Application and development of artificial intelligence and computer technology in the field of intelligent robots

Table of Contents

Development of AI, computer technology and intelligent robots

In the development of AI, its core is how to produce intelligent machines^12,13. From the perspective of computer, it is to use computer programming system to simulate human intelligent behavior. In this way, the intelligent robot can achieve its desired functions, develop new functions by itself, and explore its psychological response¹⁴.

There are many contents of computer technology, as shown in Fig. 1.

The substance of computer technology can be broadly categorized into four categories, as shown in Fig. 1: computer systems, computer equipment, computer component technology, and computer assembly technology.

The development process of China’s intelligent robot industry is a typical progressive development model. At the initial stage, it is mainly through the introduction of advanced technologies and talents from various countries, gradually establishing an independent research and development system, and exploring robot operation, virtual reality, etc^15,16. China’s intelligent robot industry started to enter a phase of rapid expansion during the middle and late 1990s as a result of external environment changes like reform, opening up, and World Trade Organization entrance. China has independently produced a number of intelligent robot products with the most sophisticated level in the world after over 20 years of rapid development and innovation in the country’s intelligent robot sector.

At present, China’s independently manufactured industrial intelligent robots and service-based intelligent robots have been widely used in the industrial field, and are booming with the help of AI, machine learning and other technologies¹⁷. Especially in some special environments, such as high risk environment, high noise, high temperature, high radiation and other harsh operating conditions, the application effect is very good. It has also been widely used in the service field, and has gradually entered the civil field.

Most of the defects of previous intelligent robots are shown in Fig. 2.

As shown in Fig. 2, in the case of low environmental stability, the intelligent robot has failures in various situations. For example, it is difficult to adjust its own emotions in a strange environment and needs help from others. Its brain structure is not complete, and its knowledge reserve is not large. It can only carry out short-term memory and cannot absorb previous experience. Therefore, there is still a great possibility of mistakes. Good service awareness, similar thinking ability and basic thinking ability are lacking, which makes it difficult to adapt to the changing situation. The communication with people is not very harmonious, and the mechanical system cannot communicate well with people.

The relationship between AI and robot intelligent pathfinding is close and complementary. AI provides robots with the ability to handle complex tasks, make decisions, and adapt to the environment, especially in dynamic and unknown environments. Intelligent pathfinding, as the core issue of robot autonomous navigation, relies on algorithms and technologies in AI, such as deep learning, fuzzy logic, genetic algorithms, etc., to resolve uncertainties in path planning, optimize path selection, and avoid obstacles. AI enables robots to perform real-time analysis and decision-making based on environmental data, improve pathfinding efficiency and accuracy, and thus achieve more efficient automated operations. Therefore, AI is not only the technical basis for intelligent robot pathfinding, but also a key factor in improving robot autonomy and flexibility.

Application of navigation

FNN can effectively improve the accuracy of robot navigation when faced with complex and dynamic environments¹⁸. Intelligent pathfinding usually requires the robot to find an optimal path in an unknown or changing environment, and fuzzy neural networks can handle uncertainty and fuzzy information in the environment through the combination of fuzzy logic and neural networks. FNN simulates the decision-making process of the human brain and combines fuzzy rules to help robots make more flexible judgments when perceiving environmental data, optimize path selection, avoid obstacles, and improve navigation effects. In intelligent pathfinding, FNN not only enhances the robot’s decision-making ability, but also adjusts navigation strategies in real time according to different environmental conditions, thereby improving the robot’s adaptability and autonomy in complex environments.

Fuzzy rules are used to construct the neural network, which ensures that the weight of the network has obvious physical meaning, and maintains the adaptability of the neural network. On this basis, combined with the experience accumulated by human navigation experts, the network weight is adjusted adaptively to achieve better real-time intelligent robot navigation control effect.

First, the inputs and outputs of neurons in fuzzy neural networks are formally described. Then, the input of each neuron is defined as: ${f^{(w)}}(x_m^{(w – 1)},\cdots,x_m^{(w – 1)};u_{ko}^w,\cdots,u_{ko}^w)$, and the output of the node is: $x_k^{(w)} = {g^{(w)}}({f^{(w)}})$. The superscript w represents the layer number.

The first layer is the input layer, which contains m input values. The node is used as the input node to represent the variables of the input language. Its function is to transfer the input value to the next layer, namely $f_o^{(1)} = {x_o}$. The connection weight of the first layer is 1.

The second layer is fuzzification, whose function is to fuzzify the input variables. This paper applies Gaussian function to radial basis function, namely:

$$F = \frac{{{{(x_o^{(2)} – {v_{ok}})}^2}}}{{\delta _{ok}^2}}$$

(1)

In Formula (1), ${v_{ok}}$ and ${\delta _{ok}}$ are respectively expressed as the center and width of the Gaussian function of the k-th term of the o-th input language variable. If there is accurate input at the input level, then the corresponding membership can be obtained at the second level.

The third node is the rule node that represents fuzzy logic rules. This level of connection is the prerequisite or prerequisite for establishing fuzzy logic, that is, the calculation of the scope of application of each rule. Rule nodes are used for fuzzy processing:

$$f = \min (x_1^{(2)},x_2^{(2)},\ldots,x_m^{(2)})$$

(2)

The connection right of this layer is 1.

The network connection of the fourth layer defines a generated rule node, and each rule corresponds to the input result. It is standardized. The number of nodes is the same as the third layer, that is:

$$f = \sum\limits_{o = 1}^n {x_o^{(3)}} ,{\bar \alpha _k} = \frac{{x_o^{(3)}}}{f}$$

(3)

The fifth layer is the output layer after defuzzification to generate all intelligent robot control rules, namely:

$${y_l} = \sum\limits_{o = 1}^n {{u_{lk}}} {\bar \alpha _k}$$

(4)

The application of neural networks in intelligent robot navigation is mainly reflected in its ability to make autonomous decisions by learning complex environmental patterns. Traditional navigation methods rely on fixed rules and preset paths, while neural networks can make real-time adjustments based on dynamic changes in the environment and sensor inputs, and have greater adaptability and flexibility. Especially in FNN, fuzzy logic is used to process uncertain data inputs in the environment, enabling robots to make effective navigation decisions under complex or ambiguous environmental conditions. For example, SOM and BPNN can learn environmental features and generate suitable navigation paths through neural networks, thereby improving the robot’s positioning accuracy and path planning capabilities.

In addition, the advantages of neural networks in navigation are also reflected in their powerful self-learning capabilities. Through training on a large amount of historical data, neural networks can automatically optimize their weight parameters, continuously accumulate knowledge from experience, and gradually improve the accuracy and efficiency of navigation. The design based on ART neural networks can effectively handle pattern recognition problems that robots may encounter in complex environments, and further enhance the stability and reliability of the navigation system. This makes neural networks particularly suitable for application in dynamic, unknown or rapidly changing environments, improving the performance and autonomy of robots in actual tasks.

3.3 Application of path planning

The path planning of intelligent robots has been extensively studied by numerous academics, who have produced a number of methods¹⁹. The efficacy of intelligent robot path planning is guaranteed by the AI technique. Specifically, the use of GA in intelligent robot path design has raised the intelligent robot’s level to a new level and made its motion trajectory essentially consistent with the anticipated outcomes²⁰. Researchers are thoroughly examining the path application method of intelligent robots while also consistently implementing and enhancing GA. They have proposed two genetic operations, crossover operation and mutation operation. Intelligent robots’ job efficiency can be maximized by utilizing a range of evolutionary methods.

(1)

Structure of neural network motion controller.

This research adopts a multilayer feedforward neural network and uses neural networks as the motion controller for intelligent robots. The neural network motion controller’s operating duration is depicted in Fig. 3.

Based on this, the neural network is continuously trained and merged with the input sensor data to efficiently avoid obstacles during each cycle of operation. The intelligent robot can constantly approach the target point and finish the path planning task since the input data also includes the angle information between the target distance and the target location.

It is assumed that the position coordinate of the intelligent robot in the workspace at the moment ${t_o}$ (the current moment) is $({x_o},{y_o})$, and the next action behavior is determined by the neural network control system. In this paper, the output of the neural network at this moment is regarded as the rotational driving force of the left and right wheels of the intelligent robot (assuming ${f_z}$, ${f_r}$), and the current moving direction angle is ${\theta _o}$. In this way, under the action of the driving force, the incremental $\Delta {\theta _o}$ of the moving direction angle of the intelligent robot is:

$$\Delta {\theta _o} = {f_z} – {f_r}$$

(5)

The angular increment of the motion direction is strictly limited here to ensure that there is no sharp turning during the motion and the generated trajectory is smooth. The movement direction angle increment $\Delta {\theta _o}$ is rigidly specified, and its value is clamped between (− 0.2, 0.2) by a function, so that the movement direction angle at the next moment can be obtained as:

$${\theta _{o + 1}} = {\theta _o} + \Delta \theta$$

(6)

For the moving speed of the intelligent robot at the next moment, the following formula can be used to calculate:

$${b_{o + 1}} = {f_z} + {f_r}$$

(7)

Therefore, the position coordinates of the working space of the intelligent robot at the next moment are:

$$({x_{o + 1}},{y_{o + 1}}) = ({x_o} + {b_{o + 1}} \times ( – {\sin _{o + 1}}),{y_o} + {b_{o + 1}} \times \cos {\theta _{o + 1}})$$

(8)

In order to determine the intelligent robot’s motion trajectory, the location data at each time point is simultaneously simultaneously saved in a vector array and represented by a red rectangular block.

In path planning, neural networks help intelligent robots find the best path in complex or dynamic environments by learning and optimizing path features in different environments. Specifically, the role of neural networks is reflected in the following aspects: First, using neural networks for environmental perception, through real-time data input obtained by sensors (such as the location, distance, current location of the robot, etc.), neural networks can identify obstacles, open areas and possible paths in the environment. Using structures such as CNN, robots can automatically learn and extract important features from perceived map data. These features help robots determine the best driving direction in real time and avoid collisions with obstacles. In addition, neural networks can also accumulate historical data through training, identify and optimize common path planning patterns, and enable robots to quickly choose appropriate routes in different environments.

Secondly, when combining genetic algorithms for path optimization, neural networks play a key decision-making role. Genetic algorithms search the path planning space by simulating the natural selection mechanism to find the shortest and safest path. However, genetic algorithms usually require a lot of computing resources, so neural networks can help improve the efficiency of the algorithm. After the neural network training is completed, it can quickly evaluate the feasibility of each path and automatically provide efficient initial populations or optimized path selection for the genetic algorithm. In this way, the neural network can speed up the path planning process and ensure that the planned path is not only short but also smooth, avoiding obstacles and reducing turning angles, thereby improving the efficiency and stability of the robot in practical applications. In a dynamic environment, the neural network also has a real-time feedback mechanism that can dynamically adjust the path according to changes in sensor input during the robot’s driving process, ensuring that the robot can adapt to sudden obstacles in a timely manner and flexibly adjust its route.

(2)

Genetic manipulation.

1)

chromosome coding

In the neural network motion controller, the connection weight between neurons is encoded as an individual represented by a real number sequence. Its length is equal to the weight in the neural network. Any floating value with (− 1, 1) attribute before the start of gene evolution cycle.

2)

fitness function design

The path planning of intelligent robot is to find a safe and effective path without conflict between a specific starting point and a target in a work environment with certain obstacles. Generally, there are many roads that meet this condition. In practice, it is often necessary to select the best route according to a certain standard, that is, short route, short planning time, smooth route, less conflicts, etc. On this basis, this paper synthesizes the above four performance indicators and gives a new fitness function:

$$f = (7000 – t) + (1200 – d) + {q_v}/30 + {q_s}/20$$

(9)

Here, t is the time required for each intelligent robot to reach the destination. d is the distance between the intelligent robot and the target point in the static state. When the steering angle is lower than the specified value without collision with obstacles, the value increases by 1 accordingly.

3)

selection operation

GA completes the selection of individuals in the group through selection operation. Some individuals are selected from their parents in a certain way and passed on to their offspring. In the competition selection method, M are randomly selected from the group (M is 4 here). Then, the best one is selected as the parent, and the selection is repeated. At the same time, because the optimized individuals are interfered by the crossover operation or mutation operation in the process of evolution, this paper adopts an individual with higher adaptability to make the number of excellent individuals in the group more, so as to optimize the way of evolution.

4)

cross operation

The primary strategy for creating new people in GA is the crossover operation. The two-point crossover approach is applied in this research. Among them, the selection of the location of the intersection point is not arbitrary. However, the weight of each neuron is regarded as an independent whole, and the location between these whole is regarded as the intersection position.

It is supposed that the number of neurons is M, and each small rectangular block in the chromosome is a neuron. Each neuron contains ${T_1},{T_2},{T_3},\ldots,{T_M}$ weight value. In this paper, a function is used to calculate the $M – 1$ possible intersection positions and store the return value of this function. Then, the two intersection positions are determined by randomly selecting the storage positions.

5)

mutation operation

Non-uniform variation is adopted, and the absolute value of variation does not exceed the given maximum value.

(3)

Algorithm design and implementation

1)

Initialization of the intelligent robot group: The starting point position is set to (25, 25), the target position is (570570570), and the initial movement direction angle is 0 (the rotation driving force of the left and right wheels is 0.16). During the movement, the rotational driving force of the left and right wheels is arbitrary (− 0.2, 0.2). The intelligent robot starts from the starting point. The connection weight of the original neural network motion controller is a randomly generated floating point number in the range of (− 1,1).
2)

During the operation, if the number of predetermined operation frames for each generation is not exceeded, the obstacle information sensed by the sensor, the distance information of the target point and the angle information of the target point and the current moving direction are used for input by the motion controller of the neural network. According to the output of the neural network, the position of the intelligent robot is continuously updated to complete the search of the obstacle avoidance path of the target point, otherwise go to 3).
3)

By solving the fitness degree of each weighted individual, the optimal average fitness degree of the current generation is found.
4)

On this basis, the newly generated weight individuals are replaced with new weight nodes. Then, the newly generated next-generation weight individual is used to replace the old individual as the new connection weight of the neural network motion controller.
5)

Step 1) to step 4) are carried out circularly until the evolution algebra reaches a predetermined value
6)

when the intelligent robot reaches the destination, the planned route is recorded

In path planning, the specific steps of GA include the following important links. First, initialize the population* and generate a certain number of random paths. Each path is represented by a chromosome, and each gene of the chromosome represents a node or a path segment on the path. Then, fitness evaluation is performed. Each path is evaluated by the fitness function. Common evaluation criteria include the total length of the path, whether obstacles are avoided, the smoothness of the path, and energy consumption. The higher the fitness value, the better the quality of the path. Next, the selection operation is performed. Some individuals are selected for mating according to the fitness value. The selection method can be roulette selection, tournament selection, etc., aiming to retain excellent individuals.

After selecting excellent individuals, a crossover operation is performed, that is, a new path is generated by exchanging part of the parent path information. For example, a single-point crossover method is used to randomly select a point on the path, and then the part of the parent path before and after the point is exchanged. Then a mutation operation is performed to increase the diversity of the population by randomly changing a part of the path (such as exchanging the order of path nodes, reversing path segments, etc.), so as to avoid the algorithm from falling into a local optimum. Finally, the population is updated, a new child path is generated and compared with the parent path, and a path with higher fitness is selected to enter the next generation. If the termination condition is met (such as reaching the maximum number of iterations or the fitness of the solution is good enough), the iteration is stopped and the optimal path is output.

The algorithm combining FNN and GA has significant advantages in intelligent robot navigation and path planning. This paper uses FNN to process uncertainty and fuzzy information in the environment to achieve high-precision navigation; at the same time, GA is used to optimize path planning to ensure the global optimality of the path. The output of FNN is used as the initial population of GA, and the path optimized by GA is fed back to FNN to dynamically adjust the navigation strategy. This collaborative optimization mechanism not only improves navigation accuracy, but also significantly shortens path planning time, solving the limitations of a single technology in complex environments. However, this combined algorithm also has certain disadvantages: the training process of FNN is relatively complicated and requires a large amount of sample data and computing resources; although GA can find the global optimal solution, its convergence speed is slow, especially when the problem scale is large, it may take a long time to search and optimize. Therefore, this combined algorithm needs to balance accuracy and computational efficiency in practical applications to achieve the best effect.

link