Robot navigation control system based on Ultra-VLN network

Abstract

The invention discloses a robot navigation control system based on an Ultra-VLN network, and belongs to the technical field of natural language processing, robot technologies and sensors, and the robot navigation control system comprises the steps that 1, an acquisition unit obtains an environment image, a human voice instruction and a current posture of a robot; 2, the processing unit converts the human voice instruction into a target location position coordinate, encodes the human voice instruction, an environment image and a robot posture, obtains an encoding sequence, and inputs the encoding sequence into an Ultra-VLN network; 3, decoding the coding sequence by the Ultra-VLN network, and generating an optimal robot action sequence according to the position coordinates of the target point; 4, the calculation unit calculates the cost consumed by the robot for executing the action sequence; and 5, the control unit controls the robot to execute the generated action sequence. According to the invention, the efficiency of responding to the human task instruction and executing the autonomous operation by the robot in a complex scene can be effectively improved.

Description

Technical Field

[0001] The present invention discloses a robot navigation and control system based on the Ultra-VLN network, which aims to improve the robot's autonomous navigation and adaptive control capabilities. It belongs to the intersection of natural language processing, robotics, and sensor technology, and is particularly suitable for automated production, intelligent manufacturing and other fields. Background Art

[0002] With the continuous development of artificial intelligence (AI) technology, robots are increasingly being used in automated production and intelligent manufacturing. Traditional robot navigation and control systems typically rely on pre-set maps and complex sensor systems. These systems still have limitations in terms of environmental changes, obstacle avoidance, and dynamic task scheduling. Therefore, improving robots' adaptive capabilities and autonomous navigation in dynamic environments is a current research hotspot.

[0003] Natural language can express the user's intentions intuitively and conveniently, so researchers have gradually invested in the study of how to use voice or text commands to control robots. Robots that understand and execute human voice commands have also become a hot research direction. Although the vision-language navigation method combined with deep learning technology has been successfully applied in the field of robot navigation and has achieved remarkable results, this type of method can perform target positioning and path planning by combining image information and language commands, enabling robots to perform tasks more intelligently and accurately. However, existing vision-language based robot navigation systems still face some challenges, especially in terms of feedback control and environmental perception. Existing technologies generally ignore how to combine the robot's vision, voice commands and environmental perception capabilities to work together, and most navigation systems lack the ability to adapt to dynamic environments. Summary of the Invention

[0004] In order to address the shortcomings of the above-mentioned existing technologies, the present invention proposes a robot navigation and control system based on the Ultra-VLN network, so that the robot can better understand human voice commands and make adaptive adjustments based on environmental changes, thereby significantly improving the autonomous navigation and control capabilities, especially enhancing the application potential in scenarios such as automated production and intelligent manufacturing.

[0005] In order to achieve the above-mentioned object, the present invention adopts the following technical solutions:

[0006] The present invention provides a robot navigation and control system based on an Ultra-VLN network. The robot comprises: a differential moving base and a depth sensor; the differential moving base comprises an inertial sensor and a collision sensor; the robot navigation and control system comprises: an acquisition unit, a processing unit, a calculation unit, and a control unit;

[0007] The human voice instructions acquired by the acquisition unit After sending it to the processing unit for preprocessing, the instruction text is obtained , the processing unit then Extract the word vector of the target location that the robot will move to , and transform the word vector into Converted to the location coordinates of the target location ;

[0008] The acquisition unit obtains the posture of the robot at the current time t collected by the inertial sensor ; At the same time, obtain the environment image set at the current time t collected by the depth sensor and send it to the processing unit together;

[0009] The processing unit processes the environment image set to obtain the obstacle image sequence L at the current time t, and records the obstacle 2D image of any k-th frame in L as , and Each obstacle in the class is labeled to obtain M obstacle feature image vectors with class labels. ,in, express The mth obstacle feature image with category label in , Indicates the total number of obstacle feature images;

[0010] The processing unit calculates the posture of the robot at the current time t After preprocessing, the preprocessed posture features at the current time t are obtained ;

[0011] The computing unit includes an Ultra-VLN network consisting of a feature extraction layer Col, a cross attention layer Ate, and a decoding layer Trans, and 、 、 Processing is performed to obtain the action sequence that the robot will perform at the next moment t+1 ;

[0012] The control unit acts according to the sequence of actions and posture features , use the control strategy function C to calculate the robot's posture at the next moment t+1 , thereby calculating the next time t+1 execution The cost at the current moment t , and according to the consideration Control the robot to execute at the next moment t+1 , and record the robot's position at the next moment t+1 ;

[0013] The control unit determines Is it equal to If it is equal, the control ends. Otherwise, the acquisition unit continues to collect the posture and environment image set of the next moment and sends it to the processing unit for processing until the robot moves to the target position. until.

[0014] The robot navigation control system based on the Ultra-VLN network described in the present invention is also characterized in that the processing unit uses formula (1-1) to process human voice instructions. Perform preprocessing:

[0015] (1-1)

[0016] In formula (1-1), It represents the pre-processing function of voice commands, and vect represents the function of converting voice commands into command text. It's a voice command The word vector representation of the i-th word in, N represents The total number of words in .

[0017] Furthermore, the feature extraction layer Col in the computing unit is composed of parallel sub-units, among which any The subunits are Convolutional layers Conv, Normalization layers Stand are constructed in sequence;

[0018] Will Input to the feature extraction layer Col and pass through each sub-unit The feature extraction process of the convolution layer Conv is performed to obtain the obstacle convolution feature set at the current time t { |j=1,2,…,J;m=1,2,3..M}, where, Indicates the The convolution feature of the mth obstacle with category label in the kth frame at the current time t output by the sub-unit;

[0019] Will Input to the feature extraction layer Col and pass through each sub-unit respectively The feature extraction process of the convolution layer Conv is performed to obtain the instruction text convolution feature set { |j=1,2,…,J}, where Indicates the The convolutional features of the instruction text output by the sub-units;

[0020] Will Input to the feature extraction layer Col, after any sub-unit The feature extraction process of the convolution layer Conv is used to obtain the posture convolution feature of the current time t ;

[0021] { |j=1,2,…,J;m=1,2,3..M} pass through each subunit respectively Normalization layer Stand processing is performed to obtain the normalized feature set of obstacles at the current time t { ;m=1,2,3..M}, where Indicates the The normalized feature of the mth obstacle with category label in the kth frame at the current time t output by the sub-unit;

[0022] { |j=1,2,…,J} pass through each subunit respectively Normalization layer Stand processing is performed to obtain the normalized feature set of the instruction text { },in, Indicates the Normalized features of the instruction text output by each sub-unit;

[0023] Through any subunit Normalization layer Stand processing is performed to obtain the normalized feature of the posture at the current time t ;

[0024] Will , and Input into the cross attention layer Ate, and use formula (1-2) to get any i-th cross attention feature at the current time t , thus obtaining the cross-attention feature set at the current time t { }:

[0025] (1-2)

[0026] In formula (1-2), Indicates the The obstacle attention features of the kth frame at the current time t are output by the cross attention heads, Indicates the The instruction text attention features output by the cross attention head, 、 、 There are three adjustment factors, represents the total number of cross-attention heads;

[0027] The decoding layer Trans uses formula (1-3) to generate the robot's action sequence at the next time t+1 :

[0028] (1-3)

[0029] In formula (1-3), d represents dimension.

[0030] Furthermore, the control unit uses formula (1-4) to predict the robot's execution at the current time t The cost :

[0031] (1-4)

[0032] In formula (1-4), Represents the robot's posture from the current moment t Switch to the next moment t+1 posture During the process, the heat energy generated by the friction of its own mechanical parts, is a penalty function that represents the risk of collision with obstacles when the robot performs an action, and has:

[0033] (1-5)

[0034] In formula (1-5), Represents the distance the robot moves from the current moment t to the next moment t+1, Represents the roughness of the road where the robot moves, represents the risk function of the robot colliding with an obstacle during its movement at the current time t, 、 are the weight coefficients for adjusting frictional heat energy and the risk of collision with obstacles, respectively.

[0035] Furthermore, the control unit compares the current time t execution Execute the action sequence at the previous moment t-1 ,like , then the control module executes at the current time t On the contrary, the control module uses formula (1-6) to calculate the posture of the robot at the next moment t+1 , and have:

[0036] (1-6)

[0037] In formula (1-6), C represents the control strategy function, is the posture feature of the robot after preprocessing at the current time t Regarding the time-varying dynamic matrix, is the sequence of actions that the robot will perform at the current time t and the posture at the next moment t+1 The transformation matrix between .

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1. This paper utilizes a robot navigation framework based on an Ultra-VLN network. This network uses deep learning algorithms to simultaneously process visual information and natural language commands, enabling precise target positioning, path planning, and obstacle avoidance. Compared to existing vision-language navigation methods, the Ultra-VLN network's advantage lies in its deep fusion of image data and natural language text data, enabling robots to more accurately understand commands and autonomously execute tasks.

[0040] 2. This invention can simultaneously process multimodal data from visual sensors, voice commands, and environmental perception systems. Using deep reinforcement learning algorithms, the system can efficiently schedule and execute autonomous tasks based on both images and human voice commands. This solution addresses the lack of coordination between voice and visual perception in existing systems, improving the accuracy and responsiveness of autonomous robot navigation.

[0041] 3. This invention adds real-time feedback on robot movements and environmental changes, ensuring the robot can dynamically adjust its state based on environmental changes. This control strategy significantly improves the robot's adaptability in complex and dynamic environments, ensuring efficient and stable task execution. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is the architecture diagram of the robot navigation control system based on the Ultra-VLN network in the present invention;

[0043] Figure 2 This is a diagram of the Ultra-VLN network architecture in the robot navigation control system based on the Ultra-VLN network of the present invention;

[0044] Figure 3 It is a system flow chart of the robot navigation control system based on the Ultra-VLN network in the present invention. DETAILED DESCRIPTION

[0045] In this embodiment, see Figure 1A robot in a robot navigation control system based on an Ultra-VLN network includes a differential mobile base, a built-in motor of the mobile base is used to drive the robot to move and turn, a depth sensor is used to obtain a depth image of the environment, and an inertial sensor is used to obtain the posture of the robot.

[0046] like Figure 2 The figure shows a robot navigation and control system based on an Ultra-VLN network, comprising a preprocessing module, a parsing module, and a control module. The acquisition unit receives depth image data of the environment and objects within it from a depth sensor, as well as robot posture data from an inertial sensor, at a frequency of 20 Hz. The processing unit, a high-performance NVIDIA processor, is primarily responsible for real-time processing of the depth image data and robot posture data.

[0047] In this embodiment, the implementation of the preprocessing module specifically includes the following steps:

[0048] Step 1.1: The acquisition unit and processing unit scan and annotate the 2D image of the environment, thereby extracting rich environmental features:

[0049] Step 1.1.1. The acquisition unit receives the current time t environment image scanned and acquired by the depth sensor, and the processing unit extracts and preprocesses the obstacle image in the environment image to obtain the image sequence of the obstacle at the current time t At the current time t, the 2D image of any obstacle in the image sequence is recorded as . Where k is the index of the image frame, representing the image sequence The kth image in .

[0050] Step 1.1.2: The processing unit processes the current environment image at time t The M obstacles in the image are labeled by type, and a feature image vector with M type labels is obtained. Where m = 1, 2, …, M represents the index of the obstacle; in this way, the robot can capture and identify obstacles during movement, thereby timely updating the avoidance strategy and rationally planning the path to avoid the risk of collision and reduce energy loss.

[0051] Step 1.2: The acquisition unit collects the robot's posture at the current time t And human voice instructions S, where the voice instructions S are target information conveyed by humans through voice, instructing the robot to perform a specific task or move in a specific direction. is the yaw angle of the robot at the current time t;

[0052] Step 1.3: The processing unit calculates the posture of the robot at the current time t , and human voice commands are pre-processed separately, thereby reducing the load of subsequent data processing on the Ultra-VLN network.

[0053] Step 1.3.1: The processing unit calculates the robot's posture at the current time t Perform preprocessing as shown in formula (2-1):

[0054] (2-1)

[0055] In formula (2-1), is the preprocessed posture feature of the robot at the current moment t, Represents the preprocessing function of the yaw angle, wrap represents the angle interval [−π,π], The first step is to correct the deviation. This step can ensure the accuracy and rationality of the robot's posture data and provide a reliable basis for subsequent motion control and navigation decisions.

[0056] Step 1.3.2: The processing unit pre-processes the human voice command S. The processing method is shown in formula (2-2):

[0057] (2-2)

[0058] In formula (2-2), Represents the preprocessing function of the voice command, which is used to remove noise, standardize the voice signal, and perform necessary voice-to-text conversion. The vect(*) function can convert the voice command Convert to command text Specifically, it can be realized by formula (2-3):

[0059] (2-3)

[0060] In formula (2-3), It is the original voice command input. The vect function can convert the voice command into Each word in is converted into its corresponding word vector for subsequent analysis and processing of the instruction. It's a voice command The word vector representation of the i-th word in .

[0061] Step 1.4: Further, the processing unit can separate and extract the instruction text The word vector of the target location that the robot wants to move to , and obtain the location coordinates of the target location through the mapping function In this way, the robot can accurately identify the target location based on human voice commands and provide a basis for subsequent path planning. Specifically, it can be achieved by formula (2-4):

[0062] =M( ) (2-4)

[0063] In formula (2-4), M is a mapping function that converts the word vector of the target location into spatial coordinates. It can be expressed as .

[0064] In this embodiment, the implementation of the parsing module specifically includes the following steps:

[0065] The Ultra-VLN network is established, which consists of three modules in sequence: feature extraction layer Col, cross attention layer Ate, and decoding layer Trans;

[0066] Among them, the feature extraction layer Col is composed of parallel sub-units, among which any The subunits are Convolutional layers Conv, Normalization layers Stand are constructed in sequence to extract multi-layer features of input image data, text data, and robot posture data.

[0067] Step 2.1: The feature image at the current time t Input to the feature extraction layer Col and pass through each sub-unit respectively The feature extraction process of the convolution layer Conv is performed to obtain the image convolution feature set at the current time t { | =1,2,…, ;m=1,2,3..M}, where Indicates the The basic unit outputs the image convolution feature at the current time t;

[0068] The command text Input to the feature extraction layer Col and pass through each sub-unit respectively The feature extraction process of the convolution layer Conv is performed to obtain the instruction text convolution feature set { | =1,2,…, },in, Indicates the The convolutional features of the instruction text output by the basic units;

[0069] The posture feature of the robot at the current time t Input to the feature extraction layer Col, after any sub-unit The feature extraction process of the convolution layer Conv is used to obtain the posture convolution feature of the current time t ;

[0070] The image convolution feature set at the current time t is { | =1,2,…, ;m=1,2,3..M} pass through each subunit respectively Normalization layer Stand processing is performed to obtain the normalized feature set of the image at the current time t { ;m=1,2,3..M}. Among them, Indicates the The normalized features of the image at the current time t are output by the basic units. This step ensures that the image features are processed evenly across different scales and ranges, and helps maintain the stability of the image features when processed by the cross-attention layer.

[0071] Instruction text convolution feature set { | =1,2,…, } Pass through each subunit separately Normalization layer Stand processing is performed to obtain the normalized feature set of the instruction text { ;m=1,2,3..M}. Among them, Indicates the The command text is normalized into features output by each basic unit. This process ensures the consistency of the command text features in different dimensions, improves the alignment of the command text features with the image features, and helps improve the efficiency of the Ultra-VLN network in processing different types of features.

[0072] The posture characteristics of the robot at the current time t Through any subunit Normalization layer Stand processing is performed to obtain the normalized feature of the posture at the current time t It reflects the posture of the robot at the current time t, and provides a guarantee for the subsequent positioning and navigation decision-making of the robot in the autonomous navigation process.

[0073] The above feature extraction and normalization process effectively extracts fine-grained features from images, voice command text, and robot posture, providing foundational support for feature processing in the subsequent cross-attention and decoding layers. This ensures that the Ultra-VLN network fully utilizes features and provides precise input for the robot's autonomous navigation in complex environments.

[0074] Step 2.2: Normalize the image feature set at the current time t obtained in step 2.1 , instruction text normalized feature set , the normalized feature of the posture at the current time t Input into the cross attention layer Ate to form the cross attention feature set at the current time t Among them, any cross attention feature at the current moment t , as shown in formula (2-5):

[0075] (2-5)

[0076] In formula (2-5), Indicates the current time t Image normalization features, Indicates the The normalized features of the instruction text, represents the normalized feature of the posture at the current time t, 、 、 is a regulation factor and is not all 0. Through the cross-attention mechanism, the association between image features, instruction text features, and posture features is effectively mined and strengthened, thereby providing rich information interaction for the subsequent decision-making process.

[0077] Step 2.3: The cross attention feature set of the current moment t obtained in step 2.2 Input to the decoding layer Trans to generate the robot's action sequence at the next time t+1 After decoding, the robot’s action sequence at the next moment t+1 is generated, as shown in Equation (2-6). This process uses the contextual information and temporal associations carried in the cross-attention feature set to infer the robot’s optimal action sequence at the next moment.

[0078] (2-6)

[0079] In formula (2-6), d represents the cross attention feature set dimension.

[0080] In this embodiment, the control module is implemented as follows: Figure 3 As shown, the specific steps include:

[0081] Step 3.1: The control module receives the action sequence of the robot at the next time t+1 generated in step 2.3 ;

[0082] Step 3.2: The control module receives the action sequence , calculate and update the robot's posture at the next moment according to the control strategy function C , specifically, it can be expressed by formula (2-7):

[0083] (2-7)

[0084] In formula (2-7), C represents the control strategy function, is the posture feature of the robot after preprocessing at the current time t Regarding the time-varying dynamic matrix, is the sequence of actions that the robot will perform at the current time t and the posture at the next moment t+1 The transformation matrix between .

[0085] The dynamic matrix in formula (2-7) , transformation matrix They can be expressed by formula (2-8) and formula (2-9) respectively:

[0086] = (2-8)

[0087] In formula (2-8), it is assumed that the initial position of the robot is taken as the coordinate origin O. The axis represents the direction of the robot's autonomous navigation. The axis represents the vertical direction of the machine. Assume that the robot's forward speed at the current time t is ,but: It's a robot along axis speed, It is along axis speed, The robot moves forward relative to The deflection angle of the axis.

[0088] = (2-9)

[0089] Through formula (2-8) and formula (2-9), the robot can be made to calculate the posture characteristics of the robot at the current time t. The action sequence executed at the next time t+1 Achieve autonomous transformation to the posture at the next moment t+1 .

[0090] Step 3.3: The calculation unit calculates the robot's execution action according to the control strategy function C in step 3.2. The cost , as shown in formula (2-10):

[0091] (2-10)

[0092] In formula (2-10), Represents the robot's posture from the current time t Execute an action Change to the posture at the next moment t+1 During the process, the heat energy generated by the friction of the mechanical parts of the machine itself. In order to quantify the cost more accurately, a penalty function is introduced. .in, Perform actions on behalf of the robot The possible risk of power exhaustion can be specifically expressed by formula (2-11):

[0093] (2-11)

[0094] In formula (2-11), L represents the distance the robot moves, Represents the roughness of the road where the robot moves, represents the risk function of the robot running out of power during movement, 、 are weight coefficients for adjusting the risk of frictional heat and power exhaustion. Penalty function It can comprehensively consider the negative effects that may occur during the robot's movement, such as heat energy loss caused by friction and mechanical energy loss caused by collision with obstacles, so that the robot can adopt more efficient and stable strategies when performing tasks.

[0095] Step 3.4: When the robot performs the action again in step 3.3 The cost When the calculation unit compares the current time t, the action is executed Execute the action at the previous moment t-1 ,like > , the control module continues to execute steps 3.2 to 3.3. Otherwise, the robot will change to the posture of the next moment t+1 ;

[0096] Step 3.5: After assigning t+1 to t, the control module continues to execute steps 3.1 to 3.4 until the robot moves to the target position. This process is repeated continuously, and in conjunction with the control strategy and cost evaluation mechanism, it ensures that the robot can optimize its movements while minimizing energy consumption and avoiding potential risks.

[0097] Through the above steps, the robot navigation and control system based on the Ultra-VLN network can achieve efficient dynamic navigation and real-time control to ensure that the robot reaches the predetermined target position smoothly and accurately in a complex environment.

Claims

1. A robot navigation and control system based on an Ultra-VLN network, the robot comprising: Differential moving base, depth sensor; the differential moving base includes an inertial sensor and a collision sensor; it is characterized in that the robot navigation control system includes: an acquisition unit, a processing unit, a calculation unit, and a control unit; The human voice instructions acquired by the acquisition unit After sending it to the processing unit for preprocessing, the instruction text is obtained , the processing unit then Extract the word vector of the target location that the robot will move to , and transform the word vector into Converted to the location coordinates of the target location ; The acquisition unit obtains the posture of the robot at the current time t collected by the inertial sensor ; At the same time, obtain the environment image set at the current time t collected by the depth sensor and send it to the processing unit together; The processing unit processes the environment image set to obtain the obstacle image sequence L at the current time t, and records the obstacle 2D image of any k-th frame in L as , and Each obstacle in the class is labeled to obtain M obstacle feature image vectors with class labels. ,in, express The mth obstacle feature image with category label in , Indicates the total number of obstacle feature images; The processing unit calculates the posture of the robot at the current time t After preprocessing, the preprocessed posture features at the current time t are obtained ; The computing unit includes an Ultra-VLN network consisting of a feature extraction layer Col, a cross attention layer Ate, and a decoding layer Trans, and 、 、 Processing is performed to obtain the action sequence that the robot will perform at the next moment t+1 ; The control unit acts according to the sequence of actions and posture features , use the control strategy function C to calculate the posture of the robot at the next moment t+1 , thus calculating the execution time at the next moment t+1 The cost at the current moment t , and according to the said consideration Control the robot to execute at the next moment t+1 , and record the robot's position at the next moment t+1 ; The control unit determines Is it equal to If it is equal, the control ends. Otherwise, the acquisition unit continues to collect the posture and environment image set of the next moment and sends it to the processing unit for processing until the robot moves to the target position. until.

2. The robot navigation control system based on Ultra-VLN network according to claim 1, characterized in that: The processing unit uses formula (2-1) to process human voice instructions Perform preprocessing: (2-1) In formula (2-1), It represents the pre-processing function of voice commands, and vect represents the function of converting voice commands into command text. It's a voice command The word vector representation of the i-th word in, N represents The total number of words in .

3. The robot navigation control system based on Ultra-VLN network according to claim 1, characterized in that: The feature extraction layer Col in the computing unit is composed of parallel sub-units, among which any The subunits are Convolutional layers Conv, Normalization layers Stand are constructed in sequence; Will Input to the feature extraction layer Col and pass through each sub-unit The feature extraction process of the convolution layer Conv is performed to obtain the obstacle convolution feature set at the current time t { |j=1,2,…,J;m=1,2,3..M}, where, Indicates the The convolution feature of the mth obstacle with category label in the kth frame at the current time t output by the sub-unit; Will Input to the feature extraction layer Col and pass through each sub-unit respectively The feature extraction process of the convolution layer Conv is performed to obtain the instruction text convolution feature set { |j=1,2,…,J}, where Indicates the The convolutional features of the instruction text output by the sub-units; Will Input to the feature extraction layer Col, after any sub-unit The feature extraction process of the convolution layer Conv is used to obtain the posture convolution feature of the current time t ; { |j=1,2,…,J;m=1,2,3..M} pass through each subunit respectively Normalization layer Stand processing is performed to obtain the normalized feature set of obstacles at the current time t { ;m=1,2,3..M}, where Indicates the The normalized feature of the mth obstacle with category label in the kth frame at the current time t output by the sub-unit; { |j=1,2,…,J} pass through each subunit respectively Normalization layer Stand processing is performed to obtain the normalized feature set of the instruction text { },in, Indicates the Normalized features of the instruction text output by each sub-unit; Through any subunit Normalization layer Stand processing is performed to obtain the normalized feature of the posture at the current time t ; Step 3.2, , and Input into the cross attention layer Ate, and use formula (3-1) to get any i-th cross attention feature at the current time t , thus obtaining the cross-attention feature set at the current time t { }: (3-1) In formula (3-1), Indicates the The obstacle attention features of the kth frame at the current time t are output by the cross attention heads, Indicates the The instruction text attention features output by the cross attention head, 、 、 There are three adjustment factors, represents the total number of cross-attention heads; Step 3.3: The decoding layer Trans generates the action sequence of the robot at the next time t+1 using formula (3-2) : (3-2) In formula (3-2), d represents dimension.

4. The robot navigation control system based on Ultra-VLN network according to claim 1, characterized in that: The control unit uses formula (4-1) to predict the robot's execution at the current time t The cost : (4-1) In formula (4-1), Represents the robot's posture from the current moment t Switch to the next moment t+1 posture During the process, the heat energy generated by the friction of its own mechanical parts, is a penalty function that represents the risk of collision with obstacles when the robot performs an action, and has: (4-2) In formula (4-2), Represents the distance the robot moves from the current moment t to the next moment t+1, Represents the roughness of the road where the robot moves, represents the risk function of the robot colliding with an obstacle during its movement at the current time t, 、 are the weight coefficients for adjusting frictional heat energy and the risk of collision with obstacles, respectively.

5. The robot navigation control system based on Ultra-VLN network according to claim 1, characterized in that: The control unit compares the current time t and performs Execute the action sequence at the previous moment t-1 ,like , then the control module executes at the current time t On the contrary, the control module uses formula (5-1) to calculate the posture of the robot at the next moment t+1 , and have: (5-1) In formula (5-1), C represents the control strategy function, is the posture feature of the robot after preprocessing at the current time t Regarding the time-varying dynamic matrix, is the sequence of actions that the robot will perform at the current time t and the posture at the next moment t+1 The transformation matrix between .

Images