Page 68 - Read Online

Page 63 Li et al. Intell Robot 2021;1(1):58-83 I http://dx.doi.org/10.20517/ir.2021.08

Figure 3. An example of the bio-inspired neural network.

3. PATH PLANNING
A basic path planning problem can be defined as: given a work environment with obstacles, the target, and
the initial robot position, a collision-free path should be generated from the initial position to the target. The
bio-inspired neurodynamics model has been wildly used in real-time path planning without any learning pro-
cedures and any prior knowledge of the dynamic environment. The key point of the neurodynamics-based
path planning approach is to represent the work environment as one-to-one corresponding to the neurons in
the neural network. The dynamics of each neuron in the neural network are characterized by Equation (2). An
example of a neural network is shown in Figure 3. Thus, the neural activity for the -th neuron is obtained by
( )
∑
d + + −
= − + ( − ) [ ] + [ ] − ( + )[ ] (9)
d
=1
where represents the activity value of those neighboring neurons; is the number of neighboring positions
of the -th neuron; [ ] is a linear-above-threshold function defined as [ ] = max { , 0}; and [ ] is defined
+
+
−

as [ ] = max {− , 0}. In the bio-inspired neural network, the excitatory input is consisted of two parts,
−

[ ] and ∑ =1 [ ] , where [ ] is the external input from targets, and ∑ =1 [ ] is the internal input
+
+
+
+

through the propagation of the positive activity from its neighborhoods. The inhibitory input has only

external input [ ] , which is from the obstacle, and only has local effects (no negative activity propagation).
−
Thus, the target has maximum and positive neural activity, which could globally propagate through the neural
network to attract the robot, while the obstacles have only local effects without propagating. The path selection
rule of the individual robot can be defined as: the next move position of the robot is the maximum neural
activity of its current neuron’s neighbors. After robots move to the next position, the next position becomes
a new current position until the current position is the location of the target. The robot would never choose
the position of an obstacle to be the next movement due to the negative neural activity of obstacles. Thus, the
robot is able to avoid collisions and move to the target. It is important to note that the path planning process is
without any learning procedures and any prior knowledge of the dynamic environment. Due to the real-time
performance and computational efficiency, bio-inspired neural network path planning approaches have been
developed for various robot systems. In this section, based on the different types of robots, three categories
are divided: mobile robots, cleaning robots, and underwater robots.
3.1. Mobile robots
Path planning of mobile robots has received a lot of interest because mobile robots have been participating in
human life. In this section, two main challenges of mobile robot path planning are focused: real-time collision-