Page 14 - Read Online

Page 207 Zhu et al. Intell Robot 2022;2(3):200­222 I http://dx.doi.org/10.20517/ir.2022.13

neural network algorithm.

The genetic algorithm (GA) and ant colony algorithm (ACO) were widely used in the early times for under-
water path planning. The GA method imitates the natural selection and evolution procedure to provide the
optimal solution through iterations, which has been involved in the path planning and obstacle avoidance
under the underwater environment of dynamic currents effect [48,49] . The ACO method belongs to the swarm
intelligence algorithm, where it is designed based on the swarm behavior of ant groups while chasing food, and
theantbehavior-basedintelligentmethodhasbeenproved toworkwellintheUUVglobalpathplanning [50,51] .
The swarm intelligence methods have been broadly applied in UUV path planning in recent years due to their
simple implementation, fast convergence speed, and satisfactory robustness when modeling based on different
swarming animal groups [52] . The swarm intelligence algorithms provide outstanding performance in the path
planning of UUVs, yet the local minimum problem can be produced by this intelligent method, which finally
leads to premature execution before reaching the destination.

FuzzylogicperformswellintheUUVpathplanningandobstacleavoidanceowingtoitsexpertiseinprocessing
the information uncertainty as the underwater environment is of high uncertainty and incompleteness [53,54] .
Kim et al. used the fuzzy logic-based algorithm to deduce the turning direction and angle of the UUV to avoid
collisions and complete the path planning [55] . Ali developed a fuzzy ontology modeling method to realize
the UUV path planning [56] . The fuzzy logic-based intelligent algorithm does not need to establish accurate
mathematical models as it is derived from the human cognitive experience. Thus, fuzzy logic can retune
itself during the navigation and overcome the local minimum problem. However, the fuzzy logic rule relies
heavily on experts’ experience and approximations, and unverified errors cannot be thoroughly avoided. The
complexity of the dynamic environmental factors also challenges the adaptiveness of the fuzzy logic design [57] .

The application of neural networks in vehicle path planning has obtained wide attention in recent decades [58] .
Ghatee applied the Hopfield neural network in the optimization of path planning distances [59] . Li et al. pro-
posedabio-inspiredneuralnetworkforvehiclepathplanning,wherebothoptimalplanningpathsandcollision
avoidance are realized with high efficiency [60] . The bio-inspired neural network helps to derive the optimal
path that is composed of the continuous coordinates of the vehicle movement, based on a grid-based map and
itscorrespondingneuralnetwork model, whereeachgrid representsa singleneuron, as shownin Figure5. The
bio-inspired neural network algorithm continuously updates the state of neurons by transmitting the informa-
tion through the network to give an instant reaction and reduces the complexity by limiting the searching area
to a certain range. Therefore, the bio-inspired neural network path planning utilizes the preserved information
in the neurons to update its planning design while adjusting the network on time such that it is well suited to
the dynamic underwater environment, providing an efficient and high adaptive approach for the UUV path
planning [61] .

In recent years, the application of reinforcement learning (RL) in UUV path planning has grown quickly. The
RL method updates the vehicle’s states and converges to the optimal path planning solution by making actions
according to rewards set based on the environment. RL-based path planning combined with APF for inter-
vention AUVs has been developed to remove sea urchins at an affordable cost [62] . AUV path planning in a
complex and changeable environment is achieved through the combination of RL and deep learning [63] . Wang
et al. proposed a multi-behavior critic RL algorithm for AUV path planning to overcome problems associated
with oscillating amplitudes and low learning efficiency in the early stages of training, and they reduced the
time consumed by the RL algorithm convergence for avoiding obstacles [64] . However, the slow convergence
issue of RL-based path planning methods still needs further investigation.

The methods that are commonly used in the point-to-point path planning of the UUV are summarized in
Table 2, where their implement theory, benefits, and drawbacks are described. Details of various intelligent