Safe Predictive Mobile Robot Navigation in Aware Environments
Michael Arndt and Karsten Berns
Robotics Research Lab, Dept. of Computer Sciences, University of Kaiserslautern, Kaiserslautern, Germany
Keywords:
Mobile Robot Path Planning, Safety, Human Motion Prediction, Human Aware Planning, Ambient Intelli-
gence, Smart Environment
Abstract:
It is a common goal to improve safety and performance of mobile indoor robots by predicting the movements
of people in the surroundings. In contrast to many related works which exclusively employ sensors mounted on
mobile robots, this work shows a method to achieve this goal in a smart environment where external sensors are
used to sense people’s positions. By using probabilistic models and filters, the evolution of the environment’s
state is predicted and optimal paths with respect to safety and performance are planned. Experiments in
reality and in a simulation environment show the applicability in real-world scenarios and the advantages over
classical path planning approaches.
1 INTRODUCTION
It has been acknowledged by many researchers that
safety and human awareness are important topics for
mobile robots employed in human environments. As
(Alami et al., 2006, sec. I) state, safety is an impor-
tant key to successfully introduce robots into such do-
mains. The concept of human aware motion planning
has been researched a lot in the past (Sisbot et al.,
2005; Sisbot et al., 2007; Bennewitz et al., 2005;
Guzzi et al., 2013) to achieve robot motion that does
interfere with human motion as little as possible.
At the same time, so called smart, aware, am-
bient or ubiquitous technologies are becoming more
and more important and present in everyday life (Au-
gusto et al., 2010; Lukowicz et al., 2012) and also
the field of robotics tries to make use of that tech-
nology (Saffiotti and Broxvall, 2005; Coradeschi and
Saffiotti, 2006; Sanfeliu et al., 2008; Amato et al.,
2012).
Previous work by the authors has already shown
that the risk, a mobile robot imposes on human in-
habitants of smart environments, can be decreased by
making use of environmental sensor information. By
assessing the current situation, as perceived by wire-
less sensor nodes, distributed in the environment, the
path planner of a mobile robot is able to find paths that
are safe according so a safety metric, but also efficient
subject to the length of the path (Arndt and Berns,
2014). However, the previous method yields subopti-
mal results, as it only evaluates possible paths accord-
ing to the current situation at the time of planning.
While it provides good results for setups that remain
static after the path has been planned, this assump-
tion is of course not true in general, as environments
inhabited by people are typically highly dynamic.
In reality, people frequently change their loca-
tions, thus the environments evolve during the traver-
sal of a path of the mobile robot (i.e. while it is driv-
ing), thus the environment is time variant. This fact
will now be taken into account within this work by
making use of prediction.
When planning with respect to time, the dimen-
sionality of the planning problem increases. This
leads to an increase in complexity, which makes the
search for exact solutions often unfeasible (Linde-
mann and LaValle, 2005). However, the method pre-
sented later in this work is able to find exact solutions
in typical application scenarios with acceptable com-
putation times.
The rest of the paper is structured in the following
way: First, work related to this one is introduced in
the next section. Afterwards, a brief overview about
the probabilistic state estimation methods that are em-
ployed, is given, followed by a description of the
essence of this work, the prediction of future states
and the planning using the outcome. Section 5 de-
scribes the experiments that have been conducted in
reality as well as in a simulation environment to val-
idate the proposed algorithms in detail. The exper-
imental results are followed by a conclusion and an
outlook over possible future works.
15
Arndt M. and Berns K..
Safe Predictive Mobile Robot Navigation in Aware Environments.
DOI: 10.5220/0005509500150023
In Proceedings of the 12th International Conference on Informatics in Control, Automation and Robotics (ICINCO-2015), pages 15-23
ISBN: 978-989-758-123-6
Copyright
c
2015 SCITEPRESS (Science and Technology Publications, Lda.)
2 RELATED WORK
Path planning in time variant, or dynamic environ-
ments has been known for quite a while and has al-
ready been researched in different domains. This sec-
tion aims to give an overview about some techniques
and algorithms that can be found in the literature.
The work by (Alvarez et al., 2004) has presented a
solution for a predictive path planner for autonomous
underwater vehicles in ocean environments. Due to
the fact that littoral waters are highly variable and
routes for the underwater vehicles should be energy
efficient, forecasts are used to predict the conditions
of the ocean in advance (Alvarez et al., 2004, sec. I)
to achieve better plans. In their work, the authors
acknowledge the high dimensionality of the tempo-
ral planning problem and thus the difficulty of finding
computationally feasible exact solutions. Instead, in
their work, they make use of genetic algorithms and
accept possibly non-optimal solutions.
Regarding mobile robots in indoor environments,
the tracking of people (by the robot itself) is an im-
portant foundation of planning in general (Bruce and
Gordon, 2004) and human aware planning in particu-
lar. The authors acknowledge that people usually fol-
low trajectories between points of interest and do not
just walk randomly (Brownian motion) or with a con-
stant velocity through the environment. In that work,
they have used training data (trajectories of people in
the environment) to learn goal locations of people in
the environment. Later, in their motion model, they
calculate paths from people’s current locations to the
possible goal locations to be able to predict their mo-
tion and improve their tracking, compared to a simple
Brownian motion model.
The work by (Bennewitz et al., 2005) has simi-
lar goals as the one by Bruce et al. and is also highly
related to this work. It describes a way to learn mo-
tion patterns of people from laser scanner data and to
derive hidden Markov models (HMMs) from the pat-
terns. They later use the HMMs to estimate current
as well as also future positions of people. The authors
evaluated their approach in different settings. Exper-
imental results show that the HMMs can be used to
estimate the position of a single person and also mul-
tiple persons in the environment, even during times of
occlusions (Bennewitz et al., 2005, 5.2.1 - 5.2.2).
Especially related to this work is their experiment
to improve the navigation behavior of the robot using
the motion patterns (Bennewitz et al., 2005, 5.3). An
A
search is performed for the least-cost path in the
time-space configurations of the robot with predicted
people trajectories, leading to a significant reduction
in the time needed for the robot to perform a naviga-
tion task with prediction enabled.
An approach with a two-step prediction method
was presented by (Foka and Trahanias, 2010). They
describe an approach to enhance the path planning ca-
pabilities of a mobile indoor robot by trying to predict
future motion of humans in the environment. Their
solution makes predictions on two levels. The short-
term prediction only tries to predict the position of a
sensed human in the next time step. More interesting
in the context of this work is the long-term predic-
tion, which works with the help of having defined so-
called hot points in the environment. Those are points
where people will most likely end their paths through
the environment, e.g. stairs, elevators or seats. These
points can be either manually defined or automatically
learned. The connections between tracked people and
hot points in their field of view will be interpreted as
possible paths they take and those position in the map
will be penalized for the robot, so it is less likely that
robot and person will meet.
Another very interesting work by (Luber et al.,
2010) employs the concept of social forces to predict
the future actions of people, especially in highly pop-
ulated environments. The social forces model the be-
havior of people in crowds and are used to improve
predictions during multi-target tracking in their work.
The authors have proven the concept by evaluating the
number of data association errors in their multi-target
tracker during occlusion of targets. Using the social
force model instead of just a constant velocity model
enabled them to reduce the number of data association
errors up to 50%, depending on the scenario. How-
ever, in that work, their predictions are only used for
improving the tracking (by improving the data asso-
ciation step), and not for the long-term path planning,
in contrast to this work.
3 STATE ESTIMATION
The state of the smart environment is estimated us-
ing Bayesian filters. The whole process is described
in detail in the previous work (Arndt and Berns,
2012b, Section 2), so this section will just give a brief
overview about the process and it explains modifica-
tions and improvements relevant to this work. For the
full details, the reader is referred to the previous work.
People in the environment are detected using
AmICA wireless sensor nodes (Wille et al., 2010)
equipped with passive infrared (PIR) sensors. These
sensor-events are fed to a probabilistic filter which
tracks the positions of people in the environment.
The internal state space of the probabilistic filter
is an undirected graph G(V,E) similar to the one used
ICINCO2015-12thInternationalConferenceonInformaticsinControl,AutomationandRobotics
16
by (Liao et al., 2003). Let the vertices of this graph
be v
i
V and the edges e
i j
E. The people that
are tracked are assumed to be moving on these edges
(edge e
i j
is the one connecting vertices v
i
and v
j
).
While the possibility of having complex models to
describe people’s movements in the environment has
been mentioned previously (Arndt and Berns, 2012b),
the authors have only used simple ones until recently.
For this work, the probabilities for moving from one
edge to another edge have been specified for interest-
ing vertices of the motion graph. Formally, the prob-
abilities p(e
jl
|e
i j
) for transferring from edge e
i j
to
edge e
jl
are now initialized non-uniformly for some
vertices. This allows to model paths that are much
less likely to be traversed by humans as others.
For the experiments in this work, the graphs as
well as the transition probabilities have been manu-
ally defined. However, it should be noted, that there
are various methods available in the literature to infer
the layout of such a graph and the respective transi-
tion probabilities from observations of the environ-
ment. For example, (Bennewitz et al., 2005) define
resting places between which trajectories are learned
and (Ikeda et al., 2012) use a very similar sub-goal
concept to learn positions at which people take deci-
sions on their further paths.
4 PREDICTIVE PATH PLANNING
4.1 State Prediction
To understand how the evolution of the environment is
predicted, it should first be recapped how probabilis-
tic filters are used to track objects. They generally
work in two steps: first, the current state is predicted
using models for motion and afterwards the belief is
updated with new incoming sensor information.
This separation into two steps is the reason why
these filters are very well suited for the long term pre-
diction of the evolution of the environment. If the
update step is simply skipped, the filter will contin-
uously predict the state of the environment. For this
case, the well-known recursive Bayesian update rule
can then be re-written as in (1). In the formula, η is
the normalizing constant while x
t1
and x
t
represent
the previous and the current state of the environment,
respectively.
belief(x
t
) = η
Z
p(x
t
|x
t1
) belief(x
t1
)dx
t1
(1)
In the implementation of this work, this process
is decoupled from the on-line tracking of the people.
When a new path has to be planned, a copy of the cur-
rent belief of the on-line tracker is created and only
this copy is then iteratively updated to “see” into the
future, leaving the rest of the tracking framework un-
affected.
4.2 Evaluation of Risk
To find safe paths in the environment, the notion of
safety is important. This work will define the risk im-
posed on a human being by the robot exactly as in
(Arndt and Berns, 2014, sec. 3.1), as it has shown to
be a useful metric. To recap, the function that gives
the cost in terms of risk is a parameterized piecewise
polynomial according to (2).
f
a,b
: [0;] [0;1]
f
a,b
(d) =
(
1 (a
1
d)
b
d a
0 d > a.
(2)
The function assigns a risk-cost for each distance
d between the human and the robot. It will be max-
imal at d = 0, i.e. when robot and person coincide
and monotonically decrease till zero. The parame-
ters a and b are used to control the exact shape of the
curve. Parameter a reflects the distance from which
on the cost will be zero, i.e. it defines which distance
is assumed to be perfectly safe. The second parame-
ter b > 0 defines the steepness of the cost curve. For
the experiments later in this work, the cost function
f
8,2
(d) has been used. An illustration of that instance
of the function is given in Figure 1.
0 2 4
6
8 10
0
0.5
1
Distance [m]
Cost
f
8,2
(d)
Figure 1: Plot of the cost function for a = 8 and b = 2.
As outlined in Section 3, the state estimation of
people’s positions is done on a graph structure. The
risk function is thus applied to the belief (e.g. rep-
resented by particles) and the outcome is a separate
graph structure, the cost graph, with edge weights
representing the risks associated with driving along
the corresponding edge (Arndt and Berns, 2014, sec.
3.3.2).
In a last step, the edge weights of this risk cost
graph must be fused with the ones of the classic dis-
SafePredictiveMobileRobotNavigationinAwareEnvironments
17
tance cost graph, to obtain a resulting graph that al-
lows to find paths that have minimal costs in terms
of both safety and distance. This fusion takes two
weights α and β as arguments which can be used to
adapt the preference for safe or short paths, see (Arndt
and Berns, 2014, sec. 3.4) for details.
4.3 Planning
Path planning for the mobile robot takes place on the
same graph which is also used to constrain the human
motion and where the tracking of people takes place.
This has shown to be very convenient, as only a single
representation of the state space has to be maintained.
It was noted before, that the complexity of tem-
poral path planning is high and thus not computation-
ally feasible for all applications. However, for typical
graphs, our research has shown that it is perfectly fea-
sible to do the prediction for all simple (without hav-
ing cycles) paths from start to goal and to finally select
the one with the least predicted costs. This subsection
goes more into detail of how the candidate paths are
evaluated for the resulting cost.
For each candidate path
1
(v
1
,..., v
n
) between the
start vertex v
1
and the goal vertex v
n
, the algorithm
AnalyzePath (see Figure 2) will be called with the
current belief at the time of planning. This algorithm
further needs the two parameters l
step
and v
robot
which
define the step length for discretization and the as-
sumed robot velocity, respectively. The step length
is important, because it defines the granularity of the
planning process in the time-dimension.
Roughly speaking, the algorithm will simulate the
robot driving a distance of l
step
and then predict the
future at the next point of time. Setting l
step
to a
smaller value increases the time resolution of the pre-
diction, but will of course also slow the algorithm
down. Reasonable values for the two parameters are
e.g. l
step
= 1m and v
robot
= 0.5
m
s
, leading to a time of
two seconds between predicted steps.
As long as the end of the path has not yet been
reached, the filter is advanced and the next segment
(the distance l
step
) will be analyzed with the algo-
rithm AnalyzeSegment (see Figure 3). This algo-
rithm takes the current edge e
i j
, the current linear po-
sition t on the vertex, the distance to travel d and the
current accumulated cost c as arguments. It traverses
the current edge for the specified distance d. If it can
stay on the edge, the fraction of the traversed distance
of the edge’s weight is summed up to the cost c and
the control is returned to the higher level algorithm.
If, however, the end of the current edge is reached it
1
Every simple path from start to goal qualifies as a can-
didate path.
will first check if the end of the path has been reached
in which case it returns and also the AnalyzePath al-
gorithm will terminate. If the path is not ending, the
algorithms calls itself recursively for the next edge till
the requested distance has been traversed or the end of
the path has been reached. At the end, when the single
costs of all candidate paths are known, the one with
the lowest overall costs (as returned in Algorithm 2
line 13) is selected as the resulting path.
To make these steps more intuitively accessible,
the rest of this section is dedicated to an example of
the algorithm applied to a real-world path-planning
task. Figure 4 shows the output of the “classic” path
planner that uses probabilistic information (as de-
scribed in (Arndt and Berns, 2014)), but does not do
any prediction about the evolution of the state.
From first glimpse, this path does indeed look
quite reasonable, as it aims to avoid the region of high
probability for a person being present (red patches on
the heatmap). If, however, the prediction is enabled,
the path looks different and at first “worse” (i. e. more
expensive), see Figure 5 which depicts the situation
at the time of planning. In this and the following fig-
ures, the blue circle will indicate the predicted robot
position.
In the subsequent figures 6 and 7, the filter is ad-
vanced as the robot (hypothetically) moves along the
planned path. Based on the motion model of the prob-
abilistic filter, the belief takes certain “paths” of the
graph. It can be easily seen that the chosen path
Require: l
step
> 0 {the step length for discretization}
Require: v
robot
> 0 {the assumed robot velocity}
Require: (v
0
,..., v
n
), n > 1 {the path to be ana-
lyzed}
Require: belief(x) {the initial belief}
Returns: Total cost of the path
1: AnalyzePath((v
0
,..., v
n
),belief(x)) :
2: InitializeFilter(belief(x)) {initialize the proba-
bilistic filter with the initial belief}
3: time delta = l
step
/v
robot
4: {start on the beginning of first edge of the path}
5: t 0
6: e e
v
0
v
1
7: c 0 {start with zero costs}
8: end reached false
9: while ¬end reached do
10: belief(x) AdvanceFilter(time delta)
{advance probabilistic filter for the simulated
amount of time}
11: (end reached,e,t,c)
AnalyzeSegment(e,t,l
step
,c)
12: end while
13: return c
Figure 2: Function to analyze a whole path.
ICINCO2015-12thInternationalConferenceonInformaticsinControl,AutomationandRobotics
18
Require: (v
0
,..., v
n
) {the whole path}
Require: d > 0 {remaining distance to travel}
Require: e
i j
E : w
e
i j
0 {all edge weights ac-
cording to the current belief}
Require: Successor(e) {function that finds the suc-
cessor of edge e on the current path}
1: AnalyzeSegment(e
i j
,t,d, c) :
2: l = EdgeLength(e
i j
)
3: t
0
= t + (d/l) {calculate the new position on
edge}
4: if t
0
> 1 then
5: {new position is beyond current edge}
6: c c + (1 t) · w
e
i j
{sum up the fraction tra-
versed on current edge to the costs}
7: if ¬∃Successor(e
i j
) then
8: return (true, e
i j
,t,c) {there is no successor
of edge e
i j
, the path is ending}
9: end if
10: e
0
= Successor(e
i j
) {find the next edge}
11: t
0
= 0
12: d
0
= d (1t)·l {calculate remaining length}
13: if d
0
< 0 then
14: return (false,e
0
,t
0
,c) {no more distance to
travel, end recursion}
15: end if
16: return AnalyzeSegment(e
0
,t
0
,d
0
,c)
17: else
18: c c + (d/l) · w
e
i j
{sum up the fraction tra-
versed on current edge to the costs}
19: return (false,e
i j
,t
0
,c) {as we are still on the
same edge e
i j
, just use new linear parameter
t
0
}
20: end if
Figure 3: Function to analyze a segment of a path.
mainly follows edges of the graph that have very little
probability of a person being present at the predicted
time in the future. If the original shortest path found
without prediction (Figure 4) is re-evaluated using the
predicted situation depicted in Figure 7, it can be real-
ized that the classic “shortest path” does not look that
good after all.
5 EXPERIMENTS
5.1 Experimental Setup
All of the conducted experiments are set in the facili-
ties of the authors’ research group, a common Univer-
sity building. It was equipped with a total of six Am-
ICA wireless sensor network nodes to provide ambi-
ent sensor data. A schematic view of the environment
Figure 4: Shortest path found by Dijkstra’s algorithm with-
out predicting the future states of the environment.
Figure 5: Evaluation of the path found with predictive path
planning at t = 0s.
Figure 6: Evaluation of the path found with predictive path
planning at t = 10s.
Figure 7: Evaluation of the path found with predictive path
planning at t = 36s (goal reached).
can be found in Figure 8. For each of the experiment
runs, the mobile indoor robot ARTOS was placed at a
fixed start position and then given the instruction to
plan a path to a fixed goal position with and without
enabled state prediction. The path was then traversed.
For the cost function in Equation 2, the parameters
have been set to a = 8 and b = 2.
Experiments were conducted in two sets. The first
one was done with the real robot in the real environ-
ment to prove the applicability of the proposed ap-
proach. While these experiments are of great impor-
tance, as no simulation environment can replace the
reality, they are unfortunately unable to provide sta-
tistically relevant data, as the count of runs is limited.
For this reason, a second set of experiments was con-
ducted in a simulated version of the same environ-
SafePredictiveMobileRobotNavigationinAwareEnvironments
19
A
B
C
Figure 8: Layout of the environment within which the ex-
periments have been conducted. The graph for tracking,
prediction and path planning has been overlaid. Important
locations that are referenced in the text: A: start position of
the robot; B: goal position; C: meeting room.
ment. This simulation was based on the SimVis3D
framework (Wettach et al., 2010) and handled every-
thing from simulation of vehicle physics to the behav-
ior of the passive infrared motion sensors as well as
basic radio communication of the sensor nodes (Arndt
and Berns, 2012a).
5.2 Experimental Results
5.2.1 Reality
In this section, two experiment runs with the real
robot in the real smart environment will be presented.
During the first run, a person (the author) was being
present in the meeting room all the time, “blocking”
the path through that room. As the models indicate
that, if someone is present in that room, the person
is unlikely to leave it soon, the robot predicts the sit-
uation and thus takes the path through the hallway.
Figure 9 shows the initial situation at the time of plan-
ning. The current belief as estimated by the particle
filter shows a high probability in the area of the meet-
ing room, which reflects the situation well. The posi-
tion of the robot itself, as predicted or localized by its
own localization system, respectively, is marked by a
dot and the resulting best path is visualized by arrows.
Figure 10 shows the robot while traversing the
hallway. It is important to note that the first part of
the illustration (Subfigure 10(a)) represents the pre-
Figure 9: Experiment run one, initial situation at t = 0 with
resulting best path (according to the prediction).
diction at the time of planning while the second part
(Subfigure 10(b)) is a representation of the actual sit-
uation at the time the real robot passes that point. It
can be seen that the belief in reality differs from the
predicted belief, but this is due to the fact that the real
belief is updated using new sensor values while the
prediction does not have these information available.
Nevertheless, the similarities are clearly visible. Fig-
ure 11 shows an actual photo of the robot at approxi-
mately the marked position.
(a) Prediction
(b) Reality
Figure 10: Experiment run one, robot traversing the hall-
way.
Figure 11: ARTOS following the path through the hallway,
situation roughly corresponding to Figure 10.
Fast forwarding further, the robot can be seen
reaching its goal position in Figure 12. Again, this
figure can be consulted to evaluate the accuracy of
the prediction. By comparing it to reality, it can be
seen that both beliefs about the situation differ more
strongly than in in the previous illustration, but they
are still close to each other. In both of them, the prob-
ability for the person still being present in the meeting
room is rather large.
ICINCO2015-12thInternationalConferenceonInformaticsinControl,AutomationandRobotics
20
(a) Prediction
(b) Reality
Figure 12: Experiment run one, robot reaching the goal.
Figure 13: Experiment run two, initial situation at t = 0 with
resulting best path (according to the prediction).
For the second experiment run in reality, a person
was walking through the southern part of the hallway
at the time of planning (the exact ground truth is not
known and also not indicated in the figures). For this
situation, the models indicate that it is very likely that
the person will continue its path through the hallway
and consequently, the predicted evolution of the state
leads the robot’s path planner to choose a path through
the meeting room to minimize the chances of collision
and maximize the safety of the path. The initial situa-
tion shows the belief focused in the hallway, although
not as highly concentrated as for the previous run, see
Figure 13.
The situation after some time has passed is illus-
trated in Figure 14. The belief in reality has changed
compared to the initial situation, but not as drastically
as predicted by the planner. Still, the prediction is not
wrong, as it states that the person is most likely lo-
cated in the hallway.
Finally, in Figure 15 the situation when the robot
reaches the goal is depicted. Just as in the previous
run, the prediction further deviates from the reality,
(a) Prediction
(b) Reality
Figure 14: Experiment run two, robot traversing the meet-
ing room.
(a) Prediction
(b) Reality
Figure 15: Experiment run two, robot reaching the goal.
slowly evolving into a uniform distribution over the
whole state space.
Regarding the computational feasibility it should
be noted that during the experiments, everything
2
was
calculated on-line and on-board of the robot. A WiFi
connection to the robot was only used for visualiza-
tion and initialization purposes. The calculation of the
best path took about six seconds for the shown graph
2
Including the reception of frames from the wireless
sensor network, the processing of the probabilistic filters as
well as the planning with prediction
SafePredictiveMobileRobotNavigationinAwareEnvironments
21
on an (at the time of writing) eight-year old ultra-low
voltage processor
3
.
5.2.2 Simulation
As already outlined at the beginning of this section,
the simulation experiments were used to conduct a
large number (compared to the manually conducted
experiments in reality) of runs with prediction en-
abled and also disabled. These runs were then statis-
tically analyzed to compare the traditional path plan-
ning with the proposed temporal path planning in
terms of safety.
Due to the use of the simulation environment, the
ground truth of the person’s position was available
and thus the evaluation of the safety, computed just
as in (Arndt and Berns, 2014, sec. 4.2.2), was much
easier. Four different trajectories for people have been
simulated, two of them represent the same situations
as with the real robot: walking through the southern
hallway (situation 1) and staying in the meeting room
(2). The other two are different as they do not be-
have like modeled in the filter, but take paths that are
less likely: starting in the meeting room and leaving
it through the door at the west (3) or the door at the
east (4), respectively.
For all the classes, the mean duration of the traver-
sal d, the mean distance between robot and person x
as well as the mean risk-cost c are given in Table 1.
The standard deviations of the corresponding values
are suffixed with a subscript σ. A total sum of 113
experiment runs have been used to calculate these val-
ues.
Table 1: Comparison of the effects of prediction using sim-
ulated experiment runs.
Sit. Pred. d [s] d
σ
x [m] x
σ
c c
σ
1 Off 74 27.8 10.2 1.8 0.20 0.10
On 132 23.7 12.3 1.5 0.09 0.07
2 Off 64 3.4 7.8 0.8 0.27 0.06
On 63 4.6 9.9 0.6 0.16 0.04
3 Off 62 3.5 9.5 0.3 0.17 0.01
On 66 3.5 9.8 0.1 0.17 0.01
4 Off 63 5.0 8.5 1.2 0.23 0.12
On 69 4.0 10.1 0.1 0.09 0.01
In all regarded situations, the active prediction in-
creases the mean distance x between robot and per-
son and thus decreases the risk-costs c associated with
that path, although the effect is not that distinct in sit-
uation three. However, it must also be noted that the
reduced risk often comes at the cost of an increased
3
Intel Core Duo U2500 at 1.20GHz
duration of traversal, although not always as promi-
nent as in situation one.
6 CONCLUSIONS AND FUTURE
WORKS
Building on previous work that defines metrics for
safety and how to find a trade-off between short and
safe paths, a method to predict the evolution of the
environment of a mobile robot, using external sensor
information from a smart environment, has been in-
troduced. The theoretical concept has been validated
with a real mobile robot in a smart environment. Re-
sults of the experiments show that (given good mod-
els) the evolution of the environment can indeed be
predicted to a degree that allows to optimize the path
planning in mobile robots. The simulation experi-
ments aim to analyze the concept in a more statistical
fashion, because one may of course argue that the ex-
periments in reality lack diversity. Nevertheless, the
results of these experiments also validate the princi-
ple.
The accuracy of the prediction could in the fu-
ture be increased using more sophisticated models
that also take into account the intentions people have
while moving through their environment, similar to
the goal-orientation described by (Bruce and Gordon,
2004). More complex models could then also be
trained using learning algorithms (as outlined by e.g.
(Foka and Trahanias, 2010)) instead of being hand-
crafted.
Additionally, it would be very interesting to de-
crease the computational complexity of the proposed
approach. While being perfectly feasible for the
examined examples, there might be situations with
much larger graphs or search-spaces in which it might
be needed to switch to heuristics or to find more ef-
ficient algorithms to be still able to plan paths in the
spatio-temporal domain.
REFERENCES
Alami, R., Albu-Schaeffer, A., Bicchi, A., Bischoff, R.,
Chatila, R., Luca, A. D., Santis, A. D., Giralt, G.,
Guiochet, J., Hirzinger, G., et al. (2006). Safe and
dependable physical human-robot interaction in an-
thropic domains: State of the art and challenges. In
Proc. IROS, volume 6.
Alvarez, A., Caiti, A., and Onken, R. (2004). Evolutionary
path planning for autonomous underwater vehicles in
a variable ocean. Oceanic Engineering, IEEE Journal
of, 29(2):418–429.
ICINCO2015-12thInternationalConferenceonInformaticsinControl,AutomationandRobotics
22
Amato, G., Broxvall, M., Chessa, S., Dragone, M., Gen-
naro, C., L
´
opez, R., Maguire, L., Mcginnity, T.,
Micheli, A., Renteria, A., O’Hare, G. M., and Pec-
ora, F. (2012). Robotic ubiquitous cognitive net-
work. In Novais, P., Hallenborg, K., Tapia, D. I., and
Rodr
´
ıguez, J. M. C., editors, Ambient Intelligence -
Software and Applications, volume 153 of Advances
in Intelligent and Soft Computing, pages 191–195.
Springer Berlin Heidelberg.
Arndt, M. and Berns, K. (2012a). Mobile robots in smart
environments: The current situation. In Levi, P.,
Zweigle, O., H
¨
außermann, K., and Eckstein, B., ed-
itors, Autonomous Mobile Systems 2012, Informatik
aktuell, pages 39–47. Springer Berlin Heidelberg.
10.1007/978-3-642-32217-4 5.
Arndt, M. and Berns, K. (2012b). Optimized mobile in-
door robot navigation through probabilistic tracking
of people in a wireless sensor network. In Proceedings
of the 7th German Conference on Robotics (Robotik
2012), pages 355–360, Munich, Germany. VDI Ver-
lag, Berlin.
Arndt, M. and Berns, K. (2014). Risk-reduced mobile robot
path planning in smart environments. In Proceedings
for the joint conference of ISR 2014 and ROBOTIK
2014, pages 582–589. VDE VERLAG GMBH.
Augusto, J. C., Nakashima, H., and Aghajan, H. (2010).
Ambient intelligence and smart environments: A state
of the art. In Handbook of Ambient Intelligence and
Smart Environments, pages 3–31. Springer.
Bennewitz, M., Burgard, W., Cielniak, G., and Thrun, S.
(2005). Learning motion patterns of people for com-
pliant robot motion. The International Journal of
Robotics Research, 24(1):31–48.
Bruce, A. and Gordon, G. (2004). Better motion predic-
tion for people-tracking. In Proceedings of the In-
ternational Conference on Robotics and Automation
(ICRA), New Orleans, USA.
Coradeschi, S. and Saffiotti, A. (2006). Symbiotic robotic
systems: Humans, robots, and smart environments.
Intelligent Systems, IEEE, 21(3):82–84.
Foka, A. F. and Trahanias, P. E. (2010). Probabilistic au-
tonomous robot navigation in dynamic environments
with human motion prediction. International Journal
of Social Robotics, 2(1):79–94.
Guzzi, J., Giusti, A., Gambardella, L., Theraulaz, G., and
Caro, G. D. (2013). Human-friendly robot navigation
in dynamic environments. In Robotics and Automa-
tion (ICRA), 2013 IEEE International Conference on,
pages 423–430.
Ikeda, T., Chigodo, Y., Rea, D., Zanlungo, F., Shiomi, M.,
and Kanda, T. (2012). Modeling and prediction of
pedestrian behavior based on the sub-goal concept. In
Proceedings of Robotics: Science and Systems, Syd-
ney, Australia.
Liao, L., Fox, D., Hightower, J., Kautz, H., and Schulz, D.
(2003). Voronoi tracking: Location estimation using
sparse and noisy sensor data. In Intelligent Robots
and Systems, 2003. (IROS 2003). Proceedings. 2003
IEEE/RSJ International Conference on, volume 1,
pages 723 – 728.
Lindemann, S. R. and LaValle, S. M. (2005). Cur-
rent issues in sampling-based motion planning. In
Dario, P. and Chatila, R., editors, Robotics Research.
The Eleventh International Symposium, volume 15 of
Springer Tracts in Advanced Robotics, pages 36–54.
Springer Berlin Heidelberg.
Luber, M., Stork, J., Tipaldi, G., and Arras, K. (2010). Peo-
ple tracking with human motion predictions from so-
cial forces. In Robotics and Automation (ICRA), 2010
IEEE International Conference on, pages 464–469.
Lukowicz, P., Pentland, S., and Ferscha, A. (2012). From
context awareness to socially aware computing. Per-
vasive Computing, IEEE, 11(1):32–41.
Saffiotti, A. and Broxvall, M. (2005). PEIS ecologies:
Ambient intelligence meets autonomous robotics. In
Proc of the Int Conf on Smart Objects and Ambient
Intelligence (sOc-EUSAI), pages 275–280, Grenoble,
France.
Sanfeliu, A., Hagita, N., and Saffiotti, A. (2008). Network
robot systems. Robotics and Autonomous Systems,
56(10):793 – 797. Network Robot Systems.
Sisbot, E., Alami, R., Simeon, T., Dautenhahn, K., Walters,
M., and Woods, S. (2005). Navigation in the presence
of humans. In Humanoid Robots, 2005 5th IEEE-RAS
International Conference on, pages 181–188.
Sisbot, E., Marin-Urias, L., Alami, R., and Simeon, T.
(2007). A human aware mobile robot motion planner.
Robotics, IEEE Transactions on, 23(5):874–883.
Wettach, J., Schmidt, D., and Berns, K. (2010). Simulat-
ing vehicle kinematics with simvis3d and newton. In
2nd International Conference on Simulation, Model-
ing and Programming for Autonomous Robots, Darm-
stadt, Germany.
Wille, S., Wehn, N., Martinovic, I., Kunz, S., and
G
¨
ohner, P. (2010). Amica - design and im-
plementation of a flexible, compact and low-
power node platform. Technical report, University
of Kaiserslautern. http://ems.eit.uni-kl.de/uploads/
tx
uniklwehn/AmICATechReport2010.pdf.
SafePredictiveMobileRobotNavigationinAwareEnvironments
23