Page 21 - Read Online

Lei et al. Intell Robot 2022;2(4):313­32 I http://dx.doi.org/10.20517/ir.2022.18 Page 327

Table 1. Comparison of path length and approximated travel time in broiler barn. The robot is assumed to have a linear velocity of 0.4
m/s
Models Path length (m) Approximated travel time (min)
Proposed model 1182 49
BCD 16,408 683

4.1. Path planning for the detection robot
The proposed DCPP was initially compared with the zigzag and spiral methods. The start and end points
for the broiler mortality robot were the same for the three methods, and our proposed method did not have
redundant return paths. The total path length of the three methods was 912 m for the proposed method, 920
m for the zigzag method, and 917 m for the spiral method, respectively. The overall path based on the DCPP
is depicted in Figure 9B.

With the assistance of DCPP, IPP is shown in Figure 9A and C. The cases are executed between two feed-
ing/drinking lines, indicated as black bars on the left and right of the figures. The robot accounts for the
overall DCPP direction and historical dead bird distribution data depicted as heat maps with red color areas
indicating a higher possibility of broiler mortality appearance. Cyan areas indicate multiple branches subdi-
vided by IPP. A path indicated as a red line designates a maximal information gain. Once the robot meets a
random obstacle (indicated as the black object in Figure 10), such as welfare enrichment (e.g., perches, straw
bales, and pecking stones) [38] , the information gain of IPP is optimized with the historical data and DCPP
direction while avoiding the obstacle. Not only is the trajectory planned with previous data, but it is also up-
dated with new data [Figure 11], resulting in a real-time efficient path for the robot. In Figure 11, the trajectory
changes from roaming between two obstacles to roaming atop the upper obstacle after the broiler mortality
distribution data are updated based on the YOLO V4 dead bird detection and localization.

Limited by the working environment in the broiler barn, it is assumed that the detection range of the robot
is 50 cm and the linear speed is 0.4 m/s. Through the traditional CCPP algorithm, such as BCD method, the
path length of the CCPP is 16,408 m, which requires approximately 683 min to complete. However, DCPP
with IPP is fine-tuned on the basis of the information gained on the paths generated by DCPP. In light of
various information distribution in different environments, the final generated global trajectory length in each
grid is only 12%-37% longer than DCPP. The final path length of the DCPP with IPP is 1182 m, which requires
approximately49mintocomplete. ComparedwiththetraditionalCCPP,ourproposedalgorithmhasashorter
pathlengthandthetraveltimealsomeetsthedemandsofbroilerbarns. ThefinalresultsareoutlinedinTable1.

4.2. Path planning of the removal robot
With the AI-based advancement of sensing techniques, locations of broiler mortality are obtained in an over-
all trajectory for the second robot. The HMTR receives the mortality location information and generates a
collision-free route to reach the targets in a reasonable and efficient sequence so that the total traveling dis-
tance is minimized. To validate the adaptability and efficiency of our algorithms in various number of targets,
three datasets were selected for simulation and comparative studies [39] . In each dataset, we iteratively per-
formed 30 executions to compute the mean and standard deviation of path length. Table 2 summarizes the
qualitative comparison results between the features of our algorithm and state-of-art models, such as genetic
algorithm (GA), particle swarm optimization (PSO), self-organizing maps (SOM), and imperialist competi-
tive algorithm (ICA). The SOM algorithm is similar to a typical artificial neural network algorithm, except
it utilizes a competitive learning process instead of back-propagation that utilizes gradient descent. The ICA
algorithm is a biologically inspired algorithm, which simulates the social-political process of imperialism and