Robotic SLAM - a Review from Fog Computing and Mobile

Edge Computing Perspective

Swarnava Dey Arijit Mukherjee

TCS Research and Innovation TCS Research and Innovation
Kolkata, India Kolkata, India
[email protected] [email protected]

ABSTRACT for performing complex robotic tasks including multi-robot

Offloading computationally expensive Simultaneous Local- Simultaneous Localization and Mapping (SLAM), robotic
ization and Mapping (SLAM) task for mobile robots have at- vision etc. In Cloud Robotics, as elaborated in [10], [13],
tracted significant attention during the last few years. Lack mobile robots use remote cloud server over the Internet as re-
of powerful on-board compute capability in these energy source for offloading intensive computation. A robotic clus-
constrained mobile robots and rapid advancement in com- ter, introduced first in [16], is an ad-hoc group of robots
pute cloud access technologies laid the foundation for de- sharing each other’s computing resources, for parallelization
velopment of several Cloud Robotics platforms that enabled of compute intensive algorithms. Though computation of-
parallel execution of computationally expensive robotic al- floading in cloud robotics rendered complex tasks such as
gorithms, especially involving multiple robots. In this work map merging, cooperative navigation etc. feasible in multi-
the Cloud Robotics concept is extended to include the cur- robot scenario, its major disadvantage is dependence on per-
rent emphasis of computing at the network edge nodes along sistent connection to external network. On the other hand
with the Cloud. The requirements and advantages of using robotic clusters are self contained with persistent availabil-
edge nodes for computation offloading over remote cloud or ity in a multi-robot scenario albeit limited in both compute
local robot clusters are discussed with reference to the ETSI and memory capacity compared to servers. The efficacy of
‘Mobile-Edge Computing’ initiative and OpenFog Consor- such a robotic cluster is yet to be proved for compute and
tium’s ‘OpenFog Architecture’. A Particle Filter algorithm memory intensive tasks. As an example in a multi-robot
for SLAM is modified and implemented for offloading in a SLAM, merging local maps created by multiple robots into
multi-tier edge+cloud setup. Additionally a model is pro- a global one, will require at least one of the robots to be
posed for offloading decision in such a setup with experi- powerful enough to merge the whole map in memory along
ments and results demonstrating the efficacy of the proposed with routine sensing and actuation tasks.
dynamic offloading scheme over static offloading strategies. Cloud robotics in its current form assumes continuous con-
nectivity to the back-end cloud infrastructure. However, in
real world situations, especially in disaster scenarios, such
CCS Concepts assumption does not hold. To handle the intermittent cloud
•Computer systems organization → Embedded sys- connectivity and meet the requirements of latency sensi-
tems; Redundancy; Robotics; •Networks → Network reli- tive applications, researchers in Mobile Cloud Computing
ability; (MCC) proposed Fog Computing [3] that extends Cloud
Computing to include resources at the edge of the network.
Keywords Along with different MCC frameworks, industry led stan-
dards like Mobile Edge Computing (MEC) [9] from ETSI,
Mobile Edge Computing; Fog Computing; SLAM; offload- OpenFog Architecture are currently under standardization
ing; robotics; and it seems running robotic middleware in LTE base sta-
tions, routers, gateways etc. is a possibility in near future.
1. INTRODUCTION In field robotics, a mobile robot builds a map of an un-
Mobile robots are generally constrained devices in terms known environment and tries to locate itself on that map.
of processing power, storage capacity and energy. Cloud This process is known as SLAM and it is computationally
Robotics and Robotic Clusters are the two prevalent ap- expensive given the noise in observation and robot poses.
proaches used to augment mobile robot’s processing power Algorithm for Grid Mapping with Rao Blackwellized Parti-
cle Filter (RBPF) SLAM [7] is an improvement over stan-
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
dard RBPF SLAM. This algorithm produces more accurate
for profit or commercial advantage and that copies bear this notice and the full cita- robot localization and maps if executed on powerful com-
tion on the first page. Copyrights for components of this work owned by others than puting resources. Data parallel execution of this SLAM is
ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or re- also studied extensively in recent past [12], [6].
publish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from [email protected]. The current work tries to leverage a multi-tier cloud +
MOBIQUITOUS ’16 Adjunct Proceedings, November 28-December 01, edge computing setup to offload computationally expensive
2016, Hiroshima, Japan portions of the above mentioned SLAM algorithm. The edge

c 2016 ACM. ISBN 978-1-4503-4759-4/16/11. . . $15.00 server (e.g. MEC or a router with computation capability)
DOI: http://dx.doi.org/10.1145/3004010.3004032

153
takes the offloading decision based on computation, commu- SLAM to cloud infrastructure. In [2], authors present a
nication load and energy usage with an objective of minimiz- framework for offloading images to the OpenStack cloud us-
ing the overall execution time of a SLAM algorithm. ing ROS [22] topics based message passing for performing
The main contribution of this paper is demonstration of compute intensive operations like feature extraction, neigh-
the efficacy of a Fog/MEC computing platform for perfor- borhood matching from image database and getting back
mance enhancement of SLAM for mobile robots. A Particle location update for the robots. RoboEarth [26] is a cloud
Filter algorithm for SLAM is modified and implemented for robotics framework that provides mechanisms for knowledge
offloading in a multi-tier edge+cloud setup. Additionally sharing among robots. It can store multirobot workflows, as-
a model is proposed for offloading decision in such a setup sign subtasks and monitor execution of those tasks on mo-
with experiments and results demonstrating the efficacy of bile robots. Rapyuta [11], the cloud engine for RoboEarth
the proposed dynamic offloading scheme over static offload- is a popular open source cloud robotics platform. It repre-
ing strategies. sents a mobile robot through a lightweight virtual machine
in cloud server. In a recent work [12] on parallelizing Rao-
2. RELATED WORK Blackwellized PF SLAM [7], an Apache Storm based stream
processing framework is enhanced to support offloading of
Limited processing power, energy and memory in mobile compute intensive scan matching segment to a set of data
robots has triggered the development of parallel SLAMs and parallel distributed cloud servers. Detailed profiling is per-
use of external resources for executing scalable robotic tasks. formed to find the effects of I/O overhead and parallel be-
The State of the Art (SoA) survey for the current work is havior of the algorithm. Based on the analyses authors sug-
presented in the following sequence: gested a cut-off on the number of times the scan matching
• parallel multithreaded SLAM in a single machine, is done, without compromising much on the quality of the
generated map.
• distributed SLAM using cloud computing or cluster The cloud robotics platforms are less suitable in scenarios
computing, where connectivity to external network is intermittent and
applications have low latency requirements, such as disaster
• distributed SLAM with computation offloading in edge situations as mentioned before. A robotic cluster [16], an
and cloud server. ad-hoc group of robots sharing each other’s computing re-
sources, was introduced as a low cost, scalable alternative to
Parallel execution of different SLAM algorithms are at-
solely cloud based execution of robotic workflows. In [6], au-
tempted either using parallel programming constructs or
thors used a robotic cluster to implement a RBPF SLAM [7].
leveraging data parallelism. In 2010, authors of [5] used
The authors developed a distributed message queue based
GPU based Compute Unified Device Architecture (CUDA)
framework to offload the most compute intensive segment
to track features and compute visual odometry from stereo
of the SLAM running in one robot to other robots. De-
camera and implement a real time Visual SLAM (VSLAM).
tailed benchmarking was performed both with and without
Monte Carlo localization uses the motion model on past lo-
maintaining offloading context in the server robots taking
cations to find best matches on a known map using a Particle
standard datasets of odometry and laser based observation.
Filter (PF). In 2011, authors of [20] used GPS with vehicle
In this work the SLAM executes in only one of the robots
speed to enhance the algorithm and executed that on GPU
and other robots are used as compute resources. In gen-
using OpenMP to get large performance improvement. An-
eral, the major drawback of a robotic cluster based mul-
other CUDA based parallelization of a PF based SLAM [25],
tirobot mapping is the lack of memory and computation
presented in 2013, splits the PF SLAM such that the com-
resources for final merging of individual maps generated by
putation heavy particle weight computation is performed in
robots. Though SLAM with computation offloading in edge
the GPU and sensing, map update etc. operations are ex-
and cloud server setup is not attempted earlier, such prior
ecuted in the CPU. To reduce data transfer between GPU
works [4, 18, 19] exist in mobile cloud offloading. In [4]
and CPU a map partitioning scheme is used, resulting in a
authors consider a joint optimization involving cloud server
28 times increase in performance of the weight calculation
and an computationally capable access point.
step. In 2015, Parallel Tracking and Mapping (PTAM), a
In the current work, usage of an intermediate server entity
VSLAM variant was extended to stereo PTAM (S-PTAM)
along with cloud is advocated for offloading compute inten-
to overcome several limitations of PTAM in [21]. Among
sive segments of a SLAM algorithm. Such edge servers are
the different subtasks in S-PTAM like feature extraction
active area of research in MEC, Fog Computing.
from camera image, matching image features with existing
map, pose refinement etc., keypoint detection and descrip-
tor extraction are computationally heavy. The parallel im- 3. BACKGROUND
plementation separates out the tracking and image analysis In this section the building blocks of a edge computing
phase to produce real time response and results comparable setup and the SLAM algorithm to be offloaded in such a
to SoA. Though the parallel algorithms summarized above setup are discussed.
are mostly implemented on a single machine, those form the
basis for distributed SLAM frameworks, as offloading to ex- 3.1 Mobile Edge Computing
ternal nodes requires identification of the parallel segments MEC [9] is a key emerging technology under the process of
of a program. standardization by ETSI that allows hosting of low latency
Leveraging cloud nodes for parallel execution of robotic services at the edge of a RAN. Of the different sites where
algorithms is a well researched topic and the reader is redi- MEC services can be deployed, LTE eNodeB (the transceiver
rected to a 2015 survey [14] for detailed information. In that directly communicates with mobile handsets) and the
recent past several researchers have attempted offloading multi-RAT access points seems most suitable for hosting

154
robot algorithms. Multi-RAT access points [23] can aggre- and based on a threshold the particles are resampled
gate data in specific indoor / outdoor locations using hetero- to replace particles with low weight.
geneous access technologies (Wifi, bluetooth, cellular etc.)
helping the robots to communicate with the SLAM/robotic • An occupancy grid map, with grid positions repre-
algorithms hosted in the access point. The perceived ben- sented as occupied / not-occupied, is updated with
efits of offloading computationally intensive segments of a robot trajectory and observations.
robotic SLAM on MEC servers are: (i) latency reduction The parallel nature of the algorithm is shown in Fig. 1 where
in processing of robot motion and observations with com- solid arrows indicate forward flow and dashed arrows return
putation outsourced in MEC server, (ii) memory availabil- control. Despite being an old algorithm (2007), it is quite
ity in MEC server for global map merging, storing robot popular among researchers owing to its data parallel nature,
task workflows etc., without involving core cloud, (iii) use of rendering it suitable for implementing in parallel processing
specialized communication primitives provided by LTE cells frameworks.
(refer to Connected Cars use case in [9]) and (iv) availability
of standardized MEC platform API for service development 4. OFFLOADING FRAMEWORK
with guaranteed SLA.
The computation offloading system presented in this sec-
3.2 Fog Computing tion is modelled around the algorithm presented in section 3.3.
Like MEC, OpenFog Consortium Architecture Working As multiple particles are required to build accurate maps
Group is also envisaging the usage of resources at the edge and mobile robots are resource constrained, some particles
of the network in Fog Computing [8]. The edge servers can needs to be offloaded to external resources. The system
be routers, gateways or even MEC hosting infrastructure. model, numerical model and the implementation of such an
When standardized, expected benefits in terms of offload- offloading framework is presented in the following subsec-
ing robotic workflows are: (i) resource discovery and regis- tions.
tration can be useful for finding out capabilities of heteroge- Once particles are created and initialized the same set of
neous robots as well as edge compute resources, (ii) current computations, as shown in the parallel block of Fig. 1 are
workload based pooling of compute resources from edge and performed for each particle. For the current work the com-
end user computing resources and (iii) further offloading of putation happening inside each particle for processing a new
robotic tasks to core cloud or other Fog cluster scan it is considered as an unit of work. Another assump-
Though MEC and Fog promise to come up with several tion made here is about the particle processing time and the
new features suitable for field robotics, the mere usage of link data transfer time between resources. The particle process-
nodes in the network edge to augment mobile robot’s com- ing time is assumed to be same for all particles operating
putation is a value add and is hitherto untapped in available on the same dataset, executing in the same category of re-
cloud robotic frameworks. source (e.g. edge or cloud or a particular type of robot).
Particle processing time is known beforehand through ex-
3.3 Algorithm for Grid Mapping with RBPF ecution profiling on each resource type. The data trans-
SLAM fer time is calculated from bandwidth between resources.
The SLAM considered here for partitioning and offloading Though variable bandwidth can be handled [24], here the
is an improvement over RBPF SLAM and is detailed in [7]. bandwidth between robot and edge and between edge and
A simplified flow of the SLAM along with the parallel seg- cloud are assumed to be constant. For offloaded execution,
ments are presented here. SLAM problem requires a mobile the communication overheads associated are: (i) sending the
robot to build a map of its surroundings and at the same odometry and observations from robot to edge and edge to
time localize itself on that map. The robot uses its mo- cloud, (ii) transfer of updated location cloud to edge and
tion model, observations and earlier poses to predict its new edge to robot and (iii) particle migration (including map)
pose and correct the pose from noise and uncertainty. In after resampling from each of the resource to other resource.
this SLAM particle filters are used to predict current poses 4.1 Assumptions & System Model
of the robot, with multiple PFs tracking the current pose of
The edge executes the offloading decision and decides how
the robot simultaneously. PFs are Bayes filters capable of
many particles will be processed in the robot, cloud and
handling nonlinear probability distribution of mobile robot
itself. The decision algorithm executes in three different
motion and the concentration of particles in a particular area
scenarios - (i) during initial SLAM algorithm run for de-
denotes that the chances of the robot taking that particular
ciding the number of particles to be assigned to each re-
pose is more. The major steps of this SLAM are:
source, (ii) on each resampling as some particles are invali-
• As the particle filter uses a nonlinear probability distri- dated and new particles are added, resulting in migration of
bution function (PDF) for estimating the current robot some particles with associated maps and (iii) finally, in case
pose, a new pose must be sampled from the PDF. How- of a connectivity failure to core cloud. All the above scenar-
ever, the non-parametric nature of the PDF requires ios are applicable for each robot and our work is in progress
the algorithm to sample the new pose from a proposal for defining a joint decision making strategy involving the
normal probabilistic motion model based distribution, robots as well.Map for each particle of a robot is built and
for each particle. updated in the robot/edge /cloud and the map of the best
• The ratio of sampled pose and the target pose is used particle is used as the selected map.
to assign a weight to a particle . 4.2 Numerical Model for Offloading Decision
• If the variance of the weight of the particles are more, Let the following constants and variables be considered for
a scattered representation of robot’s pose is denoted representing the formal model for the offloading decision:

155
 Observations
Last Particle 1 1.Use the best
Particle 2
odometry particle to up-
date Map
1. Match predicted pose with last map 2. Location up-
2. Sample new pose from motion & observation date to robot
Create/ Set params to
3. Update particle weight 3. Process next
Recreate/ particles Particle k
4. Update map reading
Invalidate
Particles Parallel
Last robot Pose, Map Particle n Resampling
& Particle Weight

Figure 1: The Improved Grid Mapping RBPF SLAM algorithm with parallel segment.

Pr , Pe , Pc denote a particle’s scan processing time in

robot, edge , cloud respectively. Ore denote the time to send
T = min((x1 ∗ Er + x2 ∗ We + x3 ∗ Wc )
observation and odometry data from robot to edge and Oec (1)
from edge to cloud for further offloading. Lre denote the +ρ ∗ max(x1 ∗ Pr , x2 ∗ Te , x3 ∗ Tc ))
time to send location update from edge to robot and Lec
such that:
from cloud to edge . For each particle’s scan processing
time at edge is denoted by Te = Pe + Ore + Ler and at cloud
by Tc = Pc + Ore + Oec + Lec + Lce The system & numerical x1 + x2 + x3 ≤ X (2)
is shown in Fig. 2.
x1 ∗ ω ≤ E thr (3)

x1 ∗ Er ≤ E thr (4)

x1 ∗ Eo ≤ E thr (5)

max(x1 ∗ Pr , x2 ∗ Te , x3 ∗ Tc ≤ T res (6)

and
xi ≥ 0i = 1, 2...X(integers) (7)
The objective function is a weighted sum of cost to offload
and time to process the particles, where ρ is the weight. The
first constraint stipulates that the summation of assigned
particles in the resources must be less than the total num-
ber of particles used to run that SLAM algorithm. The
second constraint stipulates that the maximum number of
particles that can be assigned to the robot is bounded by an
Figure 2: System & Numerical model for SLAM offloading energy threshold (E thr ) where ω is a factor relating energy
consumed for processing a particle at robot. The third and
For energy consumption and execution cost at edge or fourth constraints are essentially same as the second one,
cloud, the approach of a prior work by Chen et. al. [4] but take into account explicit energy usage for particle pro-
is followed. For each particle, let Er denotes the process- cessing and data transfer by the robot. The fifth constraint
ing energy in the robot and there is no other cost involved. relates to the real time response required (T res ) for the robot
Eo denotes the energy overhead to transfer observation and to get location update to ensure moving at a certain speed.
odometry and receive location update from robot. Ce and The no cloud connectivity scenario can be modelled in
Cc denote the scan processing cost per particle in the edge the same way as above and is comparatively simple. The
and the cloud respectively. For each particle the weighted objective function is provided below omitting the details due
sum of processing cost and data transfer energy overhead to space constraint:
at edge is denoted by We = Eo + α ∗ Ce and at cloud by
T = min((x1 ∗ Er + x2 ∗ We ) + ρ ∗ max(x1 ∗ Pr , x2 ∗ Te ))
Wc = Eo + β ∗ Cc where α and β are the relative weights.
x1 , x2 and x3 denote the number of particles assigned to As stated earlier this is a work in progress and the exper-
robot, edge and cloud and X is the total number of particles iments are not conducted on real robots. The details of the
used for running the SLAM algorithm. simulated experimental setup is given in the next section.
Formally, the edge server solves the following integer linear
program shown below under the initial SLAM algorithm run
scenario given in section 4.1 for getting the values x1 , x2 , x3 : 5. RESULTS & DISCUSSION

156
5.1 Experimental Environment Fig. 5 and Fig. 6 with maximum of 120 particles used for
Virtual Machines (VM) of different configurations are used running the SLAM algorithm. The mean execution time
for representing different category of resources - robot VM of the parallel block of Fig. 1 is plotted against particles
with dual core 1.2GHz CPU and 512 MB memory, edge VM used. In robot only execution at most 20 particles could
with two 2.6GHz CPU cores and 4GB memory, cloud VM
with four 2.6GHz CPU cores and 8GB memory. Odometry
and observation are streamed from flat files in robot VMs
using standard SLAM datasets [15] available from here [1].
Only a subset of the offloading optimization formulated in
equation 1 is evaluated
min(max(x1 ∗ Pr , x2 ∗ Te , x3 ∗ Tc ))
setting the bandwidth between robot and edge as 100 Mb/s
and between edge and cloud as 15 Mb/s. Execution time of
particles are profiled beforehand in all category of resources.
The execution profiling in edge is done along with running
the PuLP [17] modeller, used for solving proposed optimiza-
tion.

140
Number of particles used for SLAM

Figure 5: Evaluation on ACES Building dataset

140
Number of particles used for SLAM

Figure 3: Evaluation on MIT Killian Court dataset

140
Number of particles used for SLAM

Figure 6: Evaluation on Freiburg Indoor Building 079

dataset

be used. Execution for most particle sizes was possible for

static offloading to the edge server , showing benefit of of-
floading when cloud is not connected. The proposed offload-
ing scheme almost always outperformed the static offloading
strategies.

6. CONCLUSIONS
In this work the issue of computation offloading to a Mo-
140
bile edge / Fog computing setup is addressed taking mobile
Number of particles used for SLAM
robot’s SLAM algorithm as an example. In a setup con-
sisting of robot, edge and cloud server a dynamic offload-
Figure 4: Evaluation on Intel Research Lab dataset ing scheme was proposed. An integer linear programming
problem was formulated and partially evaluated to imple-
The subset of the optimization given in section 5.1 is eval- ment the dynamic offloading decision strategy. Throughout
uated on four different datasets as shown in Fig. 3, Fig. 4, the presentation of the offloading system, several challenges,

157
benefits of using MEC / Fog computing are highlighted in [13] B. Kehoe, S. Patil, P. Abbeel, and K. Goldberg. A
context of cloud robotics. Experiments conducted on vir- survey of research on cloud robotics and automation.
tual machines showed that the proposed offloading scheme IEEE Transactions on Automation Science and
almost always outperformed the static offloading strategies. Engineering, 12(2):398–409, April 2015.
The major contribution of the work is demonstration of per- [14] B. Kehoe, S. Patil, P. Abbeel, and K. Goldberg. A
formance enhancement of robotic SLAM using servers at survey of research on cloud robotics and automation.
network edge. To realize the possibility of running robotic IEEE Transactions on Automation Science and
middleware in LTE base stations, routers, gateways etc. sev- Engineering, 12(2):398–409, 2015.
eral research challenges need to be addressed. Working with [15] R. Kümmerle, B. Steder, C. Dornhege, M. Ruhnke,
field robots on recent SLAM algorithms and refining the of- G. Grisetti, C. Stachniss, and A. Kleiner. On
floading strategy will be explored as a continuation of this measuring the accuracy of slam algorithms.
work. Autonomous Robots, 27(4):387–407, 2009.
[16] A. Marjovi, S. Choobdar, and L. Marques. Robotic
clusters: Multi-robot systems as computer clusters: A
7. REFERENCES topological map merging demonstration. Robotics and
[1] Slam benchmarking. 2016. Autonomous Systems, 60(9):1191–1204, 2012.
[2] P. Benavidez, M. Muppidi, P. Rad, J. J. Prevost, [17] S. Mitchell, M. OSullivan, and I. Dunning. Pulp: a
M. Jamshidi, and L. Brown. Cloud-based realtime linear programming toolkit for python. The University
robotic visual slam. In Systems Conference (SysCon), of Auckland, Auckland, New Zealand, http://www.
2015 9th Annual IEEE International, pages 773–777. optimization-online. org/DB FILE/2011/09/3178.
IEEE, 2015. pdf, 2011.
[3] F. Bonomi, R. Milito, P. Natarajan, and J. Zhu. Fog [18] A. Mukherjee, H. S. Paul, S. Dey, and A. Banerjee.
Computing: A Platform for Internet of Things and Angels for distributed analytics in iot. In Internet of
Analytics, pages 169–186. Springer International Things (WF-IoT), 2014 IEEE World Forum On,
Publishing, Cham, 2014. pages 565–570. IEEE, 2014.
[4] M.-H. Chen, B. Liang, and M. Dong. A semidefinite [19] T. Nishio, R. Shinkuma, T. Takahashi, and N. B.
relaxation approach to mobile cloud offloading with Mandayam. Service-oriented heterogeneous resource
computing access point. In 2015 IEEE 16th sharing for optimizing service latency in mobile cloud.
International Workshop on Signal Processing In Proceedings of the first international workshop on
Advances in Wireless Communications (SPAWC), Mobile cloud computing & networking, pages 19–26.
pages 186–190. IEEE, 2015. ACM, 2013.
[5] B. Clipp, J. Lim, J. M. Frahm, and M. Pollefeys. [20] K. Par and O. Tosun. Parallelization of particle filter
Parallel, real-time visual slam. In Intelligent Robots based localization and map matching algorithms on
and Systems (IROS), 2010 IEEE/RSJ International multicore/manycore architectures. In Intelligent
Conference on, pages 3961–3968, Oct 2010. Vehicles Symposium, pages 820–826. IEEE, 2011.
[6] B. D. Gouveia, D. Portugal, D. C. Silva, and [21] T. Pire, T. Fischer, J. Civera, P. De Cristóforis, and
L. Marques. Computation sharing in distributed J. J. Berlles. Stereo parallel tracking and mapping for
robotic systems: A case study on SLAM. IEEE Trans. robot localization. In Intelligent Robots and Systems
Automation Science and Engineering, 12(2):410–422, (IROS), 2015 IEEE/RSJ International Conference on,
2015. pages 1373–1378. IEEE, 2015.
[7] G. Grisetti, C. Stachniss, and W. Burgard. Improved [22] M. Quigley, J. Faust, T. Foote, and J. Leibs. Ros: an
techniques for grid mapping with rao-blackwellized open-source robot operating system.
particle filters. IEEE transactions on Robotics, [23] R. Wang, H. Hu, and X. Yang. Potentials and
23(1):34–46, 2007. challenges of c-ran supporting multi-rats toward 5g
[8] O. C. A. W. Group. Openfog architecture overview. mobile networks. IEEE Access, 2:1187–1195, 2014.
2016. [24] R. Wolski, S. Gurun, C. Krintz, and D. Nurmi. Using
[9] e. a. Hu, Y. C. Mobile edge computingâĂŤa key bandwidth data to make computation offloading
technology towards 5g., etsi white paper 11. 2015. decisions. In Parallel and Distributed Processing, 2008.
[10] G. Hu, W. P. Tay, and Y. Wen. Cloud robotics: IPDPS 2008. IEEE International Symposium on,
architecture, challenges and applications. IEEE pages 1–8. IEEE, 2008.
Network, 26(3):21–28, May 2012. [25] H. Zhang and F. Martin. Cuda accelerated robot
[11] D. Hunziker, M. Gajamohan, M. Waibel, and localization and mapping. In Technologies for
R. D’Andrea. Rapyuta: The roboearth cloud engine. Practical Robot Applications (TePRA), 2013 IEEE
In Robotics and Automation (ICRA), 2013 IEEE International Conference on, pages 1–6. IEEE, 2013.
International Conference on, pages 438–444. IEEE, [26] O. Zweigle, R. van de Molengraft, R. d’Andrea, and
2013. K. Häussermann. Roboearth: Connecting robots
[12] S. Kamburugamuve, H. He, G. Fox, and D. Crandall. worldwide. In Proceedings of the 2Nd International
Cloud-based parallel implementation of slam for Conference on Interaction Sciences: Information
mobile robots. In Proceedings of the International Technology, Culture and Human, ICIS ’09, pages
Conference on Internet of Things and Cloud 184–191, New York, NY, USA, 2009. ACM.
Computing, ICC ’16, pages 48:1–48:7, New York, NY,
USA, 2016. ACM.

158