Abstract

1 Introduction

2 Theoretical Background

3 Related Works

4 Methodology

5 Experiments and Results

6 Conclusions

Declarations

References

Abstract

     RoboCup 3D Soccer Simulation is a robot soccer competition based on a high-fidelity simulator with autonomous humanoid agents, making it an interesting testbed for robotics and artificial intelligence. Due to the recent success of Deep Reinforcement Learning (DRL) in continuous control tasks, many teams have been using this technique to develop motions in Soccer 3D. This article focuses on learning humanoid robot behaviors: completing a racing track as fast as possible and dribbling against a single opponent. Our approach uses a hierarchical controller where a model-free policy learns to interact model-based walking algorithm. Then, we use DRL algorithms for an agent to learn how to perform these behaviors. Finally, the learned dribble policy was evaluated in the Soccer 3D environment. Simulated experiments show that the DRL agent wins against the hand-coded behavior used by the ITAndroids robotics team in 68.2% of dribble attempts.

Introduction

     RoboCup is an international academic competition created to foster robotics and artificial intelligence research [27]. It has an ambitious long-term goal of having a team of humanoid robots beating the human soccer World Cup champions by 2050. There are many leagues with different game rules and constraints on robot designs to accelerate progress towards this objective.

RoboCup 3D Soccer Simulation (Soccer 3D) is a league of RoboCup based on a robot soccer simulator with a high-fidelity simulation model of the Nao humanoid robot. The particular contributions to RoboCup reside in being a research environment for high-level multi-agent cooperative decision-making, and humanoid robot control [44]. A simulation environment is convenient for machine learning algorithms due to their need for large amounts of data [36]. Dealing with real robots is time-consuming due to the need to recharge batteries or reallocate robots manually to set up experiments. Moreover, experience collection may be largely accelerated by running many simulations in parallel and executing in faster than real-time. Unfortunately, transferring behaviors learned in simulation to real robots is challenging due to the so-called reality gap. Still, some works have succeeded in doing so, usually by executing a final fine-tuning process on the real robot [36].

Conclusions

     Our main objective was to learn high-level soccer behaviors using reinforcement learning in this work. We addressed the problem with state-of-the-art model-free deep reinforcement learning algorithms, namely DDPG, TRPO, and PPO. Therefore, we learned behaviors while dealing with the complex dynamics of a humanoid robot.

     To facilitate, we used a hierarchical approach where the agent learns to command a model-based walking engine based on the Zero Moment Point (ZMP) concept. The walking engine receives the desired velocities in forward, lateral and rotational directions and outputs the joint angles.

     We developed a DRL framework for integrating DRL algorithms with the RoboCup 3D Soccer Simulation environment to accomplish our objective. In our results, PPO achieved the best performance, which was expected, and effectively learned humanoid robot behaviors.

Advanced composite structural components in aeronautics are characterized by very high production costs because of their dimensions, complex shapes and expensive forming equipment. For these components, such as horizontal stabilizers and wings, a defect occurrence is often critical because large part of inner surfaces, made of long and tapered narrow tunnels, are not reachable for repair operations. In these cases, the part is rejected with a relevant economic loss and high production costs. For this reason, aircraft constructors plan huge investments for defects avoidance during the forming processes of CFRP and to develop effective, robust and reliable repair tools and methods. Mobile robotics can play an important role, with specific systems capable of moving into narrow channels of wings structures (i.e. multi spar boxes) and repair it in accordance to technical standards. This paper describes an innovative mobile robot architecture for bonded repair scarfing operations on CFRP components. Targeting and responding to the demanding machining requirements, the functional-oriented design approach clearly highlights the advantages of a modular robotic solution. The mobile robotic architecture can be also applied in other fields with similar challenging manufacturing operations for further inspection, detection and machining operations.

Introduction

Working in narrow space is a typical issue for the repair process of aeronautical structural components. Carbon Fiber Reinforced Plastic (CFRP) components, as wings and horizontal stabilizers, are often realized in the form of multi spar boxes assemblies. A typical example is the Boeing 787 Dreamliner horizontal stabilizer (tailplane) which has a multi-spar boxes structure and its cross-section is composed by a plurality of longitudinal tapered narrow tunnels. During the manufacturing process of these components some defects may occur. In order to reduce the high cost of discards, the necessity for an efficient and robust repair process arises. In literature, different technological processes and techniques for CFRP aircraft structures repair are available [1]. This paper deals with the application of such techniques in narrow spaces, as the multi spar boxes structures. Currently, the CFRP repair techniques are typically performed manually [2] and sometimes the damage is not accessible by the operator. In large CFRP multi-spar boxes structural components, indeed, human operators are able to repair internal damages only if localized in proximity of the wing tip or root. In all the other cases there are no fix strategies currently available and the components, often very expensive, must be discarded [3]. Thus, the position of the damage is crucial, as it causes a large number of discards in aeronautics manufacturing, with consequent increase of the production costs. Due to the lack of solutions of applying fixing techniques in narrow multi spar boxes, in this paper a strategy for the automation of such a process is described. Mobile robotics is identified as the key enabling technology that makes it possible. As a further advantage, some studies, remarked the importance of automated machining of CFRP and the advantages in terms of accuracy, quality and reliability compared with a manual operation [4]. Nevertheless, there are no available solutions for the repair of internal defects. The proposed approaches, indeed, consider either new specific devices to be manually fixed in proximity of the damage [5,6], or the use of conventional robots, too big for this purpose [7]. In [8] a robotic inspection cell is presented, composed of three industrial manipulators performing both, photogrammetry acquisitions and ultrasound-based NDT inspection. The robots can move along linear guides and the inspection of large components is enabled; however, there is no possibility for internal defects detection.

1- Introduction

2- Automobile Internet of Things

3- IoT Requirements for Robotic First Aid System

4- Robotic First Aid System IoT Architecture

5- Conclusion

References

Abstract

Health emergencies occur in passenger cars where victims do not have immediate access to either layperson or professional, proper medical services, resulting in deterioration of their health or death. Installation of a robotic first aid system for passenger car occupants has been proposed. This research is part of a larger work of designing the system and identifies the existing hardware and software automotive infrastructure that the robotic first aid system can functionally use. These resources are reviewed in categories of sensing, connectivity, data analysis and information exchange, which form the internet of things. Consequently, the system’s design for internet of things mechanical requirements are identified in sensing. Finally, a preliminary framework for the system’s internet of things infrastructure is established to facilitate the development of a smart robot.

Introduction

We live in an invisible jungle of communication where millions of devices interact with each other at any one time, their signals travelling at colossal speed through land, air and even water – all because of the internet. As envisioned by its pioneers, the internet today connects computers in automobiles, homes, industrial and business spaces, and wearable devices, forming the Internet of Things (IoT). The past decade has seen rapid development of IoT systems because gadgets operate more efficiently, effectively and at lower costs towards a common goal when connected to share information and facilities, both hardware and software [1], [2]. One of the most significant research areas in the field of IoT is human healthcare. Vehicle safety is a top priority for automakers, consumers, governments and other parties concerned with road transport safety [3]–[5]. Despite their efforts, several studies report that morbidity and mortality rates resulting from road-traffic accidents (RTAs) are on the rise in some regions of the world, particularly low- and middle-income countries (LICs and MICs), and road traffic injuries are a leading cause of preventable death [6]–[9]. Additionally, vehicle occupants are susceptible to non-traumatic medical emergencies including cardiovascular and respiratory complications [10]–[12]. It is well documented that quick and correct medical assistance is the key to survival of any major medical emergency, such as an injury, a stroke or a heart attack [13], [14]. First aid given to victims of RTAs is known to preserve life, prevent further harm y and at lower costs towards a common goal when connected to share information and facilities, both hardware and software [1], [2]. One of the most significant research areas in the field of IoT is human healthcare. and promote recovery in most cases [15]–[18]. Unfortunately, a number of studies have found that there are high chances of victims not getting proper help for reasons including unavailability of bystanders, responders fearful of worsening the condition and entrapment of victims in the wreckage, thus inaccessible [16], [19]– [25]. Evidence suggests that professional emergency medical service (EMS) post-crash response also influences the recovery of patients [26], [27], and the underlying assumption is that EMS response times should be reduced continuously. A more recent report confirms that if trauma care systems for severely injured RTA victims in LICs and MICs countries were improved to match those in high-income countries (HICs), half a million lives could be saved each year [28]. However, it has been shown that even for areas with effective EMSs, the greatest delays in responding to emergencies come not from EMS response time but from the reaction of patients and bystanders [29]. For such scenarios, it has been suggested that policy concentrates on public first aid education – the same public with a low response rate as mentioned earlier!

1- Introduction

2- Camera, image and color models

3- Image processing techniques

4- Application 1: Wheel rim detection

5- Application 2: Automatic 3-D crankshaft verification system

6- Application 3: Wheel alignment measuring system

7- Application 4: Stereo visual odometry for mobile robotics

8- Conclusion

References

Abstract

Modern applications of industrial automation and robotics are increasingly relying on image processing techniques. This paper shows ways in which image processing can be applied to solve actual problems in robotics and instrumentation. The paper starts by presenting the fundamentals of camera models, digital acquisition of images and selected processing techniques, followed by examples of applications of such knowledge. Four examples of image processing applications are shown: rim detection in automotive wheel images, dimensional verification of crankshafts, measurement of wheel alignment angles of a car and a stereo visual odometry algorithm for mobile robotics. The examples not only illustrate the uses of different image processing techniques, but may also inspire the development of new robotic and industrial automation products.

Introduction

The engineering community is experiencing a dramatic growth in application of image processing as a tool to perform non-intrusive precision measurement, autonomous robot navigation or reliable verification of industrial automation processes. Image acquisition and processing techniques are familiar to physics, computer, electrical and mechanical engineering sciences for a relatively long time. Specialized literature on this field is usually dedicated to specific aspects such as the geometry of camera projections and computer vision [1–5], image acquisition and processing [6], or either the scientific and industrial applications of image processing and analysis [7]. This paper is both a review and a tutorial of applied image processing in industrial automation, aimed at researchers, postgraduate students and engineers. Digital images can be associated with geometric properties of objects, such as shapes and dimensions, which enables a myriad of applications and developments. In order to exploit such possibilities, the paper provides insight into some of the key techniques associated with image acquisition, processing and analysis. Theoretical aspects on the geometry of perspective projection, camera calibration, epipolar geometry and stereo image correspondence are further elaborated with regards to technical details. Such a mathematical framework is the core for the use of cameras as measuring devices in photogrammetry. Four applications of image processing are shown through examples in actual robotics and instrumentation scenarios: rim detection in automotive wheel images; dimensional verification of crankshafts; measurement of the wheel alignment angles of a car, and a stereo visual odometry algorithm for mobile robotics. The examples are innovative concepts that can inspire a wider range of image-related applications. Each example considers four aspects: (1) a motivation; (2) an explicit description of the system setup and calibration; (3) mathematical rationale and references to first principles that explicit the assumptions behind each application; (4) empirical results and discussion.

In some existing omnidirectional mobile robots, the movement of each wheel is generally controlled by one driven motor cooperates with one steering motor, such design will increase the energy consumption. This paper investigates an omnidirectional mobile robot based on a special coupling system, so as to solve the above problem. We propose a unique robot’s steering method by designing the parallelogram connecting rod as the coupling system, which makes each wheel only need to be equipped with one motor to complete omnidirectional motion and always keep unchanged robot’s attitude. In theory, we conduct the structural analysis, dynamic analysis and kinematic analysis of the mobile robot and explain why our system is under-actuated. In experiments, the square trajectory and the serpentine trajectory are exploited to verify the rationality of design and theoretical derivation by comparing the simulation results with the experimental results. By performing mobile robot walking experiments under different ground condition respectively, the stability and adaption of the under-actuated omnidirectional mobile robot is also favourably confirmed. After detailed theoretical analysis and the support of experiments, it is carried out that the under-actuated mechanism proposed in this paper can indeed realize the movement of the robot in all directions. As well as the attitude of the robot always keeps constant, as we expected, enables to realize the steering of robot flexibly even in a narrow space.

Introduction

In recent years, with the rise of artificial intelligence technology, the robots have been applied to several fields. Such as the climbing robots for detecting tasks [1], [2], mobile robots that can be used in home and industrial environments [3]. Particularly, mobile robot as an important branch of robotics has received a great attention by most countries [4]. The earliest appeared is the wheeled robot developed by Stanford, which is primarily capable of self-service reasoning, planning and control in complex environments [5]. As time goes by, the relevant researches on mobile robots are also deepening with the assistance of advanced technology. A mass of omnidirectional mobile robots have been applied widely due to the emergence of omnidirectional wheel technology. In order to meet the use of mobile robots in different scenarios, the choice of wheel mechanism is particularly important [3]. Among them, the design scheme of Mecanum wheel is proposed in 1973 and is subsequently exploited [6], [7]. In [8], presents an omni-directional mobile robot using four custom-made mecanum wheels, each wheel is independently powered using a precisian gear DC motors. The KUKA company in Germany takes full advantage of its own robot development and launches the YouBot robot into people’s view [9], [10]. Subsequently, KUKA introduces the omniMove large mobile robot platform, which has flexible mobility and powerful load capacity despite its bulk [11].

I- Introduction

II- Proposed Method

III- Experiment

IV- Results

V- Discussion and Conclusion

References

Abstract

Robot-assisted surgery is an emerging technology that has undergone rapid growth with the development of robotics and imaging systems. Innovations in vision, haptics, and accurate movements of robot arms have enabled surgeons to perform precise minimally invasive surgeries. Real-time semantic segmentation of the robotic instruments and tissues is a crucial step in robot-assisted surgery. Accurate and efficient segmentation of the surgical scene not only aids in the identification and tracking of instruments but also provides contextual information about the different tissues and instruments being operated with. For this purpose, we have developed a light-weight cascaded convolutional neural network to segment the surgical instruments from high-resolution videos obtained from a commercial robotic system. We propose a multi-resolution feature fusion module to fuse the feature maps of different dimensions and channels from the auxiliary and main branch. We also introduce a novel way of combining auxiliary loss and adversarial loss to regularize the segmentation model. Auxiliary loss helps the model to learn low-resolution features, and adversarial loss improves the segmentation prediction by learning higher order structural information. The model also consists of a light-weight spatial pyramid pooling unit to aggregate rich contextual information in the intermediate stage. We show that our model surpasses existing algorithms for pixelwise segmentation of surgical instruments in both prediction accuracy and segmentation time of high-resolution videos.

INTRODUCTION

ROBOT-ASSISTED minimally invasive surgery (RMIS) has revolutionized the practice of surgery by optimizing surgical procedures, improving dexterous manipulations and enhancing patient safety [1]. Recent developments in the field of robotics, vision and smaller instruments have impacts on minimally invasive intervention. The common extensively used surgical robotic system is the Da Vinci Xi robot [2]–[5] enable remote control laparoscopic surgery with long kinematic chains. The Raven II [6] is a robust surgical system consists of spherical positioning mechanisms. Remarkable recent surgical tools with complex actuation systems utilized micro-machined super-elastic tool [7] and concentric tubes [8]. However, with the reduction in size and complex actuation mechanisms, control of the instruments and cognitive representation of the robot kinematics are forthwith remarkably challenging in a surgical scenario. In addition, there are factors that complicate the surgical environment such as shadows and specular reflections, partial occlusion, smoke, and body fluid as well as the dynamic nature of background tissues. Hence, real-time surgical instruments detection, tracking, and isolation [9]–[12] from tissue are the key focus in the field of RMIS.

Recently, MANET (Mobile Ad Hoc Network) researchers have shown increased interest in using mobile robot technology for their testbed platforms. Despite the existence of review papers that discuss the usage of mobile robot technology pertaining to a MANET testbed from the perspective of a MANET researcher, said ndings are rather lacklustre as it is not the sole purpose of said reviews. Hence, this review aims to comprehensively discuss and analyse MANET testbeds that were facilitated with mobile robot technology for previous undertaken research. To enable readers to keep abreast with mobile robot technology used in previous research, whilst presenting the advantages and disadvantages of said methods, this review will rst supercially discuss prior robot based MANET testbed facilities, before presenting technical analysis overview and critical analysis. Additionally, suggestions to heighten mobility mechanisms by using mobile robots to be more practical, easy and inexpensive are also included in this paper. The technical and critical content of this review is expected to be a source of reference for other MANET researchers interested in the most suitable mobile robots to ensure real mobility in their MANET testbeds.

Introduction

Most research that involves mobile ad-hoc network (MANET) evaluation utilise simulation methods. Recently, it has become increasingly clear that the current practice of utilizing network simulators could only provide summarized assumptions in modeling the characteristics of the real systems. As the researcher needs to observe the eect and inuence of the MAC and the physical network layer due to mobility and topology changes against the links and communication quality of the MANET, a MANET testbed with real mobility is therefore vital. In terms of accuracy, the results obtained from this approach are not available using methods such as network simulation and emulation. As of now, only a few experiments have been conducted in the MANET testbeds as compared to simulation based MANET experiments [121]. There are various methods of mobility implementation that are used in MANET testbeds, which can be divided into two main types, namely real mobility and emulated mobility (sometimes called virtual mobility). Both methods have their own advantages and disadvantages. The distinctive feature of emulated mobility is its non-physical node mobility. In other words, although the real implementation of the MANET testbed is performed on the data-link layer and the application layer above, the node mobility is not conducted physically. Physical node movement and topology changes are carried out through the use of emulation [22]. There are several dierent methods of emulated mobility such as instance migration [23], on connections [22], RF matrix switches [7, 2427] and the use of virtual machine technology and virtual networks [28]. Among the wireless testbed platforms that have used emulated mobility methods Open-Access Research Testbed for Next-Generation Wireless Networks (ORBIT)3 [23], Ad-hoc Protocol Evaluation (APE) testbed with emulated mobility4 [2931], the Carnegie Mellon University Wireless Emulator (CMUWE)5 [3240], Castadiva6 [4143], Mobi-emu7 [44], Emulab8 [45, 46], MOBNET [47, 48], MobiNet [49], the Resilience Evaluation Framework for Ad Hoc Networks (REFRAHN) [50, 51], MeshTest9 [7, 2427] and WISEBED10 [52]. Emulated mobility methods have the advantage of being repeatable and reproducible that are almost the same as the network simulators. Mobility mechanisms that are conducted are therefore more manageable and predictable. However, emulated mobility methods are unable to represent the actual MAC layer and physical layer and hence the result obtained using emulated mobility are less accurate than the use of real mobility [22, 44]. Real mobility in MANET testbed experiments is vital in order to obtain accurate and realistic result. However, implementing real mobility in the testbeds is a dicult task. If this is done correctly, real mobility is able to provide very accurate experimental result on the impact of mobility on MANETs [1, 10, 53]. Previous research on MANETs have reported the use of various approaches to provide real mobility mechanisms in their MANET testbeds. Some of the methods included the use of cars [5459], taxis [60], trains [61], [52, 6266], bicycles [67, 68], humans [1, 812, 6981], remote control cars [82], and multiple mobile robots (to be discussed in the next section).

Introduction

According to the findings obtained in the context of a Global Initiative on Neurology and Public Health carried out by the World Health Organization (WHO), many of the neurological disorders are chronic and progressive, constitute a global public health problem [1], and affect especially the elderly people. In addition, a higher life expectancy makes the population of people over 60 increasingly higher [2]. The main patient groups served by the rehabilitation service in the United Kingdom are for neurological pathologies, as a survey reported [3]. 70% of respondents provided neurological rehabilitation services for people with stroke, multiple sclerosis, traumatic brain injury, degenerative neurological diseases, and other neuromuscular conditions. Other services that were represented were those that provided rehabilitation to people with severe single-incident brain injury (10%), spinal injury (9%), amputees (5%), musculoskeletal disability (4%), learning disabilities (1%), and pain (1%). In Spain, a similar situation is detected where musculoskeletal and articular disability (50%), neurological diseases (15%), traumatic injuries (29%), and others (6%) were treated in the rehabilitation services [4]. This situation, together with the need for rehabilitation and assistance for people with disabilities, means that robotic care and rehabilitation may play an important role in the years ahead. Nowadays, research on the use of robotic systems in different fields related to healthcare is widespread [5–7]. In the field of rehabilitation, scientific literature shows various classifications of such systems according to their level of interaction [8], the extremities that are treated [9–12], the modularity of the rehabilitation robots [13, 14], control strategies [15, 16], and the effectiveness of treatment [17–20]. However, no analysis has been done of the rehabilitation process as such, and the contribution of robotics in the different stages of the rehabilitation cycle or process has not been studied. In this paper, a systematic literature review is conducted to identify the contribution of robotics for upper limb neurorehabilitation highlighting its relation with the rehabilitation cycle and to clarify the prospective research directions in the development of an autonomous rehabilitation process.

Mirror therapy has been used in rehabilitation for multiple indications since the 1990s. Current evidence supports some of these indications, particularly for cerebrovascular accidents in adults and cerebral palsy in children. Since 2000s, computerized or robotic mirror therapy has been developed and marketed. Objectives. To map the extent, nature, and rationale of research activity in robotic or computerized mirror therapy and the type of evidence available for any indication. To investigate the relevance of conducting a systematic review and meta-analysis on these therapies. Method. Systematic scoping review. Searches were conducted (up to May 2018) in the Cochrane Library, Google Scholar, IEEE Xplore, Medline, Physiotherapy Evidence Database, and PsycINFO databases. References from identifed studies were examined. Results. In sum, 75 articles met the inclusion criteria. Most studies were publicly funded (57% of studies; n = 43), without disclosure of confict of interest (59% of studies; n = 44). Te main outcomes assessed were pain, satisfaction on the device, and body function and activity, mainly for stroke and amputees patients and healthy participants. Most design studies were case reports (67% of studies; n = 50), with only 12 randomized controlled trials with 5 comparing standard mirror therapy versus virtual mirror therapy, 5 comparing second-generation mirror therapy versus conventional rehabilitation, and 2 comparing other interventions. Conclusion. Much of the research on second-generation mirror therapy is of very low quality. Evidence-based rationale to conduct such studies is missing. It is not relevant to recommend investment by rehabilitation professionals and institutions in such devices.

Introduction

Mirror therapy was originally described by Ramachandran and Rogers-Ramachandran, who suggested its use in amputees with phantom limb pain [1]. Tey introduce an inexpensive new device: a mirror was placed vertically on a table so that the mirror refection of the patient’s intact hand was superimposed on the felt position of the phantom [1]. Tis standard mirror therapy has been used in rehabilitation for multiple indications since the 1990s [2]. A good level of evidence supports some of these indications, particularly for cerebrovascular accidents in adults [3, 4] and cerebral palsy in children [5]. Cost is very low, because a simple little and not specifcally dedicated mirror can be used [1]. Dedicated mirror boxes cost about $65 each [6]. Since 2000s, virtual reality or robot has been developed and marketed to treat various diseases as a more technologically sophisticated version of the standard mirror therapy introduced in 1996 [7, 8]. Robotic devices and virtual reality are increasingly used and assessed in rehabilitation and research [9, 10]. Tis second-generation devices are probably much more expensive than standard mirror therapy: they ofen present a technological complexity that requires investment, constant maintenance, and highly qualifed operators [11]. Low cost virtual reality device costs about $252 to purchase [12]. Low cost robotic device for robotic gait rehabilitation was estimated to cost $25,000, which is less than 10% of the price of device currently available in Brazil for the same indication [13]. For some indications, virtual reality such as robotics has no greater efectiveness than more conventional techniques [14, 15]. Studies evaluating the impact on various outcomes of these mirror therapy devices exist [8, 16, 17], but no review summarizes the available data. Te purpose of this review was as follows: (1) to map the extent, nature, and rationale of research activity in robotic or computerized mirror therapy; (2) to summarize the main sources and types of evidence available about the efectiveness of these therapies for any indication; (3) to investigate the relevance of conducting a systematic review and metaanalysis on these therapies.

Stroke, caused by death of brain cells as a result of blockage of a blood vessel supplying the brain (ischemic stroke) or bleeding into or around the brain (hemorrhagic stroke), is a serious medical emergency [1]. Stroke can result in death or substantial neural damage and is a principal contributor to long-term disabilities [1, 2]. According to the World Health Organization estimates, 15 million people suffer stroke worldwide each year [3]. Although technology advances in health care, the incidence of stroke is expected to rise over the next decades [4]. The expense on both caring and rehabilitation is enormous which reaches $34 billion per year in the US [5]. More than half of stroke survivors experience some level of lasting hemiparesis or hemiplegia resulting from the damage to neural tissues. These patients are not able to perform daily activities independently and thus have to rely on human assistance for basic activities of daily living (ADL) like feeding, self-care, and mobility [6]. The human hands are very complex and versatile. Researches show that the relationship between the distal upper limb (i.e., hand) function and the ability to perform ADL is stronger than the other limbs [7–9]. The deficit in hand function would seriously impact the quality of patients’ life, which means more demand is needed on the hand motor recovery. However, although most patients get reasonable motor recovery of proximal upper extremity according to relevant research findings, recovery at distal upper extremity has been limited due to low effectivity [10]. There are two main reasons for challenges facing the recovery of the hand. First, in movement, the hand has more than 20 degree of freedom (DOF) which makes it flexible, thus being difficult for therapist or training devices to meet the needs of satiety and varied movements [11]. Second, in function, the area of cortex in correspondence with the hand is much larger than the other motor cortex, which means a considerable amount of flexibility in generating a variety of hand postures and in the control of the individual joints of the hand. However, to date, most researches have focused on the contrary, lacking of individuation in finger movements [12, 13]. Better rehabilitation therapies are desperately needed.

In recent decades, robotics has attracted more and more attention from researchers since they have been widely used in scientific researches and engineering applications, such as space exploration, under water survey, industrial and military industries, welding, painting and assembly, and medical applications, and so on [1,2]. Much effort has been contributed to robotics, and different types of robot manipulators have thus been developed and investigated, such as serial manipulators consist of redundant manipulators [3] and mobile manipulators [4], parallel manipulators [5], and cable-driven manipulators [6]. A redundant manipulator is often designed as a series of links connected by motor-actuated joints that extends from a fixed base to an end-effector while a mobile manipulator is often designed as a robotic device composed of a mobile platform and a redundant manipulator fixed to the platform [7]. Different from these serial manipulators, a parallel manipulator is a mechanical system that usually uses several serial chains to support a single platform, or end-effector. Besides, a cable-driven manipulator is a special parallel manipulator, in which the moving platform is driven by cables instead of rigid links [8]. Using these manipulators to save labors and increase accuracies are becoming common practices in various industrial fields. As a consequence, many approaches have been proposed, investigated and employed for the control of robot manipulators [9]. Among them, thanks to many advantages in parallel distributed structure, nonlinear mapping, ability to learn from examples, high generalization performance, and capability to approximate an arbitrary function with sufficient number of neurons, the neural-network-based approach is a competitive way to control movements of robot manipulators [1]. Generally speaking, neural networks can be classified into different types according to different criterions. For example, in terms of the structure of the network, they can be classified into two categories: feedforward neural networks and recurrent neural networks [10,11]. A feedforward neural network is an artificial neural network with no cycles or feedback signal inside while a recurrent neural network allows bi-directional information flow, which means the information inside flows from a successive node to a previous one (or called feedback) or forms a closed cycle within a single node. In this paper, we make a relatively comprehensive review of research progress on controlling these robot manipulators by means of neural networks. The overall organization of the paper is as follows. After the introduction, we present preliminaries on the control of robot manipulators based on neural networks in Section 2. Section 3 presents and reviews different types of robot manipulators in detail with the corresponding schematics being illustrated. In addition, Section 4 revisits applications of different neural networks to the control of robot manipulators. Moreover, two possible future research directions on control of robot manipulators using neural networks are pointed out in Section 5. Finally, Section 6 concludes the paper with final remarks.

Industrial robots are base-stationary, reprogrammable and multifunction manipulators that are designed to move material, parts, tools, or specialized devices through variable programmed motions to perform a variety of tasks. The industrial robots have replaced human beings in dangerous, monotonous, or strenuous tasks that humans do not want to do. These activities frequently take place in spaces that are poorly ventilated, poorly lighted, or filled with noxious or toxic fumes. The trend for future industrial robots has been toward the electric-motor powered, servo-controlled industrial robots that are typically floor-standing machines [1]. These industrial robots have proved to be the most costeffective because they are the most versatile. As a fundamental and essential subject of industrial robots, the trajectory tracking control has attracted considerable attention over the last few years. A variety of control approaches for high-precision trajectory tracking control of industrial robots have been presented in the literature. The proportional integral derivative (PID) controllers have been widely used for motion tracking control of industrial robots due to the simple structure and model free control design [2]. However, the PID controllers should be carefully tuned according to different operating conditions, which is usually time-consuming. A cooperative target tracking control algorithm was proposed in [3] for motion target tracking of a group of mobile robots. A distributed Kalman filter was also designed to estimate the target position. The effectiveness of the control algorithm was verified based on simulation and experimental results. An adaptive trajectory tracking controller was designed in [4] for robot motion tracking with unknown parameters and uncertain dynamics. The back-stepping controller was designed by using the learning ability of neural networks which avoid the knowledge of the robot dynamics. Simulations and experiments on a commercial robot platform were used to verify the performances of the designed control algorithm with classical back-stepping controller. A model predictive control scheme incorporating neural-dynamic optimization was presented in [5] for robotic trajectory tracking. The model predictive control approach was iteratively transformed as a constrained quadratic programming problem, and then a primal-dual neural network was used to solve this problem over a finite receding horizon.