constraints of AUV and its working environment.
platform in real marine environments.
Guerrero-González et al.
Ocean Engineering 235 (2021) 109355
13
C. Cheng et al.
Table 2
Comparison of main local path planning methods for unknown and dynamic obstacle avoidance.
Algorithm
Advantages and disadvantages
Reference
Improvement
Rapidly-exploring
Solve high dimensional space
Tan et al.
(
2005
)
Add the generation of sub nodes
Random Trees
Excellent ability to explore
Hernández et al.
(
2011
)
Reduce exploration space
The path is usually suboptimal
Hernández et al.
(
2015
)
Optimization path
Poor stability
Li et al.
(
2017
)
Add liveness attributes to nodes
Poor real-time performance
Yu et al.
(
2017
)
Optimization path
Taheri et al.
(
2019
)
/
Fu et al.
(
2019
)
Control the randomness of sub nodes
Artificial Potential
Simple structure
Ding et al.
(
2005
)
A third virtual force is introduced
Field
Good real-time performance
Saravanakumar and Asokan
(
2013
)
Consider multi-point potential field force
Fast path planning speed
Cheng et al.
(
2015
)
Consider AUV and current velocity
Easy to fall into a local minimum
Villar et al.
(
2016
)
Virtual obstacles are introduced
Oscillations near obstacles
Ge et al.
(
2018
)
Considering the cooperation of multi-AUV
Noguchi and Maki
(
2019
)
/
Khalaji and Tourajizadeh
(
2020
)
An adjustable avoidance gain is added
Fan et al.
(
2020
)
/
Lin et al.
(
2020
)
/
Fuzzy Logic
Strong flexibility
Khanmohammadi et al.
(
2007
)
/
Algorithm
Good real-time performance
Xu and Feng
(
2009
)
Combine with finite state automata
Small amount of calculation
Yang et al.
(
2009
)
Adjust the value fuzzy membership function
No precise mathematical models
Abbasi et al.
(
2010
)
Consider the speed of the obstacle
Lack of systematicness
Yang and Zhu
(
2011
)
Add acceleration braking module
Rely on expert knowledge
Fang et al.
(
2015
)
Combine with BK triangle product
Sun et al.
(
2018b
)
Optimization of the value fuzzy membership function
Li et al.
(
2019b
)
Consider changes in distance from obstacles
Neural Network
Strong learning ability
Duan et al.
(
2001
)
Combine with fuzzy logic control
Simple learning rules
Guerrero-González et al.
(
2011
)
/
Have storage capacity
Li and Guo
(
2012
)
Dynamic neural network
Nonlinear mapping capability
Yan and Zhu
(
2011
)
Bio-inspired neural network
The sample is not easy to obtain
Ding et al.
(
2014
)
Combine with leader–follower formation control
Long training time
Zhu et al.
(
2014
)
Combine with map planning method
Ni et al.
(
2017
)
Dynamic change of bio-inspired neural network
Wu et al.
(
2018
)
Add the lateral inhibition effect
Cao and Peng
(
2018
)
Combine with potential field
Sun et al.
(
2018a
)
Glasius bio-inspired neural network
Reinforcement
High generalization ability
Kawano and Ura
(
2002b
)
Add teaching Method
Learning
Strong robustness
Kawano and Ura
(
2002a
)
Establish hierarchical reinforcement learning
Without prior obstacle information
Chen et al.
(
2009
)
Combine with neural network
Strong learning ability
Huang et al.
(
2014
)
/
Dimensional disasters
Gore et al.
(
2019
)
/
Handcrafted state features
Noguchi and Maki
(
2019
)
/
Bhopale et al.
(
2019
)
Balance exploration and exploitation
Sun et al.
(
2020
)
Design Hierarchical Deep Q Network
Cao et al.
(
2020
)
Establish hierarchical reinforcement learning
Deep Reinforcement
High generalization ability
Cao et al.
(
2019
)
Asynchronous Advantages actor–critic
Learning
Reduce dimensions
Wu et al.
(
2019
)
Proximal Policy Optimization
Automatic learning state features
Hou et al.
(
2020
)
Deterministic Policy Gradient
Training takes a long time
Havenstrøm et al.
(
2021
)
Proximal Policy Optimization
the neural network can drive AUV to change its direction and ef-
fectively guide it to move between obstacles. In addition,
Hernández
et al.
(
2011
) used SPARUS AUV to test the performance of the HRRT
algorithm for obstacle avoidance in the Underwater Robotics Lab of
the University of Girona. Obstacles of different shapes are made with
insulating board and placed in the water with a depth of 3 m. The
results show that the HRRT algorithm can generate a path in less than
100 ms which meets certain real-time requirement, but cannot guide
AUV to move along the generated path autonomously.
Fan et al.
(
2020
)
used the underwater robot named ‘‘Intelligence Ocean I’’ to verify the
effectiveness of an improved artificial potential field method. In their
method, a distance correction factor is added to the exclusion function
to solve the problem of local minimum, and relative speed method is
introduced for dynamic obstacle detection and avoidance. However,
obstacle avoidance in 3D environments are not considered in their
experiments.
On the other hand, multi-AUV cooperation can not only improve
the efficiency of performing underwater tasks, but also present more
intelligent behaviors compared to a single AUV (
Li et al.
,
2014b
).
The task distribution (
Li and Zhang
,
2017
), collaborative search (
Ge
et al.
,
2018
), formation control (
Yuan et al.
,
2017
) and other aspects
of multiple AUVs have achieved good results, though most of them are
also only tested on simulation platforms (
Wu et al.
,
2018
;
Cao et al.
,
2020
). For path planning of multi-AUV in real environment, it is more
challenging because it needs to take other AUVs into account. More
importantly, the communication between AUVs is limited and usually
only underwater acoustic communication can be used for multi-AUV
cooperation, compared to other robotic platforms on land or in the
air.
Cao et al.
(
2019
) carried out multi-AUV target search experiments
with deep reinforcement learning model for obstacle avoidance in
a pool of 10 m
×
20 m. In their experiments, a remote-controlled
submersible was placed in the pool as an obstacle. Their results show
that two AUVs completed the collision-free search to the target point
with the pre-trained deep reinforcement learning model, achieving
similar performance to the simulation experiments. However, due to
the lacking of underwater acoustic communication for multi-AUV, the
experiments are only carried out on the surface of the pool with WiFi
wireless communication. Data acquisition methods of wireless sensor
networks based on mobile sink
Mehto et al.
(
2020
) might be useful
to overcome the limitation of acoustic communication for multi-AUV
communication, since AUV is a relatively resource poor device which
is similar to mobile sink.
Do'stlaringiz bilan baham: