Cloud Robotics: Current Status and Open Issues - AMiner

2169-3536 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. Seehttp://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2016.2574979, IEEEAccess

1

Abstract—With the development of cloud computing, big data and other emerging technologies, the integration of cloud technology and multi-robot systems makes it possible to design multi-robot systems with improved energy efficiency, high real-time performance and low cost. In order to address the potential of clouds in enhancing robotics for industrial systems, this paper describes the basic concepts and development process of cloud robotics and the overall architecture of these systems. Then, the major driving forces behind the development of cloud robotics are carefully analyzed from the point of view of cloud computing, big data, open source resources, robot cooperative learning, and network connectivity. Subsequently, the key issues and challenges in the current cloud robotic systems are proposed, and some possible solutions are also given. Finally, the potential value of cloud robotic systems in different practical applications is discussed.

Index Terms— cloud computing, big data, open source, cloud robotics, internet of things

I. INTRODUCTION

he introduction of automation equipment in industrial production in the past few decades has incurred many

improvements in the industrial sector. With the development of industrial robots, programmed robots have reached high levels of performance in real-time applications, accuracy, robustness and compatibility. However, when facing the extreme environment, such as the earthquake, exploring the unknown space, operating fast and accurate grasp of the real demand, due to the unknown conditions, preprogrammed robotics cannot meet actual application needs. As network technology developed during the latter part of the 1990s, researchers developed and improved the control of robotic network interfaces and their robustness, and the field of “networked robotics” [1] appeared.

A robotic network refers to a group of robots connected through a wired or wireless communication network [2]. An individual robot in networked robotics is regarded as a node. With sensing data and information shared among nodes, the operators can transmit command data remotely and receive measurement feedback, thus ensuring that a specific operation is carried out. The development of networked robotics has allowed them to be utilized in a variety of applications, such as long-distance medical surgery, disaster relief and other

Jiafu Wan, Shenglong Tang, Di Li and Shiyong Wang are with the School of

Mechanical and Automotive Engineering, South China University of Technology, China (e-mail: [email protected], [email protected], [email protected], [email protected])

Hehua Yan is with the School of Electrical Engineering, Guangdong Mechanical & Electrical College, China (e-mail: [email protected])

Athanasios V. Vasilakos is with the Department of Computer Science, Electrical and Space Engineering, Luleå University of Technology, Sweden (e-mail: [email protected])

Corresponding author: Hehua Yan, email: [email protected]

specialized cases. However, like the single robot, robotics networks also face some inherent physical limitations: Due to the limitations of the robot’s volume and other factors, there are obvious limitations to the computing and storage capacity of individual robots. This leads to a limited capacity of the traditional networked robotics when facing high complexity processing tasks. Also, the corresponding performance improvement of individual robots has obvious limitations as well.

With the development of cloud computing, big data [3] and other emerging technologies [4, 5], the integration of cloud technology and multi-robot systems allows for the design of multi-robot systems with high performance and high complexity. In 2009, the European Union project RoboEarth [6, 7], led by Holland’s Eindhoven University, and resulted in the development of a World Wide Web equivalent for robots. RoboEarth can be regarded as a web-based large database, the implementation of which allows robots to share information learned from each other, and finally update the database to achieve a virtuous closed loop. At the 2010 Humanoids conference, James J. Kuffner proposed the concept of “cloud robotics” [8] and elaborated the potential advantages of robot clouds for the first time. The concept of cloud robotics soon caused extensive discussion and research, with researchers in Singapore presenting the construction of the DAvinCi framework [9], Japanese researchers building the business platform Rapyuta [10,11] and the development of the open source software package ROS (Robot Operating System) with efforts of Willow Garage’s team [12] accelerating the development of cloud robotics. Kehoe et al. [13] presented a survey that introduced the most relevant work and organized it around the potential benefits of introducing cloud technologies for robotics and automation. To explore the benefits of cloud-based technologies for sensing, Mohanarajah et al. [14] presented a cloud-based collaborative visual Simultaneous Localization and Mapping (SLAM) system consisting of low-cost robots and remote Amazon servers, which allows real-time map estimation through parallel computing and implements a highly sophisticated system using low-cost components, thus lowering the technological threshold for further research. Additionally, Agüero et al. [15], through their participation at an open competition, presented a software framework for cloud-hosted robot simulation aimed at advancing robotic systems dealing with disaster response.

However, some technical challenges cannot be ignored. With the introduction of cloud technologies, the selection of the types of computation distributions and communication modes that should be applied in different scenarios is critical for overall performance. In addition, extracting relevant patterns from data in the cloud is a typical requirement in different applications, which poses the challenge of data format conversion when dealing with data uploading and downloading.

Cloud Robotics: Current Status and Open Issues

Jiafu Wan, Shenglong Tang, Hehua Yan, Di Li, Shiyong Wang and Athanasios V. Vasilakos

T



2

Another important aspect is cloud security, especially the storage of important data in the cloud; this increases the requirements on various aspects of the systems. Finally, to ensure real-time performance, choosing the service quality guarantee methods and corresponding effect analyses is challenging.

In this paper, there are three main contributions as follows: The major enabling technologies of cloud robotics are

analyzed, including big data, cloud computing, open source resources, cooperative robot learning, and network connectivity.

The most typical applications (e.g., SLAM, grasping and navigation) are introduced.

The key problems and challenges of current cloud robotics system are stated: resource allocation and scheduling, data interaction between the robots and the cloud platform, cloud security, and service quality guarantee method and effect analysis.

The remaining content is organized in the following fashion: First, we describe the overall structure of the robotic cloud, and then analyze several major elements of its ecology and conduct an analysis of current key issues to be solved. Then, the robotic cloud’s current applications and main results are summarized, while in the end of the article we discuss prospects for the future development of the cloud robotics.

II. SYSTEM ARCHITECTURE OF CLOUD ROBOTICS

Networked robotics can be seen as transition state between preprogrammed robots to cloud-enabled robots [16]. As previously mentioned, cloud robotics aim at transferring the high complexity of the computing process to the cloud platform through communication technology. This greatly reduces the computational load on individual robots. Figure 1 describes the main architecture of the robotic cloud.

Cloud Infrastructure

DataBase

Server

Proxy Server

simultaneous localization and mapping Navigation

Grasping

Fig. 1. System architecture of cloud robotics

The architecture of cloud robotics is mainly composed of two

parts: the cloud platform and its related equipment and the bottom facility. The bottom facilities usually include all types of mobile robots, unmanned aerial vehicles, machinery and other equipment. Accordingly, the cloud platform is composed of a large number of high-performance servers, proxy servers, massive spatial databases and other components.

Multi-robot cooperative works, such as SLAM and navigation [17] networks, are typical applications of cloud robots. The term “networked robotics” refers to the communication mode of cloud robotics and multi-robot systems are composed as cooperative computing networks using wireless communication technologies. The major advantages of cooperative computing networks are the following: 1) a collaborative computing network can gather computing and storage resources, and can dynamically allocate these resources according to specific work requirements; and 2) because of the exchange of information, decisions can be made cooperatively between machines. The nodes’ computing and storage capabilities’ deficiency in networked robotic systems may lead to large delays, while in cloud robotic systems the nodes can collaborate with spare nodes by transferring computing or storage tasks. Nodes which are not directly connected to the cloud resource can connect to the cloud through other nodes that have already established links to the cloud. The application of this mechanism greatly expands the manageable complexities of tasks accomplished through multi-robot cooperative work and enhances the efficiency of specific work tasks.

Other tasks that do not involve a need for additional robots but require complex operations, such as grasping, are also hot research areas. Due to the fact that the target object is usually unknown, several researchers have so far proposed the incorporation of a large number of sensing devices on the grabbing body [18], aiming to improve the accuracy of grasping. However, in actual industrial environments, a demand for high accuracy and fewer sensors is more in line with production practices [19]. At the same time, due to the limitation of physical space and materials, the equipment and storage facilities are limited. This leads to a bottleneck of development and research. With the introduction of big data, with a small amount of sensor and other characteristic data uploaded, researchers can determine a match in the cloud. Then the characteristic data of the grab motion are downloaded and sent to the mechanical equipment for execution of the operation. An example of the applications of cloud robotics including SLAM and Grasping is shown in Figure 2.

In summary, the main features of the cloud robotics architecture are as follows. 1) In the cloud infrastructure, where computing tasks are dynamic and resources are elastic and available on-demand. 2) The cloud robotics’ “brain” is in the cloud. The results of processing can be obtained through networking technologies, while tasks are processed individually. 3) Computing work can be delegated to the cloud, which leads to a smaller robot load and greatly extended battery life.

III. MAJOR DRIVING FORCES BEHIND THE DEVELOPMENT OF

CLOUD ROBOTICS

Cloud robotic systems’ structure involves a number of cloud, networking and embedded systems’ technologies, as well as all kinds of wireless communication protocols. In this section, we mainly discuss four major driving forces: cloud computing, big data, open source, robot cooperative learning, and network connectivity.



3

Machine-to-Machine

Machine-to-Cloud

Product

Task for SLAM

RFID Reader

Task for Grasping Fig. 2. Implementation of cloud robotics in industrial environment

A. Cloud Computing for Robotics

As we know, cloud computing which is suitable for sampling-based analyses, has been widely used in large-scale parallel computing applications with great success. For example, Ref. [20] proposed a novel architecture which combines mobile cloud computing and vehicular cyber-physical systems. In the robotics field, due to the high computing performance of cloud-based systems, when the computing task is uploaded to the cloud, the computational load of computing equipment is reduced to a great extent. For example, grasping based on cloud computing for unknown object shapes has achieved great robustness. Ben Kehoe et al. [21] used PiCloud, a commercial cloud computing platform, and the combination of big data with grasping techniques allowed for a 90% reduction in sampling size.

In the field of multi-robot operations, cloud computing has greatly accelerated the speed of development of robotic and automation equipment. In order to achieve robot navigation, L. Riazuelo et al. [22] used a cloud platform for SLAM, achieving very good results. In addition, the combination of cloud computing and robotics in video, image analysis, data mining and other fields is an important direction of the future development of cloud robotics.

However, what we cannot ignore is that the introduction of cloud computing results in the uploading and downloading of data necessary for computing tasks. Current mainstream communication protocols, such as Zigbee, Bluetooth and Wi-Fi have developed rapidly, but the working environment has a great influence on the fluency of wireless communications. A particular example is the interaction involved in the exchange of dynamic information, which can easily cause network delays in the process of information transmission. For real-time sensitive applications, the tradeoff strategy of cloud computing and local computing to be implemented according to different environments an active area of research.

B. Big Data for Robotics

Big data can provide automation systems with access to massive resources in a cloud infrastructure, which might be a key solution to cope with the challenge of limited onboard resources [69]. The previously mentioned grasping problem is a continuing challenge for robotics. With the support of cloud resources, grasping can be achieved through several strategies. One of the advantages of big data that it provides an index of pictures, maps and object data from a global database for the terminal (this paper mainly refers to various types of machine equipment). These data include images, videos, maps, sensor networks, and so on. The most typical example is RoboEarth. As a robotic database, RoboEarth is attracting a large number of researchers to use and share due to its open source characteristics. This positive closed loop continuously expands RoboEarth with massive object and map data. These data provide important technical support for the development of robotic navigation and grasping systems. However, on one hand, big data greatly expands the possibility of overall improvement, but on the other, large datasets are often mixed with unexpected data called “dirty” data. New technologies are required to cope with this serious challenge.

C. Open Source for Robotics

With the development of cloud technology, the spirit of open source is infiltrating into the cloud robotics field. Among them, the most representative is ROS and RoboEarth.

ROS is a well-known framework, whose open source license allows the improvement of the code reuse rate and development efficiency of this robot operating system. There are two major parts to ROS: system maintenance and distribution.

(1) Main: This is the core part, designed at the Willow Garage research laboratory in collaboration with other developers, which provides maintenance. It provides some of the basic tools for distributed computing, as well as the entire ROS of the core part of the program.



4

(2) Universe: the global scope of the code, maintained by the international ROS community organization. It includes libraries of code written in several programming languages, such OpenCV and PCL. In general, the Universe comprises algorithms, frameworks and hardware drivers.

The computation graph is a kind of point-to-point networking form for ROS data processing. When the program runs, all the processes and their data will be depicted through a point-to-point network. This level mainly includes several important concepts: nodes, messages, subjects, services.

A node in this case is a process that performs a task. The use of "nodes" makes an operation based on ROS more vivid: When many nodes are running at the same time, it is easy to draw the point-to-point communications involved as a graph. The nodes communicate with each other by sending messages. Every message has a strict data structure. Not only the basic standard data types (e.g., integer, floating-point, and Boolean) but also basic array type are supported. Messages can also contain arbitrary nested structures and arrays (similar to the structure of the C language struct). Messages are exchanged in a “publish” or a “subscribe” manner. A node can publish a message on a given topic, and is concerned with a particular type of data only if it has subscribed to the particular topic. On the whole, publishers and subscribers are not aware of each other's existence. Although the topic-based publish / subscribe model is very flexible, it is not suitable for synchronous transmission modes. In ROS, another Point to Point (P2P) communication model has been proposed, which is defined by a pair of strict specifications (one for the request and one for the response) and a string. This is similar to the web server, and when other nodes send data requests to the service node, the service node responds. Figure 2 depicts the two communication methods between two nodes through the message and service mode.

ROS has numerous nodes, messages, services, tools, and library files that require an effective structure to manage the code. On the level of the ROS file system, there are some important concepts: the package, the stack and the repository. ROS software is organized in packages. Packages contain nodes, dependency libraries, configuration files, third party software, etc. The objective of a package is to provide a structure that is easy to use in order to facilitate the reuse of the software. The corresponding stack is a collection of packages, which provides a complete set of features.

RoboEarth is a dedicated website providing services for robotic systems. It is a huge network database system. Robotic systems can learn from the experience of other robots’ behavior and operating environments and ultimately share information through the database. It is apparent that any individual single robot is isolated, whose function and behavior has basically been determined through factory settings; this causes these kind of robotics to lack learning abilities. The RoboEarth project is dedicated to building an open source Internet database, so that the robots from different locations around the world can access and update this information. RoboEarth stores a large amount data for target recognition, navigation, tasks, intelligent services and other information required by robotics.

The overall architecture of the RoboEarth is shown in Figure 3. In general, RoboEarth is divided into the client, the cloud engine and the database layers. The computing tasks will be

transferred to the cloud through the network using a unified data format, which means that all uploaded information from any robot has the same structure. After unifying data format, computing tasks will be delegated to the engine cloud for cloud computing. The engine cloud may need to interact with the upper database level, and will finally return the results to the underlying robot. Some applications such as grasping may directly access the upper layer of the database for model matching, in which case the feature data will be returned to the underlying robot, in order for the latter to carry out specific operations.

External APIs + DBs

----------------------------------------------------------

----------------------------------------------------------

RoboEarth DataBasa

RoboEarth Cloud Engine

Clients

Software Componence Maps

ActionRecipes Objects

Fig. 3. Architecture of RoboEarth

D. Robot Cooperative Learning

A multi-robot system is a fundamental field of robotics. With the support of other robots in a team, a multi-robot system can optimize the individual shortcomings and significantly improve the total accuracy and complexity of tasks. Cooperating learning, one of the most important topics in multi-robot systems, still faces a lot of challenging problems.

To deal with various uncertainties in an unknown environment, teams of researchers came out with Reinforcement Learning (RL) [23] decade ago. RL has been applied to many scenarios and obtained a significant effect. However, researchers soon found that traditional RL, such as Q-learning [24], could not cope with high-real-time-demand tasks from more complicated operation [25]. The Hierarchical Reinforcement Learning (HRL) decomposition, such as HAMs and MAXQ, optimizes the traditional methods, to some extent, but cannot avoid to explore the unnecessary part of environments [26].

Dawn of cloud technologies completely offers a new sight of solution. Researchers caught in algorithm mire can take advantage of cloud technologies to accelerate information sharing for robots and automation by facilitating the exchange of data on associated control policies, trajectories, current conditions and others. For grasping, accuracy and stability can be greatly augmented by learning from previous grasp operations on an object. Similarly, the data shared, along with the resulting performance and outcomes, improves the operating performance in collective robot learning systems.



5

In SLAM, navigation and other applications, besides traditional cooperating learning, namely shared tracks, control strategies and other information, transferring data to the cloud significantly releases the computing and storage pressure of individual robot. As a consequence, the additional onboard resource gives individual flexible response and smooth operating.

E. Network Connectivity

Connectivity develops with multi-robot system as well. Previous researchers have greatly contributed to this issue [27].

Traditional inter communication within robotics is a typical Machine to Machine (M2M) communication architecture. Collaborative learning requires that the global task priority is greater than the priority of the node task, and this applies to dynamic packet interaction and control as well. These constraints result in higher requirements on the control algorithm.

It is worth noting that inter-machine communications usually use proactive and ad hoc routing protocols [28]. In unfamiliar environments, networked robots may choose a proactive routing protocol to determine a path. Proactive routing protocols include periodic packet switching and routing table updates, which makes the consumption of computing resources and memory resources particularly large.

A classic protocol for mobile robotic network is shown in Figure 4 [29]. The source node sends a RREQ (Routing Request Frame) to the surrounding nodes, the intermediate nodes update the routing tables to the source node, and maintain a reverse routing path to the source node, and then forward the RREQ. An intermediate node or a destination node which is aware of a valid path to the destination node generates a RREP (routing response frame). With the establishment of the reverse node, the RREP will finally find its way back to the source node. Now the source node is able to send packets of data to the destination node. From the above description, we can see that this protocol can lead to increased delays, which will create an adverse effect in the dynamic network topologies. That will be a continuous study area in multi-robot systems.

A

B

E

G

D

C

F

RREQRREQ

RREQRREQ

RREQ

RREQ

RREQ

RREQRREQ

RREP

RREP

RREP

Source

Destination

Fig. 4. Classic routing protocol example for mobile robotic network

With the introduction of cloud infrastructure, robotics under

clouds have another choice to deal with computing or storage task. However, latency is a challenging problem in traditional network robotics. In the cloud scenario, expenditure of time is divided into four parts: 1) pre-processing time for sending and

receiving specific data; 2) time consumed in sending and receiving data which has a lot to do with data size and network bandwidth; 3) time consumed in cloud for data processing; and 4) network delay which is inevitable. To make the introduction of cloud practical, the sum of above elements should be less than time consuming without cloud.

Wang et al [30] proposed a hierarchical auction-based mechanism for cloud robotic systems. In this study, researchers establish a hierarchical architecture as Figure 5. The relay nodes help maintain the connection of the network while the child nodes depend on the relay nodes to extend the network which establish the whole network. The key of proposed hierarchical auction-based mechanism is that all child nodes of one relay node compete for the opportunity of transmission, namely bidding. This depends on individual spatial distance, signal power, signal to noise ratio, etc. There are many nodes. This means several times of auction processes from lower layer to higher layer. With the last round of auction done, the cloud will allocate the resource according to the rank of auction result. Based on this algorithm, the simulation and experiment results show the high performance of the timely responses.

Child Node

Relay Node

Child Node

Relay Node

Proxy

Relay Node

Fig. 5. Hierarchical auction-based mechanism for cloud robotic systems

IV. TYPICAL APPLICATIONS OF CLOUD ROBOTICS

With the advent of industry 4.0 [31, 32], the introduction of cyber-physical systems [33, 34, 70] and cloud technologies makes it possible to significantly improve robotic performance. In recent years, the study of cloud robotics has achieved rapid development through the support of relevant cloud technologies. Table I shows the main work in this domain. In this section, we will introduce some classic applications of the cloud robotics, including SLAM, grasping and navigation.

TABLE I

Article SLAM Grasping Navigation

Ref.[10] ✓ ✓ ✓ Ref.[13] ✓ ✓ ✓

Ref.[14] ✓

Ref.[21] ✓

Ref.[22] ✓ ✓ ✓ Ref.[40] ✓

Ref.[59] ✓



6

A. SLAM

SLAM is a fundamental topic since 1990s when Durrant- Whyte first clearly defined SLAM [35] as exploring unknown environment problems. Extended Kalman filter (EKF) [36], the most famous SLAM method, has been applied to common uses. Athough applicable, EKF cannot deal with real-time requirements. The appearance of particle filter (PF) [37] offers another insight in SLAM. Since PF doesn’t have to directly solve probability density function and has significant advantage in status filtering and parameter estimation, PF become another effective method for SLAM.

However, high-complexity tasks lead to more complicated state estimation which is a serious challenge of the effectiveness for PF. Klein and Murray [38] proposed Parallel Tracking and Mapping (PTAM) which introduced multi-thread solution and obtained the real-time performance. But, PTAM is not able to handle large areas. Strasdat et al [39] proposed Scalable Visual SLAM (ScaViSLAM) aimed at handling large scale environment exploring. Unfortunately, this method is prone to failure when tracking failures happens.

Although the SLAM-based algorithms are becoming more and more accurate, the superiority of the current generation of algorithms for large scale maps cannot make up for the limited computational and storage resources onboard. Therefore, before the advent of the cloud robotics, the formation of large-scale maps often took a long time, which was far from meeting the real application needs.

The emergence of cloud technology allows SLAM to break the bottleneck caused by limited on-board computing and storage equipment. Riazuelo et al. [40] proposed a framework to allocate expensive map optimization and storage in the cloud where the system provides the interface to a map database.

In addition, besides RoboEarth, the relevant technical teams are also committed to continuously improving the SLAM-related code and platform construction. For example, the Kinect@Home project is working on collecting data to complete a RGB-D data set [41]. In this project, Kinect users can use this platform to carry on the 3D mapping component through their home network browser, while at the same time Kinect collects the data which the user uploaded. Another interesting project, Tango Project, launched by Google, allows telephones with specific hardware generate a map, which is to a certain extent accurate and continuous.

There are many similar platforms and projects around the world. Using a cloud platform, a significant portion of computing, map fusion and filtered state estimates can be completed in the cloud, which provides strong support for the formation of the map.

B. Grasping

The grasping of unknown objects is of great practical significance of robotics in the industrial sector. Grasping problem was first proposed by Ferrari, Canny [42] and Mirtich et al. [43] to cope with geometric analysis of polyhedral objects while the initial solution cannot deal with non-polyhedral objects. In 2003 Miller [44] first introduced object models and then Saxena et al. [45] proposed object features into grasping. However, these propositions do not concentrate on knowledge learning which has apparent drawbacks when dealing with totally unknown objects. With the rapid development of the correlation discipline, neural networks was introduced in Grasping by Luo et al. [46], and RL

was proposed by Baier-Löwenstein et al. [47]. Grasping with learning significantly improves its robustness and accuracy.

However, the fact that the unknown characteristics of objects implies a lot of data preprocessing and a large amount of computations, lets along introducing learning ability. A large number of previous studies on unknown object grasping have been carried out. For example, in [48, 49], researchers used contact sensors to improve the quality of grasping of unknown objects. However, many robots are not equipped with adequate sensors. Dogar and Srinivasa [50] proposed that a pushing operation can reduce uncertainty over the object's shape, which leads a better grasp.

With the introduction of relative sensing devices, grasping has made progress, to some extent, but it still cannot meet the requirements for adequate response speed and precision. Under the support of big data, cloud computing capabilities and storage capacity, a mechanical hand can send feature data obtained through a small number of sensors to the cloud using a specific data format. After feature analysis and model matching, data with operating characteristics will be downloaded to the manipulator so that the mechanical hand can grasp the object. It is worth mentioning that after the completion of the grasp, the relevant data can be stored in the cloud for sharing [51] in cases of similar need.

C. Navigation

Another classic application is navigation. There are two types of navigation problems: the local navigation problem and the global navigation problem. The formal has been discussed since 1980s and achieved significant progress while the later problem is still a study area with insights offered worldwide [52]. The global navigation problem copes with the larger scale in which the robot does not know the state of its destination from initial position. Previously, many approaches [53] have been proposed to deal with navigation in which a map is given. Note that in real world robotic scenario, the environment is completely unknown so that the key is to find a collision-free path with high quality.

To deal with Non-map-based scenario, researchers come out with several famous approaches such as fuzzy logic [54], sampling-based method [55]. Moreover, the traditional sensor-based approaches [56] and neural network models [57] with optimized algorithm also play a role in the global navigation problem.

However, due to the restricted onboard resources, these methods usually suffer from reliability problems, either local navigation problems or global navigation problems are vulnerable to large scale computing and storage tasks which is challenging in large navigation areas especially the later.

Cloud-related technologies provide a cloud-based navigation system for the future, and are a very promising solution that avoids these two problems. Cloud computing is not only able to provide sufficient storage space to deal with large amounts of data for mapping, but also features enough computing capacity to allow map searching and establishment. Through cloud computing, available commercial maps (e.g., Google maps) can also be exploited to provide flexible, reliable, and autonomous navigation.



7

V. KEY ISSUES AND CHALLENGES

This section focuses on key issues and challenges in cloud robotics. The collaborative learning mentioned before is actually only one of the aspects of computing resource allocation and scheduling. According to the real-time requirements of different scenarios, establishing a balance between real-time demands and processing performance is a core issue of resource allocation and scheduling. Cloud storage means the remote storage of data. Whether it is business privacy or scientific research, the importance of data also imposes further requirements on cloud security. Robotic cloud platforms need to interact using these data, which means that the data to be analyzed must be uploaded in a certain format. In addition, processed information also needs to be returned from the cloud platform to the robot in a certain form. Finally, the introduction of service quality assurance systems and effects’ analysis, in order to maintain the network flow for a given bandwidth enhances real-time performance and computational efficiency.

A. Resource Allocation and Scheduling

Uploading computational tasks with high complexity to the cloud is one of the most notable characteristics of cloud robotics. Considering different working equipment, interface settings, and network environments, for a given computational task, the choice of uploading, self-processing or assigning the task to the nearest node has an important impact on overall performance.

A number of studies on resource allocation and scheduling has examined the problem in detail. For example, Gouveia et al. [58] put forward the stateless and stateful resource allocation frameworks for the SLAM problem, aimed at computation sharing in distributed robotic systems. Without depending on an external network, the architectures proposed enable an isolated multi-robot system to share computational resources, which is enlightening for resource allocation and scheduling.

It is worth noting that, as with inter-machine communications, the amount of computations on the cloud makes the emergence of delays more likely. New algorithms and techniques are needed to counter changes in network delays in real time. Although wireless technology has made significant progress, once connection problems between the robots and cloud services appear, serious delays are almost inevitable. Therefore, when designing a new algorithm, we need to design a load distribution algorithm with an "any time" characteristic. Once it is apparent that a task that cannot be properly uploaded to the cloud, a mechanism for dynamic allocation of computing tasks should be activated, thereby reducing the delay time.

Considering the large size of a data stream in SLAM, navigation and other applications, the most critical aspect is the computation-communication tradeoff. As mentioned above, the four major time consumption (pre-processing time for sending and receiving data, time consumed for sending and receiving data, time consumed in cloud for processing and lastly network delay) should be less than the onboard time consumption. Taking SLAM as an example, applications such as stereo image processing [59], a typical task with great computation, may delegate the majority of tasks (navigation assistance, etc.) to the

cloud, which leads to less bandwidth consumption and latencies compared with other computational modes. With the increasing complexity of robot cooperative learning added to the system, transferring tasks to robotics using appropriate data structures may be taken into consideration. In any case, carefully chosen computation offloading schemes will usually outperform local computational capabilities.

B. Data Interaction between Robot and Cloud Platform

Devices and sensors from different manufacturers may output data with different structures. In fact, even different models of a product from the same manufacturer may result in considerable differences in the output data’s structure. The diversity of data structures presents high requirement on the compatibility of the cloud input interface. To solve this problem, current mainstream cloud platforms often provide multiple interfaces for various types of data formats.

However, due to the limited number of interfaces, the data to be uploaded must be properly preprocessed, and the robustness and real-time performance of the data exchange is greatly affected by data formats requiring translation from one form to another. Additionally due to their invariance, cloud devices can only handle and store data of specific structures. This means that the input interfaces need to transform the corresponding data structures into a unified format. Lastly, at the output part, when the processing of uploaded data is complete, the ready-to-output data must also be transformed to specific formats. Output data formats need to be chosen according to the compatibility of the underlying devices and the real-time requirements of the entire data exchange. Figure 7 shows the flow chart of SLAM with data interaction.

START

Update surrounding information

Upload now ?

Data format conversion

Upload to cloud

Process in cloud

Complete now ?

Download to robotics

END

N

Y

N

Y

Fig. 6. Flow chart of SLAM with data interaction

RoboBrain [60], the premier project in the robotics area,

provides access for individual robots to search the internet for concepts. While, compatible with millions of robotic



8

applications, the corresponding solutions of RoboEarth have paved the way to similar cloud platforms. To deal with demands from action models, RoboEarth provides an action ontology which describes different kinds of actions in terms of a taxonomic structure and further annotates each of these classes with its semantic properties. Additionally, there are similar ontologies for classes of object and environment models [61]. Inspired by RoboEarth, emerging cloud platforms may design targeted classes and interfaces of compatibility and robustness to deal with different scenarios. Fortunately, the technology of software-defined networks [62-64] is bringing new hope for improving the data interaction between the robots and the cloud platform.

C. Cloud Security

The introduction of cloud technology has greatly expanded the complexity of multi-robot operations. At the same time, it also introduced new technical challenges: the privacy and security issues brought by the cloud technology. These hidden dangers also affect data generated by computing devices and sensors used in cloud robotics. Commercial science and technology solutions have suffered from serious data leakage incidents, especially during the upload of photos and video to the cloud [65]. In scientific research and industrial practice, key data stored in the cloud may be far from the hacker to steal, leading to the loss of key data.

To deal with these challenges, relevant management rules and legal provisions must be established [66, 67]. On the technical front, identity management and access control systems are two of the most important aspects of cloud security. Current solutions proposed include a combination of multi-identity and personal identity authentication methods with layered encryption. Additionally, data security is at the core of cloud security, which is composed of isolation protection of static and dynamic data storage. More effective integrated verification algorithms are key for data integrity. Also, another aspect of cloud security which deals with providing software as a service (SaaS), has virtualization technology as its basis. Therefore, network, storage and server virtualization, are essential subsystems, which means that security mechanisms such as VM security isolation, access control, VM resource constraints, etc. need continuous improvement.

D. Service Quality Guarantee Methods and Effects’ Analysis

Network latency is what researches cannot ignore. In fact, the real-time demand, to great extent, affects the overall performance. The introduction of a service quality assurance system and effects’ analysis for a given bandwidth can ensure adequate network flow to great extent. The core of service quality systems and effect analysis is to balance the limited network resource and the real-time demand. In order to meet the needs of users with different service quality demands, the network is expected to provide different levels of quality according to the requirements of users: real time requirements require high priority data processing, while tasks with less than real-time requirements are granted lower priority. Using a correspondingly differentiated service model similar to that for QoS, higher bandwidth priority is given to tasks with high real-time requirement, according to the characteristics of the

cloud robotics system. As in SLAM, allocating farther nodes with more bandwidth can balance global real-time performance. Additionally, for the evaluation of network congestion, transmission data errors, and the loss of data packets, the establishment of uniform standards and specific reliability analysis approaches are of great significance for the evaluation of algorithms and the exchange of scientific research.

VI. CONCUSSIONS

This paper described the origin and development of cloud robotics, from the allocation of resources and communication mode perspectives and analyzed cloud robotics communication architecture. Four major elements of robot clouds are summarized and the key technologies to promote development are analyzed. Finally, some aspects of cloud robotic systems in different practical application are discussed. Cloud robotics, through the sharing of computing resources with the cloud platform, push multi-robot systems and collaborative learning to new heights. By delegating high complexity computational loads to the cloud platform, the robots no longer require high performing onboard equipment, thus greatly reducing the cost of the application of multi-robot systems. As Ermacora et al.[68] proposed smart city with cloud robotic service, with the development of cloud computing, big data and other fields, the robotic clouds in SLAM, grasping, navigation and other applications will achieve better performance.

ACKNOWLEDGEMENTS

This work was supported by the National Natural Science Foundation of China (Nos. 61572220, 61262013 and 51575194), the Fundamental Research Funds for the Central Universities (No. 2015ZZ079), the Natural Science Foundation of Guangdong Province, China (Nos. 2015A030313746, 2015A030308002, 2016A030313734 and 2016A030313735), the National Key Technology R&D Program of China (No. 5642015BAF20B01), the Science and Technology Planning Project of Guangdong Province, China (Nos. 2012A090100012, 2013B010134010, 2014B090921003, and 2014A050503009), and the Science and Technology Planning Project of Guangzhou City (201508030007).

REFERENCES

[1] Accessed online at http://www.nytimes.com/2014/08/18/technology/fo r-big-data-scientists.

[2] F. Li, J. Wan, P. Zhang, D. Li, D. Zhang and K. Zhou, “Usage-Specific Semantic Integration for Cyber-Physical Robot Systems,” ACM Transactions on Embedded Computing Systems, vol. 15, no. 5, Article 50, 2015.

[3] Y. Zhang, M. Chen, S. Mao, L. Hu, and V. Leung, “CAP: Crowd Activity Prediction Based on Big Data Analysis,” IEEE Network, vol. 28, no. 4, pp. 52-57, 2014.

[4] J. Wan, H. Yan, Q. Liu, K. Zhou, R. Lu and D. Li, “Enabling Cyber-Physical Systems with Machine-to-Machine Technologies,” International Journal of Ad Hoc and Ubiquitous Computing, vol. 13, no. 3/4, pp. 187–196, 2013.

[5] M. Chen, Y. Zhang, Y. Li, M. Hassan, A. Alamri, “AIWAC: Affective Interaction through Wearable Computing and Cloud Technology,” IEEE Wireless Communications Magazine, vol. 22, no. 1, pp. 20-27, Feb. 2015.

[6] Accessed online at http://www.roboearth.org/ [7] L. Riazuelo, M. Tenorth, D. Di Marco, et al. “ Roboearth semantic



9

mapping: A cloud enabled knowledge-based approach, ” IEEE Transactions on Automation Science and Engineering, vol. 12, no. 2, pp. 432-443, 2015.

[8] James J. Kuffner, “Cloud-enabled robots,” Proc. Of IEEE-RAS International Conference on Humanoid Robotics, Nashville, TN, USA, 2010.

[9] R. Arumugam, V. R. Enti, et al. “DAvinCi: A cloud computing framework for service robots,” IEEE International Conference on Robotics and Automation (ICRA) 2010, pp. 3084-3089, May, 2010.

[10] Mohanarajah, Gajamohan, et al. “Rapyuta: A cloud robotics platform,” IEEE ransactions on Automation Science and Engineering, vol. 12, no. 2, pp. 481-493, 2015.

[11] Hunziker, Dominique, et al. "Rapyuta: The roboearth cloud engine." 2013 IEEE International Conference on Robotics and Automation (ICRA), pp. 438-444, May, 2013.

[12] Quigley, Morgan, et al. “ROS: an open-source Robot Operating System,” IEEE International Conference on Robotics and Automation (ICRA) workshop on open source software, vol. 3, no. 3, pp. 2, 2009.

[13] B. Kehoe, S. Patil, P. Abbeel, and K.Goldberg, “A survey of research on cloud robotics and automation,” IEEE Transactions on Automation Science and Engineering, vol. 12, no. 2, pp. 398-409, 2015.

[14] G. Mohanarajah, V. Usenko, M. Singh, R. D'Andrea, and M. Waibel, “Cloud-based collaborative 3D mapping in real-time with low-cost robots,” IEEE Transactions on Automation Science and Engineering, vol. 12, no. 2, pp. 423-431, 2015.

[15] C. E. Aguero, N. Koenig, et al., “Inside the virtual robotics challenge: Simulating real-time robotic disaster response,” IEEE Transactions on Automation Science and Engineering, vol. 12, no. 2, pp. 494-506, 2015.

[16] G. Hu, W. P. Tay, and Y. Wen, “Cloud robotics: architecture, challenges and applications,” Network, IEEE, vol. 26, no.3, pp. 21-28, 2012.

[17] Rekleitis, M. Ioannis, Gregory Dudek, and Evangelos E. Milios, “Multi-robot cooperative localization: a study of trade-offs between efficiency and accuracy,” IEEE/RSJ International Conference on Intelligent Robots and Systems 2002, vol. 3, pp. 2690-2695, 2002.

[18] Felip, Javier, and Antonio Morales, “Robust sensor-based grasp primitive for a three-finger robot hand,” Intelligent Robots and Systems, 2009. IEEE/RSJ International Conference on, pp. 1811-1816, 2009.

[19] J. Liu, Q. Wang, J. Wan and J. Xiong, “Towards Real-time Indoor Localization in Wireless Sensor Networks,” Proc. of the 12th IEEE Int. Conf. on Computer and Information Technology, Chengdu, China, October, 2012, pp. 877–884.

[20] J. Wan, D. Zhang, Y. Sun, K. Lin, C. Zou and H. Cai, “VCMIA: A Novel Architecture for Integrating Vehicular Cyber-Physical Systems and Mobile Cloud Computing,” Mobile Networks and Applications, vol. 19, no. 2, pp. 153–160, 2014.

[21] B. Kehoe, D. Warrier, S. Patil, and K. Goldberg, “Cloud-based grasp analysis and planning for toleranced parts using parallelized Monte Carlo sampling,” IEEE Transactions on Automation Science and Engineering, vol. 12, no. 2, pp. 455-470, 2015.

[22] L. Riazuelo, M. Tenorth, et al., “Roboearth semantic mapping: A cloud enabled knowledge-based approach,” IEEE Transactions on Automation Science and Engineering, vol. 12, no. 2, pp. 432-443, 2015.

[23] Ying Wang and Clarence W. de Silva, “An object transportation system with multiple robots and machine learning,” in Proc. of the 2005 American Control Conference (ACC 2005), Portland, OR, 2005.

[24] Y. Wang and C. W. De Silva, “Extend single-agent reinforcement learning approach to a multi-robot cooperative task in an unknown dynamic environment,” IEEE International Joint Conference on Neural Networks 2006, pp. 4999-5005, 2006.

[25] Y. Cai, S. X. Yang, and X. Xu, “A combined hierarchical reinforcement learning based approach for multi-robot cooperative target searching in complex unknown environments,” IEEE Symposium on Adaptive Dynamic Programming And Reinforcement Learning (ADPRL) 2013, pp. 52-59, April, 2013.

[26] T. G. Dietterich, “The MAXQ method for hierarchical reinforcement learning,” in Proceedings of the 1998 International Conference on Machine Learning, Madison, USA.

[27] X. Li, D. Li, J. Wan, A. Vasilakos, C. Lai and S. Wang, “A Review of Industrial Wireless Networks in the Context of Industry 4.0,” Wireless Networks, DOI: 10.1007/s11276-015-1133-7, 2015.

[28] J. Liu, J. Wan, Q. Wang, P. Deng, K. Zhou and Y. Qiao, “A Survey on Position-based Routing for Vehicular Ad hoc Networks,” Telecommunication Systems, vol. 62, no. 1, pp. 15-30, 2016.

[29] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, “Wireless sensor networks: a survey,” Computer networks, vol. 38, no. 4, pp.

393-422, 2002. [30] L. Wang, M. Liu, et al., “Hierarchical auction-based mechanism for

real-time resource retrieval in cloud mobile robotic system,” 2014 IEEE International Conference on Robotics and Automation (ICRA), pp. 2164-2169, 2014.

[31] S. Wang, J. Wan, D. Li, and C. Zhang, “Implementing smart factory of industrie 4.0: an outlook,” International Journal of Distributed Sensor Networks, vol. 2016, Article ID 3159805, 10 pages, 2016. DOI: 10.1155/2016/3159805.

[32] S. Wang, J. Wan, D. Zhang, D. Li, and C. Zhang, “Towards Smart Factory for Industry 4.0: A Self-organized Multi-agent System with Big Data Based Feedback and Coordination,” Computer Networks, Elsevier Computer Networks, vol. 101, pp. 158–168, 2016.

[33] Z. Shu, J. Wan, D. Zhang, D. Li, “Cloud-Integrated Cyber-Physical Systems for Complex Industrial Applications,” Mobile Networks and Applications, DOI: 10.1007/s11036-015-0664-6, 2015.

[34] J. Wan, H. Yan, D. Li, K. Zhou and L. Zeng, “Cyber-Physical Systems for Optimal Energy Management Scheme of Autonomous Electric Vehicle,” The Computer Journal, vol. 56, no. 8, pp. 947-956, 2013.

[35] H. F. Durrant Whyte, D. Rye and E. Nebot, “Localization of autonomous guided vehicles," in Proceedings of International Symposium on Robotics Research, New York, Springer-Verlag, pp. 613-625, 1995.

[36] A. 1. Davision, D. W. Murray, “Simultaneous localization and map-building using active vision,” IEEE Transactions. on Pattern Analysis and Machine Intelligence, vol. 24, no. 7, pp. 865-880,2002

[37] T. P. Jensfel, o. Wijk, and D. Austn, et al, “Feature based condensation for mobile robot localization,” in Proceedings of IEEE International Conference on Robotics and Automation, San Francisco, pp. 2531-2537, 2000.

[38] G. Klein and D. Murray, “Parallel tracking and mapping for small ar workspaces,” 6th IEEE and ACM International Symposium on Mixed and Augmented Reality 2007 (ISMAR 2007), pp. 225–234, 2007.

[39] H. Strasdat, A. J. Davison, J. M. M. Montiel, and K. Konolige, “Double window optimization for constant time visual slam.” 2011 IEEE International Conference on Computer Vision 2011, pp. 2352–2359, 2011.

[40] L. Riazuelo, J. Civera, and J. M. M. Montiel, “C2tam: A cloud framework for cooperative tracking and mapping,” Robotics and Autonomous Systems, vol. 62, no. 4, pp. 401-413, 2014.

[41] R. Goransson, A. Aydemir, and P. Jensfelt, “Kinect@Home: Crowdsourced RGB-D data,” in Proc. IROSWorkshop on Cloud Robot, 2013.

[42] C. Ferrari and J. Canny, “Planning optimal grasps,” 1992 IEEE International Conference on Robotics and Automation, pp. 2290 - 2295, 1992.

[43] B. Mirtich and J. Canny, “Easily Computable Optimum Grasps in 2D and 3D,” Proceedings of the 1994 IEEE International Conference on Robotics and Automation, pp. 739 - 747, 1994.

[44] T. Miller, S. Knoop, P. K. Allen, and H. I. Christensen, “Automatic Grasp Planning Using Shape Primitives,” Proceedings of IEEE International Conference on Robotics and Automation, pp. 1824–1829, 2003.

[45] A. Saxena, J. Driemeyer, and A. Y. Ng, “Robotic Grasping of Novel Objects,” International Symposium on Experimental Robotics, 2006.

[46] M. Luo, W. Lu, B. Sun, and T. Mei, “Autonomous Grasping of a Space Robot Multisensory Gripper,” Proceedings of the 2006 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 1014-1019, October 9 - 15, 2006, Beijing, China.

[47] T. Baier-Löwenstein and J. Zhang, “Learning to Grasp Everyday Objects using Reinforcement-Learning with Automatic Value Cut- Off,” Proceedings of the 2007 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 1551-1556, San Diego, CA, USA, Oct. 29 - Nov 2, 2007.

[48] H. Dang and P. K. Allen, “Stable grasping under pose uncertainty using tactile feedback,” Autonomous Robots, vol. 36, no. 4, pp. 309–330, Aug. 2013.

[49] J. Felip and A. Morales, “Robust sensor-based grasp primitive for a three-finger robot hand,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., pp. 1811–1816, Oct. 2009

[50] M. R. Dogar, and S. S. Srinivasa, “Push-grasping with dexterous hands: Mechanics and a method,” IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 2010, pp. 2123-2130, October, 2010.

[51] C. Goldfeder and P. K. Allen, “Data-driven grasping,” Autonomous Robots, vol. 31, no. 1, pp. 1–20, Apr. 2011.



10

[52] M. A. Batalin, G. S.Sukhatme, and M.Hattig, “Mobile robot navigation using a sensor network,” 2004 IEEE International Conference on Robotics and Automation, vol. 1, pp. 636-641, April, 2004.

[53] F. DcUaen, D. Fox, W. Burpd, and S. Thrun, “Monte carlo localization for mobile robots,” 1999 IEEE International Conference on Robotics and Automation, vol. 2, pp. 1322-1328, 1999.

[54] J. Jochen. Steil, and W. Heiko, “Recent Trends in Online Learning for cognitive robot,” In Proceedings of European Symposium on Artific ial Neural Networks, pp. 77-87, 2006.

[55] Ismail A L T, Sheta A, and Al-Weshah M, “A mobile robot path planning using genetic algorithm in static environment,” Journal of Computer Science, vol.4, no.4, pp.341, 2008.

[56] S. Sanjeev, “QCQP-Tunneling: Ellipsoidal Constrained Agent Navigation,” In Proceedings of the 2nd IASTED International Conference on Robotics, pp 294-301, 2011.

[57] W. Hwang, and J. Weng, “Incremental hierarchical discriminating regression for indoor visual navigation,” In IEEE International Conference on Image Processing, vol.1, pp.810-813, 2001.

[58] B. D. Gouveia, D. Portugal, D. C. Silva, and L. Marques, “Computation sharing in distributed robotic systems: a case study on SLAM,” IEEE Transactions on Automation Science and Engineering, vol. 12, no. 2, pp. 410-422, 2015.

[59] J. Salmeron-Garcia, P. Inigo-Blasco, F. Diaz-del-Rio, and D. Cagigas-Muniz, “A tradeoff analysis of a cloud-based robot navigation assistant using stereo image processing,” IEEE Transactions on Automation Science and Engineering, vol. 12, no. 2, pp. 444-454, 2015.

[60] G. Kho, C. Hung and H. Cunningham, “Robo Brain: Massive Knowledge Base for Robots,” 2014.

[61] M. Tenorth, A. C. Perzylo, R. Lafrenz, and M. Beetz, “Representation and exchange of knowledge about actions, objects, and environments in the roboearth framework,” IEEE Transactions on Automation Science and Engineering, vol. 10, no. 3, pp. 643-651, 2013.

[62] Z. Shu, J. Wan, D. Li, J. Lin, A. Vasilakos and M. Imran, “Security in Software-Defined Networking: Threats and Countermeasures,” Mobile Networks and Applications, DOI: 10.1007/s11036-016-0676-x, 2016.

[63] J. Wan, S. Tang, Z. Shu, D. Li, S. Wang, M. Imran and A. Vasilakos, “Software-Defined Industrial Internet of Things in the Context of Industry 4.0,” IEEE Sensors Journal, DOI: 10.1109/JSEN.2016.2565621, 2016.

[64] M. Chen, Y. Qian, S. Mao, W. Tang, and X. Yang, “Software-Defined Mobile Networks Security,” ACM/Springer Mobile Networks and Applications, DOI: 10.1007/s11036-015-0665-5, 2016.

[65] M. Chen, Y. Zhang, Y. Li, S. Mao, and V. Leung, “EMC: Emotion-aware Mobile Cloud Computing in 5G,” IEEE Network, vol. 29, no. 2, pp. 32-38, Mar. 2015.

[66] M. Chen, “NDNC-BAN: Supporting Rich Media Healthcare Services via Named Data Networking in Cloud-assisted Wireless Body Area Networks,” Information Sciences, vol. 284, no. 10, pp. 142-156, Nov. 2014.

[67] M. Chen, Y. Hao, Y. Li, C. Lai, and D. Wu, “On The Computation Offloading at Ad Hoc Cloudlet: Architecture and Service Models,” IEEE Communications, vol. 53, no. 6, pp. 18-24, June 2015.

[68] G. Ermacora, A. Toma, R. Antonini, and Rosa, S, “Leveraging open data for supporting a cloud robotics service in a smart city environment,” Intelligent Autonomous Systems 13, Springer International Publishing, pp. 527-538, 2016.

[69] W. Yuan, P. Deng, T. Taleb, J. Wan and C. Bi, “An Unlicensed Taxi Identification Model Based on Big Data Analysis,” IEEE Transactions on Intelligent Transportation Systems, vol. 17, no. 6, pp. 1703–1713, 2016.

[70] J. Wan, H. Yan, D. Li, K. Zhou and L. Zeng, “Cyber-Physical Systems for Optimal Energy Management Scheme of Autonomous Electric Vehicle,” The Computer Journal, vol. 56, no. 8, pp. 947-956, 2013.

Cloud Robotics: Current Status and Open Issues - AMiner

Documents

Transcript of Cloud Robotics: Current Status and Open Issues - AMiner