Centaur robot body intelligence control method, device, equipment, medium, robot and system for special scene

Abstract

The application provides a Centaur robot body intelligence control method and device for special scenarios, electronic equipment, computer readable storage medium, Centaur robot and system. The method acquires a multi-view depth image matrix, leg joint pose parameters, arm joint pose parameters and inertial navigation parameters of the Centaur robot at the current time, inputs a body strategy mapping network for feature implicit alignment, generates a heterogeneous feature vector, and performs a progressive capability self-construction process to generate leg-arm coordinated linkage parameters, and then generates chassis leg driving torque instructions and double-arm end operation torque instructions to control the Centaur robot.

Description

Technical Field

[0001] This application relates to the field of robot embodied intelligent control technology, and more specifically, to a method, device, electronic device, computer-readable storage medium, centaur robot and system for special scenarios. Background Technology

[0002] In specialized environments such as mines, chemical plants, and underground utility tunnels, robots typically need to navigate and move through narrow passages, areas with dense obstacles, slopes, or complex terrain. They also need to perform tasks such as switching operations, valve adjustments, sample handling, or equipment inspection using their robotic arms. These scenarios place comprehensive demands on robots regarding mobility stability, operational continuity, and adaptability to environmental changes. Particularly when changes in robot chassis posture and end-effector operations occur simultaneously, the control system must coordinate the relationship between robot body movement and end-effector operations within a short control cycle.

[0003] In existing technologies, a common approach to robot control employs a hierarchical control architecture. This approach typically begins with a perception layer acquiring environmental data and generating a local environmental representation. Next, a navigation and planning layer generates a movement path based on this representation. Subsequently, a motion control layer controls the chassis movement according to the movement path, and a separate robotic arm control layer generates the robotic arm's actions based on the task objectives. This approach, with different control layers handling environmental perception, motion planning, chassis control, and robotic arm control respectively, exhibits relatively clear modular boundaries in its structure.

[0004] However, in specialized scenarios, robots face rapidly changing environments, and there is a strong coupling between motion navigation and dual-arm operation. Traditional hierarchical control architectures require the sequential transmission of intermediate results between different control layers, easily leading to long control response chains. Furthermore, chassis motion control and robotic arm operation control often generate control results separately, making it difficult to express the coordinated relationship between the robot's motion trend and the end-effector's intent within the same control process. When robots perform tasks in scenarios involving dynamic base disturbances, complex ground support changes, or near-obstacle operations, traditional hierarchical control schemes exhibit low adaptability to environmental changes, easily resulting in insufficient integration between motion control and end-effector control. Summary of the Invention

[0005] This application provides a method, device, electronic device, computer-readable storage medium, centaur robot and system for special scenarios to at least alleviate the technical problems of traditional hierarchical control architecture, such as long response links, insufficient connection between mobile control and end-effector control, and low adaptability to environmental changes.

[0006] A method for embodied intelligent control of centaur robots for special scenarios includes: Obtain the multi-view depth image matrix, leg joint pose parameters, arm joint pose parameters, and inertial navigation parameters of the centaur robot at the current moment; The multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters are input into the embodied policy mapping network to perform implicit feature alignment processing on the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters, thereby generating heterogeneous feature vectors that characterize the body motion trend and end-effector operation intention of the Centaur robot. In the embodied strategy mapping network, a progressive capability self-construction process is performed with the heterogeneous feature vector as the processing object to sequentially generate basic leg movement parameters, pure arm operation parameters, and leg-arm coordinated linkage parameters. Based on the leg-arm coordination parameters, the system generates chassis leg driving torque commands and end-effector operating torque commands for the centaur robot, and then controls the centaur robot through the chassis leg driving torque commands and the end-effector operating torque commands.

[0007] A centaur robot embodied intelligent control device for special scenarios includes: The parameter acquisition module is used to acquire the multi-view depth image matrix, leg joint pose parameters, arm joint pose parameters, and inertial navigation parameters of the Centaur robot at the current moment. The implicit feature alignment module is used to input the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters into the embodied policy mapping network to perform implicit feature alignment processing on the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters, thereby generating a heterogeneous feature vector characterizing the body motion trend and end-effector operation intention of the Centaur robot. The capability self-construction module is used in the embodied policy mapping network to perform a progressive capability self-construction process with the heterogeneous feature vector as the processing object, so as to sequentially generate basic leg movement parameters, pure arm operation parameters, and leg-arm coordinated linkage parameters. The torque command control module is used to generate the chassis leg driving torque command and the end effector torque command of the Centaur robot according to the leg-arm coordinated linkage parameters, and then control the Centaur robot through the chassis leg driving torque command and the end effector torque command.

[0008] An electronic device, including a memory and a processor; The memory stores computer programs; When the processor executes the computer program, it performs any of the methods described above.

[0009] A computer-readable storage medium storing a computer program; When the computer program is executed by a processor, it is used to perform any of the methods described above.

[0010] A centaur robot, comprising: The fuselage itself; A chassis leg actuator is connected to the fuselage body and is used to receive chassis leg drive torque commands and execute chassis leg movements. A dual-arm end effector is connected to the main body of the machine and is used to receive dual-arm end-operation torque commands and perform dual-arm end-operation operations. A multi-view depth acquisition component is disposed on the main body of the robot and is used to acquire a multi-view depth image matrix of the Centaur robot at the current moment. A leg joint pose acquisition component is disposed on the chassis leg actuator and is used to acquire the leg joint pose parameters of the Centaur robot at the current moment. An arm joint pose acquisition component is disposed on the dual-arm end effector and is used to acquire the arm joint pose parameters of the Centaur robot at the current moment. An inertial navigation component is disposed on the main body of the robot and is used to collect the inertial navigation parameters of the Centaur robot at the current moment. The embodied intelligent control device is connected to the multi-view depth acquisition component, the leg joint pose acquisition component, the arm joint pose acquisition component, the inertial navigation component, the chassis leg actuator, and the dual arm end effector. The embodied intelligent control device is used to input the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters into the embodied strategy mapping network to perform feature implicit alignment processing on the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters, thereby generating a heterogeneous feature vector characterizing the centaur robot's body motion trend and end-effector operation intention; The embodied intelligent control device is also used in the embodied strategy mapping network to perform a progressive capability self-construction process with the heterogeneous feature vector as the processing object, so as to sequentially generate basic leg movement parameters, pure arm operation parameters and leg-arm coordinated linkage parameters. The embodied intelligent control device is also used to generate the chassis leg driving torque command and the dual arm end effector torque command according to the leg-arm coordinated linkage parameters, so as to control the chassis leg actuator through the chassis leg driving torque command and control the dual arm end effector through the dual arm end effector torque command.

[0011] A centaur robot embodied intelligent control system for special scenarios includes: The multimodal parameter acquisition subsystem is used to acquire the multi-view depth image matrix, leg joint pose parameters, arm joint pose parameters, and inertial navigation parameters of the Centaur robot at the current moment. An embodied strategy mapping subsystem, connected to the multimodal parameter acquisition subsystem, is used to receive the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters. It then inputs these parameters into the embodied strategy mapping network to perform implicit feature alignment processing on the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters, thereby generating heterogeneous feature vectors characterizing the centaur robot's body motion trends and end-effector operation intentions. A progressive capability self-construction subsystem, connected to the embodied strategy mapping subsystem, is used in the embodied strategy mapping network to perform a progressive capability self-construction process with the heterogeneous feature vector as the processing object, so as to sequentially generate basic leg movement parameters, pure arm operation parameters, and leg-arm coordinated linkage parameters. The torque command output subsystem is connected to the progressive capability self-construction subsystem and is used to generate the chassis leg drive torque command and the end-effector operation torque command of the Centaur robot according to the leg-arm coordinated linkage parameters. An execution control subsystem, connected to the torque command output subsystem, is used to receive the chassis leg drive torque command and the dual arm end effector torque command, and then control the centaur robot through the chassis leg drive torque command and the dual arm end effector torque command.

[0012] The technical advantages of the technical solution provided in this application are: This application presents a centaur robot embodied intelligent control method for special scenarios. Addressing the problem of long response chains caused by the hierarchical transmission of intermediate results between perception, navigation, planning, chassis control, and robotic arm control in traditional hierarchical control architectures, this method first acquires the multi-view depth image matrix, leg joint pose parameters, arm joint pose parameters, and inertial navigation parameters of the centaur robot at the current moment. Then, these parameters from different sources are input into an embodied policy mapping network for implicit feature alignment, enabling environmental spatial information, leg motion state, arm operation state, and body inertial state to enter the same control processing chain. Compared to traditional hierarchical control schemes that process environmental information, chassis information, and robotic arm information separately, this application directly represents the centaur robot's body motion trend and end-effector intentions through heterogeneous feature vectors. This reduces the control delay caused by repeated conversions and transmission of intermediate results between different control layers, resulting in a higher continuity of the control response chain.

[0013] Simultaneously, this application employs a progressive capability self-construction process within the embodied policy mapping network, using heterogeneous feature vectors as the processing object, and sequentially generates basic leg movement parameters, pure arm operation parameters, and leg-arm coordinated linkage parameters. This processing method does not output chassis movement control and dual-arm operation control as two separate, unrelated control results. Instead, it first generates the basic control parameters corresponding to leg movement capabilities, then the control parameters corresponding to arm operation capabilities, and finally the leg-arm coordinated linkage parameters. Compared to traditional solutions that generate control results separately for chassis motion control and robotic arm operation control, this application can express the progressive relationship between movement capabilities, operation capabilities, and leg-arm coordinated capabilities during the control parameter generation process, resulting in a more complete connection between movement control and end-effector control.

[0014] Furthermore, this application generates chassis leg drive torque commands and end-effector manipulation torque commands for the Centaur robot based on leg-arm coordination parameters, and controls the Centaur robot through these commands. Since both the chassis leg drive torque commands and the end-effector manipulation torque commands originate from the leg-arm coordination parameters, they share a consistent collaborative control basis in their generation source. Compared to the problem of insufficient connection between chassis control results and robotic arm control results in traditional hierarchical control schemes due to different processing links, this application enables chassis leg drive and end-effector manipulation to form control outputs around the same leg-arm coordination parameters, thereby exhibiting high adaptability to environmental changes under conditions such as dynamic base disturbances, complex ground support changes, or near-obstacle operation in special scenarios. Attached Figure Description

[0015] Figure 1 This application provides an embodiment of a centaur robot's embodied intelligent control scenario for special applications. Figure 2 This application provides an embodiment of a centaur robot embodied intelligent control method for special scenarios. Figure 3 This application provides an embodiment of a centaur robot embodied intelligent control device for special scenarios. Figure 4 An electronic device is described in an embodiment of this application; Figure 5 This application provides an embodiment of a computer-readable storage medium. Figure 6 This application provides an embodiment of a mobile robot. Figure 7 This application provides an embodiment of a centaur robot embodied intelligent control system for special scenarios. Detailed Implementation

[0016] like Figure 1 As shown, this application provides an embodiment of a centaur robot's embodied intelligent control scenario for special situations; as... Figure 2 As shown in the figure, a method for embodied intelligent control of a centaur robot for special scenarios according to an embodiment of this application includes: Obtain the multi-view depth image matrix, leg joint pose parameters, arm joint pose parameters, and inertial navigation parameters of the centaur robot at the current moment; The multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters are input into the embodied policy mapping network to perform implicit feature alignment processing on the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters, thereby generating heterogeneous feature vectors that characterize the body motion trend and end-effector operation intention of the Centaur robot. In the embodied strategy mapping network, a progressive capability self-construction process is performed with the heterogeneous feature vector as the processing object to sequentially generate basic leg movement parameters, pure arm operation parameters, and leg-arm coordinated linkage parameters. Based on the leg-arm coordination parameters, the system generates chassis leg driving torque commands and end-effector operating torque commands for the centaur robot, and then controls the centaur robot through the chassis leg driving torque commands and the end-effector operating torque commands.

[0017] Optionally, acquiring the multi-view depth image matrix, leg joint pose parameters, arm joint pose parameters, and inertial navigation parameters of the centaur robot at the current moment includes: Extract the first hardware timestamp corresponding to the multi-view depth image matrix; The second hardware timestamps corresponding to the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters are extracted respectively. Determine whether the difference between the first hardware timestamp and each of the second hardware timestamps is within a preset alignment range; If the differences between the first hardware timestamp and each of the second hardware timestamps are all within the preset alignment interval, then the multi-view depth image matrix is ​​used as the anchoring reference, and temporal state interpolation processing is performed on the leg joint pose parameter, the arm joint pose parameter, and the inertial navigation parameter to obtain multimodal parameters that are time-synchronized with the multi-view depth image matrix. The multimodal parameters include the time-synchronized multi-view depth image matrix, the leg joint pose parameter, the arm joint pose parameter, and the inertial navigation parameter.

[0018] Preferably, the specific implementation process for acquiring the multi-view depth image matrix, leg joint pose parameters, arm joint pose parameters, and inertial navigation parameters of the Centaur robot at the current moment is as follows: When the Centaur robot enters the current control cycle, the parameter acquisition and carrying record corresponding to the current control cycle is first read. The parameter acquisition and carrying record includes the current control cycle marker, the acquisition trigger marker of the multi-view depth acquisition component, the acquisition trigger marker of the leg joint pose acquisition component, the acquisition trigger marker of the arm joint pose acquisition component, and the acquisition trigger marker of the inertial navigation component. The parameter acquisition and carrying record is used to include the reading source of the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters into the same current control cycle, so that the subsequently generated multimodal parameters can correspond to the environmental space state, chassis leg support state, dual arm end effector state, and body posture state of the Centaur robot within the same current control cycle. Subsequently, based on the acquisition trigger flag of the multi-view depth acquisition component in the parameter acquisition record, corresponding depth image frames are read from multiple acquisition perspectives, and the depth image frames are matrix-encapsulated according to the acquisition perspective, pixel row position, and pixel column position to form the multi-view depth image matrix. The matrix layers in the multi-view depth image matrix correspond to different acquisition perspectives, the matrix rows in the multi-view depth image matrix correspond to the pixel height position under the same acquisition perspective, the matrix columns in the multi-view depth image matrix correspond to the pixel width position under the same acquisition perspective, and the elements at the intersection of the matrix rows and the matrix columns correspond to the depth distance values ​​of the Centaur robot's surrounding environment. The depth distance values ​​are used to form the environmental spatial state. Further, based on the acquisition trigger flag of the leg joint pose acquisition component in the parametric acquisition and bearing record, the angle parameters, angular velocity parameters, and joint code position parameters of each leg joint are read, and the fields are encapsulated according to the installation order of each leg joint in the chassis leg actuator to form the leg joint pose parameters. These leg joint pose parameters are used to form the chassis leg support state. Based on the acquisition trigger flag of the arm joint pose acquisition component in the parametric acquisition and bearing record, the angle parameters, angular velocity parameters, and joint code position parameters of each arm joint are read. The end-effector attitude parameters are associated with each other and their fields are encapsulated according to the kinematic chain sequence of each arm joint in the dual-arm end-effector to form the arm joint pose parameters. The arm joint pose parameters are used to form the dual-arm end-effector operation state. Based on the acquisition trigger flag of the inertial navigation component in the parameter acquisition and bearing record, the angular velocity parameters, linear acceleration parameters, and attitude angle parameters of the fuselage body are read and their fields are encapsulated according to the coordinate axis direction of the fuselage body to form the inertial navigation parameters. The inertial navigation parameters are used to form the fuselage body attitude state.Through the unified recording of the parameter acquisition data, the multi-view depth image matrix can describe the spatial distance distribution of the Centaur robot's surrounding environment, the leg joint pose parameters can describe the chassis leg support state of the chassis leg actuators, the arm joint pose parameters can describe the end-effector operation state of the dual-arm end-effectors, and the inertial navigation parameters can describe the body posture state. The above four types of parameters then enter the extraction process of the first and second hardware timestamps together, so that when the Centaur robot performs movement and dual-arm operations on mine slopes, underground pipe gallery bends, or near-field operation positions of chemical equipment, it will not mix parameters formed in different current control cycles into the same subsequent processing link.

[0019] Preferably, in the specific technical implementation of extracting the first hardware timestamp corresponding to the multi-view depth image matrix, after the multi-view depth image matrix is ​​formed, the frame acquisition time field carried by each depth image frame in the multi-view depth image matrix is ​​first read, and the frame acquisition time field, along with the corresponding acquisition view identifier, frame sequence number identifier, and current control cycle marker, is written into the depth frame time source record. The depth frame time source record is used to record when depth image frames from different acquisition views in the multi-view depth image matrix are formed by the multi-view depth acquisition component, so that the multi-view depth image matrix not only has spatial distance meaning, but also visual acquisition time meaning that can be used for subsequent time alignment. Subsequently, the depth frame time source record is subjected to time consistency determination processing. First, the maximum time span between the frame acquisition time fields within the same current control cycle is compared, then the view polling order marker corresponding to the multi-view depth acquisition component is read, and the maximum time span and the view polling order marker are combined to perform visual time merging processing on each frame acquisition time field to form the first hardware timestamp. The maximum time span is used to characterize the degree of acquisition dispersion between different depth image frames within the same current control cycle. The view polling order marker is used to characterize the acquisition sequence relationship of different acquisition views in the multi-view depth acquisition component. The visual time merging process merges multiple frame acquisition time fields into the first hardware timestamp based on the maximum time span and the view polling order marker. The technical essence of the first hardware timestamp is the visual time anchor point of the multi-view depth image matrix within the current control cycle. The first hardware timestamp originates from the frame acquisition time field carried when the depth image frame is generated on the hardware side and is formed through the visual time merging process. Therefore, the first hardware timestamp is not the data reception time, nor the software transfer time, but a time reference that can represent the visual time anchor point of the multi-view depth image matrix within the current control cycle. When the Centaur robot moves in the bend of the underground pipe gallery, the depth image frames of the forward acquisition view, the lateral acquisition view, and the downward acquisition view usually have different view polling orders. When the Centaur robot extends its arms at a close operating position of chemical equipment, the depth image frames are also affected by changes in occlusion by nearby objects and the frame buffer reading order. If the first hardware timestamp is not formed based on the time source record of the depth frame, it will be impossible to determine the visual time anchor point corresponding to the multi-view depth image matrix when performing time-series state interpolation processing on the leg joint pose parameter, the arm joint pose parameter, and the inertial navigation parameter.Based on the introduction of the first hardware timestamp, the multi-view depth image matrix can be used as an anchoring reference to participate in subsequent difference judgment and temporal state interpolation processing. The first hardware timestamp is subsequently used to calculate the difference with each of the second hardware timestamps, and further used as the interpolation target time in the temporal state interpolation processing, so that the spatial distance distribution of the multi-view depth image matrix can establish a correspondence with the motion state of the leg joint pose parameter, the arm joint pose parameter and the inertial navigation parameter in the time dimension.

[0020] Preferably, the specific implementation process of extracting the second hardware timestamps corresponding to the leg joint pose parameters, arm joint pose parameters, and inertial navigation parameters is as follows: After the leg joint pose parameters, arm joint pose parameters, and inertial navigation parameters are encapsulated according to the current control cycle, the leg acquisition time field carried by the leg joint pose parameter, the arm acquisition time field carried by the arm joint pose parameter, and the inertial acquisition time field carried by the inertial navigation parameter are read respectively. The leg acquisition time field, the arm acquisition time field, and the inertial acquisition time field are then written into the body parameter time source record along with their corresponding data source identifier, sampling sequence number identifier, and current control cycle marker. The body parameter time source record is used to retain the hardware sampling time source of the leg joint pose parameters, arm joint pose parameters, and inertial navigation parameters, so that the above three types of parameters can use their original time coordinates when participating in subsequent difference judgment and timing state interpolation processing. Subsequently, the time source records of the body parameters are grouped according to the data source identifier to form leg time source sub-records, arm time source sub-records, and inertial time source sub-records. The leg time source sub-record includes the leg acquisition time field, the data source identifier corresponding to the leg acquisition time field, the sampling sequence number identifier, and the current control cycle marker. The arm time source sub-record includes the arm acquisition time field, the data source identifier corresponding to the arm acquisition time field, the sampling sequence number identifier, and the current control cycle marker. The inertial time source sub-record includes the inertial acquisition time field, the data source identifier corresponding to the inertial acquisition time field, the sampling sequence number identifier, and the current control cycle marker. Based on the leg time source sub-records, the hardware sampling time of the leg joint pose parameters is read to form a second leg hardware timestamp; based on the arm time source sub-records, the hardware sampling time of the arm joint pose parameters is read to form a second arm hardware timestamp; based on the inertial time source sub-records, the hardware sampling time of the inertial navigation parameters is read to form a second inertial hardware timestamp.The second hardware timestamp includes the leg second hardware timestamp, the arm second hardware timestamp, and the inertial second hardware timestamp. The second hardware timestamp differs from the first hardware timestamp in that the first hardware timestamp is used to characterize the visual time anchor point of the multi-view depth image matrix, while the second hardware timestamp is used to characterize the hardware sampling time of the body motion parameters. The body motion parameters include the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters. The leg second hardware timestamp corresponds to the hardware sampling time of the leg joint pose parameters, the arm second hardware timestamp corresponds to the hardware sampling time of the arm joint pose parameters, and the inertial second hardware timestamp corresponds to the hardware sampling time of the inertial navigation parameters. Since the leg joint pose parameters are typically related to the chassis leg support state of the chassis leg actuator, the arm joint pose parameters are typically related to the end-effector operation state of the dual-arm end-effector, and the inertial navigation parameters are typically related to the fuselage body attitude state, the sampling rhythms of these three types of body motion parameters are not the same. If the leg joint pose parameters, arm joint pose parameters, and inertial navigation parameters are directly spliced ​​together simply according to the order of data arrival, the subsequently formed multimodal parameters will be mixed with body motion parameters from different hardware sampling times. After each second hardware timestamp is formed by recording the time source of the body parameters, the second hardware timestamp can be compared with the first hardware timestamp to determine the difference, and further used as the original time coordinate in the temporal state interpolation process, so that the leg joint pose parameters, arm joint pose parameters, and inertial navigation parameters can be reconstructed to the visual time anchor point corresponding to the first hardware timestamp.

[0021] Preferably, the specific implementation process for determining whether the difference between the first hardware timestamp and each of the second hardware timestamps is within a preset alignment interval is as follows: After the first hardware timestamp and each of the second hardware timestamps are formed, the alignment interval configuration record is read. The alignment interval configuration record is configured when the Centaur robot completes startup initialization or task scene switching. The alignment interval configuration record includes the visual acquisition cycle, leg joint sampling cycle, arm joint sampling cycle, inertial sampling cycle, current control cycle length, and the maximum time deviation range allowed to participate in the timing state interpolation processing. The preset alignment interval is formed based on the visual acquisition cycle, the leg joint sampling cycle, the arm joint sampling cycle, the inertial sampling cycle, the current control cycle length, and the maximum time deviation range. The technical essence of the preset alignment interval is to establish a reconstructable time tolerance range between different parameter sampling beats, so that subsequent timing state interpolation processing is only performed within the range that has time reconstruction significance. The visual acquisition cycle is used to constrain the visual acquisition rhythm of the multi-view depth image matrix. The leg joint sampling cycle is used to constrain the sampling rhythm of the leg joint pose parameters. The arm joint sampling cycle is used to constrain the sampling rhythm of the arm joint pose parameters. The inertial sampling cycle is used to constrain the sampling rhythm of the inertial navigation parameters. The maximum time deviation range is used to limit the difference boundary between the first hardware timestamp and each of the second hardware timestamps that can participate in the temporal state interpolation processing. In the scenario where the Centaur robot moves on a mine slope, the chassis leg support state of the chassis leg actuator changes with the slope undulation, and the leg joint sampling cycle has a high weight in the configuration of the preset alignment interval. In the scenario where the Centaur robot performs dual-arm operations at a close operating position in chemical equipment, the coupling degree between the dual-arm end effector's dual-arm end effector's dual-arm end effector's dual-arm body posture state is high, and the arm joint sampling cycle and the inertial sampling cycle have a high weight in the configuration of the preset alignment interval. Subsequently, the leg time difference between the first hardware timestamp and the second hardware timestamp of the leg, the arm time difference between the first hardware timestamp and the second hardware timestamp of the arm, and the inertial time difference between the first hardware timestamp and the second hardware timestamp of the inertia are calculated respectively, and the leg time difference, the arm time difference, and the inertial time difference are written into a time difference determination record. The time difference determination record, which includes the first hardware timestamp, the second hardware timestamp, and the preset alignment interval, is used to transform the time deviation relationship between the visual time anchor point and the hardware sampling time of the body motion parameters into a determineable recording structure.Further, interval hit determination processing is performed on the leg time difference, the arm time difference, and the inertial time difference respectively; if the leg time difference is within the preset alignment interval, a leg alignment admission mark is generated; if the arm time difference is within the preset alignment interval, an arm alignment admission mark is generated; if the inertial time difference is within the preset alignment interval, an inertial alignment admission mark is generated. The leg alignment admission mark, the arm alignment admission mark, and the inertial alignment admission mark together form the alignment admission determination result, which is subsequently used to determine whether to perform temporal state interpolation processing using the multi-view depth image matrix as the anchoring reference. By comparing the leg time difference, arm time difference, and inertial time difference with the preset alignment interval, this application does not use a single average time difference to replace the time deviation of various parameters. Instead, it retains the time alignment admission status of the leg joint pose parameter, arm joint pose parameter, and inertial navigation parameter separately, so that a significant lag in one type of parameter will not be masked by the smaller time difference of other parameters. Therefore, only when the leg alignment admission marker, arm alignment admission marker, and inertial alignment admission marker are all formed will the leg joint pose parameter, arm joint pose parameter, and inertial navigation parameter enter the subsequent timing state interpolation processing, thus ensuring that the subsequently formed multimodal parameters have a more consistent time source.

[0022] Preferably, if the differences between the first hardware timestamp and each of the second hardware timestamps are all within the preset alignment interval, then the multi-view depth image matrix is ​​used as the anchoring reference to perform temporal state interpolation processing on the leg joint pose parameter, the arm joint pose parameter, and the inertial navigation parameter to obtain multimodal parameters that are temporally synchronized with the multi-view depth image matrix. The specific implementation process is as follows: After the alignment admission judgment result includes the leg alignment admission mark, the arm alignment admission mark, and the inertial alignment admission mark, a temporal interpolation bearer record is established, and the multi-view depth image matrix, the first hardware timestamp, the leg joint pose parameter, the second leg hardware timestamp, the arm joint pose parameter, the second arm hardware timestamp, the inertial navigation parameter, and the second inertial hardware timestamp are written into the temporal interpolation bearer record. The temporal interpolation record is used to establish a correspondence between the visual time anchors of the multi-view depth image matrix and the original time coordinates of various body motion parameters, so that the temporal state interpolation processing can reconstruct the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters around the same first hardware timestamp, instead of letting the above three types of parameters directly enter the subsequent implicit feature alignment processing according to their respective hardware sampling times. Subsequently, using the first hardware timestamp as the interpolation target time, reference records of leg joint pose parameters located before and after the first hardware timestamp are read between adjacent sampling records of the leg joint pose parameters. The leg joint pose parameters are then subjected to time-proportional interpolation based on these reference records to form the time-synchronized leg joint pose parameters. Similarly, reference records of arm joint pose parameters located before and after the first hardware timestamp are read between adjacent sampling records of the arm joint pose parameters. The arm joint pose parameters are then subjected to time-proportional interpolation based on these reference records to form the time-synchronized arm joint pose parameters. Finally, reference records of inertial navigation parameters located before and after the first hardware timestamp are read between adjacent sampling records of the inertial navigation parameters. The inertial navigation parameters are then subjected to time-proportional interpolation based on these reference records to form the time-synchronized inertial navigation parameters. The leg joint pose parameter reference record is used to provide a sampling reference for the leg joint pose parameter before and after the first hardware timestamp; the arm joint pose parameter reference record is used to provide a sampling reference for the arm joint pose parameter before and after the first hardware timestamp; the inertial navigation parameter reference record is used to provide a sampling reference for the inertial navigation parameter before and after the first hardware timestamp; the time ratio interpolation processing reconstructs the time position of the above three types of parameters according to the time position of the first hardware timestamp between the sampling references.For the angle parameters in the leg joint pose parameters and the arm joint pose parameters, the time-proportional interpolation process combines the angle change direction and angular velocity change trend between adjacent sampling records, so that the time-synchronized leg joint pose parameters and the time-synchronized arm joint pose parameters can retain the joint motion direction. For the linear acceleration parameters, angular velocity parameters, and attitude angle parameters in the inertial navigation parameters, the time-proportional interpolation process performs segmented reconstruction according to the sampling time interval, so that the time-synchronized inertial navigation parameters can correspond to the body attitude state at the first hardware timestamp. Finally, the time-synchronized multi-view depth image matrix, the time-synchronized leg joint pose parameters, the time-synchronized arm joint pose parameters, and the time-synchronized inertial navigation parameters are encapsulated into the multimodal parameters. The multimodal parameters retain the spatial distance distribution of the multi-view depth image matrix, the chassis leg support state of the leg joint pose parameters, the end-effector operation state of the arm joint pose parameters, and the body posture state of the inertial navigation parameters. Furthermore, all four types of parameters correspond to the same first hardware timestamp. Therefore, the multimodal parameters can serve as source temporal data for subsequent input to the embodied policy mapping network. This ensures that subsequent implicit feature alignment processing deals with the environmental spatial state, chassis leg support state, end-effector operation state, and body posture state under the same visual temporal anchor point, thus better adapting to the control requirements of the Centaur robot when facing complex ground support changes, body posture disturbances, and simultaneous end-effector operations.

[0023] Optionally, the step of inputting the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters into an embodied policy mapping network to perform implicit feature alignment processing on the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters includes: The multi-view depth image matrix in the multimodal parameters is input into the visual encoding branch of the embodied policy mapping network to extract global spatial geometric mapping features; The leg joint pose parameter, the arm joint pose parameter, and the inertial navigation parameter in the multimodal parameters are input into the ontology state encoding branch of the embodied policy mapping network to extract its own kinematic representation features. In the cross-mapping nodes of the embodied policy mapping network, the self-kinematic representation features are used as query guide vectors to perform addressing and focusing extraction processing on the global spatial geometric mapping features to generate spatial-body alignment features. Then, based on the spatial-body alignment features, heterogeneous feature vectors representing the body motion trend and end-effector intention of the Centaur robot are generated.

[0024] Preferably, the specific implementation process of inputting the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters into the embodied policy mapping network to perform feature implicit alignment processing on the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters is as follows: After the multimodal parameters have been formed through pre-sequence temporal state interpolation, a feature input carrying record is first established, and the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters in the multimodal parameters are written into the feature input carrying record. The feature input carrying record is used to include the environmental space state, chassis leg support state, dual arm end effector state, and fuselage body attitude state under the same first hardware timestamp into the same input caliber, so that the data subsequently entering the embodied policy mapping network are no longer independent parameters from scattered sources, but multimodal parameters with the same temporal relationship. Subsequently, the feature input carrying records are configured with input branching according to parameter type to form visual branch input records and ontology branch input records. The visual branch input records receive the multi-view depth image matrix and use the multi-view depth image matrix as the input source of the visual coding branch. The ontology branch input records receive the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters and use the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters as the input source of the ontology state coding branch. In this application, the embodied policy mapping network does not simply concatenate the multi-view depth image matrix and the ontology motion parameters, which include the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters. Instead, it first compresses the spatial distance distribution corresponding to the multi-view depth image matrix into global spatial geometric mapping features that can participate in subsequent addressing and focusing extraction processing through the visual encoding branch. Then, it converts the ontology motion parameters into self-kinematic representation features that can be used to guide the focusing and reading of the global spatial geometric mapping features through the ontology state encoding branch. Finally, it establishes an implicit correspondence between the self-kinematic representation features and the global spatial geometric mapping features through the cross mapping nodes. The implicit alignment processing here refers to not requiring each pixel position in the multi-view depth image matrix to establish an explicit coordinate correspondence with each joint state in advance, but rather forming a spatial-ontology alignment feature that can simultaneously express the environmental spatial state and the ontology motion state through the continuous processing of the visual encoding branch, the ontology state encoding branch, and the cross mapping node within the embodied policy mapping network. The ontology motion state is composed of the chassis leg support state, the dual arm end effector state, and the fuselage body posture state.Through the above input branching configuration, the multi-view depth image matrix, the leg joint pose parameter, the arm joint pose parameter, and the inertial navigation parameter each have their corresponding input destinations. Furthermore, the visual branch input record and the ontology branch input record subsequently participate in the formation process of the global spatial geometric mapping feature and the self-kinematic representation feature, respectively, thereby avoiding processing link mixing when different types of parameters enter the embodied policy mapping network.

[0025] Preferably, in the specific technical implementation of inputting the multi-view depth image matrix from the multimodal parameters into the visual encoding branch of the embodied policy mapping network to extract global spatial geometric mapping features, the visual encoding branch first reads the multi-view depth image matrix from the visual branch input record and performs view dimension normalization on the multi-view depth image matrix to form a view-normalized depth matrix. The view-normalized depth matrix retains the meaning of matrix layers, matrix rows, and matrix columns in the multi-view depth image matrix, wherein the matrix layers correspond to different acquisition views, the matrix rows correspond to pixel height positions, the matrix columns correspond to pixel width positions, and the elements at the intersection of the matrix rows and the matrix columns correspond to depth distance values; the view dimension normalization process is used to arrange the depth distance values ​​under different acquisition views into a unified matrix reading order, so that the spatial distance distributions corresponding to the forward acquisition view, the lateral acquisition view, and the downward acquisition view can be continuously read by the same visual encoding branch. Subsequently, the visual encoding branch performs local geometric texture extraction processing on the viewpoint-normalized depth matrix to form local geometric texture features. These features express local boundaries, local obstacles, and local ground height changes in areas such as underground utility tunnel bends, mine slope undulations, or close-range operation positions of chemical equipment. These local boundaries, obstacles, and ground height changes collectively constitute the local geometric changes in the environmental spatial state. To enable these local geometric texture features to express cross-viewpoint spatial continuity, the visual encoding branch further performs cross-viewpoint spatial association processing on them to form cross-viewpoint spatial association features. These features inherit from the local geometric texture features and incorporate the depth variation relationships of adjacent spatial regions under different acquisition perspectives into the same feature scope. Finally, the visual encoding branch performs global geometric compression processing on these cross-viewpoint spatial association features to generate global spatial geometric mapping features. The technical essence of the global spatial geometric mapping feature is that the multi-view depth image matrix, after undergoing view dimension normalization, local geometric texture extraction, cross-view spatial correlation, and global geometric compression, forms the environmental spatial representation. The elements in the global spatial geometric mapping feature no longer correspond to single pixel positions, but rather to the geometric occupancy trend, obstacle distribution trend, and passable space change trend of the surrounding space region of the Centaur robot under multiple acquisition views. The geometric occupancy trend, obstacle distribution trend, and passable space change trend all originate from the cross-view spatial correlation feature and are incorporated into the global spatial geometric mapping feature after global geometric compression. The global spatial geometric mapping feature subsequently enters the cross-mapping node and undergoes addressing and focusing extraction processing using the query guidance vector corresponding to its own kinematic representation feature.Through this step-by-step processing method of visual encoding branches, the multi-view depth image matrix is ​​not merely passively input as image data, but is converted into global spatial geometric mapping features that can be technically correlated with the motion state of the body. This enables subsequent spatial-body alignment features to carry the Centaur robot's perception basis for near-distance obstacles, slope support, and work space boundaries in special scenarios. Among these features, the near-distance obstacles, slope support, and work space boundaries all participate in subsequent addressing and focusing extraction processing through the geometric occupancy trend, the obstacle surface distribution trend, and the passable space change trend.

[0026] Preferably, in the specific technical implementation of inputting the leg joint pose parameters, arm joint pose parameters, and inertial navigation parameters from the multimodal parameters into the ontology state encoding branch of the embodied strategy mapping network to extract its own kinematic representation features, the ontology state encoding branch reads the leg joint pose parameters, arm joint pose parameters, and inertial navigation parameters from the ontology branch input record, and performs leg kinematic chain regularization processing on the leg joint pose parameters to form leg kinematic chain representation features. The leg kinematic chain representation features inherit the angle parameters, angular velocity parameters, and joint encoded position parameters of each leg joint, and express the chassis leg support state according to the installation order of each leg joint in the chassis leg actuator, so that the motion trend of the chassis leg actuator in the support, swing, or load-bearing transition process in the current control cycle can be read by the subsequent feature implicit alignment processing. The support, swing, or load-bearing transition process together constitutes the kinematic chain change content in the chassis leg support state. Simultaneously, the body state encoding branch performs arm kinematic chain regularization processing on the arm joint pose parameters to form arm kinematic chain representation features. These features incorporate the angle parameters, angular velocity parameters, and end-effector attitude parameters of each arm joint, and express the end-effector operation state according to the kinematic chain sequence of each arm joint within the dual-arm end-effector mechanism. This allows the operational trends of the dual-arm end-effector mechanism when approaching the close-range operation area, adjusting the end-effector attitude, or avoiding close-range obstacles to be guided by subsequent queries. Approaching the close-range operation area, adjusting the end-effector attitude, or avoiding close-range obstacles together constitute the operation chain changes within the dual-arm end-effector operation state. Subsequently, the body state encoding branch performs fuselage attitude encoding processing on the inertial navigation parameters to form fuselage attitude representation features. These features incorporate the angular velocity parameters, linear acceleration parameters, and attitude angle parameters from the inertial navigation parameters, and express the direction and degree of disturbance of the fuselage body attitude state within the current control cycle. Furthermore, the ontological state encoding branch performs ontological state fusion encoding processing on the leg kinematic chain representation features, the arm kinematic chain representation features, and the fuselage posture representation features to extract the self-kinematic representation features. The technical essence of the self-kinematic representation features is a description of the ontological motion state composed of the chassis leg support state, the arm end-effector operation state, and the fuselage posture state. The elements in the self-kinematic representation features correspond to the support change information of the leg kinematic chain, the operation change information of the arm kinematic chain, and the posture perturbation information of the fuselage, respectively, and all of these elements use the same first hardware timestamp as a time reference.The support change information of the leg kinematic chain originates from the leg kinematic chain representation features, the operation change information of the arm kinematic chain originates from the arm kinematic chain representation features, and the attitude perturbation information of the fuselage body originates from the fuselage attitude representation features. The support change information of the leg kinematic chain, the operation change information of the arm kinematic chain, and the attitude perturbation information of the fuselage body are collectively incorporated into the self-kinematic representation features after the body state fusion encoding process. These self-kinematic representation features not only serve as the output result of the body's motion state but are also converted into a query guidance vector in the cross-mapping node. The query guidance vector differs from a regular concatenation vector. It indicates which spatial regions within the global spatial geometric mapping features require more focused reading. For example, when the leg kinematic chain representation features indicate that the chassis leg actuator is traversing a slope support area, the query guidance vector will assign higher spatial focus weights to the spatial regions corresponding to the downward and lateral acquisition perspectives. Similarly, when the arm kinematic chain representation features indicate that the dual-arm end effectors are approaching a close-range operation area, the query guidance vector will assign higher spatial focus weights to the close-range obstacle surface region in the forward acquisition perspective. Thus, the self-kinematic representation features can transform the body's motion state into reading cues for the global spatial geometric mapping features, rather than existing independently of them.

[0027] Preferably, in the cross-mapping nodes of the embodied policy mapping network, the specific implementation process of using the self-kinematic representation features as a query guidance vector to perform address-focusing extraction processing on the global spatial geometric mapping features to generate spatial-ontology aligned features is as follows: The cross-mapping node first reads the global spatial geometric mapping features and the self-kinematic representation features, and performs query guidance transformation processing on the self-kinematic representation features to form the query guidance vector. The query guidance vector inherits the support change information of the leg kinematic chain, the operation change information of the arm kinematic chain, and the attitude perturbation information of the fuselage body from the self-kinematic representation features, and converts the body motion state into guidance information for reading the spatial region of the global spatial geometric mapping features. Subsequently, the cross-mapping node performs spatial address association processing on the global spatial geometric mapping features according to the query guidance vector to form a spatial address association record. The spatial address association record is used to record the correspondence between the body motion state in the query guidance vector and the spatial regions in the global spatial geometry mapping feature. For example, the support change information of the leg kinematic chain corresponds to the ground height change region in the downward acquisition view, the operation change information of the arm kinematic chain corresponds to the close-range operation region in the forward acquisition view, and the attitude disturbance information of the fuselage body corresponds to the obstacle surface distribution region in multiple acquisition views. The ground height change region, the close-range operation region, and the obstacle surface distribution region all originate from the spatial regions in the global spatial geometry mapping feature and establish a correspondence with the query guidance vector through the spatial address association record. Subsequently, the cross-mapping node performs focus weight allocation processing on the spatial address association record to form a spatial focus weight record. The spatial focus weight record inherits the spatial address association record and is used to represent the contribution degree of each spatial region to subsequent feature reading in the current control cycle. Further, the cross-mapping node performs addressing focus extraction processing on the global spatial geometry mapping feature based on the spatial focus weight record to generate the spatial-body alignment feature. The addressing focus extraction process does not randomly extract local features from the global spatial geometric mapping features, nor does it crop according to a fixed spatial region. Instead, it reads a spatial region from the global spatial geometric mapping features that is more strongly associated with the chassis leg support state, the dual arm end effector state, and the fuselage body attitude state, based on the body motion state expressed by the query guidance vector.The technical essence of the spatial-void alignment feature is that it is a fusion representation result formed by the global spatial geometric mapping feature guided by its own kinematic representation feature. The spatial-void alignment feature retains both the geometric occupancy trend, obstacle distribution trend, and passable space change trend in the environmental spatial state, as well as the read preference of the vocal body's motion state for the aforementioned spatial regions. This fusion representation result is subsequently used as the spatial-void alignment feature in the generation process of heterogeneous feature vectors, enabling the spatial-void alignment feature to continue transmitting the correspondence between the environmental spatial state and the vocal body's motion state to subsequent processing. For Centaur robots operating in scenarios involving complex ground support changes and simultaneous dual-arm end-effector manipulation, the spatial-void alignment feature allows the subsequently generated heterogeneous feature vectors to no longer reflect only the environmental spatial state or only the vocal body's motion state, but rather to express the correspondence between spatial regions and the vocal body's motion state within the same current control cycle.

[0028] Preferably, the specific implementation process of generating heterogeneous feature vectors representing the centaur robot's body motion trend and end-effector intention based on the spatial-body alignment features is as follows: After the spatial-body alignment features are formed at the cross-mapping nodes, the embodied policy mapping network first performs trend semantic separation processing on the spatial-body alignment features to form candidate features for body motion trends and candidate features for end-effector intentions. The candidate features for body motion trends incorporate elements from the spatial-body alignment features related to the chassis leg support state, the body posture state, and the trend of changes in the passable space, and are used to describe the centaur robot's body motion trend in the current control cycle; the candidate features for end-effector intentions incorporate elements from the spatial-body alignment features related to the end-effector state of the two arms, the near-field operation area, and the trend of obstacle distribution, and are used to describe the centaur robot's end-effector intentions in the current control cycle. Subsequently, motion trend compression processing is performed on the candidate features of fuselage motion trends to form fuselage motion trend features. These features express the motion adaptation relationship between the chassis leg actuators and the environmental spatial state within the current control cycle, such as slope support, obstacle avoidance, or attitude adjustment within narrow passages. Operation intention compression processing is performed on the candidate features of end-effector operation intentions to form end-effector operation intention features. These features express the operation adaptation relationship between the dual-arm end-effectors and the environmental spatial state within the current control cycle, such as close-range extension, end-effector attitude adjustment, or operation area avoidance. Slope support, obstacle avoidance, and attitude adjustment within narrow passages all enter the fuselage motion trend features through the candidate features, and close-range extension, end-effector attitude adjustment, and operation area avoidance all enter the end-effector operation intention features through the candidate features. Further, the fuselage motion trend features and the end-effector operation intention features are concatenated and encapsulated according to a predetermined feature arrangement caliber to generate the heterogeneous feature vector. The predetermined feature arrangement caliber is configured during the initialization of the embodied policy mapping network to specify the arrangement order, field boundaries, and reading method of the fuselage motion trend features and the end-operation intention features in the heterogeneous feature vector, so that the subsequent progressive capability self-construction process can read the fuselage motion trend features and the end-operation intention features from the heterogeneous feature vector respectively.The technical essence of the heterogeneous feature vector is to encapsulate the environmental spatial state and the body motion state after implicit feature alignment into a unified control input representation. Specifically, the fuselage motion trend feature in the heterogeneous feature vector originates from the expression of the chassis leg support state, the fuselage body attitude state, and the trend of changes in the passable space by the space-body alignment feature. Similarly, the end-effector intention feature in the heterogeneous feature vector originates from the expression of the dual-arm end-effector operation state, the near-field operation area, and the obstacle surface distribution trend by the space-body alignment feature. Based on the heterogeneous feature vector, the subsequent progressive capability self-construction process can read the fuselage motion trend feature and the end-effector intention feature from the same data carrier, ensuring that the generation processes of basic leg movement parameters, pure arm operation parameters, and leg-arm coordinated linkage parameters share a common source of the environmental spatial state and the body motion state.

[0029] Preferably, the structural layers of the embodied policy mapping network can be configured according to different data forms of the multi-view depth image matrix and the ontology motion parameters, and an implicit correspondence between the two types of features is established within the same current control cycle through the cross-mapping nodes. Specifically, the visual encoding branch includes a view dimension normalization layer, a multi-scale depth convolution layer, a cross-view geometric fusion layer, and a global geometric compression layer; the view dimension normalization layer reads the multi-view depth image matrix from the visual branch input record, and arranges the multi-view depth image matrix in a fixed reading order according to the acquisition view, pixel row position, and pixel column position to form a view dimension normalized depth matrix, so that the depth distance values ​​under different acquisition views can enter the continuous calculation process of the same visual encoding branch; the multi-scale depth convolution layer performs local neighborhood convolution processing on the view dimension normalized depth matrix to form local geometric texture features. Local geometric texture features are used to express local boundaries, local obstacle surfaces, and local ground height changes in underground pipe gallery bends, mine slope undulations, or close-range operation positions of chemical equipment. The cross-view geometric fusion layer performs cross-view adjacency fusion processing on the local geometric texture features to form cross-view spatial association features. The cross-view spatial association features are used to incorporate the depth change relationships of adjacent spatial regions in the forward acquisition view, lateral acquisition view, and downward acquisition view into the same feature carrying caliber. The global geometric compression layer performs channel merging and spatial region compression processing on the cross-view spatial association features to form the global spatial geometric mapping features. Accordingly, the body state encoding branch includes a leg kinematic chain encoding layer, an arm kinematic chain encoding layer, a fuselage attitude encoding layer, and a body state fusion encoding layer. The leg kinematic chain encoding layer reads the leg joint pose parameters and sequentially encodes the angle parameters, angular velocity parameters, and joint coded position parameters according to the installation order of each leg joint in the chassis leg actuator to form leg kinematic chain representation features. The arm kinematic chain encoding layer reads the arm joint pose parameters and sequentially encodes the angle parameters, angular velocity parameters, and end-effector attitude associated parameters according to the kinematic chain order of each arm joint in the dual arm end actuator to form arm kinematic chain representation features. The fuselage attitude encoding layer reads the inertial navigation parameters and encodes the angular velocity parameters, linear acceleration parameters, and attitude angle parameters for attitude change to form fuselage attitude representation features. The body state fusion encoding layer performs fusion encoding processing on the leg kinematic chain representation features, the arm kinematic chain representation features, and the fuselage attitude representation features to form the body kinematic representation features.Subsequently, the cross-mapping node includes a query guidance transformation layer, a spatial address association layer, a focus weight allocation layer, and an addressing focus extraction layer. The query guidance transformation layer converts the self-kinematic representation features into the query guidance vector. The spatial address association layer establishes a spatial address association record in the global spatial geometric mapping features based on the query guidance vector. The focus weight allocation layer forms a spatial focus weight record based on the spatial address association record. The addressing focus extraction layer reads the associated spatial region in the global spatial geometric mapping features based on the spatial focus weight record to form the spatial-ontology alignment feature. Through the above structural layers, the embodied policy mapping network can first extract the global spatial geometric mapping features and the self-kinematic representation features respectively, and then the cross-mapping node guides the reading of the spatial region of the global spatial geometric mapping features with the self-kinematic representation features, so that the spatial-ontology alignment feature can simultaneously express the environmental spatial state and the ontology motion state.

[0030] Preferably, the training samples of the embodied policy mapping network can be constructed based on the historical operating data of the Centaur robot in mine slopes, underground pipe gallery bends, and close-range operation positions of chemical equipment. Specifically, the multimodal parameter records formed within the historical control cycle are first read. The multimodal parameter records include the multi-view depth image matrix after time-series state interpolation processing, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters. The first hardware timestamp corresponding to the multimodal parameter records, the historical drive state of the chassis leg actuators, the historical operation state of the dual arm end effectors, and the historical attitude state of the robot body are retained. Subsequently, the multimodal parameter records are segmented according to application scenarios to form sample segments of mine slopes, underground utility tunnel curves, and close-range operation of chemical equipment. The mine slope sample segment provides the correspondence between slope support, local ground height changes, and chassis leg support status; the underground utility tunnel curve sample segment provides the correspondence between narrow passages, lateral obstacle distribution, and fuselage attitude adjustment; and the close-range operation of chemical equipment sample segment provides the correspondence between the close-range operation area, the end effector status of the dual arms, and the fuselage body attitude status. Further, based on the historical drive and operation states in each scene segment, fuselage motion trend training labels and end effector intention training labels are configured for the multimodal parameter records. The fuselage motion trend training labels express the motion trend of the chassis leg actuators within the corresponding historical control cycle, and the end effector intention training labels express the operation trend of the dual arm end effectors within the corresponding historical control cycle. For spatial region relationships that are inconvenient to label directly, spatial region association training labels can be formed based on the temporal correspondence between the historical driving state, the historical operation state, and the historical attitude state of the fuselage body. These spatial region association training labels are used to indicate spatial regions in the global spatial geometric mapping features that are strongly correlated with the chassis leg support state, the dual arm end-effector operation state, and the fuselage body attitude state. Subsequently, the multimodal parameter records, the fuselage motion trend training labels, the end-effector operation intention training labels, and the spatial region association training labels are encapsulated into embodied strategy training samples. When the embodied strategy training samples enter the training process, the visual encoding branch first generates the global spatial geometric mapping features based on the multi-view depth image matrix; the body state encoding branch then generates the self-kinematic representation features based on the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters; and the cross-mapping node then generates the spatial-body alignment features based on the global spatial geometric mapping features and the self-kinematic representation features.Through this training sample construction method, the embodied strategy training samples do not simply pair images and joint data, but incorporate the environmental spatial state, body motion state, fuselage motion trend training annotations, end-effector operation intention training annotations, and spatial region association training annotations under the same first hardware timestamp into the training basis. This enables the embodied strategy mapping network to learn which spatial regions affect the support changes of the chassis leg actuators in special scenarios, which spatial regions affect the operation changes of the dual arm end-effectors, and how the fuselage body attitude state participates in the spatial region reading process.

[0031] Preferably, the training process of the embodied policy mapping network can be completed using a method of multiple training losses constrained together. At the start of training, the multi-view depth image matrix from the embodied policy training samples is input into the visual encoding branch to obtain the global spatial geometric mapping features; the leg joint pose parameters, arm joint pose parameters, and inertial navigation parameters from the embodied policy training samples are input into the ontology state encoding branch to obtain the self-kinematic representation features; the global spatial geometric mapping features and the self-kinematic representation features are input into the cross-mapping node to obtain the spatial-ontology alignment features; then, trend semantic separation processing is performed on the spatial-ontology alignment features to obtain candidate features for fuselage motion trends and candidate features for end-effector intentions, further forming fuselage motion trend features, end-effector intention features, and the heterogeneous feature vector. During training, global spatial geometric representation loss, ontology state representation loss, spatial region association loss, fuselage motion trend loss, and end-effector intention loss can be set. The global spatial geometric representation loss is used to constrain the global spatial geometric mapping features generated by the visual encoding branch, so that the global spatial geometric mapping features can retain the geometric occupancy trend, obstacle surface distribution trend, and passable space change trend in the multi-view depth image matrix; the ontology state representation loss is used to constrain the self-kinematic representation features generated by the ontology state encoding branch, so that the self-kinematic representation features can retain the chassis leg support state corresponding to the leg joint pose parameters, the dual arm end-effector operation state corresponding to the arm joint pose parameters, and the fuselage body attitude state corresponding to the inertial navigation parameters; the spatial region association loss is used to constrain the spatial address association record and the spatial focus weight record generated by the cross mapping node, so that the spatial focus weight record and the spatial region association training label form a consistent spatial region reading relationship; the fuselage motion trend loss is used to constrain the fuselage motion trend features, so that the fuselage motion trend features and the fuselage motion trend training label maintain a high consistency; the end-effector operation intention loss is used to constrain the end-effector operation intention features, so that the end-effector operation intention features and the end-effector operation intention training label maintain a high consistency. The aforementioned training losses are weighted and fused according to pre-configured training weights to form the training target loss. The training weights are configured according to the task emphasis of the scene segment during training initialization. For example, in the mine slope sample segment, the training weights corresponding to the fuselage motion trend loss and the spatial region association loss can be increased. In the chemical equipment close-range operation sample segment, the training weights corresponding to the end-operation intention loss and the spatial region association loss can be increased.Subsequently, the trainable parameters in the visual encoding branch, the ontology state encoding branch, and the cross-mapping node are iteratively updated according to the training objective loss until the decrease in the training objective loss is lower than the preset training convergence range in multiple consecutive training rounds, or the deviation between the fuselage motion trend feature and the fuselage motion trend training label, and the deviation between the end-operation intention feature and the end-operation intention training label, both reach the preset training completion conditions. After training, the embodied policy mapping network retains the post-trained parameters in the visual encoding branch, the ontology state encoding branch, and the cross-mapping node, and uses the post-trained parameters to perform implicit feature alignment processing on the multimodal parameters in the current control cycle during actual control, thereby generating the heterogeneous feature vector that can enter the progressive capability self-construction process.

[0032] Optionally, the step of generating heterogeneous feature vectors characterizing the centaur robot's body motion trends and end-effector intentions based on the spatial-void alignment features includes: The spatial-ontology alignment features are input into the intent decoupling network layer of the embodied policy mapping network to perform feature stripping processing on the spatial-ontology alignment features, resulting in the first sub-feature and the second sub-feature. The first sub-feature is mapped to a continuous probability distribution of foot contact with the ground to form a body motion trend feature that characterizes the body motion trend of the centaur robot. The second sub-feature is mapped to a six-degree-of-freedom pose reachability distribution within the workspace to form an end-operation intention feature characterizing the end-operation intention of the Centaur robot. The fuselage motion trend features and the end-operation intention features are spliced ​​together to form the heterogeneous feature vector.

[0033] Preferably, the specific implementation process of inputting the spatial-void alignment features into the intent decoupling network layer of the embodied policy mapping network to perform feature stripping processing on the spatial-void alignment features to obtain the first sub-feature and the second sub-feature is as follows: After the spatial-void alignment features have been generated at the cross-mapping node, the spatial-void alignment features are first read, and an intent decoupling bearer record is established; the intent decoupling bearer record carries the environmental spatial state and the body motion state in the spatial-void alignment features, wherein the environmental spatial state includes geometric occupancy trend, obstacle surface distribution trend, and passable space change trend, and the body motion state includes chassis leg support state, dual-arm end effector state, and fuselage body attitude state. Subsequently, the spatial-void alignment features are input into the alignment feature sorting layer in the intent decoupling network layer. The alignment feature sorting layer sorts the spatial region elements and motion state elements in the spatial-void alignment features according to the feature source order within the same current control cycle to form the alignment feature sorting result. The alignment feature processing result is not a new control output, but rather arranges the spatial region reading preferences, chassis leg support status, dual-arm end-effector operation status, and fuselage body attitude status already included in the spatial-body alignment features into a feature carrier form that can be subsequently branched for reading. The spatial region reading preferences originate from the spatial focus weight records formed when the cross-mapping nodes perform addressing and focusing extraction processing on the global spatial geometric mapping features, and are further used in the alignment feature processing result to distinguish the degree of contribution of different spatial regions to the fuselage motion trend and end-effector operation intent. Subsequently, the shared semantic compression layer in the intent decoupling network layer performs shared semantic compression processing on the alignment feature processing result to form shared semantic compressed features. The shared semantic compressed features are used to retain the environmental spatial states that both the fuselage motion trend and the end-effector operation intent depend on, such as the distribution of near-field obstacles, changes in slope support areas, and trends in traversable space in narrow passages. Furthermore, the dual-channel gated stripping layer in the intent decoupling network layer reads the shared semantic compressed features and performs feature stripping processing on the shared semantic compressed features based on the kinematic chain changes corresponding to the chassis leg support state and the operational chain changes corresponding to the end-effector operation state, to obtain the first sub-feature and the second sub-feature. The first sub-feature and the second sub-feature are two sub-level representation results formed by the same spatial-void alignment feature through different reading directions. The difference between the two is that the first sub-feature inherits elements that are more strongly associated with the chassis leg support state, the fuselage body attitude state, and the trend of changes in the passable space, and is subsequently used to form the fuselage motion trend feature; the second sub-feature inherits elements that are more strongly associated with the end-effector operation state, the close-range operation area, and the obstacle surface distribution trend, and is subsequently used to form the end-effector operation intent feature.Through the alignment feature organization layer, shared semantic compression layer, and dual-channel gated stripping layer of the intent decoupling network layer, both the first sub-feature and the second sub-feature can be traced back to the spatial-ontology alignment feature, and the two enter different subsequent mapping processes, thereby avoiding mixing the fuselage movement trend and the end-operation intent in the same indistinguishable feature carrier.

[0034] Preferably, in the specific technical implementation of mapping the first sub-feature to a continuous foot-to-ground probability distribution to form a body motion trend feature characterizing the centaur robot's body motion trend, the intent decoupling network layer first writes the first sub-feature into a body motion trend mapping record. This record is used to carry elements in the first sub-feature related to the chassis leg support state, the body posture state, and the changing trend of the passable space. Subsequently, the foot-to-ground state reading layer in the intent decoupling network layer performs foot-to-ground association reading processing on the first sub-feature to form a foot-to-ground association representation feature. This feature carries the chassis leg support state from the first sub-feature and expresses the support relationship between each foot and the ground space area according to the installation order of each leg joint in the chassis leg actuator. Further, the terrain contact discrimination layer in the intent decoupling network layer reads the foot-end association representation features and, combined with elements from the traversable space change trend in the first sub-feature, performs continuous mapping processing on the ground contact state of each foot end within the current control cycle to form a continuous foot-end ground contact probability distribution. This continuous foot-end ground contact probability distribution is not a single binary result of ground contact or non-ground contact, but rather expresses the trend of each foot end contacting the ground space area at different time positions or different support phases in the form of continuous probabilities. Each probability element in the continuous foot-end ground contact probability distribution corresponds to the probability of a foot end contacting the ground space area within the current control cycle, and the order of the probability elements in the continuous foot-end ground contact probability distribution is consistent with the installation order of the leg joints in the chassis leg actuator. Subsequently, the fuselage motion trend integration layer in the intent decoupling network layer reads the continuous foot-end ground contact probability distribution and, combined with the fuselage body attitude state in the first sub-feature, performs fuselage motion trend integration processing on the continuous foot-end ground contact probability distribution to form the fuselage motion trend feature. The technical essence of the aforementioned fuselage motion trend feature is to convert the content related to support phase, traversable ground area, and fuselage body attitude disturbance in the first sub-feature into a motion trend representation that can be read by the subsequent progressive capability self-construction process. This motion trend representation continues to participate in the generation process of basic leg movement parameters in the form of the fuselage motion trend feature. For scenarios where the Centaur robot is moving on a mine slope or turning in a bend in an underground utility tunnel, the continuous foot contact probability distribution can reflect the contact changes of each foot during slope support, ground height changes, or attitude adjustments in narrow passages. The fuselage motion trend feature further combines the continuous foot contact probability distribution with the fuselage body attitude state, enabling the subsequent generation of basic leg movement parameters to read the motion trend source related to the chassis leg actuators.Therefore, the first sub-feature is processed by foot-end association reading, continuous mapping, and fuselage motion trend integration to form the fuselage motion trend feature, and a portion of the information biased towards the moving support in the space-body alignment feature is transmitted to the subsequent control parameter generation process.

[0035] Preferably, the specific implementation process of mapping the second sub-feature to a six-DOF pose reachability distribution within the workspace to form an end-effector intention feature characterizing the end-effector intention of the Centaur robot is as follows: After the second sub-feature has been formed by the dual-channel gating stripping layer, the second sub-feature is first written into the end-effector intention mapping carrying record. The end-effector intention mapping carrying record is used to carry elements in the second sub-feature related to the end-effector operation state, near-field operation area, and obstacle distribution trend. Subsequently, the end-effector posture reading layer in the intention decoupling network layer performs end-effector posture reading processing on the second sub-feature to form an end-effector posture representation feature. The end-effector posture representation feature carries the end-effector operation state in the second sub-feature and expresses the end-effector position change, end-effector orientation change, and end-effector posture adjustment trend in the current control cycle according to the kinematic chain sequence of each arm joint in the end-effector. Furthermore, the workspace meshing mapping layer in the intent decoupling network layer reads the end-effector pose representation features and, combined with the near-field operating region and obstacle distribution trend in the second sub-feature, performs pose candidate region mapping processing on the workspace to form a workspace pose candidate record. The workspace pose candidate record expresses candidate operating space positions that the dual-arm end-effector can reach or approach within the current control cycle. Each candidate operating space position in the workspace pose candidate record has a source relationship with the near-field operating region, the obstacle distribution trend, and the dual-arm end-effector operating state. Subsequently, the pose accessibility mapping layer in the intent decoupling network layer performs six-degree-of-freedom pose accessibility mapping processing on the workspace pose candidate record to form a six-degree-of-freedom pose accessibility distribution within the workspace. The six-degree-of-freedom (DOF) pose accessibility distribution within the workspace is used to express the accessibility of the dual-arm end effector under different candidate operational spatial positions and different candidate postures within the workspace. Specifically, the position component of the six-DOF pose accessibility distribution corresponds to changes in spatial position within the workspace, and the posture component corresponds to changes in the end-effector orientation of the dual-arm end effector. Further, the end-effector operation intent integration layer in the intent decoupling network layer reads the six-DOF pose accessibility distribution within the workspace and, in conjunction with elements related to the near-field operation area and obstacle surface distribution trends in the second sub-feature, performs operation intent integration processing on the six-DOF pose accessibility distribution within the workspace to form the end-effector operation intent feature.The technical essence of the end-effector operation intent feature is to convert the content related to the operational accessibility, operation direction, and near-field environmental constraints of the dual-arm end effector in the second sub-feature into an operational intent representation that can be read by the subsequent progressive capability self-construction process. This operational intent representation continues to participate in the generation process of pure arm operation parameters in the form of the end-effector operation intent feature. For the scenario where the Centaur robot performs dual-arm operations in a near-field operation position in chemical equipment, the six-degree-of-freedom pose accessibility distribution in the workspace can reflect the accessibility changes of the dual-arm end effector when approaching the near-field operation area, adjusting the end-effector posture, and avoiding near-field obstacles. The end-effector operation intent feature further combines the six-degree-of-freedom pose accessibility distribution in the workspace with the obstacle surface distribution trend, so that the source of the operational intent related to the dual-arm end effector can be read when generating pure arm operation parameters.

[0036] Preferably, the specific implementation process of concatenating the fuselage motion trend feature and the end-effector intention feature to form the heterogeneous feature vector is as follows: After the fuselage motion trend feature and the end-effector intention feature are formed respectively, a feature arrangement configuration record is first read. The feature arrangement configuration record is configured during the initialization of the embodied strategy mapping network. The feature arrangement configuration record includes the field start position, field end position, field start position, field end position, and reading order of the subsequent progressive capability self-construction process of the fuselage motion trend feature, the end position, and the end position of the end-effector intention feature. The technical essence of the feature arrangement configuration record is to configure the field boundaries in the same heterogeneous feature vector for the fuselage motion trend feature and the end-effector intention feature, so that the fuselage motion trend feature can be read when generating basic leg movement parameters, the end-effector intention feature can be read when generating pure arm operation parameters, and both the fuselage motion trend feature and the end-effector intention feature can be read simultaneously when generating leg-arm coordinated linkage parameters. Subsequently, the body motion trend features are encapsulated into a motion trend field to form a body motion trend field. This field incorporates the continuous foot-to-ground probability distribution, chassis leg support state, and body posture state from the body motion trend features. Similarly, the end-effector intention features are encapsulated into an intention field to form an end-effector intention field. This field incorporates the six-DOF pose accessibility distribution within the workspace, the dual-arm end-effector state, and obstacle distribution trends from the end-effector intention features. Further, based on the feature arrangement and configuration record, the body motion trend field and the end-effector intention field are concatenated to form the heterogeneous feature vector. The body motion trend field in the heterogeneous feature vector expresses the centaur robot's body motion trend, and the end-effector intention field expresses the centaur robot's end-effector intention. Although both originate from the same space-body alignment feature, they have already undergone mapping processing with different technical meanings through the first and second sub-features in the intention decoupling network layer. Therefore, the heterogeneous feature vector is no longer a simple vector after compressing the spatial-ontology alignment feature. Instead, according to the reading requirements of the subsequent progressive capability self-construction process, the fuselage motion trend feature and the end-operation intention feature are encapsulated in the same data carrier with clear field boundaries.For Centaur robots in scenarios where complex ground support changes and dual-arm end-effector operations occur simultaneously, the heterogeneous feature vector can simultaneously retain the source of the body motion trend related to the mobile support and the source of the end-effector intention related to the operation reachability. This allows the subsequent generation process of basic leg movement parameters, pure arm operation parameters, and leg-arm coordinated linkage parameters to use the two sub-level representation results derived from the same space-body alignment feature, avoiding the disconnect between chassis leg control and dual-arm end-effector operations at the feature input level.

[0037] Preferably, the difference between the first sub-feature and the second sub-feature, and their relationship with the heterogeneous feature vector, can be further understood as follows: the first sub-feature is not an arbitrary segment of data extracted from the spatial-void alignment feature, but rather a representation of the moving support direction extracted from the shared semantic compression feature by the dual-channel gating stripping layer in the intent decoupling network layer based on the chassis leg support state, fuselage body attitude state, and the trend of change in passable space; the second sub-feature is not the remaining ordinary data in the spatial-void alignment feature, but rather a representation of the end-operation direction extracted from the shared semantic compression feature by the dual-channel gating stripping layer based on the end-operation state of the dual arms, the near-field operation area, and the trend of obstacle surface distribution. The moving support direction representation participates in the formation process of the fuselage motion trend feature in the form of the first sub-feature, and the end-operation direction representation participates in the formation process of the end-operation intent feature in the form of the second sub-feature. The first and second sub-features share the spatial-void alignment feature in their origin and participate in the formation of the body motion trend feature and the end-effector operation intention feature, respectively. Therefore, they are sub-features with the same origin and direction, rather than independent features that are not related to each other. Further, the first sub-feature, after foot-end association reading processing, continuous mapping processing, and body motion trend integration processing, forms the body motion trend feature. This feature is subsequently used in the heterogeneous feature vector to express the movement support trend of the Centaur robot's chassis leg actuators. The second sub-feature, after end-effector posture reading processing, pose candidate region mapping processing, six-DOF pose reachability mapping processing, and operation intention integration processing, forms the end-effector operation intention feature. This feature is subsequently used in the heterogeneous feature vector to express the operation reachability trend of the Centaur robot's dual-arm end-effectors. The movement support trend originates from the body motion trend feature and is read when generating basic leg movement parameters; the operation reachability trend originates from the end-effector operation intention feature and is read when generating pure arm operation parameters. Through this homogeneous, directional, and reassembled encapsulation process, the heterogeneous feature vectors can retain the consistency of origin between the fuselage motion trend and the end-effector's operational intent, while also preserving the differences in their readings during subsequent control parameter generation. Specifically, for underground utility tunnel bends, mine slopes, or close-range operation positions of chemical equipment, the first sub-feature enables the heterogeneous feature vectors to express the motion trend of the chassis leg actuators when facing changes in ground height and fuselage attitude disturbances. The second sub-feature enables the heterogeneous feature vectors to express the operational intent of the dual-arm end-effectors when facing the close-range operation area and obstacle distribution trends. Together, these two features enable the subsequent progressive capability self-construction process to obtain both movement and operation information from the same control input representation.

[0038] Optionally, in the embodied policy mapping network, a progressive capability self-construction process is performed using the heterogeneous feature vector as the processing object to sequentially generate basic leg movement parameters, pure arm operation parameters, and leg-arm coordinated linkage parameters, including: In the first strategy phase, the network connection weights related to bi-arm control in the embodied strategy mapping network are locked to generate the basic leg movement parameters corresponding to the quadrupedal movement pattern based on the heterogeneous feature vector. In the second strategy phase, the network connection weights related to chassis leg control in the embodied strategy mapping network are locked, and the basic leg movement parameters are kept in a static state, so as to generate the pure arm operation parameters corresponding to the static base operation mode based on the heterogeneous feature vector.

[0039] Preferably, the specific implementation process of performing a progressive capability self-construction process in the embodied strategy mapping network, using the heterogeneous feature vector as the processing object to sequentially generate basic leg movement parameters, pure arm operation parameters, and leg-arm coordinated linkage parameters, is as follows: After the heterogeneous feature vector has been formed by splicing the fuselage motion trend feature and the end-effector operation intention feature, a stage control bearing record is first established, and the heterogeneous feature vector, the fuselage motion trend feature, the end-effector operation intention feature, the current control cycle marker, and the structural layer identifier in the embodied strategy mapping network are written into the stage control bearing record. The stage control bearing record is used to incorporate the reading method of the same heterogeneous feature vector in different strategy stages into the same stage control caliber, so that the first strategy stage and the second strategy stage are not isolated strategy processing segments, but a progressive processing process of sequentially establishing chassis leg control capabilities and dual-arm control capabilities around the same heterogeneous feature vector. The structural layer identifiers in the embodied strategy mapping network include at least the view dimension normalization layer, multi-scale deep convolutional layer, cross-view geometric fusion layer, and global geometric compression layer in the visual encoding branch; the leg kinematic chain encoding layer, arm kinematic chain encoding layer, fuselage attitude encoding layer, and ontology state fusion encoding layer in the ontology state encoding branch; the query guidance transformation layer, spatial address association layer, focus weight allocation layer, and addressing focus extraction layer in the cross mapping node; and the alignment feature organization layer, shared semantic compression layer, dual-channel gating stripping layer, foot state reading layer, terrain contact discrimination layer, fuselage motion trend integration layer, end effector attitude reading layer, workspace gridding mapping layer, pose accessibility mapping layer, and end effector operation intent integration layer in the intent decoupling network layer. These structural layer identifiers are subsequently used to distinguish between network connection weights related to chassis leg control and network connection weights related to dual arm control, enabling these weights to enter different locked or updatable states in the first and second strategy stages, respectively. The difference between the first strategy stage and the second strategy stage is that the first strategy stage uses the fuselage motion trend characteristics as the main reading content, and by locking the network connection weights related to the dual arm control, the body strategy mapping network first forms the basic leg movement parameters corresponding to the quadrupedal motion mode; the second strategy stage uses the end-operation intention characteristics as the main reading content, and by locking the network connection weights related to the chassis leg control and keeping the basic leg movement parameters in a static state, the body strategy mapping network generates the pure arm operation parameters corresponding to the static base operation mode without changing the chassis leg support benchmark that has already been formed.Through the stage control bearing record, the basic leg movement parameters generated in the first strategy stage enter the static state configuration process of the second strategy stage. The static state configuration process is used to convert the basic leg movement parameters into a chassis leg support reference that can be read by the second strategy stage. The pure arm operation parameters generated in the second strategy stage, together with the basic leg movement parameters, provide a source for the generation of subsequent leg-arm coordinated linkage parameters.

[0040] Preferably, in the specific technical implementation of locking the network connection weights related to dual-arm control in the embodied strategy mapping network during the first strategy stage to generate the basic leg movement parameters corresponding to the quadrupedal movement mode based on the heterogeneous feature vector, the heterogeneous feature vector and the fuselage movement trend features are first read from the stage control bearing record, and a first strategy stage weight locking record is established. The first strategy stage weight locking record is used to record the network connection weights related to dual-arm control that are in a locked state and the network connection weights related to chassis leg control that are in an updatable state during the first strategy stage. The network connection weights related to dual-arm control specifically include the network connection weights in the arm kinematic chain encoding layer used to process the arm joint pose parameters, the network connection weights in the end-effector posture reading layer used to read the second sub-feature, the network connection weights in the workspace meshing mapping layer used to form workspace pose candidate records, the network connection weights in the pose accessibility mapping layer used to form the six-degree-of-freedom pose accessibility distribution in the workspace, and the network connection weights in the end-effector intention integration layer used to form end-effector intention features. The network connection weights related to dual-arm control are configured to be locked in the first strategy phase. Specifically, during the first strategy phase, weight lock records are written with weight update masking flags to the arm kinematic chain encoding layer, the end effector posture reading layer, the workspace mesh mapping layer, the pose accessibility mapping layer, and the end effector operation intent integration layer. This ensures that the training error or stage evaluation result does not change the network connection weights related to dual-arm control during backpropagation. Simultaneously, the network connection weights related to chassis leg control remain updatable. These weights include network connection weights in the leg kinematic chain encoding layer used to process leg joint pose parameters, network connection weights in the foot state reading layer used to read the first sub-feature, network connection weights in the terrain contact discrimination layer used to form a continuous foot contact probability distribution, and network connection weights in the fuselage motion trend integration layer used to form fuselage motion trend features. Subsequently, the embodied strategy mapping network reads the continuous foot-to-ground probability distribution, the chassis leg support state, and the fuselage body posture state based on the fuselage motion trend features in the heterogeneous feature vector, and performs quadrupedal motion pattern mapping processing on the continuous foot-to-ground probability distribution, the chassis leg support state, and the fuselage body posture state to generate the basic leg movement parameters.The basic leg movement parameters are used to express the foot support sequence, foot swing trend, body posture adjustment direction, and the source of changes in chassis leg support state of the chassis leg actuator within the current control cycle. These parameters all originate from the body motion trend characteristics' expression of the continuous foot contact probability distribution, chassis leg support state, and body posture state. For scenarios involving mine slopes or underground utility tunnel curves, the first strategy phase allows the embodied strategy mapping network to first establish a correspondence between the chassis leg actuator and the passable ground space. This ensures that the basic leg movement parameters can reflect the quadrupedal movement pattern of the Centaur robot under complex ground support changes, without being interfered with by changes in the operation of the end effectors of the dual arms. The basic leg movement parameters subsequently enter the static state configuration process of the second strategy phase and serve as the chassis leg support reference when the pure arm operation parameters are generated in the second strategy phase.

[0041] Preferably, in the second strategy phase, in the specific technical implementation of locking the network connection weights related to chassis leg control in the embodied strategy mapping network and keeping the basic leg movement parameters in a static state to generate the pure arm operation parameters corresponding to the static base operation mode based on the heterogeneous feature vector, the basic leg movement parameters generated in the first strategy phase are first read from the phase control bearing record, and a second strategy phase weight locking record is established. The second strategy phase weight locking record is used to record the network connection weights related to chassis leg control that are in a locked state and the network connection weights related to dual arm control that are in an updatable state in the second strategy phase. The network connection weights related to chassis leg control are configured to be in a locked state in the second strategy phase, specifically by writing weight update masking flags to the leg kinematic chain encoding layer, the foot end state reading layer, the terrain contact discrimination layer, and the fuselage motion trend integration layer in the second strategy phase, so that the chassis leg support state corresponding to the basic leg movement parameters does not change with training error or phase evaluation results in the second strategy phase. Subsequently, the basic leg movement parameters are written into the static base state record. This static base state record is used to convert the foot support sequence, foot swing trend, fuselage attitude adjustment direction, and chassis leg support state change sources from the basic leg movement parameters into base constraint sources in the static base operation mode. Here, "static state" does not mean that the chassis leg actuators completely stop participating in data, but rather that the basic leg movement parameters, as an established chassis leg support benchmark, enter the second strategy stage, enabling the embodied strategy mapping network to read stable base constraint sources when generating the pure arm operation parameters. The chassis leg support benchmark originates from the basic leg movement parameters, and the base constraint source originates from the static base state record. The base constraint source is used to limit the disturbance of the pure arm operation parameters on the basic leg movement parameters during the static base operation mapping process. Meanwhile, the network connection weights related to dual-arm control remain updatable during the second strategy phase. The embodied strategy mapping network reads the six-degree-of-freedom pose reachability distribution, dual-arm end-of-arm operation state, and obstacle distribution trend in the workspace based on the end-of-arm operation intention features in the heterogeneous feature vector. It then performs static base operation mapping processing on the six-degree-of-freedom pose reachability distribution, dual-arm end-of-arm operation state, and obstacle distribution trend in the workspace in conjunction with the static base state record to generate the pure arm operation parameters.The pure arm operation parameters are used to express the end-effector's position adjustment direction, end-effector orientation adjustment direction, source of arm joint kinematic chain changes, and source of near-distance obstacle avoidance within the current control cycle. These parameters all originate from the end-effector's operational intent characteristics, which express the six-degree-of-freedom pose accessibility distribution within the workspace, the end-effector's operational state, and the obstacle surface distribution trend. For near-distance operation positions in chemical equipment, the second strategy stage enables the embodied strategy mapping network to centrally establish a mapping relationship between the end-effector and the six-degree-of-freedom pose accessibility distribution within the workspace, given that the chassis leg support reference for the chassis leg actuator has already been formed. This allows the pure arm operation parameters to reflect the operational trend of the end-effector in static base operation mode. The pure arm operation parameters subsequently enter the third strategy stage together with the basic leg movement parameters to provide the arm operation basis and leg movement basis for the leg-arm coordinated linkage parameters.

[0042] Preferably, the progressive relationship between the first strategy stage and the second strategy stage can be further realized through stage training constraints. Specifically, in the first strategy stage, after the embodied strategy mapping network reads the heterogeneous feature vector, it mainly establishes the basic leg movement parameters around the fuselage motion trend features. Therefore, the training constraints corresponding to the first strategy stage include foot contact distribution loss, chassis attitude stability loss, and leg movement continuity loss. The foot contact distribution loss is used to constrain the deviation between the continuous foot contact probability distribution and the foot contact state in the historical control cycle. The chassis attitude stability loss is used to constrain the deviation between the fuselage attitude adjustment direction corresponding to the basic leg movement parameters and the fuselage body attitude state change in the historical control cycle. The leg movement continuity loss is used to constrain the change amplitude of the basic leg movement parameters between adjacent current control cycles, so that the basic leg movement parameters can express a continuous quadrupedal movement pattern. The foot contact distribution loss, the chassis attitude stability loss, and the leg movement continuity loss together form the training objective loss for the first strategy stage. This first strategy stage training objective loss is only used to update the network connection weights related to chassis leg control, and avoids the network connection weights related to arm control by locking records in the first strategy stage. The first strategy stage training objective loss is used to inversely affect the network connection weights related to chassis leg control in the leg kinematic chain encoding layer, the foot state reading layer, the terrain contact discrimination layer, and the fuselage motion trend integration layer, enabling these structural layers to form chassis leg control capabilities related to the basic leg movement parameters. In the second strategy phase, after the embodied strategy mapping network reads the heterogeneous feature vector, it mainly establishes the pure arm operation parameters around the end-effector operation intent features. Therefore, the training constraints corresponding to the second strategy phase include pose reachability distribution loss, end-effector operation deviation loss, and base perturbation suppression loss. The pose reachability distribution loss is used to constrain the deviation between the six-degree-of-freedom pose reachability distribution in the workspace and the dual-arm end-effector reachability state in the historical control cycle. The end-effector operation deviation loss is used to constrain the deviation between the end-effector position adjustment direction and end-effector orientation adjustment direction corresponding to the pure arm operation parameters and the dual-arm end-effector operation state in the historical control cycle. The base perturbation suppression loss is used to constrain the perturbation amplitude of the static base state record during the generation of the pure arm operation parameters, so that the pure arm operation parameters are formed in the static base operation mode. The pose reachability distribution loss, the end-effector deviation loss, and the base disturbance suppression loss together form the training objective loss for the second strategy stage. The training objective loss for the second strategy stage is only used to update the network connection weights related to the dual-arm control, and the network connection weights related to the chassis leg control are avoided by locking the records through the weights of the second strategy stage.The training objective loss of the second strategy stage is directed to act inversely on the network connection weights related to dual-arm control in the arm kinematic chain encoding layer, the end-effector posture reading layer, the workspace mesh mapping layer, the pose accessibility mapping layer, and the end-effector operation intent integration layer, enabling the aforementioned structural layers to form dual-arm control capabilities related to the pure arm operation parameters. Through the phased constraints of the training objective loss of the first and second strategy stages, the embodied strategy mapping network first forms the basic leg movement parameters corresponding to the quadrupedal movement mode, and then forms the pure arm operation parameters corresponding to the static base operation mode under the condition that the basic leg movement parameters remain in a static state. This provides a leg movement basis and arm operation basis with a source relationship for the subsequent generation of the leg-arm coordinated linkage parameters in the third strategy stage.

[0043] Optionally, the step of performing a progressive capability self-construction process in the embodied policy mapping network, using the heterogeneous feature vector as the processing object, to sequentially generate basic leg movement parameters, pure arm operation parameters, and leg-arm coordinated linkage parameters, further includes: In the third strategy phase, the locked state of the network connection weights related to dual-arm control and the network connection weights related to chassis leg control in the embodied strategy mapping network is released. The basic leg movement parameters and the pure arm operation parameters are used as prior guiding input nodes to guide the embodied policy mapping network to remap the heterogeneous feature vector under dynamic base perturbation, thereby generating the leg-arm coordinated linkage parameters.

[0044] Preferably, the specific implementation process for unlocking the network connection weights related to dual-arm control and chassis leg control in the embodied strategy mapping network during the third strategy phase is as follows: After the basic leg movement parameters have been generated in the first strategy phase and the pure arm operation parameters have been generated in the second strategy phase, the phase control bearing record is read first, and the heterogeneous feature vector, the basic leg movement parameters, the pure arm operation parameters, the first strategy phase weight locking record, the second strategy phase weight locking record, and the current control cycle marker are extracted from the phase control bearing record, thereby establishing a third strategy phase weight unlocking record. The third strategy phase weight unlocking record is used to take over the network connection weights related to dual-arm control that are locked in the first strategy phase, and the network connection weights related to chassis leg control that are locked in the second strategy phase, so that the network connection weights related to dual-arm control and the network connection weights related to chassis leg control can be restored to an updatable state in the same third strategy phase. The difference between the first strategy stage, the second strategy stage, and the third strategy stage is as follows: The first strategy stage locks the network connection weights related to the dual-arm control, enabling the embodied strategy mapping network to prioritize the establishment of the basic leg movement parameters in the quadrupedal movement mode; the second strategy stage locks the network connection weights related to the chassis leg control and keeps the basic leg movement parameters in a static state, enabling the embodied strategy mapping network to establish the pure arm operation parameters in the static base operation mode; the third strategy stage no longer places the chassis leg actuators and dual-arm end effectors in separate independent strategy processing segments, but instead unlocks the network connection weights related to the dual-arm control and the network connection weights related to the chassis leg control, allowing the basic leg movement parameters and the pure arm operation parameters to jointly participate in the remapping of the heterogeneous feature vectors. Through the weight unlocking record of the third strategy stage, the leg movement basis formed in the first strategy stage and the arm operation basis formed in the second strategy stage no longer remain in their respective independent control branches, but are incorporated into the collaborative mapping caliber under the same dynamic base disturbance state; wherein, the leg movement basis corresponds to the foot support sequence, foot swing trend, fuselage attitude adjustment direction and chassis leg support state change source already formed in the basic leg movement parameters, and the arm operation basis corresponds to the end position adjustment direction, end orientation adjustment direction, arm joint kinematic chain change source and near obstacle avoidance source already formed in the pure arm operation parameters. The leg movement basis and the arm operation basis together provide a common stage source for the subsequent generation of the leg-arm collaborative linkage parameters.

[0045] Preferably, in the specific technical implementation of unlocking the network connection weights related to the dual-arm control and the network connection weights related to the chassis leg control in the third strategy stage, the weight unlocking record in the third strategy stage first reads the structural layer identifier in the embodied strategy mapping network, and locates the network connection weights related to the chassis leg control and the network connection weights related to the dual-arm control based on the structural layer identifier. The network connection weights related to chassis leg control include network connection weights in the leg kinematic chain encoding layer for processing the leg joint pose parameters, network connection weights in the foot state reading layer for reading the first sub-feature, network connection weights in the terrain contact discrimination layer for forming a continuous foot contact probability distribution, and network connection weights in the fuselage motion trend integration layer for forming fuselage motion trend features; the network connection weights related to dual arm control include network connection weights in the arm kinematic chain encoding layer for processing the arm joint pose parameters, network connection weights in the end-effector posture reading layer for reading the second sub-feature, network connection weights in the workspace meshing mapping layer for forming workspace pose candidate records, network connection weights in the pose accessibility mapping layer for forming a six-degree-of-freedom pose accessibility distribution within the workspace, and network connection weights in the end-effector operation intent integration layer for forming end-effector operation intent features. Subsequently, the third strategy phase weight unlock record removes the weight update masking flags written for the network connection weights related to the dual-arm control in the first strategy phase weight lock record, and removes the weight update masking flags written for the network connection weights related to the chassis leg control in the second strategy phase weight lock record, thus forming a weight joint update admission record. This weight joint update admission record is used to characterize that both the network connection weights related to the chassis leg control and the network connection weights related to the dual-arm control can be adjusted according to the phase evaluation results in the third strategy phase. Furthermore, the weight joint update admission record also includes the query guidance transformation layer, spatial address association layer, focus weight allocation layer, and addressing focus extraction layer in the cross-mapping node within the updateable scope, enabling the query guidance vector, the spatial address association record, the spatial focus weight record, and the spatial-body alignment feature to re-adapt to the basic leg movement parameters and the pure arm operation parameters under dynamic base disturbance conditions.Therefore, the third strategy stage does not simply restore the training state of all structural layers, but rather focuses on the collaborative mapping relationship between the basic leg movement parameters, the pure arm operation parameters, and the heterogeneous feature vectors. This allows the network connection weights related to chassis leg control, the network connection weights related to dual arm control, and the network connection weights corresponding to the query guidance transformation layer, spatial address association layer, focus weight allocation layer, and addressing focus extraction layer in the cross-mapping nodes to enter an adjustable state. The network connection weights corresponding to the query guidance transformation layer, the spatial address association layer, the focus weight allocation layer, and the addressing focus extraction layer jointly participate in the reformation process of the spatial-ontology alignment features in the third strategy stage, thereby avoiding the basic leg movement parameters and the pure arm operation parameters only combining at the output end without participating in the feature source reconstruction of the heterogeneous feature vectors.

[0046] Preferably, the specific implementation process of using the basic leg movement parameters and the pure arm operation parameters as prior guidance input nodes is as follows: After the weighted joint update admission record is formed, the basic leg movement parameters and the pure arm operation parameters are read, and a prior guidance load record is established; the prior guidance load record is used to include the foot support sequence, foot swing trend, fuselage attitude adjustment direction, and chassis leg support state change source in the basic leg movement parameters, and the end position adjustment direction, end orientation adjustment direction, arm joint kinematic chain change source, and near obstacle avoidance source in the pure arm operation parameters into the same prior guidance caliber. The prior guidance input node is not an additional execution component independent of the embodied policy mapping network, but rather a feature injection position set by the embodied policy mapping network in the third policy stage. This feature injection position connects the fuselage motion trend integration layer, the end-operation intent integration layer, and the cross-mapping node, enabling the basic leg movement parameters to be written back to the read link corresponding to the fuselage motion trend feature, enabling the pure arm operation parameters to be written back to the read link corresponding to the end-operation intent feature, and allowing both the basic leg movement parameters and the pure arm operation parameters to participate in the query guidance transformation process in the cross-mapping node. Subsequently, leg prior encoding processing is performed on the basic leg movement parameters to form leg prior guidance features. These leg prior guidance features inherit the foot support sequence, foot swing trend, fuselage attitude adjustment direction, and chassis leg support state change source from the basic leg movement parameters, and are used to indicate the read direction of the fields related to the fuselage motion trend in the heterogeneous feature vector during the third policy stage. Simultaneously, arm prior encoding is performed on the pure arm operation parameters to form arm prior guidance features. These features inherit the end-effector position adjustment direction, end-effector orientation adjustment direction, source of arm joint kinematic chain change, and source of near-obstacle avoidance from the pure arm operation parameters, and are used to indicate the reading direction of fields related to end-effector operation intent in the heterogeneous feature vector during the third strategy stage. Further, the leg prior guidance features and the arm prior guidance features are written into the prior guidance input node, enabling the prior guidance input node to simultaneously provide the basis for leg movement and arm operation when subsequently remapping the heterogeneous feature vector. Through this prior guidance input node, the third strategy stage no longer relearns the relationship between leg movement and arm operation from a blank state, but rather rereads the heterogeneous feature vector based on the values ​​formed in the first and second strategy stages, focusing on leg-arm coordination.The leg prior guidance features and the arm prior guidance features subsequently participate together in the formation process of the collaborative query guidance vector, so that the prior guidance input node not only serves as the input position of the stage parameter, but also as one of the sources of the collaborative query guidance vector in the formation process of the collaborative space-ontology alignment features.

[0047] Preferably, the specific implementation process of guiding the embodied strategy mapping network to remap the heterogeneous feature vector under dynamic base perturbation state, thereby generating the leg-arm collaborative linkage parameters, is as follows: After the prior guidance input node has written the leg prior guidance features and the arm prior guidance features, the body posture state, chassis leg support state, and dual-arm end effector operation state within the current control cycle are read, and a dynamic base perturbation state record is established. The dynamic base perturbation state record is used to express the state changes generated by the Centaur robot performing dual-arm operations when the mobile base is not completely stationary. The dynamic base perturbation state record includes changes in body posture adjustment direction caused by the support phase transition of the chassis leg actuator, changes in body posture state caused by changes in the operation posture of the dual-arm end effectors, and changes in chassis leg support state caused by near-distance obstacle avoidance. Subsequently, the embodied strategy mapping network reads the heterogeneous feature vector, the prior guidance input node, and the dynamic base perturbation state record, and reconstructs the query guidance vector in the cross-mapping node to form a collaborative query guidance vector. The collaborative query guidance vector inherits the leg prior guidance features, the arm prior guidance features, and the dynamic base perturbation state record, and is used to redetermine the spatial region in the global spatial geometric mapping features that is simultaneously related to leg movement and arm operation. Further, the spatial address association layer re-forms a collaborative spatial address association record based on the collaborative query guidance vector, the focus weight allocation layer re-forms a collaborative spatial focus weight record based on the collaborative spatial address association record, and the addressing focus extraction layer performs re-addressing focus extraction processing on the global spatial geometric mapping features based on the collaborative spatial focus weight record to form a collaborative space-body alignment feature. The collaborative space-body alignment feature continues to be written back to the reading process of the heterogeneous feature vector, so that the fuselage motion trend feature and the end-effector operation intention feature in the heterogeneous feature vector re-establish a correspondence under the dynamic base perturbation state. Subsequently, the embodied strategy mapping network performs leg-arm collaborative linkage mapping processing on the re-mapped heterogeneous feature vector to generate the leg-arm collaborative linkage parameters. The leg-arm coordination linkage parameter is used to express the linkage control source of the chassis leg actuator and the dual arm end effector within the same current control cycle. The linkage control source includes the correspondence between foot support change and end position adjustment, the correspondence between fuselage attitude adjustment direction and end orientation adjustment direction, and the correspondence between near obstacle avoidance source and chassis leg support state change source.Therefore, the third strategy stage can incorporate the dynamic base disturbance state record into the remapping process of the heterogeneous feature vector based on the already formed basic leg movement parameters and pure arm operation parameters, so that the leg-arm coordinated linkage parameters become the common parameter source for the subsequent generation of chassis leg driving torque commands and dual arm end operation torque commands.

[0048] 7. The centaur robot embodied intelligent control method for special scenarios according to claim 6, characterized in that the progressive capability self-construction process is constrained by a state evaluation node, the state evaluation node uses a multi-dimensional performance evaluation criterion as the evaluation basis, the multi-dimensional performance evaluation criterion including: Obtain the speed deviation between the actual and expected speed of the centaur robot in the current control cycle; Record the electromechanical actuation power consumption of each joint of the centaur robot in the current control cycle; Obtain the end-effector spatial distance between the actual end-effector position of the centaur robot's arm and the target operating position; Measure the tilt angle of the centaur robot on the chassis plane during the current control cycle; The minimum physical gap between the physical envelope of the centaur robot and the surrounding environmental obstacle surface is detected.

[0049] Preferably, the specific implementation process for obtaining the speed deviation between the actual and expected travel speed of the Centaur robot in the current control cycle is as follows: When the progressive capability self-construction process enters the evaluation node, the current control cycle marker is first read, and the current control cycle displacement record and expected speed configuration record are read based on the current control cycle marker. The current control cycle displacement record originates from the pose change of the Centaur robot's body during the current control cycle. The pose change of the body is formed by the linear acceleration parameter, attitude angle parameter, and foot support change of the chassis leg actuator in the inertial navigation parameters. The current control cycle displacement record is used to express the actual spatial position change of the Centaur robot during the current control cycle and is used to subsequently form the actual travel speed, rather than being estimated solely based on the swing state of a single foot. The expected speed configuration record is formed by the task path segmentation result configuration before the Centaur robot enters the task execution. The task path segmentation result is formed by dividing the current operation path into path segments according to the ground slope, channel width, and spatial proximity of the target operation position. The expected speed configuration record includes the path segment reading position, path segment travel direction, path segment allowable speed range, and expected travel speed corresponding to the path segment reading position corresponding to the current control cycle. The path segment reading position, path segment travel direction, and path segment allowable speed range together define the reading scope of the expected travel speed. The expected travel speed is used to express the travel state that the Centaur robot should achieve in the current control cycle. Subsequently, the actual displacement within the current control cycle is read based on the displacement record of the current control cycle, and time-normalized by combining the actual displacement with the length of the current control cycle to form the actual travel speed. The actual displacement incorporates the spatial position changes in the displacement record of the current control cycle, and the actual travel speed incorporates the actual displacement and the length of the current control cycle, and is used to express the travel speed generated by the chassis leg actuators driving the main body within the current control cycle. Further, the actual travel speed and the expected travel speed are written into the speed evaluation carrying record, and the difference between the actual travel speed and the expected travel speed is determined in the speed evaluation carrying record according to the same direction caliber corresponding to the travel direction of the path segment, to generate the speed deviation. The technical essence of the speed deviation is the deviation of the Centaur robot's actual travel state from its expected travel state within the current control cycle. The speed deviation is subsequently incorporated into the multi-dimensional performance evaluation criteria and used to constrain whether the control parameters formed in the first strategy stage, the second strategy stage, and the third strategy stage match the travel requirements of the current control cycle.For scenarios involving mine slopes or underground utility tunnel curves, if the actual travel speed deviates significantly from the expected travel speed, it usually indicates that the basic leg movement parameters or the leg-arm coordination parameters need further adjustment in terms of the support rhythm of the chassis leg actuators. Therefore, the speed deviation is not an isolated speed evaluation result, but rather an important input for subsequent stage evaluation of the network connection weights related to chassis leg control in the embodied strategy mapping network and the spatial reading weights in the cross-mapping nodes.

[0050] Preferably, in the specific technical implementation of recording the electromechanical actuation power consumption of each joint of the centaur robot in the current control cycle, the joint actuation state record is first read according to the current control cycle marker. This joint actuation state record originates from the response data of the chassis leg actuators and the dual-arm end effectors to the driving torque commands of the chassis legs and the operating torque commands of the dual arms within the current control cycle. The joint actuation state record includes sampled values ​​of the driving current, driving voltage, joint speed, and torque response for each leg joint, as well as the driving current, driving voltage, joint speed, and torque response for each arm joint. This joint actuation state record is used to include the energy consumption sources of the chassis leg actuators and the dual-arm end effectors within the same current control cycle in the same recording caliber, so that the subsequently generated electromechanical actuation power consumption can simultaneously reflect the movement support process and the dual-arm operation process. Subsequently, the joint actuation state records are grouped according to joint identifiers and sampling sequence numbers to form joint power calculation load records. These records retain corresponding drive current sampling values, drive voltage sampling values, joint speed sampling values, and torque response sampling values ​​for each leg and arm joint, and are used to connect multiple sampling points of the same joint within the current control cycle into the same power calculation caliber. Further, the drive current and drive voltage sampling values ​​in the joint power calculation load records undergo power conversion processing at the same sampling time to form joint power conversion results. These results are then combined with the joint speed and torque response sampling values ​​to perform actuation state correction processing, resulting in the electromechanical actuation power consumption for each joint. The joint power conversion result is used to express the joint electrical energy input intensity at the same sampling moment. The electromechanical actuation power consumption is further incorporated into the joint rotation speed sampling value and torque response sampling value based on the joint power conversion result. Therefore, the electromechanical actuation power consumption can express the electromechanical energy intensity consumed by each joint in the current control cycle to perform moving support or dual-arm operation. The electromechanical actuation power consumption is not simply a record of the current magnitude at a certain moment, but a periodic evaluation result formed after incorporating the joint actuation state record, the joint power calculation load record, and the joint power conversion result. The electromechanical actuation power consumption is subsequently incorporated into the multi-dimensional performance evaluation criteria and participates in the evaluation processing of the state assessment node together with the speed deviation, the end-effector spatial distance, the tilt angle amplitude, and the minimum physical clearance.In the first strategy stage, the electromechanical actuation power consumption can reflect whether the basic leg movement parameters lead to high energy consumption of the chassis leg actuators; in the second strategy stage, the electromechanical actuation power consumption can reflect whether the pure arm operation parameters lead to high energy consumption of the dual-arm end effectors; in the third strategy stage, the electromechanical actuation power consumption can reflect the actuation load distribution of the leg-arm coordinated linkage parameters when movement and operation occur simultaneously, thereby establishing a traceable technical correlation between the electromechanical actuation power consumption and the progressive capability self-construction process.

[0051] Preferably, the specific implementation process for obtaining the end-effector spatial distance between the actual end-effector position and the target operating position of the Centaur robot is as follows: After reading the current control cycle marker at the state evaluation node, the end-effector position evaluation and bearing record is first read. The end-effector position evaluation and bearing record includes the arm joint pose parameters, the body posture state, the installation offset of the dual-arm end-effector, and the target operating position configuration record. The installation offset of the dual-arm end-effector is configured when the Centaur robot completes assembly calibration. The installation offset of the dual-arm end-effector is used to express the installation position relationship of the dual-arm end-effector relative to the body coordinate reference. The target operating position configuration record is formed by the spatial positioning result of the current task. The spatial positioning result comes from the reading of the spatial point distribution of the near-distance operating area by the multi-view depth image matrix and the intra-segment positioning of the operating position by the task path segmentation result. The target operating position configuration record includes the target operating position that the dual-arm end-effector should approach in the current control cycle and the workspace coordinate reference where the target operating position is located. Subsequently, based on the arm joint pose parameters, the angle parameters, angular velocity parameters, and end-effector posture parameters of each arm joint are read. Combined with the installation offset of the dual-arm end effector and the fuselage attitude state, kinematic chain position recursion processing is performed on the angle parameters, angular velocity parameters, and end-effector posture parameters of each arm joint to form the actual arm end-effector position. The actual arm end-effector position expresses the actual spatial position of the dual-arm end effector relative to the workspace coordinate reference within the current control cycle. The actual arm end-effector position incorporates the arm joint pose parameters, the installation offset of the dual-arm end effector, and the fuselage attitude state; therefore, the actual arm end-effector position is not calculated independently of changes in the fuselage attitude. Further, the actual arm end-effector position and the target operating position are written into the end-effector spatial distance calculation record, and spatial difference calculation processing is performed on the actual arm end-effector position and the target operating position under the same workspace coordinate reference to generate the end-effector spatial distance. The technical essence of the end-effector spatial distance is the degree of spatial deviation of the dual-arm end effector relative to the target operating position within the current control cycle. This end-effector spatial distance is subsequently incorporated into the multi-dimensional performance evaluation criteria and used to constrain whether the pure arm operation parameters and the leg-arm coordinated linkage parameters can cause the dual-arm end effector to converge towards the target operating position. For near-field operating positions in chemical equipment or narrow operating positions in underground pipe corridors, the target operating position is usually located near an obstacle surface. The actual end-effector position is affected not only by the arm joint pose parameters but also by changes in the fuselage attitude state caused by the chassis leg effector.Therefore, by incorporating the actual end-arm position, the target operation position, and the body posture into the same calculation caliber through the end-arm spatial distance calculation record, the end-arm spatial distance can become a quantitative basis for evaluating the coordination relationship between the end-arm operation state and the body motion state, rather than only evaluating the static distance from the end-arm to the target operation position.

[0052] Preferably, in the specific technical implementation of measuring the tilt angle of the Centaur robot's chassis plane in the current control cycle, the chassis plane attitude bearing record is first read according to the current control cycle marker. The chassis plane attitude bearing record includes the inertial navigation parameters, the leg joint pose parameters, the chassis leg actuator support foot position record, and the body coordinate reference. The chassis leg actuator support foot position record is derived from the foot support sequence in the basic leg movement parameters and the joint code position parameter in the leg joint pose parameters. The chassis leg actuator support foot position record is used to express the spatial position of the foot participating in support within the current control cycle. The body coordinate reference is configured during the Centaur robot assembly calibration. The body coordinate reference is used to express the coordinate reading aperture of the body in the front-back, left-right, and up-down directions. Subsequently, the support plane fitting processing is performed on the support foot position record of the chassis leg actuator to form a chassis support plane record. This chassis support plane record incorporates the spatial positions of the supporting feet within the current control cycle and is used to express the support geometry formed by the contact area between the chassis leg actuator and the ground. Further, based on the attitude angle and angular velocity parameters in the inertial navigation parameters, the attitude change of the fuselage coordinate reference relative to the gravity direction is processed to form a fuselage attitude direction record. This fuselage attitude direction record incorporates the inertial navigation parameters and is used to express the lateral and longitudinal tilt states of the fuselage within the current control cycle. Subsequently, the chassis support plane record and the fuselage attitude direction record are written into a chassis plane tilt evaluation record. In the chassis plane tilt evaluation record, the tilt relationship of the chassis plane expressed by the chassis support plane record relative to the fuselage coordinate reference and the gravity direction is processed to extract the angle amplitude, thereby generating the tilt angle amplitude. The tilt angle amplitude is used to express the degree of deviation of the Centaur robot's chassis plane from the attitude reference direction jointly defined by the gravity direction and the body coordinate reference within the current control cycle. The attitude reference direction is jointly provided by the body coordinate reference and the attitude angle parameter in the inertial navigation parameters. The tilt angle amplitude is not a single attitude angle parameter, but an evaluation result formed by combining the recorded positions of the chassis leg actuators, the recorded chassis support plane, the recorded inertial navigation parameters, and the recorded body attitude direction. For scenarios involving movement on mine slopes, the tilt angle amplitude can reflect the impact of ground undulations on the support state of the chassis leg actuators; for scenarios where the end effectors of the dual arms extend to perform close-range operations, the tilt angle amplitude can reflect the disturbance of the end effector's operating state on the body attitude state.The tilt angle amplitude is subsequently incorporated into the multidimensional performance evaluation criteria and used to constrain the attitude adjustment effects of the basic leg movement parameters, the pure arm operation parameters, and the leg-arm coordinated linkage parameters at different strategy stages, so that the state evaluation node can evaluate the progressive capability self-construction process from the chassis plane attitude level.

[0053] Preferably, the specific implementation process of detecting the minimum physical gap between the physical envelope surface of the Centaur robot and the surrounding environmental obstacle surface is as follows: After the multi-view depth image matrix in the current control cycle has completed time synchronization, the physical gap evaluation and bearing record is read first. The physical gap evaluation and bearing record includes the multi-view depth image matrix, the body posture state, the leg joint pose parameters, the arm joint pose parameters, and the physical envelope surface configuration record. The physical envelope surface configuration record is configured when the Centaur robot completes assembly calibration. The physical envelope surface configuration record includes the outer edge sampling points of the body, the outer edge sampling points of the chassis leg actuators, and the outer edge sampling points of the dual arm end actuators. The physical envelope surface configuration record is used to express the external boundary positions that the Centaur robot can occupy under different joint poses. Subsequently, based on the leg joint pose parameters, the arm joint pose parameters, and the body posture state, the outer edge sampling points of the body, the outer edge sampling points of the chassis leg actuators, and the outer edge sampling points of the dual arm end effectors in the physical envelope configuration record are updated with poses for the current control cycle to form the physical envelope surface for the current control cycle. This physical envelope surface for the current control cycle inherits from the physical envelope configuration record and is used to represent the external boundary of the space actually occupied by the Centaur robot within the current control cycle. Simultaneously, the multi-view depth image matrix undergoes depth point spatial projection processing to form an environmental depth point record. This environmental depth point record inherits from the matrix layers, rows, columns, and corresponding depth distance values ​​in the multi-view depth image matrix and is used to represent the spatial point distribution of the Centaur robot's surrounding environment. Furthermore, the environmental depth point records undergo obstacle surface continuity processing to form a surrounding environmental obstacle surface. This surrounding environmental obstacle surface inherits from the environmental depth point records and is used to represent spatial obstacle boundaries that affect the Centaur robot's movement and dual-arm operation, such as underground pipe gallery walls, mine slope protrusions, or the outer surfaces of chemical equipment. Subsequently, the physical envelope surface of the current control cycle and the surrounding environmental obstacle surface are written into the physical gap calculation record, and the outer edge sampling points on the physical envelope surface of the current control cycle and the obstacle surface sampling points on the surrounding environmental obstacle surface are subjected to nearest distance search processing to form the minimum physical gap. The technical essence of the minimum physical gap is the minimum spatial margin between the Centaur robot's own external boundary and the surrounding environmental obstacle surface within the current control cycle. The minimum physical gap is subsequently incorporated into the multi-dimensional performance evaluation criteria and used to evaluate the spatial safety redundancy of the basic leg movement parameters, the pure arm operation parameters, and the leg-arm coordinated linkage parameters in a near-range obstacle environment.By performing a nearest-distance search process between the physical envelope of the current control cycle and the surrounding environment obstacle surface, the minimum physical gap can simultaneously reflect the distance relationship between the chassis leg actuator and the surrounding environment obstacle surface during movement, as well as the distance relationship between the dual-arm end effector and the surrounding environment obstacle surface during operation. This allows the state assessment node to evaluate the progressive capability self-construction process from the perspective of spatial gaps.

[0054] Optionally, the execution steps of the state evaluation node for the multidimensional performance evaluation criteria include: The speed deviation, the electromechanical actuation power consumption, the end-effector spatial distance, and the tilt angle amplitude are respectively mapped to cost penalty values; Map the minimum physical gap to a safety redundancy value; The cost penalty value and the safety redundancy value are subjected to evaluation value fusion processing to generate a comprehensive evaluation scalar; The comprehensive evaluation scalar is then passed back to the embodied policy mapping network to trigger an adaptive optimization iteration of the network connection weights until the comprehensive evaluation scalar reaches a preset convergence threshold.

[0055] Preferably, the specific implementation process of the execution steps of the state evaluation node for the multi-dimensional performance evaluation criteria is as follows: After the current round parameter generation is completed in the progressive capability self-construction process, the current control cycle marker is read first, and an evaluation input load record is established based on the current control cycle marker; the evaluation input load record incorporates the speed deviation, the electromechanical actuation power consumption, the end effector spatial distance, the tilt angle amplitude, and the minimum physical clearance, and simultaneously incorporates the basic leg movement parameters generated in the first strategy stage, the pure arm operation parameters generated in the second strategy stage, and the leg-arm coordinated linkage parameters generated in the third strategy stage. The technical essence of the evaluation input load record is to incorporate the centaur robot's travel state, actuation load state, end effector state, chassis posture state, and spatial clearance state within the same current control cycle into the same evaluation caliber, so that the state evaluation node can evaluate the basic leg movement parameters, the pure arm operation parameters, and the leg-arm coordinated linkage parameters formed by the progressive capability self-construction process according to the same time source, rather than evaluating each control parameter in isolation. Subsequently, the evaluation caliber configuration record is read. The evaluation caliber configuration record is configured when the Centaur robot completes the task scenario initialization. The evaluation caliber configuration record includes the evaluation boundary of the speed deviation, the evaluation boundary of the electromechanical actuation power consumption, the evaluation boundary of the end-effector spatial distance, the evaluation boundary of the tilt angle amplitude, the evaluation boundary of the minimum physical gap, the normalized conversion caliber corresponding to each of the evaluation boundaries, and the cost penalty value and the safety redundancy value participate in the reading order of the comprehensive evaluation scalar. The evaluation caliber configuration record is not an arbitrarily set numerical record, but is configured according to the control requirements corresponding to the mine slope, underground pipe gallery bend, or close-range operation position of chemical equipment. In the mine slope scenario, the evaluation caliber configuration record ensures that the speed deviation and the tilt angle amplitude have a high reading ratio in the subsequent evaluation numerical fusion processing. In the underground pipe gallery bend scenario, the evaluation caliber configuration record ensures that the minimum physical clearance and the tilt angle amplitude have a high reading ratio in the subsequent evaluation numerical fusion processing. In the close-range operation position of chemical equipment, the evaluation caliber configuration record ensures that the end space distance and the minimum physical clearance have a high reading ratio in the subsequent evaluation numerical fusion processing.The evaluation input carrying record and the evaluation caliber configuration record are jointly entered into subsequent cost mapping processing, safety redundancy mapping processing, and evaluation value fusion processing. The speed deviation, electromechanical actuation power consumption, end-point spatial distance, and tilt angle amplitude are mapped to the cost penalty value, and the minimum physical gap is mapped to the safety redundancy value. The cost penalty value and the safety redundancy value continue to participate in the evaluation value fusion processing to generate the comprehensive evaluation scalar that can be back-transmitted to the embodied policy mapping network. Through the coordination of the evaluation input bearing record and the evaluation caliber configuration record, the differences between the first strategy stage, the second strategy stage, and the third strategy stage can also be preserved in the evaluation numerical fusion processing: the first strategy stage mainly evaluates the impact of the basic leg movement parameters on the movement support of the chassis leg actuator; the second strategy stage mainly evaluates the impact of the pure arm operation parameters on the operational reachability of the dual-arm end actuator; and the third strategy stage mainly evaluates the impact of the leg-arm coordinated linkage parameters on the synchronous control of the chassis leg actuator and the dual-arm end actuator, so that the state evaluation node can form continuous constraints on the progressive capability self-construction process.

[0056] Preferably, in the specific technical implementation of mapping the speed deviation, electromechanical actuation power consumption, end-effector spatial distance, and tilt angle amplitude to the cost penalty value, the speed deviation, electromechanical actuation power consumption, end-effector spatial distance, and tilt angle amplitude are first read from the evaluation input load record, and then written into the cost mapping load record. Since the speed deviation represents the deviation between the actual travel speed and the expected travel speed, the electromechanical actuation power consumption represents the electromechanical energy intensity of each joint in the current control cycle, the end-effector spatial distance represents the spatial deviation of the actual arm end position relative to the target operating position, and the tilt angle amplitude represents the deviation of the chassis plane relative to the attitude reference direction, the speed deviation, electromechanical actuation power consumption, end-effector spatial distance, and tilt angle amplitude have different dimensions and different numerical ranges, they cannot be directly added, compared, or multiplied. Therefore, the state evaluation node first performs speed range normalization processing on the speed deviation according to the evaluation criteria configuration record to form a speed cost penalty value. This speed cost penalty value inherits from the speed deviation and expresses the degree to which the basic leg movement parameters or the leg-arm coordinated linkage parameters cause the actual travel state to deviate from the expected travel state. Subsequently, the electromechanical actuation power consumption is normalized from the power range to form a power cost penalty value. This power cost penalty value inherits from the electromechanical actuation power consumption and expresses the actuation load degree of the chassis leg actuator and the dual-arm end effector within the current control cycle. Further, the end-effector spatial distance is normalized from the end-effector distance range to form an end-effector cost penalty value. This end-effector cost penalty value inherits from the end-effector spatial distance and expresses the degree of insufficient convergence of the dual-arm end effector relative to the target operating position. Subsequently, the tilt angle amplitude is normalized within the attitude tilt range to form a tilt cost penalty value. This tilt cost penalty value is derived from the tilt angle amplitude and expresses the degree of attitude deviation of the chassis plane relative to the attitude reference direction within the current control cycle. The speed cost penalty value, the power cost penalty value, the end-effector cost penalty value, and the tilt cost penalty value together constitute the cost penalty value, which is then incorporated into the evaluation value fusion process. This approach is necessary because the speed deviation, electromechanical actuation power consumption, end-effector spatial distance, and tilt angle amplitude in the multi-dimensional performance evaluation criteria originate from different data domains such as travel speed, electromechanical actuation, end-effector spatial distance, and chassis attitude. Without first forming the cost penalty value, the subsequent comprehensive evaluation scalar will be mixed with data of inconsistent dimensions, thus hindering the stable evaluation basis for the adaptive optimization iteration of the network connection weights.Through the cost mapping record, the speed deviation, the electromechanical actuation power consumption, the end-effector spatial distance, and the tilt angle amplitude can all be entered into the subsequent evaluation value fusion processing with comparable cost penalty values. Furthermore, the speed cost penalty value, the power cost penalty value, the end-effector cost penalty value, and the tilt cost penalty value can still be traced back to the speed deviation, the electromechanical actuation power consumption, the end-effector spatial distance, and the tilt angle amplitude, respectively.

[0057] Preferably, the specific implementation process of mapping the minimum physical gap to the safety redundancy value is as follows: After the minimum physical gap has been formed by the nearest distance search process between the physical envelope surface of the current control cycle and the surrounding environment obstacle surface, the state evaluation node reads the minimum physical gap from the evaluation input bearing record and writes the minimum physical gap into the safety redundancy mapping bearing record. The physical envelope surface of the current control cycle is derived from the physical envelope surface configuration record. The physical envelope surface configuration record is configured when the Centaur robot completes the external boundary calibration of the body, chassis leg actuators, and dual arm end effectors. The physical envelope surface configuration record includes the external boundary of the body, the external boundary of the chassis leg actuators, the external boundary of the dual arm end effectors, and the update caliber of the external boundary as the joint pose changes. The physical envelope surface of the current control cycle is formed based on the physical envelope surface configuration record, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters, and is used to express the spatial coverage range of the Centaur robot's external boundary within the current control cycle. The safety redundancy mapping record incorporates the evaluation boundary of the minimum physical gap from the minimum physical gap, the physical envelope of the current control cycle, the surrounding environment obstacle surface, and the evaluation caliber configuration record. This allows the minimum physical gap to be mapped according to a spatial gap caliber that matches the current task scenario. The minimum physical gap has a different technical meaning from the velocity deviation, electromechanical actuation power consumption, end effector spatial distance, and tilt angle amplitude: the velocity deviation, electromechanical actuation power consumption, end effector spatial distance, and tilt angle amplitude typically increase the penalty value with increasing deviation, while the minimum physical gap, with its own value increasing, usually indicates a more sufficient spatial margin between the Centaur robot and the surrounding environment obstacle surface. Therefore, the state evaluation node performs gap reverse normalization on the minimum physical gap to form the safety redundancy value; this safety redundancy value incorporates the minimum physical gap and is used to express the spatial safety redundancy between the physical envelope of the current control cycle and the surrounding environment obstacle surface within the current control cycle. The gap reverse normalization process does not simply convert the minimum physical gap into another distance value. Instead, it converts the minimum physical gap into a safety redundancy value that can participate in the evaluation value fusion process together with the cost penalty value, based on the evaluation boundary of the minimum physical gap in the physical envelope configuration record, the surrounding environment obstacle surface, and the evaluation caliber configuration record.For underground utility tunnel bends, the surrounding environmental obstacles are typically formed by the tunnel walls and raised ground boundaries. The safety redundancy value reflects the space margin for the chassis leg actuators when moving in confined spaces. For close-range operation positions of chemical equipment, the surrounding environmental obstacles typically include the equipment's outer surface and obstacles near the close-range operation area. The safety redundancy value reflects the space margin between the dual-arm end effectors and the surrounding environmental obstacles during extension. The safety redundancy value, along with the cost penalty value, subsequently participates in the generation of the comprehensive evaluation scalar. When the comprehensive evaluation scalar is passed back to the embodied policy mapping network, it generates evaluation constraints on the spatial address association layer, focus weight allocation layer, and addressing focus extraction layer in the cross-mapping nodes. This ensures that the query guidance vector pays more attention to the spatial region related to the surrounding environmental obstacles in subsequent adaptive optimization iterations.

[0058] Preferably, the specific implementation process of performing evaluation value fusion processing on the cost penalty value and the safety redundancy value to generate the comprehensive evaluation scalar is as follows: After the speed cost penalty value, the power cost penalty value, the end-of-line cost penalty value, the tilt cost penalty value, and the safety redundancy value are all formed, the state assessment node establishes an evaluation value fusion carrying record and writes the speed cost penalty value, the power cost penalty value, the end-of-line cost penalty value, the tilt cost penalty value, and the safety redundancy value into the evaluation value fusion carrying record. The evaluation value fusion carrying record is used to incorporate the mobile support cost, electromechanical actuation cost, end-of-line operation cost, chassis attitude cost, and spatial safety redundancy into the same evaluation carrier, so that the subsequently formed comprehensive evaluation scalar can simultaneously express the comprehensive state of the progressive capability self-construction process in terms of mobile support, electromechanical actuation, end-of-line operation, chassis attitude, and spatial clearance. Subsequently, the evaluation fusion coefficient configuration record is read. This record is configured synchronously when the evaluation caliber configuration record is formed. The evaluation fusion coefficient configuration record includes the evaluation fusion coefficients corresponding to the speed cost penalty value, the power cost penalty value, the terminal cost penalty value, the tilt cost penalty value, and the safety redundancy value. The evaluation fusion coefficient configuration record is formed based on the path segment travel direction, the workspace coordinate reference of the target operation position, the allowable speed range of the path segment, and the evaluation boundary of the minimum physical gap in the task scenario. The path segment travel direction, the workspace coordinate reference of the target operation position, and the allowable speed range of the path segment are written into the task scenario configuration record during task scenario initialization. The task scenario configuration record and the evaluation caliber configuration record are used together to generate the evaluation fusion coefficient configuration record, giving each evaluation fusion coefficient in the record a scenario source. The function of the evaluation fusion coefficient configuration record is to adjust the contribution relationship of various evaluation objects into the comprehensive evaluation scalar in different task scenarios, rather than simply adding all evaluation objects according to a fixed ratio. For mining slope scenarios, the evaluation fusion coefficients corresponding to the speed penalty value and the tilt penalty value can be configured to a higher read ratio, so that the comprehensive evaluation scalar can better reflect the adaptation status of the chassis leg actuator to changes in slope support; for close-range operation positions of chemical equipment, the evaluation fusion coefficients corresponding to the end-effector penalty value and the safety redundancy value can be configured to a higher read ratio, so that the comprehensive evaluation scalar can better reflect the operational reachability and spatial clearance status of the dual-arm end-effector.Subsequently, based on the evaluation fusion coefficient configuration record, the cost penalty value and the safety redundancy value in the evaluation value fusion bearing record are subjected to evaluation value fusion processing to generate the comprehensive evaluation scalar. The comprehensive evaluation scalar is not a substitute name for a single evaluation object, but rather an evaluation result formed by incorporating the cost penalty value, the safety redundancy value, and the evaluation fusion coefficient configuration record. This comprehensive evaluation scalar directly enters the backpropagation process and is used to trigger the adaptive optimization iteration of the network connection weights in the embodied policy mapping network. Through the evaluation value fusion processing, the state evaluation node can convert the multi-dimensional performance evaluation criteria into a backpropagable comprehensive evaluation scalar, while still retaining the source contributions of the speed cost penalty value, the power cost penalty value, the terminal cost penalty value, the tilt cost penalty value, and the safety redundancy value to the comprehensive evaluation scalar, providing a traceable evaluation basis for subsequent adjustments to the network connection weights.

[0059] Preferably, the specific implementation process of backpropagating the comprehensive evaluation scalar to the embodied policy mapping network to trigger adaptive optimization iteration of the network connection weights until the comprehensive evaluation scalar reaches a preset convergence threshold is as follows: After the comprehensive evaluation scalar is formed, the state evaluation node establishes a backpropagation record and writes the comprehensive evaluation scalar, the speed cost penalty value, the power cost penalty value, the end-effector cost penalty value, the tilt cost penalty value, the safety redundancy value, the current control cycle flag, and the structural layer identifier in the embodied policy mapping network into the backpropagation record. The backpropagation record is used to record the evaluation source when the comprehensive evaluation scalar is backpropagated to the embodied policy mapping network, so that the subsequent adaptive optimization iteration can distinguish whether the adjustment requirement of the network connection weights is caused by the mobile support cost, electromechanical actuation cost, end-effector operation cost, chassis attitude cost, or spatial safety redundancy. Subsequently, based on the structural layer identifier, the comprehensive evaluation scalar is decomposed into evaluation transmission quantities related to chassis leg control, evaluation transmission quantities related to dual arm control, and evaluation transmission quantities related to spatial reading. Specifically, the evaluation transmission quantities related to chassis leg control are fed back to the leg kinematics chain encoding layer, foot state reading layer, terrain contact discrimination layer, and fuselage motion trend integration layer to trigger the adaptive optimization iteration of network connection weights related to chassis leg control. The evaluation transmission quantities related to dual arm control are fed back to the arm kinematics chain encoding layer, end-effector posture reading layer, workspace grid mapping layer, pose accessibility mapping layer, and end-effector operation intent integration layer to trigger the adaptive optimization iteration of network connection weights related to dual arm control. The evaluation transmission quantities related to spatial reading are fed back to the query guidance conversion layer, spatial address association layer, focus weight allocation layer, and address focus extraction layer to trigger the adaptive optimization iteration of spatial reading weights in the cross-mapping nodes. In the first strategy phase, the reverse transmission load record mainly adjusts the network connection weights related to chassis leg control based on the speed cost penalty value, the power cost penalty value, and the tilt cost penalty value. In the second strategy phase, the reverse transmission load record mainly adjusts the network connection weights related to dual-arm control based on the end cost penalty value, the power cost penalty value, and the safety redundancy value. In the third strategy phase, the reverse transmission load record simultaneously adjusts the network connection weights related to chassis leg control, the network connection weights related to dual-arm control, and the spatial reading weights in the cross-mapping node based on the cost penalty value and the safety redundancy value, so that the leg-arm coordinated linkage parameters can continue to iterate under dynamic base disturbance conditions.Further, the convergence condition threshold configuration record is read. This record is configured during task scenario initialization based on the current control cycle length, the allowed speed range of the path segment, the spatial proximity of the target operation position, the evaluation boundary of the minimum physical gap, and the allowed stage evaluation fluctuation range. The spatial proximity of the target operation position and the allowed stage evaluation fluctuation range are configured in the task scenario configuration record. The convergence condition threshold configuration record is used to explain the source and configuration method of the convergence condition threshold. Subsequently, the comprehensive evaluation scalar re-formed after each round of adaptive optimization iteration is compared with the convergence condition threshold for convergence determination. When the comprehensive evaluation scalar satisfies the convergence condition threshold, the adaptive optimization iteration of the network connection weights in the current round is stopped, and the network connection weights in the current round are written into the network connection weight update record. When the comprehensive evaluation scalar has not yet satisfied the convergence condition threshold, the adaptive optimization iteration of the network connection weights continues based on the backpropagation bearer record. The network connection weight update record subsequently participates in the remapping of the heterogeneous feature vector in the next current control cycle, so that the evaluation result of the state evaluation node can be applied to the subsequent generation process of the basic leg movement parameter, the pure arm operation parameter, and the leg-arm coordinated linkage parameter.

[0060] Optionally, the step of generating the chassis leg drive torque command and the end effector torque command of the centaur robot based on the leg-arm coordination parameters includes: The input impedance control mapping layer of the leg-arm coordinated linkage parameter is used to analyze the chassis leg support impedance stiffness and the dual arm operation compliance stiffness. A command stream sequence for the leg motor is generated based on the chassis leg support impedance stiffness, and then the command stream sequence for the leg motor is used as the chassis leg drive torque command. Based on the dual-arm operation compliance stiffness, a command stream sequence is generated for each joint motor of the robotic arm, and then the command stream sequence for each joint motor of the robotic arm is used as the dual-arm end-effector operation torque command.

[0061] Preferably, the specific implementation process of generating the chassis leg drive torque command and the dual arm end effector torque command of the Centaur robot based on the leg-arm coordinated linkage parameters is as follows: After the leg-arm coordinated linkage parameters have been generated in the third strategy stage, a torque command mapping record is first established, and the leg-arm coordinated linkage parameters, the current control cycle marker, the basic leg movement parameters, the pure arm operation parameters, the dynamic base disturbance state record, and the network connection weight update record are written into the torque command mapping record. The torque command mapping record is used to include the leg-arm coordinated linkage control formed in the third strategy stage, which originates from the basic leg movement parameters formed in the first strategy stage and the pure arm operation parameters formed in the second strategy stage, into the same torque generation caliber, so that the subsequently generated chassis leg drive torque command and dual arm end effector torque command can be traced back to the same leg-arm coordinated linkage parameter, instead of being generated separately based on the unrelated basic leg movement parameters and pure arm operation parameters. The difference between the first strategy stage, the second strategy stage, and the third strategy stage is that the first strategy stage focuses on generating the basic leg movement parameters corresponding to the quadrupedal movement mode, the second strategy stage focuses on generating the pure arm operation parameters in the static base operation mode, and the third strategy stage focuses on remapping the heterogeneous feature vector according to the basic leg movement parameters and the pure arm operation parameters in the dynamic base disturbance state, so as to form the leg-arm coordinated linkage parameters that can simultaneously express the changes in foot support, end-effector position adjustment, fuselage attitude adjustment direction, end-effector orientation adjustment direction, and near-distance obstacle avoidance source. Subsequently, the torque command is mapped to the bearing record input impedance control mapping layer. The impedance control mapping layer reads the leg-arm coordinated linkage parameters and performs field normalization processing on the leg-arm coordinated linkage parameters according to the leg control field, arm control field, and coordinated disturbance field to form an impedance mapping input record. Among them, the leg control field inherits the foot support change and fuselage attitude adjustment direction in the leg-arm coordinated linkage parameters; the arm control field inherits the end position adjustment and end orientation adjustment direction in the leg-arm coordinated linkage parameters; and the coordinated disturbance field inherits the near obstacle avoidance source and dynamic base disturbance state record in the leg-arm coordinated linkage parameters. The leg control field, the arm control field, and the coordinated disturbance field jointly participate in the formation of the impedance mapping input record.The impedance control mapping layer is not a convolutional processing layer based on local image neighborhood sliding reading, because the leg-arm coordination parameters are not a two-dimensional image array, but rather structured control parameters including foot support changes, end-effector position adjustments, fuselage attitude adjustment direction, end-effector orientation adjustment direction, and near-range obstacle avoidance sources. The impedance control mapping layer employs a combination of parameter projection, joint chain constraint reading, and segmented amplitude-limiting mapping to convert the leg-arm coordination parameters in the impedance mapping input record into stiffness adjustment parameters. These stiffness adjustment parameters include the chassis leg support impedance stiffness and the dual-arm operating compliance stiffness, which are subsequently analyzed. Through the connection between the torque command mapping load record, the impedance mapping input record, and the impedance control mapping layer, the leg-arm coordination parameters can enter the subsequent stiffness analysis processing and command stream sequence generation process, ensuring that the chassis leg drive torque command and the dual-arm end-effector operating torque command have a consistent leg-arm coordination control source within the current control cycle.

[0062] Preferably, in the specific technical implementation of inputting the leg-arm coordination linkage parameters into the impedance control mapping layer to resolve the chassis leg support impedance stiffness and dual-arm operation compliance stiffness, the impedance control mapping layer first reads the leg-arm coordination linkage parameters from the impedance mapping input record and performs parameter scaling processing on the leg-arm coordination linkage parameters to form scale-normalized leg-arm coordination linkage parameters. The parameter scaling processing is used to convert foot support changes, end-effector position adjustments, fuselage attitude adjustment directions, end-effector orientation adjustment directions, and near-distance obstacle avoidance sources into numerical values ​​that can jointly participate in stiffness resolution processing, avoiding the direct mixing of angle-type parameters, displacement-type parameters, and disturbance-type parameters in stiffness mapping. Subsequently, the leg impedance resolution channel in the impedance control mapping layer reads the foot support changes, fuselage attitude adjustment directions, and chassis leg support state change sources from the scale-normalized leg-arm coordination linkage parameters, and performs joint chain constraint reading according to the installation order of each leg joint in the chassis leg actuator to form leg impedance mapping features. The leg impedance analysis channel is used to read the scale-normalized leg-arm coordinated linkage parameters for the chassis leg actuators. The leg impedance mapping feature inherits the sources of foot support changes, fuselage attitude adjustment direction, and chassis leg support state changes from the scale-normalized leg-arm coordinated linkage parameters, and is used to express the stiffness adjustment direction of each foot to resist changes in ground height, slope support changes, and fuselage attitude disturbances in the current control cycle. Further, the impedance control mapping layer performs piecewise stiffness projection processing on the leg impedance mapping feature to form the chassis leg support impedance stiffness. The piecewise stiffness projection processing maps the leg impedance mapping feature piecewise according to the support phase of the leg joint, foot support changes, and fuselage attitude adjustment direction, so that the chassis leg support impedance stiffness can correspond to the support requirements of different leg joints in the current control cycle. The technical essence of the chassis leg support impedance stiffness is to convert the correspondence between foot support changes and fuselage attitude adjustment direction into impedance stiffness distribution results of the chassis leg actuators at different leg joints, enabling subsequent command stream sequences for the leg motors to generate differentiated torques based on different foot support phases. Simultaneously, the arm compliance analysis channel in the impedance control mapping layer reads the scale-normalized leg-arm coordinated linkage parameters, including end-effector position adjustment, end-effector orientation adjustment direction, sources of arm joint kinematic chain changes, and sources of near-distance obstacle avoidance. It then reads joint chain constraints based on the kinematic chain sequence of each arm joint in the dual-arm end-effectors to form arm compliance mapping features.The arm compliance analysis channel is used to read the scale-normalized leg-arm coordination parameters for the dual-arm end effectors. The arm compliance mapping feature inherits the end-effector position adjustment, end-effector orientation adjustment direction, source of arm joint kinematic chain change, and source of near-distance obstacle avoidance from the scale-normalized leg-arm coordination parameters, and is used to express the compliance adjustment direction required by the dual-arm end effectors when approaching the target operating position, adjusting the end-effector orientation, and avoiding surrounding environmental obstacles. Subsequently, the impedance control mapping layer performs segmented compliance projection processing on the arm compliance mapping feature to form the dual-arm operating compliance stiffness. The segmented compliance projection processing performs segmented mapping on the arm compliance mapping feature according to the kinematic chain sequence of the arm joints, end-effector position adjustment, end-effector orientation adjustment direction, and source of near-distance obstacle avoidance, so that the dual-arm operating compliance stiffness can correspond to the operating requirements of different arm joints in the current control cycle. The technical essence of the dual-arm operation compliant stiffness is to convert the correspondence between the end-effector operation trend and the spatial clearance constraint into the compliant stiffness allocation result of each joint motor of the robotic arm. This enables the subsequent command flow sequence for each joint motor of the robotic arm to perform executable torque distribution between the operation direction and the obstacle avoidance direction. The chassis leg support impedance stiffness and the dual-arm operation compliant stiffness are two types of stiffness parameters derived from the same leg-arm collaborative linkage parameter analysis. The difference is that the chassis leg support impedance stiffness is oriented towards the foot support and body attitude adjustment of the chassis leg actuator, while the dual-arm operation compliant stiffness is oriented towards the end-effector operation and near-distance obstacle avoidance of the dual-arm end-effector. The chassis leg support impedance stiffness is subsequently incorporated into the generation process of the chassis leg drive torque command, while the dual-arm operation compliant stiffness is subsequently incorporated into the generation process of the dual-arm end-effector operation torque command.

[0063] Preferably, the specific implementation process of generating a command stream sequence for the leg motor based on the chassis leg support impedance stiffness, and then using the command stream sequence for the leg motor as the chassis leg drive torque command, is as follows: After the chassis leg support impedance stiffness is formed, a leg command stream generation record is first established, and the chassis leg support impedance stiffness, the basic leg movement parameters, the foot support change in the leg-arm coordinated linkage parameters, the fuselage attitude adjustment direction in the leg-arm coordinated linkage parameters, the current control cycle marker, and the installation order of each leg joint in the chassis leg actuator are written into the leg command stream generation record. The leg command stream generation record is used to incorporate the stiffness source of the chassis leg support impedance stiffness, the motion source of the basic leg movement parameters, and the installation order of each leg joint in the chassis leg actuator into the same command generation caliber, so that the subsequently generated command stream sequence for the leg motor can correspond to the specific leg joint of the chassis leg actuator, rather than just forming torque data that cannot be sent to the leg motor. Subsequently, based on the leg command stream, a load record is generated, reading the joint angle parameters, joint angular velocity parameters, and joint coded position parameters corresponding to each leg joint. These parameters are then mapped to the chassis leg support impedance stiffness to form candidate leg joint torque values. This leg joint impedance mapping process incorporates the chassis leg support impedance stiffness and the joint angle, angular velocity, and coded position parameters corresponding to each leg joint, ensuring that the candidate leg joint torque values ​​correspond to the torque direction and amplitude source required by each leg joint in the current control cycle. The candidate leg joint torque values, based on the chassis leg support impedance stiffness, express the torque direction and amplitude source required by the corresponding leg joint in the current control cycle. Furthermore, the candidate leg joint torque values ​​are processed by joint sequence arrangement to form a leg joint torque arrangement record. The joint sequence arrangement process arranges the candidate leg joint torque values ​​according to the installation order of each leg joint in the chassis leg actuator. The leg joint torque arrangement record arranges the candidate leg joint torque values ​​according to the installation order of each leg joint in the chassis leg actuator, so that the torque distribution order between different leg joints is consistent with the actual kinematic chain order of the chassis leg actuator.Subsequently, based on the leg motor communication encapsulation interface, the leg joint torque arrangement record is encapsulated with instruction fields to form the instruction stream sequence for the leg motors. The leg motor communication encapsulation interface is configured during the Centaur robot's initialization and includes a leg motor identification field, a joint installation order field, a torque direction field, a torque amplitude field, a current control cycle marker field, and an instruction verification field. This interface describes how the instruction stream sequence for the leg motors is read by the chassis leg actuators. The instruction field encapsulation process, based on the leg motor communication encapsulation interface, writes the candidate leg joint torque values ​​from the leg joint torque arrangement record into the corresponding leg motor identification field, torque direction field, and torque amplitude field, enabling the instruction stream sequence for the leg motors to be read by the chassis leg actuators according to the joint installation order. Finally, the instruction stream sequence for the leg motors is used as the chassis leg drive torque instruction. The chassis leg drive torque command is not directly copied from the basic leg movement parameters, but is formed from the chassis leg support impedance stiffness after leg joint impedance mapping, joint sequence arrangement, and command field encapsulation. Therefore, the chassis leg drive torque command can translate the foot support change and fuselage attitude adjustment direction in the leg-arm coordinated linkage parameters into command fields readable by each leg motor. In scenarios such as mine slopes or underground utility tunnel curves, the chassis leg drive torque command enables the chassis leg actuators to output torque according to the foot support change and fuselage attitude adjustment direction expressed by the leg-arm coordinated linkage parameters during slope support, ground height changes, and narrow passage attitude adjustments.

[0064] Preferably, the specific implementation process of generating a command flow sequence for each joint motor of the robotic arm based on the dual-arm operation compliance stiffness, and then using the command flow sequence for each joint motor of the robotic arm as the dual-arm end-effector operation torque command, is as follows: After the dual-arm operation compliance stiffness is formed, an arm command flow generation load record is first established, and the dual-arm operation compliance stiffness, the pure arm operation parameters, the end-effector position adjustment in the leg-arm coordinated linkage parameters, the end-effector orientation adjustment direction in the leg-arm coordinated linkage parameters, the near-distance obstacle avoidance source in the leg-arm coordinated linkage parameters, the current control cycle marker, and the kinematic chain sequence of each arm joint in the dual-arm end-effector actuator are written into the arm command flow generation load record. The arm command flow generation load record is used to incorporate the stiffness source of the dual-arm operation compliance stiffness, the operation source of the pure arm operation parameters, and the leg-arm coordinated linkage control source of the leg-arm coordinated linkage parameters into the same arm command generation caliber, so that the subsequently generated command flow sequence for each joint motor of the robotic arm can correspond to each arm joint in the dual-arm end-effector actuator. Subsequently, based on the arm command stream, a bearing record is generated, reading the joint angle parameters, joint angular velocity parameters, and end-effector attitude parameters corresponding to each arm joint. These parameters are then mapped to the dual-arm operational compliance stiffness to form candidate arm joint torque values. This arm joint compliance mapping process incorporates the dual-arm operational compliance stiffness and the joint angle, angular velocity, and end-effector attitude parameters corresponding to each arm joint, ensuring that the candidate arm joint torque values ​​correspond to the torque direction and amplitude source required by each arm joint in the current control cycle. The candidate arm joint torque values, based on the dual-arm operational compliance stiffness, express the torque direction and amplitude source required by the corresponding arm joint in the current control cycle to approach the target operating position, adjust the end-effector orientation, and avoid surrounding obstacles. Furthermore, the candidate arm joint torque values ​​are processed by kinematic chain sequence arrangement to form an arm joint torque arrangement record. The kinematic chain sequence arrangement process arranges the candidate arm joint torque values ​​according to the kinematic chain sequence of each arm joint in the end effector of the two arms. The arm joint torque arrangement record arranges the candidate arm joint torque values ​​according to the kinematic chain sequence of each arm joint in the end effector of the two arms, so that the torque distribution sequence of the end effector of the two arms from the shoulder connection position to the end effector position is consistent with the actual joint kinematic chain.Subsequently, based on the communication encapsulation interface of each joint motor of the robotic arm, the instruction field encapsulation processing of the arm joint torque arrangement record is performed to form the instruction stream sequence for each joint motor of the robotic arm. The communication encapsulation interface of each joint motor of the robotic arm is configured when the Centaur robot completes its startup initialization. The communication encapsulation interface of each joint motor of the robotic arm includes a robotic arm joint motor identification field, a kinematics sequence field, a torque direction field, a torque amplitude field, an end-effector posture association field, a current control cycle marker field, and an instruction verification field. The communication encapsulation interface of each joint motor of the robotic arm is used to describe how the instruction stream sequence for each joint motor of the robotic arm is read in the dual-arm end-effector. The instruction field encapsulation processing, based on the communication encapsulation interface of each joint motor of the robotic arm, writes the candidate values ​​of the arm joint torque in the arm joint torque arrangement record into the corresponding robotic arm joint motor identification field, torque direction field, torque amplitude field, and end-effector posture association field, so that the instruction stream sequence for each joint motor of the robotic arm can be read by the dual-arm end-effector in kinematics sequence. Finally, the instruction stream sequence for each joint motor of the robotic arm is used as the dual-arm end-effector torque command. This dual-arm end-effector torque command is not directly copied from the pure arm operation parameters, but rather formed from the dual-arm operation compliance stiffness after arm joint compliance mapping, kinematic chain sequence arrangement, and instruction field encapsulation. Therefore, the dual-arm end-effector torque command can translate the end-effector position adjustment, end-effector orientation adjustment direction, and near-distance obstacle avoidance source in the leg-arm coordination linkage parameters into instruction fields readable by each joint motor of the robotic arm. In the near-distance operation position of the chemical equipment, the dual-arm end-effector torque command enables the dual-arm end-effector to still output torque according to the end-effector position adjustment, end-effector orientation adjustment direction, and near-distance obstacle avoidance source expressed by the leg-arm coordination linkage parameters when there is attitude disturbance on the moving base. This, together with the chassis leg drive torque command, forms the Centaur robot's movement support and end-effector operation control results within the same current control cycle.

[0065] like Figure 3 As shown in the figure, a centaur robot embodied intelligent control device for special scenarios is provided in an embodiment of this application, which includes: The parameter acquisition module is used to acquire the multi-view depth image matrix, leg joint pose parameters, arm joint pose parameters, and inertial navigation parameters of the Centaur robot at the current moment. The implicit feature alignment module is used to input the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters into the embodied policy mapping network to perform implicit feature alignment processing on the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters, thereby generating a heterogeneous feature vector characterizing the body motion trend and end-effector operation intention of the Centaur robot. The capability self-construction module is used in the embodied policy mapping network to perform a progressive capability self-construction process with the heterogeneous feature vector as the processing object, so as to sequentially generate basic leg movement parameters, pure arm operation parameters, and leg-arm coordinated linkage parameters. The torque command control module is used to generate the chassis leg driving torque command and the end effector torque command of the Centaur robot according to the leg-arm coordinated linkage parameters, and then control the Centaur robot through the chassis leg driving torque command and the end effector torque command.

[0066] like Figure 4 As shown, an electronic device according to an embodiment of this application includes a memory and a processor; The memory stores computer programs; When the processor executes the computer program, it performs the method described in any one of the claims of this application.

[0067] like Figure 5 As shown, this is a computer-readable storage medium according to an embodiment of the present application, wherein the computer-readable storage medium stores a computer program; When the computer program is executed by a processor, it is used to perform the method described in any of the claims of this application.

[0068] like Figure 6 As shown, this is an embodiment of a centaur robot according to this application, which includes: The fuselage itself; A chassis leg actuator is connected to the fuselage body and is used to receive chassis leg drive torque commands and execute chassis leg movements. A dual-arm end effector is connected to the main body of the machine and is used to receive dual-arm end-operation torque commands and perform dual-arm end-operation operations. A multi-view depth acquisition component is disposed on the main body of the robot and is used to acquire a multi-view depth image matrix of the Centaur robot at the current moment. A leg joint pose acquisition component is disposed on the chassis leg actuator and is used to acquire the leg joint pose parameters of the Centaur robot at the current moment. An arm joint pose acquisition component is disposed on the dual-arm end effector and is used to acquire the arm joint pose parameters of the Centaur robot at the current moment. An inertial navigation component is disposed on the main body of the robot and is used to collect the inertial navigation parameters of the Centaur robot at the current moment. The embodied intelligent control device is connected to the multi-view depth acquisition component, the leg joint pose acquisition component, the arm joint pose acquisition component, the inertial navigation component, the chassis leg actuator, and the dual arm end effector. The embodied intelligent control device is used to input the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters into the embodied strategy mapping network to perform feature implicit alignment processing on the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters, thereby generating a heterogeneous feature vector characterizing the centaur robot's body motion trend and end-effector operation intention; The embodied intelligent control device is also used in the embodied strategy mapping network to perform a progressive capability self-construction process with the heterogeneous feature vector as the processing object, so as to sequentially generate basic leg movement parameters, pure arm operation parameters and leg-arm coordinated linkage parameters. The embodied intelligent control device is also used to generate the chassis leg driving torque command and the dual arm end effector torque command according to the leg-arm coordinated linkage parameters, so as to control the chassis leg actuator through the chassis leg driving torque command and control the dual arm end effector through the dual arm end effector torque command.

[0069] like Figure 7 As shown in the figure, a centaur robot embodied intelligent control system for special scenarios according to an embodiment of this application includes: The multimodal parameter acquisition subsystem is used to acquire the multi-view depth image matrix, leg joint pose parameters, arm joint pose parameters, and inertial navigation parameters of the Centaur robot at the current moment. An embodied strategy mapping subsystem, connected to the multimodal parameter acquisition subsystem, is used to receive the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters. It then inputs these parameters into the embodied strategy mapping network to perform implicit feature alignment processing on the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters, thereby generating heterogeneous feature vectors characterizing the centaur robot's body motion trends and end-effector operation intentions. A progressive capability self-construction subsystem, connected to the embodied strategy mapping subsystem, is used in the embodied strategy mapping network to perform a progressive capability self-construction process with the heterogeneous feature vector as the processing object, so as to sequentially generate basic leg movement parameters, pure arm operation parameters, and leg-arm coordinated linkage parameters. The torque command output subsystem is connected to the progressive capability self-construction subsystem and is used to generate the chassis leg drive torque command and the end-effector operation torque command of the Centaur robot according to the leg-arm coordinated linkage parameters. An execution control subsystem, connected to the torque command output subsystem, is used to receive the chassis leg drive torque command and the dual arm end effector torque command, and then control the centaur robot through the chassis leg drive torque command and the dual arm end effector torque command.

[0070] Figures 2-7 For an exemplary description, please refer to the above. Figure 1 This will not be elaborated upon here.

Claims

1. A method for embodied intelligent control of a centaur robot for special scenarios, characterized in that, include: Obtain the multi-view depth image matrix, leg joint pose parameters, arm joint pose parameters, and inertial navigation parameters of the centaur robot at the current moment; The multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters are input into the embodied policy mapping network to perform implicit feature alignment processing on the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters, thereby generating heterogeneous feature vectors that characterize the body motion trend and end-effector operation intention of the Centaur robot. In the embodied strategy mapping network, a progressive capability self-construction process is performed with the heterogeneous feature vector as the processing object to sequentially generate basic leg movement parameters, pure arm operation parameters, and leg-arm coordinated linkage parameters. Based on the leg-arm coordination parameters, the system generates chassis leg driving torque commands and end-effector operating torque commands for the centaur robot, and then controls the centaur robot through the chassis leg driving torque commands and the end-effector operating torque commands.

2. The method for intelligent control of a centaur robot embodied in special scenarios according to claim 1, characterized in that, The acquisition of the multi-view depth image matrix, leg joint pose parameters, arm joint pose parameters, and inertial navigation parameters of the centaur robot at the current moment includes: Extract the first hardware timestamp corresponding to the multi-view depth image matrix; The second hardware timestamps corresponding to the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters are extracted respectively. Determine whether the difference between the first hardware timestamp and each of the second hardware timestamps is within a preset alignment range; If the differences between the first hardware timestamp and each of the second hardware timestamps are all within the preset alignment interval, then the multi-view depth image matrix is ​​used as the anchoring reference, and temporal state interpolation processing is performed on the leg joint pose parameter, the arm joint pose parameter, and the inertial navigation parameter to obtain multimodal parameters that are time-synchronized with the multi-view depth image matrix. The multimodal parameters include the time-synchronized multi-view depth image matrix, the leg joint pose parameter, the arm joint pose parameter, and the inertial navigation parameter.

3. The method for intelligent control of a centaur robot embodied in special scenarios according to claim 2, characterized in that, The step of inputting the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters into an embodied policy mapping network to perform implicit feature alignment processing on the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters includes: The multi-view depth image matrix in the multimodal parameters is input into the visual encoding branch of the embodied policy mapping network to extract global spatial geometric mapping features; The leg joint pose parameter, the arm joint pose parameter, and the inertial navigation parameter in the multimodal parameters are input into the ontology state encoding branch of the embodied policy mapping network to extract its own kinematic representation features. In the cross-mapping nodes of the embodied policy mapping network, the self-kinematic representation features are used as query guide vectors to perform addressing and focusing extraction processing on the global spatial geometric mapping features to generate spatial-body alignment features. Then, based on the spatial-body alignment features, heterogeneous feature vectors representing the body motion trend and end-effector intention of the Centaur robot are generated.

4. The centaur robot embodied intelligent control method for special scenarios according to claim 3, characterized in that, The generation of heterogeneous feature vectors characterizing the centaur robot's body motion trends and end-effector intentions based on the spatial-void alignment features includes: The spatial-ontology alignment features are input into the intent decoupling network layer of the embodied policy mapping network to perform feature stripping processing on the spatial-ontology alignment features, resulting in the first sub-feature and the second sub-feature. The first sub-feature is mapped to a continuous probability distribution of foot contact with the ground to form a body motion trend feature that characterizes the body motion trend of the centaur robot. The second sub-feature is mapped to a six-degree-of-freedom pose reachability distribution within the workspace to form an end-operation intention feature characterizing the end-operation intention of the Centaur robot. The fuselage motion trend features and the end-operation intention features are spliced ​​together to form the heterogeneous feature vector.

5. The method for intelligent control of a centaur robot embodied in special scenarios according to claim 1, characterized in that, In the embodied strategy mapping network, a progressive capability self-construction process is performed using the heterogeneous feature vectors as the processing object to sequentially generate basic leg movement parameters, pure arm operation parameters, and leg-arm coordinated linkage parameters, including: In the first strategy phase, the network connection weights related to bi-arm control in the embodied strategy mapping network are locked to generate the basic leg movement parameters corresponding to the quadrupedal movement pattern based on the heterogeneous feature vector. In the second strategy phase, the network connection weights related to chassis leg control in the embodied strategy mapping network are locked, and the basic leg movement parameters are kept in a static state, so as to generate the pure arm operation parameters corresponding to the static base operation mode based on the heterogeneous feature vector.

6. A centaur robot embodied intelligent control device for special scenarios, characterized in that, include: The parameter acquisition module is used to acquire the multi-view depth image matrix, leg joint pose parameters, arm joint pose parameters, and inertial navigation parameters of the Centaur robot at the current moment. The implicit feature alignment module is used to input the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters into the embodied policy mapping network to perform implicit feature alignment processing on the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters, thereby generating a heterogeneous feature vector characterizing the body motion trend and end-effector operation intention of the Centaur robot. The capability self-construction module is used in the embodied policy mapping network to perform a progressive capability self-construction process with the heterogeneous feature vector as the processing object, so as to sequentially generate basic leg movement parameters, pure arm operation parameters, and leg-arm coordinated linkage parameters. The torque command control module is used to generate the chassis leg driving torque command and the end effector torque command of the Centaur robot according to the leg-arm coordinated linkage parameters, and then control the Centaur robot through the chassis leg driving torque command and the end effector torque command.

7. An electronic device, characterized in that, Including memory and processor; The memory stores computer programs; When the processor executes the computer program, it performs the method according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program; When the computer program is executed by a processor, it is used to perform the method described in any one of claims 1-5.

9. A centaur robot, characterized in that, include: The fuselage itself; A chassis leg actuator is connected to the fuselage body and is used to receive chassis leg drive torque commands and execute chassis leg movements. A dual-arm end effector is connected to the main body of the machine and is used to receive dual-arm end-operating torque commands and perform dual-arm end operations. A multi-view depth acquisition component is disposed on the main body of the robot and is used to acquire a multi-view depth image matrix of the Centaur robot at the current moment. A leg joint pose acquisition component is disposed on the chassis leg actuator and is used to acquire the leg joint pose parameters of the Centaur robot at the current moment. An arm joint pose acquisition component is disposed on the dual-arm end effector and is used to acquire the arm joint pose parameters of the Centaur robot at the current moment. An inertial navigation component is disposed on the main body of the robot and is used to collect the inertial navigation parameters of the Centaur robot at the current moment. The embodied intelligent control device is connected to the multi-view depth acquisition component, the leg joint pose acquisition component, the arm joint pose acquisition component, the inertial navigation component, the chassis leg actuator, and the dual arm end effector. The embodied intelligent control device is used to input the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters into the embodied strategy mapping network to perform feature implicit alignment processing on the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters, thereby generating a heterogeneous feature vector characterizing the centaur robot's body motion trend and end-effector operation intention; The embodied intelligent control device is also used in the embodied strategy mapping network to perform a progressive capability self-construction process with the heterogeneous feature vector as the processing object, so as to sequentially generate basic leg movement parameters, pure arm operation parameters and leg-arm coordinated linkage parameters. The embodied intelligent control device is also used to generate the chassis leg driving torque command and the dual arm end effector torque command according to the leg-arm coordinated linkage parameters, so as to control the chassis leg actuator through the chassis leg driving torque command and control the dual arm end effector through the dual arm end effector torque command.

10. A centaur robot embodied intelligent control system for special scenarios, characterized in that, include: The multimodal parameter acquisition subsystem is used to acquire the multi-view depth image matrix, leg joint pose parameters, arm joint pose parameters, and inertial navigation parameters of the Centaur robot at the current moment. An embodied strategy mapping subsystem, connected to the multimodal parameter acquisition subsystem, is used to receive the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters. It then inputs these parameters into the embodied strategy mapping network to perform implicit feature alignment processing on the multi-view depth image matrix, the leg joint pose parameters, the arm joint pose parameters, and the inertial navigation parameters, thereby generating heterogeneous feature vectors characterizing the centaur robot's body motion trends and end-effector operation intentions. A progressive capability self-construction subsystem, connected to the embodied strategy mapping subsystem, is used in the embodied strategy mapping network to perform a progressive capability self-construction process with the heterogeneous feature vector as the processing object, so as to sequentially generate basic leg movement parameters, pure arm operation parameters, and leg-arm coordinated linkage parameters. The torque command output subsystem is connected to the progressive capability self-construction subsystem and is used to generate the chassis leg drive torque command and the end-effector operation torque command of the Centaur robot according to the leg-arm coordinated linkage parameters. An execution control subsystem, connected to the torque command output subsystem, is used to receive the chassis leg drive torque command and the dual arm end effector torque command, and then control the centaur robot through the chassis leg drive torque command and the dual arm end effector torque command.

Images