Page 18 - Read Online

Li et al. Intell Robot 2021;1(1):58-83 I http://dx.doi.org/10.20517/ir.2021.08 Page 68

Figure 7. CCPP in a completely known environment. A: the generated robot path; B: the neural activity landscape when the robot reaches
point C [63] .

navigation with a limited reading range. Thus, the key challenge of CCPP in unknown environments is to
design the map-building algorithm and combine it with previous coverage algorithms studies. Combing with
the sensor detection, an improved CCPP algorithm based on the neurodynamics model was developed in
unknown environments [64,65] . The robots move to the nearest unclean areas and detect the environment until
the cleaning task is finished. A real-robot platform iRobot Create 2 was used to test the proposed algorithm in
unknown environments [66] . The actual cleaning robots testing showed that the effectiveness of the proposed
algorithm, inwhichtheroboticsystemscouldcooperativelyworktogetherinalargeandcomplexenvironment.

3.3. Underwater robots
The autonomous underwater vehicle (AUV) or unmanned underwater vehicle (UUV) have been studied in a
variety of tasks such as underwater rescue, data collection, and ocean exploration. In addition, some bionic
robots are also studied, such as robotic fish [67,68] . Unlike the work environment of mobile robots or cleaning
robots, the underwater environment is more complex and uncertain. Firstly, based on the 2-D neural network
structure, a 3-D grid-based neural network is typically required to represent the underwater environment. Sec-
ondly, the effect of the ocean or river currents is necessary to consider. Finally, the robots work in underwater
environments, facing many uncertainties, such as some robots broken down. Based on different task require-
ments, three major research fields of underwater robots using the neurodynamics model are studied in this
section.

3.3.1. Navigation
For the underwater environment, the neural network architecture needs to be extended to the 3-D environ-
ment, where more complex topography of randomly distributed obstacles is involved. Figure 8 shows a typ-
ical AUV path planning in 3-D underwater environments. In 2-D neural network architecture, each neuron
connects with 8 neighborhood neurons, whereas, in the 3-D neural network, each neuron connects with 26
neighborhood neurons [69] . Thus, the computation complexity dramatically increased. In order to improve the
efficiency in the 3-D underwater environment, a dynamic bio-inspired neural network was proposed to guide
the movement of AUV in large unknown underwater environments [25] . A virtual target selection approach
was applied to search the path and avoid dead loop situations. Since the large unknown environment is par-