaddress the difficulty of distinguishing between two adjacentobjects as the algorithm avoids point cloud segmentation. As aresult, there is also no direct way to grasp a specific object ofinterest. If the problem of point cloud segmentation is solved,it means that a grasp pose is generated for a single standingobject, this approach is suitable with our work and provide areliable grasp for specific type of object.III. T HE P ROPOSED M ETHOD
The proposed approach included three stages as shown inFigure 1: • Object selection : In this stage, the robot system has toselect one object out of a multitude of different objectspresent within a given environment. The priority forobject selection can be specified by the users, in termsof the following criteria: proximity, object class, physicalproperties. If there are multiple objects in the same type,the system needs criteria to score them and grasp eachitem in a predefined order. This step uses the first CNNto detect, select then output the point cloud cluster of theselected object, which is sent to the second deep networkto generate the grasps. • Grasp Pose Selection : In this step, the input is the selectedobject’s point cloud data, which is used to estimate andgenerate grasp poses on the object. The second CNNgenerates a number of different poses, then the mostsuitable grasping pose is selected. This requires a filterapplied on candidate grasp poses, which is scored andthe highest scored pose is selected and given as outputto the next step. • Object Grasping and Sorting : A motion planner is usedto generate a trajectory that helps the robot’s gripper toreach the desired grasp pose for grasping target object.ig. 2: Flowchart of selecting an object to grasp.It is important for the generated robot trajectory tobe collision-free. The motion planner, therefore, needsa model of the environment so that it can check forcollisions when generating trajectories.
A. Object Selection
It is challenging to perform object detection directly usingpoint cloud data as input. Several visual cues such as colors orshapes, that are normally used to recognize an object can beaffected. To overcome the challenge, color images and pointcloud data are combined to make the decision. The steps forobject selection are illustrated in Figure 2. RGB images ofitems are given as input to an image detector to put a boundingbox around the selected object.YOLOv3 provides the performance metrics for object de-tection and selection, which is used to score each item in thesame class in case multiple items are available. The objectwith the highest score is given as output to the next process.The center of the bounding box s D (as shown in Figure 3)is calculated.Fig. 3: Objects detection by using YOLOv3 network.In another process, point cloud data of different objects isfiltered to remove noise, outliers, and reduce the amount ofdata by using the following: statistical filter, a voxel filter, anda working space filter. A workspace filter is used to removethe data points that do not belong to a predefined workspace.At this step, plane segmentation and target extraction is usedto separate the items as shown in Figure 4. Fig. 4: Corresponding clusters of objects in Figure 5 areseparated. The red, green, and blue bars are xyz axes of thecamera’s depth coordinate frame.Fig. 5: Four objects to be selected. The cup is chosen to begrasped.Also shown in Figure 4, the 3D centroid coordinates ofeach object c i D are calculated, where i is index for the objectcluster i . These centroids are then projected to get 2D points c i Dp in the image coordinate (using the camera calibrationparameters). At the last step, the Euclidean distances between s D and c i Dp are calculated, and the smallest distance isused to determine the corresponding cluster belonging to theselected object:ig. 6: Grasp generation & filtering on the selected item. • Calculate the distance d i for cluster i : d i = ( s D x − c i Dp x ) + ( s D y − c i Dp y ) . • Compare and pick the cluster with the smallest distance.All the coordinates in the formula are calculated from thecamera base frame.To illustrate the process, for instance, a cup is the targetobject to be grasped by the manipulator. The process todetermine the corresponding cluster is shown in Figure 5. Theyellow squares are c Dp points, and the red circle is s D ,which is calculated from the coordinates of the bounding boxof the cup returned by the trained YOLOv3 network. Thenearest yellow square to the red circle determines the thirdcluster from the left (as shown in Figure 4) belongs to theobject to be grasped (the cup). After that, the grasp detectionalgorithm can be performed on this cluster.In Figure 4, four set of green points corresponding to thepoint cloud data of four objects shown in Figure 5. The red,green, and blue bars denote x , y , and z axes of camera’s depthcoordinate frame. B. Grasp Generation and Filtering
The flowchart for grasp posture estimation and generationis illustrated in Figure 6. After having the cluster of thegrasped object, the algorithm for Grasp Pose Detection [16] isperformed. Each point in the cloud is associated with a singleviewpoint, from which the point is captured. The algorithmalso considers the geometric parameters of the robot gripperand a subset of points C G belonging to target objects.To have the subset of points that belong to objects ofinterest C G , those central points calculated earlier from thelabeled bounding boxes and spheres of points around them areused. With point cloud data and centroid points as inputs, thealgorithm samples points uniformly around C G . After that, thealgorithm calculates the surface normal and an axis of majorprincipal curvature of the object surface. Potential grasp posecandidates are generated at regular orientations orthogonalto the curvature axis. These grasp poses are then pushedforward until the fingers make contact with the point cloud.The grasps poses without any data points between the fingersare discarded. The remaining grasp poses are given as input toa four-layered CNN for pose classification between viable andnon-viable grasping poses. At this step, the algorithm is usedfor scoring grasp poses to pick the grasp pose with highestscore. The algorithm gives a high score for grasps that areat the upper part of the object (higher chance of successfulgrasps). It also considers the gripper orientations and poses ifthey are similar to the current pose of the gripper or pointstowards the robot position. This helps to minimize the robot’s movements. In Figure 7, the best pose to grasp is the red one.Fig. 7: Eight highest scored poses to grasp an object and thebest pose is in red. This posture points toward the object and issimilar to the current gripper’s pose. Moreover, the contactingpoint of grasp pose in on the upper part of the target object. C. Grasp Execution
To execute the chosen grasp on a real robot, there is a needto generate a collision-free trajectory, which transforms thecurrent gripper’s pose to the desired grasp pose. To checkfor collisions,
TrajOpt [7] needs an accurate simulated modelof the environment, which includes the robot model and thepoint cloud data of objects in the robot base coordinate. Theexecution of the chosen grasp was illustrated in Figure 8.Fig. 8: Flowchart of executing the chosen grasp.As all the objects that the camera sees are in the cameracoordinate, there is a need to calculate a transformationmatrix from this coordinate to the robot base coordinate. TheAruco tags [17] were used to calculate the transformationmatrix through the end-effector coordinate of the robot. Thesecoordinates are defined in Figure 9. • An Aruco tag is attached on the end-effector of therobot (the gripper is temporarily removed) so that thetransformation matrix between the tag’s coordinate C tag and the end-effector coordinate C e is known. • The transformation matrix between C tag and the cameracoordinate C cam can be computed using any Aruco tagsoftware package. We use aruco ros, and the detectionresult is shown in Figure 9. • The transformation matrix between the C e and C base isprovided by the forward kinematics of the robot.ig. 9: Coordinates: top - Aruco tag’s coordinate C tag in thecamera coordinate C cam ; middle - end-effector coordinate C e ,bottom - robot base coordinate C base . Color: X, Y, Z = red,green, blue.The homogeneous transformation matrix between C cam and C base is the multiplication of the above matrices: M basecam = M basee M etag M tagcam . Using M basecam , the point cloud data in C cam can be easily con-verted to C base . Using this matrix, the point cloud data fromthe camera’s coordinate can be transformed and displayedthem in the robot base coordinate using OpenRave [18] asshown in Figure 10.After transformation has been performed to C base thatincludes the robot, the objects, and the desired grasp pose;TrajOpt will generate a collision-free trajectory. The trajectoryis a series of 6-joint position tuples, which can then besuccessfully performed on the actual robot.IV. E XPERIMENT
A. Hardware
An AUBO-i5 robot from AuboRobotics 6-DOF with pay-load of 5 kg is used in the practical experimentation. The robotis controlled using a Python driver provided by the manufac-turer. A Robotiq 2-finger 85 adaptive gripper is mounted onthe arm. The gripper is position-controlled with positional andforce feedback and adjustable gripping force between 20N and Fig. 10: Simulated environment in OpenRave.Fig. 11: Hardware setup.235N. For the vision system, Asus Xtion Pro RGB-D camerahas been used. The camera outputs 640x480 RGB imagesalong with point cloud data. The camera is fixed at a positionon a table during the experiment. The whole setup is illustratedin Figure 11.
B. Camera Calibration
As grasping objects require accurate positional information,it is essential for the camera to be properly calibrated. Thereare two different calibration procedures for RGB and depthimages that have been performed. • RGB calibration: We use the procedure mentioned in thisROS tutorial . A checkerboard of size 8x6 is used in thisstep. • Depth calibration: The camera that we use is notorious forhaving depth error without calibration. We also sufferedthe same error of about 5 cm at a distance of 50 cm. http://wiki.ros.org/camera calibration/Tutorials/MonocularCalibration ig. 12: Mitigated depth error after a calibration.It can be seen in Figure 12, the tag from point clouddata is moved 5 cm in front of the actual position. Wefollow the instructions for calibrating the depth sensor byusing jsk pcl ros package . The idea is to align the depthestimation from RGB images with the depth calculatedfrom the depth sensor. We used the same size checker-board with a smaller grid in order to cover the betterour working space while performing the calibration. Aftercalibration, the error is reduced to 1 cm. C. Training YOLOv3
Object image data was collected using online images from 4categories: lotion bottles, deodorant bottles, cups, and cans. Tocome up with this list, we need to test the performance of GDPalgorithm on these objects to make sure that the algorithm cangenerate valid grasps poses. Some transparent objects are noteffectively detected by laser beams from the camera, due towhich, they were not used the experimentation. Additionally,the neural network used in GPD algorithm to generate graspposes was trained in certain objects, as a result, it might notgeneralize with novel objects. A total of 800 images for eachclass were used for training and validation. After that, thedata was manually labelled and fine-tuning was performedwith the instruction for training with custom objects from aGitHub repository . For the selected object dataset, 80% ofthe total images for training and 20% of the images datasetfor testing. The default configuration of weights and hyper-parameters has been chosen for training of object classificationmodel (from darknet53.conv.74). The final performance of thetrained object classification model is 84.39% mAP and 0.5Intersection over Union (IOU). An example of detected objectsis shown in Figure 3. The trained network is used to returnthe coordinates of rectangles to cover the detected objects.From these bounding boxes, the centroids for target objectsare calculated, as discussed in the previous section. D. Results
The trajectory generated by
TrajOpt , given the transformedpoint cloud data, the robot model, and the desired grasp pose https://jsk-recognition.readthedocs.io/en/latest/jsk pcl ros/calibration.html https://github.com/AlexeyAB/darknet ONCLUSIONS AND F UTURE W ORK
This paper presented the development of an autonomousrobotic manipulator for sorting application. Two state-of-the-art Deep Conventional Neural Networks were used to pro-cess RGB-D data for both object detection and robust grasppose generation. Combining with
Trajopt motion planning, itformed a viable solution for an autonomous sorting robot.Moreover, this paper also discussed the design of a graspfiltering, which works as an interface between the existinggrasp pose detection algorithm and the variance in sortingrobot system. This can ensure that the different algorithmscan work robustly with our gripper’s configuration.The proposed system is validated by an experiment utilizingthe detection, grasping and sorting of different object types.The experiment results on various objects show that ourproposed combination of deep learning-based object detectionmodel, grasp detection & filtering, and the manipulator controlmethod is able to provide robust and efficient object graspingand sorting of different objects. The proposed approach can beadapted to different types of manipulators, gripper mechanismsand robots.There are a few drawbacks to our research, which can beimproved in the future. The first problem is the processing timedue to large amount of data being used by motion planningand Grasp Pose Detection algorithms. Future research shouldfocus towards improving the data representation to reducethe processing data. Another potential improvement would betowards development of grasp pose filter, which is able towork efficiently with relative position change between robotjoints and camera. Moreover, combining more cameras willprovide better point cloud data to represent the object, whichin turn will improve the Grasp Pose detection in clutteredenvironment.We also plan to extend this work to multi-manipulator col-laboration in which both collaborative and distributed control a) Robot & scene loaded. (b) Planning is started. (c) Building trajectories. (d) Path planning finished.
Fig. 13: Planning for one grasp pose by OpenRave. First (a), the robot model, the point cloud data, and the desired grasp poseof the scene are loaded. The robot coordinate and the desired grasp pose are plotted with X axis (red), Y axis (green), Z axis(blue). Next (b), the point cloud data is simplified by approximated meshes for collision checking, starting the path planning.(c) Trajectories is then built. (d) The trajectory (red) is generated successfully without any collision.Fig. 14: Generated trajectory replayed on the robot. Thetrajectory consists of a set of 6-joint positions, which is thenreplayed on the real robot at 10Hz for safety. This frequencycan be up to several hundred Hz as specified by the robotmanual.[19]–[23] and deep reinforcement learning [12], [24]–[26]will be investigated to allow multiple manipulators to worktogether efficiently while avoiding collision. The multi-agentcooperative control and sensing research in our previous work[27]–[39] will be utilized.R EFERENCES[1] M. Gupta and G. S. Sukhatme, “Using manipulation primitives for bricksorting in clutter,” in . IEEE, 2012, pp. 3883–3889.[2] A. Zeng, S. Song, K.-T. Yu, E. Donlon, F. R. Hogan, M. Bauza, D. Ma,O. Taylor, M. Liu, E. Romo et al. , “Robotic pick-and-place of novelobjects in clutter with multi-affordance grasping and cross-domain imagematching,” in . IEEE, 2018, pp. 1–8.[3] J. Gu´erin, S. Thiery, E. Nyiri, and O. Gibaru, “Unsupervised roboticsorting: Towards autonomous decision making robots,” arXiv preprintarXiv:1804.04572 , 2018.[4] J. Mahler, M. Matl, V. Satish, M. Danielczuk, B. DeRose, S. McKinley,and K. Goldberg, “Learning ambidextrous robot grasping policies,”
Science Robotics , vol. 4, no. 26, p. eaau4984, 2019.[5] L. Jin, S. Li, H. M. La, and X. Luo, “Manipulability optimization ofredundant manipulators using dynamic neural networks,”
IEEE Trans-actions on Industrial Electronics , vol. 64, no. 6, pp. 4710–4720, June2017. [6] H. Nguyen and H. La, “Review of deep reinforcement learning for robotmanipulation,” in , Feb 2019, pp. 590–595.[7] J. Schulman, J. Ho, A. X. Lee, I. Awwal, H. Bradlow, and P. Abbeel,“Finding locally optimal, collision-free trajectories with sequential con-vex optimization.” in
Robotics: science and systems , vol. 9, no. 1.Citeseer, 2013, pp. 1–10.[8] A. Sehgal, H. La, S. Louis, and H. Nguyen, “Deep reinforcementlearning using genetic algorithm for parameter optimization,” in , Feb2019, pp. 596–601.[9] J. Redmon and A. Farhadi, “Yolov3: An incremental improvement,” arXiv preprint arXiv:1804.02767 , 2018.[10] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y. Fu, and A. C.Berg, “Ssd: Single shot multibox detector,” in
European conference oncomputer vision . Springer, 2016, pp. 21–37.[11] K. Simonyan and A. Zisserman, “Very deep convolutional networks forlarge-scale image recognition,” arXiv preprint arXiv:1409.1556 , 2014.[12] H. Nguyen, H. M. La, and M. Deans, “Hindsight experience replay withexperience ranking,” in ,Aug 2019, pp. 1–6.[13] J. Redmon and A. Angelova, “Real-time grasp detection using convo-lutional neural networks,” in . IEEE, 2015, pp. 1316–1322.[14] I. Lenz, H. Lee, and A. Saxena, “Deep learning for detecting roboticgrasps,”
The International Journal of Robotics Research , vol. 34, no.4-5, pp. 705–724, 2015.[15] A. ten Pas and R. Platt, “Using geometry to detect grasp poses in 3dpoint clouds,” in
Robotics Research . Springer, 2018, pp. 307–324.[16] M. Gualtieri, A. ten Pas, K. Saenko, and R. Platt, “High precisiongrasp pose detection in dense clutter,” in
Intelligent Robots and Systems(IROS), 2016 IEEE/RSJ International Conference on . IEEE, 2016, pp.598–605.[17] S. Garrido-Jurado, R. Mu˜noz-Salinas, F. J. Madrid-Cuevas, and M. J.Mar´ın-Jim´enez, “Automatic generation and detection of highly reliablefiducial markers under occlusion,”
Pattern Recognition , vol. 47, no. 6,pp. 2280–2292, 2014.[18] R. Diankov and J. Kuffner, “Openrave: A planning architecture forautonomous robotics,”
Robotics Institute, Pittsburgh, PA, Tech. Rep.CMU-RI-TR-08-34 , vol. 79, 2008.[19] H. X. Pham, H. M. La, D. Feil-Seifer, and M. Deans, “A distributedcontrol framework for a team of unmanned aerial vehicles for dynamicwildfire tracking,” in , Sep. 2017, pp. 6648–6653.[20] A. D. Dang, H. M. La, and J. Horn, “Distributed formation controlfor autonomous robots following desired shapes in noisy environment,”in , Sep. 2016, pp. 285–290.[21] H. M. La, W. Sheng, and J. Chen, “Cooperative and active sensing inmobile sensor networks for scalar field mapping,”
IEEE Transactionson Systems, Man, and Cybernetics: Systems , vol. 45, no. 1, pp. 1–12,Jan 2015.22] H. M. La and W. Sheng, “Distributed sensor fusion for scalar field map-ping using mobile sensor networks,”
IEEE Transactions on Cybernetics ,vol. 43, no. 2, pp. 766–778, April 2013.[23] H. M. La and W. Sheng, “Dynamic target tracking and observingin a mobile sensor network,”
Robotics and Autonomous Systems , Aug 2018, pp. 1–6.[25] M. Rahimi, S. Gibb, Y. Shen, and H. M. La, “A comparison ofvarious approaches to reinforcement learning algorithms for multi-robotbox pushing,” in
Advances in Engineering Research and Application ,H. Fujita, D. C. Nguyen, N. P. Vu, T. L. Banh, and H. H. Puta, Eds.Cham: Springer International Publishing, 2019, pp. 16–30.[26] H. M. La, R. Lim, and W. Sheng, “Multirobot cooperative learning forpredator avoidance,”
IEEE Transactions on Control Systems Technology ,vol. 23, no. 1, pp. 52–63, Jan 2015.[27] A. D. Dang, H. M. La, T. Nguyen, and J. Horn, “Formation controlfor autonomous robots with collision and obstacle avoidance using arotational and repulsive force based approach,”
International Journalof Advanced Robotic Systems , vol. 16, no. 3, p. 1729881419847897,2019. [Online]. Available: https://doi.org/10.1177/1729881419847897[28] L. Jin, S. Li, H. M. La, X. Zhang, and B. Hu, “Dynamic task allocationin multi-robot coordination for moving target tracking: A distributedapproach,”
Automatica
International Journal of Advanced Robotic Systems , vol. 15,no. 3, p. 1729881418773874, 2018. [Online]. Available: https://doi.org/10.1177/1729881418773874[30] M. T. Nguyen, H. M. La, and K. A. Teague, “Collaborative andcompressed mobile sensing for data collection in distributed roboticnetworks,”
IEEE Transactions on Control of Network Systems , vol. 5,no. 4, pp. 1729–1740, Dec 2018.[31] H. X. Pham, H. M. La, D. Feil-Seifer, and M. C. Deans, “A distributedcontrol framework of multiple unmanned aerial vehicles for dynamicwildfire tracking,”
IEEE Transactions on Systems, Man, and Cybernet-ics: Systems , pp. 1–12, 2018.[32] T. Nguyen, H. M. La, T. D. Le, and M. Jafari, “Formation controland obstacle avoidance of multiple rectangular agents with limitedcommunication ranges,”
IEEE Transactions on Control of NetworkSystems , vol. 4, no. 4, pp. 680–691, Dec 2017.[33] T. Nguyen and H. M. La, “Distributed formation control of nonholo-nomic mobile robots by bounded feedback in the presence of obstacles,”in , July 2017, pp. 206–211.[34] F. Muoz, E. S. EspinozaQuesada, H. M. La, S. Salazar, S. Commuri,and L. R. GarciaCarrillo, “Adaptive consensus algorithms forreal-time operation of multi-agent systems affected by switchingnetwork events,”
International Journal of Robust and NonlinearControl , vol. 27, no. 9, pp. 1566–1588, 2017. [Online]. Available:https://onlinelibrary.wiley.com/doi/abs/10.1002/rnc.3687[35] T. Nguyen, T. Han, and H. M. La, “Distributed flocking control of mobilerobots by bounded feedback,” in , Sep. 2016, pp.563–568.[36] H. M. La and W. Sheng, “Flocking control of multiple agents in noisyenvironments,” in , May 2010, pp. 4964–4969.[37] ——, “Flocking control of a mobile sensor network to track and observea moving target,” in , May 2009, pp. 3129–3134.[38] ——, “Adaptive flocking control for dynamic target tracking in mobilesensor networks,” in , Oct 2009, pp. 4843–4848.[39] H. M. La and Weihua Sheng, “Moving targets tracking and observingin a distributed mobile sensor network,” in2009 American ControlConference