Download as pdf or txt
Download as pdf or txt
You are on page 1of 4

A simple NBV selection method for smoother

exploration of unknown environments

Adonisz Dimitriu Istvan Harmati


Department of Control Engineering Department of Control Engineering
Budapest University of Technology and Economics Budapest University of Technology and Economics
Budapest, Hungary Budapest, Hungary
d.adonisz96@gmail.com harmati@iit.bme.hu

Abstract—Next Best View (NBV) method needs to build a map of the environment in
performs well in the autonomous order to execute different tasks in the region –
exploration of unknown environments. This in some cases this task is the exploration itself.
paper proposes a novel perspective for
Frontier-based exploration [2] moves the
selecting the next best position, that
robot towards the boundaries of the known
effectively deals with corner-like segments
space. This approach has been widely
of the map and provides a smoother
investigated, and proven to be an easy and
exploration.
efficient way for dealing with unknown and
Index Terms—autonomous exploration,
partially-known terrain.
NBV, search and rescue, surveillance
Another group of exploration strategies is the
I. INTRODUCTION NBV method [3], which selects the next best
Exploration, search and rescue operations, observation point for the robot to obtain the
disaster relief, and surveillance [1] are classical most sensory information.
applications of mobile robots, that have been Information-based approaches [4] are one
researched actively since robots have the of the most studied NBV selection methods in
potentially to minimize the risks humans could the literature. It evaluates candidate positions
experience in dangerous or hardly accessible for the amount of expected new information and
environments. For these applications, the robot chooses the one that maximizes the specified
utility function. The utility or cost function can
include many objectives, such as the
information gain, path-cost, and other factors
that characterize the importance of specific task
assignments at different locations [6][7]. The
common effort is to drive the robot towards the
area that yields a large amount of new space
during exploration. This paper introduces a
method that states the opposite: positions with
less new information are preferred as the NBV.
The rest of the article is structured as
follows. Section II describes the methodology,
that is used to compare the two objectives.
Section III provides the test results of the
comparison. Section IV concludes the article.
Fig. 1: Occupancy grid map
Section V: acknowledgment.
Fig. 2: Comparison of performances in different
environments Fig. 3: Performance on 10%, 50%, 80% obstacle
concentration
II. METHODOLOGY
In the current study, an occupancy grid map the environment are preferred over large
is utilized to represent the 2D environment. In information gain areas. This approach provides
this representation, each grid has a probabilistic a smoother area coverage because agents do not
value of being occupied. Based on the need to revisit certain parts of the map (Fig. 4
occupation values, each cell is classified into 3 agains Fig. 5). For comparison, we adopt the
types of grid cells, which are occupied cells, free following information gain and position
cells, and unknown cells. An illustration of such a selection expressions from [5]:
map can be seen in Fig. 1. The robot is equipped
with a ring of sensors to detect obstacles in a Large-gain NBV Less-gain NBV
circular area (red dashed line) during the
𝐼(𝑝) = 𝐴(𝑝) ∙ 𝑒 −𝜆𝐿∗𝑐𝑝 𝐼(𝑝) = 𝐴(𝑝) ∙ 𝑒 𝜆𝑆 ∗𝑐𝑝
simulation.
𝑁𝐵𝑉𝐿 = argmax 𝐼(𝑝) 𝑁𝐵𝑉𝑆 = argmin 𝐼(𝑝)
Frontier-based exploration is motivated by 𝑝∈𝐹 𝑝∈𝐹

considering frontier cells 𝐹 (unknown cells that


where the information gain 𝐼(𝑝) consists of the
are adjacent to free cells), while information-
new expected area 𝐴(𝑝) at position 𝑝
based NBV algorithms rely on finding the next
discounted by its path-cost 𝑐𝑝 from the robot.
best position for the agent to efficiently explore
the environment. The best position is classically
defined by the information gain 𝐼(𝑝) of the
candidate positions 𝑝 – typically frontier cells –
discounted by the distance from the agent.
In the literature, many publications
emphasize utility functions that expect a large
amount of new space at the next best position
(large-gain NBV). The major disadvantage of
these approaches is that they tend to leave out
corner-like spaces in the environment, resulting
in low-resolution map segments, or in some
cases, agents must revisit these specific
locations for further exploration (Fig. 4). In this
paper, a counterintuitive approach is presented:
agents are oriented to frontier positions with
Fig.4: Large-gain NBV exploration
smaller expected information gain (less-gain
NBV). This means, that corner-like segments of
a. b. c.

d. e. f.
Fig. 5: Less-gain NBV exploration steps

The path cost – calculated with A* algorithm – is Algorithm 1: EXPLORATION ALGORITHM


the traveling distance between the robot and
Initialize position 𝑝 of the robot
the respective position. 𝜆𝐿 and 𝜆𝑆 are the tuning
parameters of the path-cost, that are while there is any frontier cell
determined experimentally. Note, that these determine frontier cells 𝐹
utility functions are arbitrarily chosen and are
for each frontier cells 𝑓 ∈ 𝐹
anti-symmetric to each other to illuminate the
qualitative difference between the two opposite 𝑐𝑓 = 𝑑𝑖𝑠𝑡(𝑝, 𝑓) – with A*
perspectives. Algorithm 1 describes the whole
𝐴(𝑓) = 𝑐𝑎𝑙𝑐𝐸𝑥𝑝𝑒𝑐𝑡𝑒𝑑𝐴𝑟𝑒𝑎(𝑓)
exploration scheme in more details.
𝐴(𝑓) ∙ 𝑒 𝜆𝐿 𝑐𝑓 , 𝑙𝑒𝑠𝑠 𝑔𝑎𝑖𝑛 𝑁𝐵𝑉
III. TEST RESULTS 𝐼(𝑓) = {
𝐴(𝑓) ∙ 𝑒 −𝜆𝐿 𝑐𝑓 , 𝑙𝑎𝑟𝑔𝑒 𝑔𝑎𝑖𝑛 𝑁𝐵𝑉
Fig. 2 reveals the overall performance of the end
two concepts in 120 different environments
argmax 𝐼(𝑓) , 𝑙𝑒𝑠𝑠 𝑔𝑎𝑖𝑛 𝑁𝐵𝑉
with growing map size and increasing obstacle 𝑓∈𝐹
concentration. Each environment was randomly NBV = {
argmin 𝐼(𝑓) , 𝑙𝑎𝑟𝑔𝑒 𝑔𝑎𝑖𝑛 𝑁𝐵𝑉
generated multiple times, with the given 𝑓∈𝐹

parameters and their results were averaged. pathToNBV = 𝑝𝑎𝑡ℎ(𝑝, 𝑁𝐵𝑉) – with A*
𝑝 ← first step of pathToNBV
𝝀𝑳 𝝀𝑺
0.2 0.2 end
TABLE I: Parameters used in the algorithms
As the figure illustrates, large-gain NBV REFERENCES
required more steps to complete the [1] J. Delmerico, E. Mueggler, J. Nitsch, and D.
exploration in all of the averaged instances. Fig. Scaramuzza, “Active autonomous aerial
3 reveals that with higher obstacle exploration for ground robot path planning,”
concentration (that means more corners on the IEEE Robot. Autom. Lett., vol. 2, no. 2, pp. 664–
map), the less-gain NBV performs increasingly 671, Apr. 2017.
better, than the large-gain NBV scheme. Table I [2] D. Holz, N. Basilico, F. Amigoni, and S. Behnke,
shows the tuning parameters of the path-cost “Evaluating the efficiency of frontier-based
utilized in the simulation. exploration strategies,” in ISR 2010 (41st
International Symposium on Robotics) and
Fig. 4 and Fig. 5 presents some phases of the ROBOTIK 2010 (6th German Conference on
different NBV exploration in a highly crowded Robotics). VDE, 2010, pp. 1–8.
environment. Note that at the bottom of the [3] A. Bircher, M. Kamel, K. Alexis, H. Oleynikova,
map, the robot bypassed a region that could and R. Siegwart, “Receding horizon” next-best-
potentally be a better choice (between b. and c.) view” planner for 3d exploration,” in 2016 IEEE
international conference on robotics and
and had to return to the location at the end (e.
automation (ICRA). IEEE, 2016, pp. 1462–1468.
and f.). This was an unavoidable move, as the
[4] F. Amigoni and V. Caglioti, “An information-
agent does not see the full map (as we do), but based exploration strategy for environment
only the discovered regions: it preferred the mapping with mobile robots,” Robotics and
next location that was more „corner-like” at that Autonomous Systems, vol. 58, no. 5, pp. 684–
particular point. It is easy to tell which direction 699, 2010.
to choose by looking at the whole map, but [5] H. H. González-Banos and J.-C. Latombe,
perceiving only the fraction of it, makes it a hard “Navigation strategies for exploring indoor
and problematic topic and is open for further environments,” The International Journal of
discussion. Robotics Research, vol. 21, no. 10-11, pp. 829–
848, 2002
IV. CONCLUSION [6] A. Visser and B. A. Slamet, “Balancing the
information gain against the movement cost for
Executing operations in a short time are crucial
multi-robot frontier exploration,” in European
in many emergency situations, hence finding the Robotics Symposium 2008. Springer, 2008, pp.
optimal trajectory is of high significance for 43–52.
mobile robots. NBV selection is the most [7] I. Arvanitakis, K. Giannousakis, and A. Tzes,
common exploration strategy because it allows “Mobile robot navigation in unknown
a wide spectrum of heuristics to be involved in environment based on exploration principles,”
the planning. Changing the perspective to less- in 2016 IEEE Conference on Control
gain NBV provides the means to effectively deal Applications (CCA). IEEE, 2016, pp. 493–498.
with corners of the map and at the same time, it
outperforms the concept of the large-gain NBV
selection in terms of exploration steps. The
proposed method achieves a smoother
exploration and can be easily adapted to other
utility functions.
ACKNOWLEDGMENT
The research presented in this paper, carried
out by BME was supported by the Ministry of
Innovation and the National Research,
Development and Innovation Office within the
framework of the Artificial Intelligence National
Laboratory Programme.

You might also like