Semi-centaur robot multi-modal gait generation and control method and device based on imitation learning, equipment, medium, robot and system

Abstract

The application provides a semi-human robot multi-modal gait generation and control method and device based on imitation learning, electronic equipment, computer readable storage medium, semi-human robot and system. The method obtains a multi-modal motion time sequence trajectory of a target alien organism, performs redirection mapping according to physical kinematic constraint conditions of a semi-human robot, and generates an initial control trajectory; a standard gait feature data is generated through gait generation network denoising and smoothing; a target control strategy is iteratively solved in combination with a dynamic stability standard and an interaction level, and a joint driving instruction is generated to execute multi-modal gait switching.

Description

Technical Field

[0001] This application relates to the field of robot gait control technology, and more specifically, to a method, apparatus, electronic device, computer-readable storage medium, centaur robot and system based on imitation learning for generating and controlling multimodal gait of a centaur robot. Background Technology

[0002] Centaur robots possess both upper limb manipulation capabilities and multi-legged mobility. In scenarios such as navigating complex terrain, transporting loads, emergency operations, and human-robot collaborative operations, they need to switch between different movement modes according to task type, ground environment, and load status. Because this type of robot involves the coupling relationship between torso posture, quadrupedal support, and multi-joint actuation, in actual operation, gait switching not only needs to meet the requirements of motion continuity, but also needs to take into account support stability and control adaptability under different task rhythms.

[0003] In existing Centaur robot control schemes, independent motion control templates are typically configured for different work states. This scheme first establishes a corresponding set of joint control parameters based on a specific work state; then, when the robot enters a different work state, the corresponding control template is invoked; finally, switching between different control templates is performed according to fixed switching conditions, and the corresponding joint control commands are executed by the underlying drive unit.

[0004] However, the above-mentioned solutions have obvious technical drawbacks. Due to the significant differences in parameter specifications and state constraint methods among different control templates, problems such as discontinuity in motion timing and abrupt changes in joint drive response can easily occur when the work state changes. At the same time, different tasks, such as light-load fast passage and heavy-load slow passage, have different requirements for gait execution frequency and stability constraints. If each task state relies on separate training or separate configuration of control templates, the training time will be long, the cross-task generalization ability will be low, and it will be difficult to meet the needs of Centaur robots for smooth multimodal gait switching and task-adaptive control in complex work scenarios. Summary of the Invention

[0005] This application provides a method, apparatus, electronic device, computer-readable storage medium, centaur robot, and system for generating and controlling multimodal gait of a centaur robot based on imitation learning, so as to at least alleviate the above-mentioned technical problems.

[0006] This application provides a method for multimodal gait generation and control of centaur robots based on imitation learning. Addressing the shortcomings of existing methods that rely on independent control templates under different operational states, leading to discontinuous motion timing and abrupt changes in joint drive responses, this method first acquires the multimodal motion timing trajectory of the target heterogeneous organism. Then, based on the physical kinematic constraints of the centaur robot, it redirects and maps the multimodal motion timing trajectory to generate the initial control trajectory of the centaur robot. Compared to traditional methods that configure control templates separately for different operational states, this method allows the external motion timing to be converted into a control framework defined by the centaur robot's own physical kinematic constraints. This means that subsequent gait generation no longer directly relies on separate parameter templates, which helps reduce the impact of shape differences and control space differences between different gait sources on the continuity of joint drive.

[0007] This application inputs the initial control trajectory into a gait generation network. Within the gait generation network, the initial control trajectory is first denoised, and then the denoised result is temporally smoothed to generate standard gait feature data. Compared to traditional methods that directly call fixed control templates and switch between them, the above process reduces abrupt changes in the initial control trajectory through denoising before generating the target control strategy, and then maintains the continuity of the gait change process through temporal smoothing. This allows the standard gait feature data to serve as a unified input basis for subsequent control strategy iterations, thereby alleviating the problem of abrupt switching caused by inconsistent parameter standards between different control templates.

[0008] This application constructs a dynamic stability criterion for constraining the vertical axial position of the centaur robot's center of mass based on a variable-axis inverted pendulum mechanism. It also introduces a partially observable Markov mechanism, defining the gait control process as an interaction hierarchy encompassing both the environmental state space and the motion space. Compared to traditional methods that primarily rely on fixed switching conditions to invoke corresponding control templates, this approach incorporates stability constraints related to the vertical axial position of the center of mass into the gait control process. Furthermore, it expresses the correspondence between the environmental state space and the motion space through an interaction hierarchy. This allows the gait generation process to not only consider motion reproduction but also incorporate stability constraints during iterative solving, thereby improving the adaptability to changes in the support stability range during multimodal gait switching.

[0009] This application, within the interaction layer, uses a proximal policy optimization mechanism and a dynamic stability criterion to iteratively solve standard gait feature data, generating a multi-band decoupled target control strategy. Compared to traditional methods that require separate training or configuration of control templates for different task states such as light-load fast passage and heavy-load slow passage, the above process uses standard gait feature data as a unified gait input basis and iteratively forms a multi-band decoupled target control strategy under the constraints of the dynamic stability criterion. This allows the same control strategy to adapt to different gait execution frequencies, reducing the training time and parameter maintenance burden caused by reconfiguring control templates for each task state.

[0010] This application generates joint drive commands based on a target control strategy to drive the joint motors of a Centaur robot, controlling the Centaur robot to perform multimodal gait switching. Compared to the traditional approach of switching between different independent control modules before outputting joint control commands, the joint drive commands in this application originate from a target control strategy formed through redirection mapping, denoising and smoothing, stability constraints, and iterative solutions. This creates a more continuous technical relationship between the drive input of the joint motors and the multimodal gait switching process, thereby improving the smoothness, task adaptability, and control stability of the Centaur robot's multimodal gait switching in complex work scenarios. Attached Figure Description

[0011] Figure 1 This application provides an embodiment of a multimodal gait generation and control scenario for a centaur robot based on imitation learning. Figure 2 This application provides an embodiment of a method for generating and controlling multimodal gait of a centaur robot based on imitation learning. Figure 3 This application provides an embodiment of a multimodal gait generation and control device for a centaur robot based on imitation learning. 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 multimodal gait generation and control system for a centaur robot based on imitation learning. Detailed Implementation

[0012] like Figure 1 As shown, this is an embodiment of the present application of a multimodal gait generation and control scenario for a centaur robot based on imitation learning; Figure 2As shown in the figure, this application provides an embodiment of a method for generating and controlling the multimodal gait of a centaur robot based on imitation learning, comprising: The multimodal motion time-series trajectory of the target heterogeneous organism is acquired, and the multimodal motion time-series trajectory is redirected and mapped according to the physical kinematic constraints of the centaur robot to generate the initial control trajectory of the centaur robot. The initial control trajectory is input into the gait generation network. The initial control trajectory is first denoised within the gait generation network, and then the denoised result is temporally smoothed to generate standard gait feature data. Based on the variable axis inverted pendulum mechanism, a dynamic stability criterion is constructed to constrain the vertical axis position of the centaur robot's center of mass. A partially observable Markov mechanism is introduced to define the gait control process as an interactive hierarchy that includes the environmental state space and the action space. Within the interaction layer, based on the near-end policy optimization mechanism and the dynamic stability criterion, the standard gait feature data is iteratively solved to generate a multi-band decoupled target control strategy. The target control strategy generates joint drive commands to drive the joint motors of the centaur robot and control the centaur robot to perform multimodal gait switching.

[0013] Optionally, before acquiring the multimodal motion temporal trajectory of the target heterologous organism, the following steps are included: Using the target heterologous organism as the motion acquisition object, walking gait data, running gait data and climbing gait data of the target heterologous organism are collected by optical capture device; Based on the acquisition time markers corresponding to each gait data, the walking gait data, the running gait data, and the climbing gait data are time-aligned to generate gait time-series data corresponding to different gait categories; The key bone contact point temporal change features are extracted from the gait temporal data. The key bone contact point temporal change features characterize the bone position change state of the target heterologous organism at different times when it comes into contact with the supporting environment. According to the gait category, the temporal change features of the corresponding key bone contact points are sequentially spliced ​​to generate the multimodal motion temporal trajectory of the target heterologous organism.

[0014] Preferably, the specific implementation process of the step "using the target heterologous organism as the motion acquisition object, and acquiring the walking gait data, running gait data, and climbing gait data of the target heterologous organism through an optical capture device" is as follows: Before the acquisition begins, motion acquisition point identification records are configured according to the skeletal connection relationship of the target heterologous organism. The motion acquisition point identification records include the installation position of each skeletal acquisition point on the target heterologous organism, the connection order between each skeletal acquisition point, and the feature extraction correspondence between each skeletal acquisition point and the temporal change characteristics of subsequent key skeletal contact points. Among them, the installation position is used to limit the spatial acquisition source of each skeletal acquisition point on the target heterologous organism, the connection order is used to limit the skeletal motion transmission relationship between each skeletal acquisition point, and the feature extraction correspondence is used to limit the field writing position of each skeletal acquisition point in the temporal change characteristics of subsequent key skeletal contact points, so that the three-dimensional position data of the skeletal acquisition points acquired by the optical capture device can maintain the same data caliber as the skeletal motion relationship of the target heterologous organism. Subsequently, the optical capture device is calibrated spatially, configuring the acquisition space origin, horizontal support reference plane, vertical reference line, acquisition frame interval, and acquisition time stamp writing method into the acquisition calibration record. The acquisition space origin is used to define the coordinate starting position of the 3D position data of the skeletal acquisition point. The horizontal support reference plane is used to define the spatial reference when the target alien organism comes into contact with the support environment. The vertical reference line is used to define the reading aperture in the height direction of the 3D position data of the skeletal acquisition point. The acquisition frame interval is used to define the time interval between adjacent acquisition frames. The acquisition time stamp writing method is used to establish a correspondence between each frame of 3D position data of the skeletal acquisition point and the corresponding acquisition time. The acquisition calibration record is used to convert each frame of 3D position data of the skeletal acquisition point output by the optical capture device into acquisition frame data with a unified spatial and temporal source. The acquisition frame data continues to serve as the data component unit of the walking gait data, the running gait data, and the climbing gait data. Subsequently, when the target alien organism performs walking gait, running gait, and climbing gait respectively, the optical capture device continuously reads the three-dimensional position data of each skeletal acquisition point in the acquisition space corresponding to the motion acquisition point identifier record, and writes the three-dimensional position data of the skeletal acquisition points, acquisition time marker, gait category identifier, and skeletal acquisition point identifier at the same acquisition time into the acquisition frame data.For walking gait, the acquired frame data records the walking gait data formed by the alternating landing of support bone acquisition points, the low-amplitude forward movement of swing bone acquisition points, and the gradual displacement of trunk bone acquisition points. The walking gait data expresses the alternating support relationship under walking gait through the support bone acquisition points, swing bone acquisition points, and trunk bone acquisition points. For running gait, the acquired frame data records the running gait data formed by the shortened contact time of support bone acquisition points, the increased forward displacement speed of swing bone acquisition points, and the large periodic fluctuations of trunk bone acquisition points. The running gait data expresses the rapid support switching relationship under running gait through the support bone acquisition points, swing bone acquisition points, and trunk bone acquisition points. For climbing gait, the acquired frame data records the climbing gait data formed by the combined movement of support bone acquisition points in the vertical and forward directions, and the alternating contact between multiple bone acquisition points and the support environment. The climbing gait data expresses the height variation support relationship under climbing gait through the support bone acquisition points and multiple bone acquisition points. The resulting walking gait data, running gait data, and climbing gait data are not simply video clips, but data sequences with acquisition time markers, gait category identifiers, skeletal acquisition point identifiers, and skeletal acquisition point 3D position data. The walking gait data, running gait data, and climbing gait data continue to participate in subsequent temporal alignment processing, so that motion changes under different gait categories can be converted into the preliminary data source of the multimodal motion temporal trajectory.

[0015] Preferably, the specific implementation process of the step "aligning the walking gait data, running gait data, and climbing gait data according to the acquisition time markers corresponding to each gait data to generate gait time sequence data corresponding to different gait categories" is as follows: First, the acquisition time markers in the walking gait data, running gait data, and climbing gait data are read respectively, and the acquisition frame data in each type of gait data is sorted according to the order of the acquisition time markers to form walking frame sequence data, running frame sequence data, and climbing frame sequence data. The walking frame sequence data is formed by sorting the acquisition frame data in the walking gait data according to the acquisition time markers, the running frame sequence data is formed by sorting the acquisition frame data in the running gait data according to the acquisition time markers, and the climbing frame sequence data is formed by sorting the acquisition frame data in the climbing gait data according to the acquisition time markers. The walking, running, and climbing frame sequence data retain their original temporal order during acquisition. However, due to differences in execution rhythm, duration, and acquisition starting point, these data cannot be directly concatenated in their original frame order. Therefore, a unified timeline index is established based on the acquisition frame intervals in the acquisition calibration record. The walking, running, and climbing frame sequence data are then mapped to this unified timeline index, ensuring that the 3D position data of the skeletal acquisition points in these data correspond to the same temporal representation. For adjacent acquisition frames with interval differences between acquisition time markers, time-position completion processing is performed based on the changes in the three-dimensional position data of the skeletal acquisition points in the adjacent acquisition frames to form completed frame data that matches the unified time axis index. This completed frame data is then written into the corresponding walking, running, or climbing frame sequence data, ensuring that the walking, running, and climbing frame sequence data, after time-position completion processing, can use the same unified time axis index. For acquisition frames that appear repeatedly under the same acquisition time marker, acquisition frames that are continuous in the spatial direction of change with the preceding and following acquisition frames are retained based on the skeletal acquisition point identifier and gait category identifier. These retained acquisition frames are then used as valid acquisition frames after filtering for duplicate acquisition frames, ensuring that the filtering of duplicate acquisition frames does not change the spatial direction of change of the same skeletal acquisition point in the preceding and following acquisition frames. After the above-mentioned timing alignment processing, the walking frame sequence data forms gait timing data corresponding to the walking gait category, the running frame sequence data forms gait timing data corresponding to the running gait category, and the climbing frame sequence data forms gait timing data corresponding to the climbing gait category.The gait time-series data includes, in its data structure, gait category identifiers, a unified timeline index, skeletal acquisition point identifiers, and 3D position data of the skeletal acquisition points. The gait category identifiers distinguish the sources of the walking gait data, running gait data, and climbing gait data. The unified timeline index converts inconsistent acquisition time markers under different gait categories into comparable and concatenable temporal positions. The skeletal acquisition point identifiers maintain the correspondence between the 3D position data of the skeletal acquisition points and the motion acquisition point identifier records. The 3D position data of the skeletal acquisition points is used to subsequently determine the spatial change state of contact between the skeletal acquisition points and the supporting environment. Therefore, when extracting key skeletal contact point temporal change features from the gait time-series data, the sequential changes in contact between the skeletal acquisition points and the supporting environment can be determined within the same temporal expression caliber.

[0016] Preferably, in the specific technical implementation of the step "extracting key skeletal contact point temporal change features from the gait temporal data, wherein the key skeletal contact point temporal change features characterize the skeletal position change state of the target heterologous organism in contact with the support environment at different times", gait temporal data corresponding to different gait categories are first read, and the support environment position reference is determined based on the horizontal support reference plane in the acquisition calibration record. The support environment position reference is used to describe the spatial reference range in which the target heterologous organism can make support contact during walking, running, and climbing. The support environment position reference is also used to subsequently calculate the vertical distance of the three-dimensional position data of the skeletal acquisition points relative to the support environment. Subsequently, based on the correspondence between each skeletal acquisition point and the skeletal connection relationship in the motion acquisition point identification record, skeletal acquisition points that can characterize the landing, push-off, support, or climbing states are selected from the gait temporal data, and the selected skeletal acquisition points are used as candidate data for key skeletal contact points. The candidate data for key skeletal contact points originates from the three-dimensional position data of skeletal acquisition points in the gait time-series data. This candidate data is further combined with the unified timeline index to participate in continuous frame determination, rather than being directly used as the temporal variation feature of the key skeletal contact points. The continuous frame determination first reads the three-dimensional position data of the same skeletal acquisition point on adjacent unified timeline indices, then calculates the vertical distance between the same skeletal acquisition point and the support environment position reference based on the support environment position reference, and compares this vertical distance with the contact determination boundary. The contact determination boundary originates from the acquisition space calibration results of the optical capture device and the range of height variation of the skeletal acquisition points of the target heterologous organism in a static support state. This boundary distinguishes between the time period when the skeletal acquisition point makes contact with the support environment and the time period when the skeletal acquisition point leaves the support environment. When the vertical distance of the same bone acquisition point relative to the support environment position reference is within the contact determination boundary in continuous acquisition frame data, and the three-dimensional position data change of the same bone acquisition point shows a low displacement state during the support phase, the position change of the same bone acquisition point on the corresponding unified time axis index is recorded as contact position change data. The contact position change data includes the contact occurrence time, contact duration, contact bone source, and contact spatial location, so that the contact position change data can express the temporal change of the bone acquisition point and the support environment forming contact. When the aforementioned same bone acquisition point leaves the contact determination boundary, and the three-dimensional position data change of the same bone acquisition point shows a lifting and forward movement state during the swing phase, the position change of the same bone acquisition point on the corresponding unified time axis index is recorded as ground-off position change data. The ground-off position change data includes the ground-off switching time, ground-off bone source, and ground-off spatial location, so that the ground-off position change data can express the temporal change of the bone acquisition point after leaving the support environment.Subsequently, the contact position change data, the ground-lift position change data, the corresponding skeletal acquisition point identifier, the corresponding gait category identifier, and the corresponding unified time axis index are correlated and encoded to form the key skeletal contact point temporal change features. The technical essence of the key skeletal contact point temporal change features is to convert the continuously changing three-dimensional position data of skeletal acquisition points in the gait temporal data into a temporal contact description that can express the contact occurrence time, contact duration, contact skeletal source, contact spatial location, ground-lift switching time, and ground-lift spatial location. The temporal contact description continues to participate in the subsequent sequence splicing according to the gait category, so that the sequence splicing process not only preserves the spatial movement trajectory of the target heterologous organism, but also preserves the changing relationship supporting the contact rhythm under different gait categories.

[0017] Preferably, the specific implementation process of the step "according to the gait category, sequentially splicing the corresponding key bone contact point temporal change features to generate the multimodal motion temporal trajectory of the target heterologous organism" is as follows: After the key bone contact point temporal change features have been formed corresponding to walking gait category, running gait category, and climbing gait category respectively, the key bone contact point temporal change features under the same category are first read according to the gait category identifier, and the key bone contact point temporal change features under the same category are sorted in order within the category according to the unified time axis index to form walking contact change sequence, running contact change sequence, and climbing contact change sequence. The walking contact change sequence is formed by arranging the key bone contact point temporal change features corresponding to the walking gait category according to the unified time axis index, the running contact change sequence is formed by arranging the key bone contact point temporal change features corresponding to the running gait category according to the unified time axis index, and the climbing contact change sequence is formed by arranging the key bone contact point temporal change features corresponding to the climbing gait category according to the unified time axis index. The walking contact change sequence preserves the temporal relationship of alternating contact generation of support skeletal acquisition points under the walking gait category; the running contact change sequence preserves the temporal relationship of rapid contact and rapid ground lift of support skeletal acquisition points under the running gait category; and the climbing contact change sequence preserves the temporal relationship of multiple skeletal acquisition points generating contact with the support environment at different spatial heights under the climbing gait category. Subsequently, the walking, running, and climbing contact change sequences are written into a multimodal trajectory splicing support record, and the multimodal trajectory splicing support record retains the starting unified time axis index, ending unified time axis index, skeletal acquisition point identifier, and contact state switching position corresponding to each gait category. The starting unified time axis index is used to record the starting temporal position of the sequence corresponding to each gait category; the ending unified time axis index is used to record the ending temporal position of the sequence corresponding to each gait category; the skeletal acquisition point identifier is used to maintain the correspondence between the sequence corresponding to each gait category and the motion acquisition point identifier record; and the contact state switching position is used to record the switching position between contact position change data and ground lift position change data. By retaining the starting unified time axis index, the ending unified time axis index, the skeleton acquisition point identifier, and the contact state switching position in the multimodal trajectory splicing record, the sequence splicing process can distinguish the boundaries of different gait categories without mistaking contact changes between different gait categories for the same continuous gait cycle.Subsequently, based on the arrangement relationship between the gait categories, the walking contact change sequence, the running contact change sequence, and the climbing contact change sequence are sequentially connected, and a category switching identifier is retained at the connection position of adjacent gait categories. The category switching identifier is used to record the continuity relationship between the end unified time axis index of the previous gait category and the start unified time axis index of the next gait category. The category switching identifier and the contact state switching position in the multimodal trajectory splicing record jointly participate in the subsequent redirection mapping process, so that when the multimodal motion temporal trajectory is redirected and mapped according to the physical kinematic constraints of the Centaur robot, the connection position of different gait categories in time can be identified, and the contact rhythm change of the target alien organism can be mapped into an initial control trajectory that the Centaur robot can execute. The resulting multimodal motion temporal trajectory includes the walking contact change sequence, the running contact change sequence, the climbing contact change sequence, and the category switching identifier. The multimodal motion temporal trajectory not only preserves the skeletal position change state of the target heterologous organism under different gait categories, but also preserves the continuity relationship between different gait categories, so that the multimodal motion temporal trajectory can serve as a unified trajectory source for subsequent redirection mapping processing.

[0018] Optionally, the step of retargeting and mapping the multimodal motion temporal trajectory according to the physical kinematic constraints of the centaur robot to generate the initial control trajectory of the centaur robot includes: Obtain the physical limits of the link length and joint angle of the centaur robot to construct the physical kinematic constraints; Based on the physical kinematic constraints, the mechanical space coordinate range of the centaur robot is determined as the target mechanical motion range for the multimodal motion temporal trajectory mapping. Based on the target's mechanical motion range, a spatial geometric transformation matrix is ​​established to transform the spatial coordinates of the target heterogeneous organism into the mechanical spatial coordinates of the centaur robot. The spatial geometric transformation matrix represents the physical limitations of the link length and the physical limits of the joint angle. The spatial geometric transformation matrix is ​​used to project the multimodal motion temporal trajectory into coordinate space to scale the multimodal motion temporal trajectory to the physical shape range of the centaur robot, thereby generating the initial control trajectory.

[0019] Preferably, the specific implementation process of the step "obtaining the physical limits of the link length and joint angle of the Centaur robot to construct the physical kinematic constraints" is as follows: Before redirecting the multimodal motion time-series trajectory, the shape calibration data and joint calibration data of the Centaur robot are read first. The shape calibration data comes from the Centaur robot shape calibration record formed when the Centaur robot completes the assembly calibration. The Centaur robot shape calibration record includes the link assembly length between the torso connection part and each limb connection part, the link connection sequence between adjacent links, and the motion direction diameter corresponding to each link. The joint calibration data comes from the joint zero-position calibration record formed when the Centaur robot completes the joint zero-position calibration. The joint zero-position calibration record includes the joint zero position corresponding to each joint motor, the joint forward rotation boundary, the joint reverse rotation boundary, and the joint rotation direction diameter. Subsequently, the physical limits of the link lengths are extracted from the Centaur robot's body calibration record. These physical limits are not single length values, but rather express the non-extendable assembly dimensional relationships between the Centaur robot's links and the reachable positional boundaries of each link within the mechanical space. The physical limits of the joint angles are extracted from the joint zero-position calibration record. These limits express the forward and reverse rotation boundaries of each joint motor around the joint zero position, and constrain the subsequently generated initial control trajectory to not exceed the angular range that the joint motors can execute. Then, the physical limits of the link lengths and the physical limits of the joint angles are written into the kinematic constraint load record, and a correspondence is established between link identifiers, joint identifiers, physical limits of link lengths, and physical limits of joint angles in the kinematic constraint load record, ensuring that the spatial extension range of each link corresponds to the physical limit of the joint angle of the connected joint motor. The kinematic constraint bearing record continues to participate in the construction of the physical kinematic constraint conditions. These physical kinematic constraint conditions are formed by the physical limitations on link lengths and joint angles in the kinematic constraint bearing record, and are used to incorporate the mechanical body boundaries and joint executable boundaries of the Centaur robot into the same constraint framework. Through this process, the physical kinematic constraint conditions can limit the spatial transformation range of the multimodal motion temporal trajectory during subsequent retargeting mapping, ensuring that changes in skeletal position originating from the target heterogeneous organism are not directly converted into control trajectories exceeding the mechanical body boundaries or joint executable boundaries of the Centaur robot. This provides a source of constraints for subsequently determining the mechanical spatial coordinate range of the Centaur robot.

[0020] Preferably, in the specific technical implementation of the step "determining the mechanical space coordinate range of the Centaur robot based on the physical kinematic constraints, as the target mechanical motion range for the multimodal motion temporal trajectory mapping", after the physical kinematic constraints have been constructed by the physical limitations of the link length and the physical limits of the joint angle, the torso reference position of the Centaur robot is first used as the spatial starting position of the mechanical space coordinate range; the torso reference position is derived from the torso connection parts in the Centaur robot's body calibration record, and the torso reference position is used to unify the spatial extension direction of each limb connection part to the same mechanical space coordinate caliber. Subsequently, according to the link connection sequence in the Centaur robot's body calibration record, the physical limitations of the link length corresponding to each link are sequentially converted into the reachable position of the link end. The reachable position of the link end is further combined with the physical limits of the joint angle in the joint zero-position calibration record, so that the reachable position of the link end under each joint angle physical limit can be limited to the spatial range that the corresponding joint motor can execute. Subsequently, boundary envelope processing is performed on multiple reachable positions of the link ends under the same limb path to form limb reachable space boundary data corresponding to that limb path. The limb path is jointly defined by the torso connection part, the corresponding limb connection part, the link connection sequence between adjacent links, and the physical limit of the joint angle of the corresponding joint motor. The limb path is used to express the mechanical connection relationship of the same limb extending from the torso connection part to the reachable position of the end link end. The limb reachable space boundary data includes forward and backward reachable boundaries, left and right reachable boundaries, and vertical reachable boundaries. The forward and backward reachable boundaries are used to limit the forward and backward mapping amplitude of the multimodal motion time-series trajectory in walking and running gaits. The left and right reachable boundaries are used to limit the left and right mapping amplitude of the multimodal motion time-series trajectory during support switching. The vertical reachable boundaries are used to limit the vertical mapping amplitude of the multimodal motion time-series trajectory in climbing gait. Next, the reachable spatial boundary data of multiple limb paths are stitched together with the torso reference position using unified coordinates to form the mechanical spatial coordinate range of the centaur robot. This mechanical spatial coordinate range expresses the overall spatial boundary that the centaur robot can reach under different limb paths and different joint angle physical limits. Further, the mechanical spatial coordinate range is written into a mechanical spatial boundary record, and the mechanical spatial coordinate range in the mechanical spatial boundary record is used as the target mechanical motion range. The target mechanical motion range is not a simple scaling ratio of the multimodal motion temporal trajectory, but rather a mechanical motion boundary derived from the physical kinematic constraints that can be executed by the centaur robot.When establishing the spatial geometric transformation matrix, the target mechanical motion range is used to define the coordinate mapping endpoint when the spatial coordinates of the target heterogeneous organism are transformed into the mechanical spatial coordinates of the centaur robot, so that the multimodal motion time-series trajectory already has a target boundary corresponding to the mechanical spatial coordinate range of the centaur robot before being mapped.

[0021] Preferably, the specific implementation process of the step "based on the target mechanical motion range, establish a spatial geometric transformation matrix for converting the spatial coordinates of the target heterogeneous organism to the mechanical spatial coordinates of the centaur robot, wherein the spatial geometric transformation matrix represents the physical limitation of the link length and the physical limitation of the joint angle" is as follows: After the target mechanical motion range is formed, the spatial coordinates of the target heterogeneous organism are first read from the multimodal motion time-series trajectory. The spatial coordinates of the target heterogeneous organism include the forward and backward coordinates, left and right coordinates, and vertical coordinates corresponding to the temporal change characteristics of the key bone contact points under different gait categories. Furthermore, the spatial coordinates of the target heterogeneous organism continue to retain the gait category identifier, unified time axis index, and bone acquisition point identifier, so that subsequent coordinate transformations can use the temporal source and bone source of the previous multimodal motion time-series trajectory. Subsequently, the spatial coordinates of the target alien organism and the target mechanical motion range are written into a coordinate transformation calibration record. This record records the spatial coordinate boundaries of the target alien organism, the mechanical spatial coordinate range of the centaur robot, and the directional correspondence between the spatial coordinates of the target alien organism and the mechanical spatial coordinates of the centaur robot. Based on this coordinate transformation calibration record, a spatial geometric transformation matrix is ​​established. The rows of this matrix correspond to the forward / backward, left / right, and vertical directions in the mechanical spatial coordinates of the centaur robot, and the columns correspond to the forward / backward, left / right, and vertical directions in the spatial coordinates of the target alien organism. The matrix element at the intersection of rows and columns represents the transformation contribution of a change in spatial coordinates from one source direction to one target mechanical direction. The transformation contribution relationship is determined based on the directional scale correspondence between the target mechanical motion range and the spatial coordinate boundary of the target alien organism, and is further constrained by the physical limitations of the link length and the physical limits of the joint angle. For example, when the spatial coordinate change of the target alien organism's climbing gait in the vertical direction is greater than the target mechanical motion range of the centaur robot in the vertical direction, the matrix element corresponding to the vertical direction in the spatial geometric transformation matrix will compress the spatial coordinate change in the vertical direction into the target mechanical motion range without changing the climbing gait sequence expressed by the unified time axis index. Furthermore, a correspondence is established between the spatial geometric transformation matrix and the physical limitations of the link length and the physical limits of the joint angle in the coordinate transformation calibration record, so that the spatial geometric transformation matrix can characterize the mechanical shape boundary and joint executable boundary of the centaur robot.Therefore, the spatial geometric transformation matrix is ​​not a coordinate conversion tool that is detached from mechanical constraints, but rather a geometric mapping data that incorporates the spatial coordinates of the target heterogeneous organism, the target mechanical motion range, the physical limitations of the link length, and the physical limits of the joint angle into the same transformation caliber. Subsequently, the spatial geometric transformation matrix continues to participate in coordinate space projection processing to convert the multimodal motion time-series trajectory into the initial control trajectory.

[0022] Preferably, the specific implementation process of the step "projecting the multimodal motion temporal trajectory into coordinate space using the spatial geometric transformation matrix to scale the multimodal motion temporal trajectory to the physical shape range of the centaur robot, thereby generating the initial control trajectory" is as follows: After the spatial geometric transformation matrix is ​​established, the temporal change features of key bone contact points at each temporal position are read sequentially according to the unified time axis index in the multimodal motion temporal trajectory, and the spatial coordinates, gait category identifier, and bone acquisition point identifier of the target alien organism are extracted from the temporal change features of the key bone contact points. Subsequently, the spatial coordinates of the target alien organism under the same unified time axis index are subjected to coordinate space projection processing with the spatial geometric transformation matrix, so that the forward / backward, left / right, and vertical coordinates in the spatial coordinates of the target alien organism are converted into the mechanical spatial coordinates of the centaur robot according to the transformation contribution relationship in the spatial geometric transformation matrix. The coordinate space projection processing continues to read the target mechanical motion range and performs boundary limiting processing between the centaur robot's mechanical space coordinates and the target mechanical motion range. When the forward / backward, left / right, or vertical coordinates of the centaur robot's mechanical space coordinates exceed the target mechanical motion range, boundary compression processing is performed on the centaur robot's mechanical space coordinates based on the corresponding directional boundaries in the target mechanical motion range to form mechanical space coordinates within the physical shape range. The mechanical space coordinates within the physical shape range still retain the corresponding unified time axis index, gait category identifier, and skeleton acquisition point identifier, enabling the mechanical space coordinates within the physical shape range to continue to express the temporal change relationship and gait category source in the multimodal motion time-series trajectory. Subsequently, based on the mechanical space coordinates within the physical shape range and the link connection sequence in the centaur robot's shape calibration record, the target joint angle data corresponding to each joint motor is determined in reverse, and the target joint angle data is checked for angle boundaries using the joint angle physical limit, ensuring that the target joint angle data is within the angle range that the joint motor can execute. The target joint angle data, after angle boundary verification, is associated and encapsulated with the mechanical space coordinates within the physical body, the unified time axis index, the gait category identifier, and the bone acquisition point identifier to form the initial control trajectory. The initial control trajectory includes the mechanical space coordinates within the physical body, the target joint angle data, the unified time axis index, the gait category identifier, and the bone acquisition point identifier. The initial control trajectory retains the gait sequence relationship in the multimodal motion time-series trajectory and transforms the spatial coordinates of the target heterogeneous organism into the physical body range of the centaur robot.Through the above coordinate space projection processing, when the initial control trajectory is subsequently input into the gait generation network, the gait generation network receives not the multimodal motion temporal trajectory that has not undergone mechanical constraint processing, but the initial control trajectory that already meets the physical kinematic constraints. This allows subsequent denoising processing, temporal smoothing processing, and standard gait feature data generation processing to be based on the same mechanical motion caliber.

[0023] Optionally, the initial control trajectory is input into the gait generation network, where the initial control trajectory is first denoised within the gait generation network, and then the denoised result is temporally smoothed to generate standard gait feature data, including: The sliding window length used to input the gait generation network is determined based on the number of consecutive control cycles corresponding to the initial control trajectory. According to the sliding window length, the initial control trajectory is divided into temporal sliding window segment data and input into the gait generation network; The temporal sliding window segmented data is sequentially input into the gait generation network. The gait generation network uses a recurrent layer with a memory gating structure to remove jump noise from the temporal sliding window segmented data, thereby generating a denoised trajectory. A smoothing window is configured at the generation end of the loop layer, and a first-order difference constraint is applied to the denoised trajectory through the smoothing window to obtain the time-smoothed trajectory. The standard gait feature data is generated based on the trajectory after time-series smoothing, and the standard gait feature data maintains the gait time-series change relationship in the initial control trajectory.

[0024] Preferably, the specific implementation process of the step "determining the sliding window length for inputting the gait generation network based on the number of consecutive control cycles corresponding to the initial control trajectory" is as follows: After the initial control trajectory has been formed by the coordinate space projection processing of the multimodal motion temporal trajectory through the spatial geometric transformation matrix, the unified time axis index, gait category identifier, bone acquisition point identifier, mechanical space coordinates within the physical shape range, and target joint angle data in the initial control trajectory are first read. The number of consecutively arranged control cycles in the initial control trajectory is then counted based on the unified time axis index to form the number of consecutive control cycles. The technical essence of the number of consecutive control cycles is to count the length of the initial control trajectory that can be continuously read in the time dimension. The number of consecutive control cycles is not a simple frame count, but rather a time range used to express whether there is a continuous control relationship between adjacent target joint angle data in the initial control trajectory. Subsequently, the gait category identifier in the initial control trajectory is read, and the differences in control cycle density between walking gait, running gait, and climbing gait are distinguished based on the gait category identifier. Specifically, the target joint angle data changes relatively smoothly for walking gait, with a higher frequency of change for running gait, and the mechanical spatial coordinates within the physical shape range of climbing gait exhibit more pronounced support switching changes in the vertical direction. Based on these differences, a sliding window length determination record is first established, and the number of consecutive control cycles, the gait category identifier, the unified time axis index, and the angle change amplitude between adjacent target joint angle data are written into the sliding window length determination record. The sliding window length determination record is used to incorporate the temporal continuity of the initial control trajectory, the source of the gait category, and the angle change amplitude between adjacent target joint angle data into the same determination criterion, so that the subsequently determined sliding window length can simultaneously cover the local gait change process and the continuity relationship between adjacent gait changes. Then, the upper limit boundary of the sliding window length is determined based on the number of continuous control cycles. This upper limit boundary limits the sliding window length to a range of continuous control cycles that the initial control trajectory can provide. The lower limit boundary of the sliding window length is determined based on the angular change amplitude between adjacent target joint angle data. This lower limit boundary limits the sliding window length to cover the complete local gait change process. After being jointly determined by the upper and lower limits of the sliding window length, the sliding window length used to input the gait generation network is obtained. This sliding window length continues to participate in the subsequent segmentation of the initial control trajectory, enabling the initial control trajectory to be divided into multiple interconnected temporal sliding window segments according to temporal continuity.

[0025] Preferably, in the specific technical implementation of the step "dividing the initial control trajectory into temporal sliding window segment data according to the sliding window length and inputting it into the gait generation network", after the sliding window length is determined, the unified time axis index in the initial control trajectory is first read, and the initial control trajectory is sequentially scanned using the unified time axis index as the movement basis for the sliding window starting point. Each sequential scan extracts a continuous segment of target joint angle data, mechanical space coordinates within the physical shape range, gait category identifier, and bone acquisition point identifier according to the sliding window length, and encapsulates the extracted continuous trajectory content into temporal sliding window segment data. The temporal sliding window segment data is not an arbitrary slice of the initial control trajectory, but a continuous control segment with the sliding window length as the time coverage range, the unified time axis index as the sequence source, and the target joint angle data and the mechanical space coordinates within the physical shape range as the main trajectory content. To preserve gait continuity between adjacent temporal sliding window segments, a segment overlap interval is established between adjacent segments. This overlap interval originates from the overlap between the end unified timeline index of the previous segment and the beginning unified timeline index of the next segment. This overlap interval is used to preserve the trend of target joint angle data changes and mechanical spatial coordinate changes within the physical shape range during subsequent denoising and temporal smoothing processes. Subsequently, the temporal sliding window segments are associated with segment boundary markers, including the beginning unified timeline index, the end unified timeline index, the gait category marker, and the skeleton acquisition point marker. This ensures that each segment can be traced back to its temporal position and gait origin within the initial control trajectory. Subsequently, the temporal sliding window segmented data is input into the gait generation network according to the unified timeline index. This ensures that the data received by the gait generation network is no longer a continuous initial control trajectory of variable length, but rather temporal sliding window segmented data with the same sliding window length, the same temporal arrangement caliber, and traceable segment boundary markers. Through this segmentation process, the gait generation network can read the local gait changes within each temporal sliding window segment and identify the gait continuity between adjacent local gait changes by utilizing the overlapping segments. This provides a unified data input format for subsequent removal of transition noise by a recurrent layer with a memory gating structure.

[0026] Preferably, the specific implementation process of the step "inputting the time-series sliding window segmented data sequentially into the gait generation network, and removing jump noise in the time-series sliding window segmented data through a loop layer with a memory gating structure within the gait generation network to generate a denoised trajectory" is as follows: After the time-series sliding window segmented data enters the gait generation network in the order of the unified time axis index, the gait generation network first reads the target joint angle data and the mechanical space coordinates within the physical body range of each segment of time-series sliding window segmented data, and forms the current trajectory change data based on the changes in similar data between adjacent unified time axis indices. The current trajectory change data is used to describe the changes in the target joint angle data of the same joint motor between adjacent control cycles, and the changes in the mechanical space coordinates within the physical body range corresponding to the same bone acquisition point identifier between adjacent control cycles. Subsequently, the loop layer with a memory gating structure within the gait generation network reads the historical gating state data corresponding to the previous unified time axis index, and performs gating fusion processing on the historical gating state data and the current trajectory change data to form the current gating state data. The historical gating state data is used to retain the gait change trends already confirmed in the preceding time-series sliding window segmented data. The current gating state data is used to express the continuous change state of the target joint angle data and the mechanical space coordinates within the physical body range under the current unified time axis index. Then, the loop layer with the memory gating structure performs jump noise judgment on the current trajectory change data based on the historical gating state data and the current gating state data: when a target joint angle data shows a sudden increase in value between adjacent unified time axis indices that is inconsistent with the historical gating state data, and the corresponding mechanical space coordinates within the physical body range do not show continuous changes in the same direction, the sudden increase in the target joint angle data is marked as jump noise; when the mechanical space coordinates within a physical body range experience a sudden increase in value within a local control cycle that is inconsistent with the historical gating state data, and the corresponding target joint angle data and gait category identifier do not show a corresponding gait switching relationship, the sudden increase in the mechanical space coordinates within the physical body range is marked as jump noise. Subsequently, the loop layer with memory gating structure replaces the jump noise according to the current gating state data, so that the target joint angle data marked as jump noise and the mechanical space coordinates within the physical shape range marked as jump noise are corrected according to the continuous change trend in the unified time axis index before and after, so as to generate the denoised trajectory.The denoised trajectory retains the gait category identifier, bone acquisition point identifier, and unified time axis index in the temporal sliding window segmented data. At the same time, it removes jump noise that does not match the continuous gait changes from the target joint angle data and mechanical space coordinates within the physical shape range, so that subsequent temporal smoothing processing can be carried out on the denoised trajectory that has reduced abrupt changes.

[0027] Preferably, in the specific technical implementation of the step "configuring a smoothing window at the generation end of the loop layer, and applying a first-order difference constraint to the denoised trajectory through the smoothing window to obtain the temporally smoothed trajectory", after the denoised trajectory has been generated by the loop layer with a memory gating structure, a smoothing window is first configured at the generation end of the loop layer. The coverage of the smoothing window is determined by the sliding window length and the segment overlap interval, so that the smoothing window can read the continuous trajectory changes within the same temporal sliding window segment data, and can also read the overlapping trajectory changes between adjacent temporal sliding window segment data. Subsequently, the target joint angle data, mechanical space coordinates within the physical shape range, unified time axis index, gait category identifier, and bone acquisition point identifier in the denoised trajectory are written into the smoothing window processing record. The smoothing window processing record is used to maintain the temporal source, gait category source, and bone acquisition point source of the data read by the smoothing window. Subsequently, a first-order difference constraint is applied to the denoised trajectory through the smoothing window. The essence of this first-order difference constraint is to compare the changes in the target joint angle data corresponding to the same joint motor between adjacent unified time axis indices, and to compare the changes in the mechanical spatial coordinates within the physical shape range corresponding to the same bone acquisition point identifier. Based on the comparison results, locally changing trajectories are smoothed out. This first-order difference constraint only constrains adjacent changes between data of the same type, avoiding direct dimensional mixing comparisons between the target joint angle data and the mechanical spatial coordinates within the physical shape range, thus preventing unreasonable direct calculations between data of different dimensions. For target joint angle data, the smoothing window reads the changes in target joint angle data within adjacent control cycles along the unified time axis index, and adjusts local target joint angle data changes that exceed the trend of adjacent target joint angle data changes within the smoothing window to target joint angle data changes that are consistent with the trend of target joint angle data changes before and after. For mechanical space coordinates within the physical shape range, the smoothing window reads the changes in mechanical space coordinates within the physical shape range in the forward / backward, left / right, and vertical directions within adjacent control cycles along the unified time axis index, and adjusts local mechanical space coordinate changes within the physical shape range that exceed the trend of mechanical space coordinate changes within adjacent physical shape ranges within the smoothing window to mechanical space coordinate changes within the physical shape range that are consistent with the trend of mechanical space coordinate changes within the physical shape range before and after. After the above first-order difference constraint processing, a time-smoothed trajectory is formed. The trajectory after temporal smoothing retains the unified time axis index, the gait category identifier, and the skeletal acquisition point identifier. At the same time, it makes the changes in target joint angle data and mechanical spatial coordinates within the physical body range between adjacent control cycles more consistent with the temporal relationship required for continuous gait execution, providing a smoothed trajectory source for the subsequent generation of the standard gait feature data.

[0028] Preferably, the specific implementation process of the step "generating the standard gait feature data based on the time-smoothed trajectory, wherein the standard gait feature data maintains the gait temporal change relationship in the initial control trajectory" is as follows: After the time-smoothed trajectory is formed, the unified time axis index, gait category identifier, skeletal acquisition point identifier, target joint angle data, and mechanical space coordinates within the physical body range are read from the time-smoothed trajectory. The time-smoothed trajectory is then ordered within the gait category according to the gait category identifier. This ordering within the gait category is used to maintain the change order of target joint angle data and the change order of mechanical space coordinates within the physical body range corresponding to walking gait, running gait, and climbing gait, ensuring that the time-smoothed trajectory does not lose the gait category source from the preceding multimodal motion time-series trajectory when generating the standard gait feature data. Subsequently, based on the unified time axis index, the target joint angle data and mechanical space coordinates within the physical body range within the same control cycle are associated and encapsulated in the same cycle to form a standard gait feature record. The standard gait feature record is used to record the correspondence between the changes in target joint angle data and the changes in mechanical spatial coordinates within the physical shape range under the same control cycle, and continues to retain the bone acquisition point identifier, so that the standard gait feature record can correspond to the source of the temporal change features of the contact point of the preceding key bone. Then, multiple standard gait feature records within consecutive control cycles are sequentially connected according to the unified time axis index to form the standard gait feature data. The standard gait feature data includes the standard gait feature record, the unified time axis index, the gait category identifier, and the bone acquisition point identifier. The technical essence of the standard gait feature data is to convert the trajectory data, which has undergone denoising and temporal smoothing, into a unified gait input basis that can be used for subsequent iterative solutions. The standard gait feature data maintains the gait temporal change relationship in the initial control trajectory because the standard gait feature data uses the unified time axis index and gait category identifier from the initial control trajectory during the generation process, and does not change the sequential relationship between walking gait, running gait, and climbing gait during denoising and temporal smoothing. Through the above processing, the standard gait feature data reduces the jump noise in the initial control trajectory and retains the temporal continuity and gait category source on which multimodal gait switching depends. This allows the gait input data under the same mechanical motion caliber to be used when iteratively solving the standard gait feature data based on the dynamic stability criterion.

[0029] 5. The method for generating and controlling multimodal gait of a centaur robot based on imitation learning according to claim 1, characterized in that the dynamic stability criterion for constraining the vertical axial position of the centaur robot's center of mass based on the variable axial inverted pendulum mechanism includes: The foot support area of ​​the centaur robot is collected in real time, and the zero moment point boundary is determined based on the foot support area. The zero moment point boundary limits the support stability range of the centaur robot during the current gait control process. The real-time three-dimensional position of the centaur robot's center of mass is obtained to extract the vertical axial position of the center of mass in the vertical direction. Based on the variable axial inverted pendulum mechanism, a linkage characterization between the vertical axial position and the acceleration of the center of mass in the horizontal direction is established, so that the vertical axial position participates in the determination of the stable state. The predicted zero-moment point location is calculated based on the aforementioned linkage characterization; Based on the geometric distance between the predicted zero-moment point position and the zero-moment point boundary, a dynamic stability criterion is constructed to constrain the vertical axial position of the centaur robot's center of mass. When the predicted zero-moment point position is located inside the zero-moment point boundary, the centaur robot is determined to be in a stable state.

[0030] Preferably, the specific implementation process of the step "real-time acquisition of the foot support area of ​​the Centaur robot, determination of the zero moment point boundary based on the foot support area, and the zero moment point boundary defining the support stability range of the Centaur robot in the current gait control process" is as follows: When the current gait control process enters the stability determination stage, the contact state data of each foot end, the spatial position data of each foot bottom, and the force state data of each foot bottom of the Centaur robot are first read, and the contact state data of each foot end, the spatial position data of each foot bottom, and the force state data of each foot bottom are written into the foot support area acquisition record. The foot contact state data is used to distinguish whether the corresponding foot is in a supporting contact state or a swing-off state. The foot bottom spatial position data is used to express the landing point spatial position of the corresponding foot during the current gait control process. The foot bottom force state data is used to exclude foot contact state data that only involves slight friction but does not form effective load. The foot support area acquisition record is used to incorporate the foot contact state data, foot bottom spatial position data, and foot bottom force state data into the same support determination criterion. Subsequently, based on the foot contact state data, target supporting feet in a supporting contact state are selected from the foot support area acquisition record, and the load-bearing effectiveness of the target supporting feet is verified based on the foot bottom force state data, so that the target supporting feet that pass the load-bearing effectiveness verification can express the effective load-bearing support source during the current gait control process. Next, the spatial position data of the bottom of each foot corresponding to the target supporting foot is projected onto the ground support reference plane to form foot support projection points. These projection points are then bounded according to their adjacency on the ground support reference plane to form the foot support region. The technical essence of the foot support region is the support range enclosed by the foot support projection points of the target supporting foot on the ground support reference plane after load-bearing effectiveness verification during the current gait control process. The foot support region is not a set of contact points of a single foot, but rather the basic data that converts the spatial positional relationships of multiple target supporting feet into a stable support range. Subsequently, the zero-moment point boundary is determined based on the outer boundary of the foot support region, and this boundary is associated with the unified time axis index, gait category identifier, and target supporting foot identifier corresponding to the current gait control process. The zero-moment point boundary is used to define the allowable distribution range of the predicted zero-moment point position in the current gait control process. Subsequently, when constructing the dynamic stability standard based on the geometric distance between the predicted zero-moment point position and the zero-moment point boundary, the zero-moment point boundary serves as the spatial determination basis for the support stability range, so that the dynamic stability standard can correspond to the effective load-bearing support source in the current gait control process.

[0031] Preferably, in the specific technical implementation of the step "obtaining the real-time three-dimensional position of the centaur robot's center of mass to extract the vertical axial position of the center of mass in the vertical direction", after the boundary of the zero-torque point has been determined by the foot support area, the torso posture data, the target joint angle data corresponding to each joint motor, and the body mass distribution calibration data of the centaur robot during the current gait control process are read, and the torso posture data, the target joint angle data, and the body mass distribution calibration data are written into the center of mass position calculation record. The torso posture data is used to express the tilt state of the centaur robot's torso structure relative to the ground support reference plane. The target joint angle data comes from the joint angle changes corresponding to the current gait control process in the standard gait feature data. The body mass distribution calibration data comes from the mass distribution record formed by the centaur robot on the torso structure, each limb link, and the end effector when completing the body mass calibration. The center of mass position calculation record is used to include the torso posture data, the target joint angle data, and the body mass distribution calibration data in the same center of mass calculation calibrator. Subsequently, based on the torso posture data, the body mass distribution calibration data is converted to the spatial posture caliber corresponding to the current gait control process. The current spatial position of each limb link relative to the torso structure is determined based on the target joint angle data, ensuring that the converted body mass distribution calibration data reflects changes in torso posture and joint angles during the current gait control process. Then, the converted body mass distribution calibration data undergoes centroid position synthesis processing to form the real-time three-dimensional position of the centroid. The real-time three-dimensional position of the centroid includes the centroid spatial positions of the centrifugal robot in the forward / backward, left / right, and vertical directions. The centroid spatial positions in the forward / backward and left / right directions are used to participate in the subsequent planar position calculation of the predicted zero-moment point position, while the centroid spatial position in the vertical direction is used to extract the vertical axial position. Subsequently, the centroid spatial position corresponding to the vertical direction is read from the real-time three-dimensional position of the centroid and written into the vertical axial position record to form the vertical axial position. The technical essence of the vertical axial position is the height of the centaur robot's overall center of mass relative to the ground support reference plane. This vertical axial position is not an isolated height reading, but rather a center of mass height state jointly determined by the torso posture data, the target joint angle data, and the body mass distribution calibration data. Subsequently, when establishing a linkage representation between the vertical axial position and the center of mass acceleration in the horizontal direction based on the variable axial inverted pendulum mechanism, the vertical axial position continues to serve as a height reference in the linkage representation, enabling the center of mass height state during the current gait control process to participate in the stability determination.

[0032] Preferably, the specific implementation process of the step "establishing a linkage representation between the vertical axial position and the centroid's acceleration in the horizontal direction based on the variable axial inverted pendulum mechanism, so that the vertical axial position participates in the stability determination" is as follows: After the vertical axial position is formed, the real-time three-dimensional position of the centroid within a continuous control cycle is read first, and the centroid's velocity change data in the horizontal direction is generated based on the positional changes of the real-time three-dimensional position of the centroid in the forward and backward and left and right directions between adjacent unified time axis indices; subsequently, the centroid's acceleration in the horizontal direction is formed based on the centroid's velocity change data in the horizontal direction within adjacent control cycles. The centroid's acceleration in the horizontal direction is used to express the horizontal inertia change state of the Centaur robot during the current gait control process, and the vertical axial position is used to express the centroid's overall centroid height state relative to the ground support reference plane. Subsequently, based on the variable-axis inverted pendulum mechanism, the vertical axial position and the centroid's acceleration in the horizontal direction are written into the linkage characterization construction record. A correspondence is established in this record between changes in vertical axial position, changes in horizontal inertia, and the foot support area. The technical essence of the variable-axis inverted pendulum mechanism is to abstract the overall centroid motion of the centaur robot under target support foot constraints into an inverted pendulum motion relationship where the horizontal response changes with the centroid height. Specifically, the higher the vertical axial position, the greater the influence range of the centroid's acceleration in the horizontal direction on the predicted zero-moment point position; conversely, the lower the vertical axial position, the greater the influence range of the centroid's acceleration in the horizontal direction on the predicted zero-moment point position. Based on this technical essence, a correlation model is performed on the vertical axial position and the centroid's acceleration in the horizontal direction to form the linkage characterization. The linkage representation does not simply record the vertical axial position and the centroid's acceleration in the horizontal direction, but rather expresses their mutual constraint relationship during the current gait control process. That is, under the same centroid acceleration in the horizontal direction, different vertical axial positions will cause different horizontal offsets in the predicted zero-moment point position; similarly, under the same vertical axial position, different centroid accelerations in the horizontal direction will cause different horizontal offsets in the predicted zero-moment point position. This linkage representation continues to participate in the subsequent calculation of the predicted zero-moment point position, making the vertical axial position no longer an additional record, but a parameter that directly affects the calculation of the predicted zero-moment point position. This allows the stable state determination to reflect the relationship between the centaur robot's centroid height state and horizontal inertia change state during walking, running, and climbing gaits.

[0033] Preferably, in the specific technical implementation of the step "calculating the predicted zero-moment point position based on the linkage characterization", after the linkage characterization has been established, the real-time three-dimensional position of the centroid, the vertical axial position, the centroid acceleration in the horizontal direction, and the linkage characterization corresponding to the current unified time axis index are read first, and the real-time three-dimensional position of the centroid, the vertical axial position, the centroid acceleration in the horizontal direction, and the linkage characterization are written into the zero-moment point position prediction record. The zero-moment point position prediction record is used to include the centroid spatial source, the centroid height state source, and the horizontal inertia change state source into the same calculation caliber, avoiding the judgment of the stable state based solely on the foot support area or solely on the centroid plane projection. Subsequently, the centroid plane position in the front-back and left-right directions is read from the real-time three-dimensional position of the centroid, and the adjustment relationship between the vertical axial position and the centroid acceleration in the horizontal direction is read according to the linkage characterization. Then, the centroid acceleration in the horizontal direction is horizontally offset according to the linkage characterization to form the zero-moment point horizontal offset data. The zero-moment point horizontal offset data is used to express the forward and backward and left and right offsets of the zero-moment point relative to the centroid plane position under the current centroid height state and horizontal inertia change state. Subsequently, the centroid plane position and the zero-moment point horizontal offset data are superimposed in the same direction to form the predicted zero-moment point position. The predicted zero-moment point position is located within the plane diameter corresponding to the ground support reference plane. The predicted zero-moment point position includes both forward and backward and left and right positions, but does not include the vertical position; the vertical axial position has already influenced the zero-moment point horizontal offset data through the linkage characterization, rather than being a quantity directly mixed and compared with the plane position. Through this processing, the predicted zero-moment point position maintains the same plane spatial diameter as the zero-moment point boundary in terms of dimensions, allowing for positional relationship determination within the ground support reference plane when subsequently calculating the geometric distance between the predicted zero-moment point position and the zero-moment point boundary. Furthermore, the predicted zero-moment point position is associated with the unified time axis index, the gait category identifier, and the foot support region, so that the predicted zero-moment point position can correspond to the support state and gait source in the current gait control process. Thus, the predicted zero-moment point position inherits the centroid plane position source from the real-time three-dimensional position of the centroid, and also forms the zero-moment point horizontal offset data based on the linkage between the vertical axial position and the centroid acceleration in the horizontal direction. This provides planar position data that can be geometrically compared with the zero-moment point boundary for the subsequent construction of the dynamic stability standard.

[0034] Preferably, the specific implementation process of the step "constructing a dynamic stability criterion for constraining the vertical axial position of the centaur robot's center of mass based on the geometric distance between the predicted zero-moment point position and the zero-moment point boundary, wherein the centaur robot is determined to satisfy a stable state when the predicted zero-moment point position is located inside the zero-moment point boundary" is as follows: After the predicted zero-moment point position is formed, the predicted zero-moment point position and the zero-moment point boundary under the same unified time axis index are read first, and the predicted zero-moment point position, the zero-moment point boundary, the foot support area, the vertical axial position, and the linkage characterization are written into the dynamic stability criterion construction record. The dynamic stability criterion construction record is used to incorporate the source of planar position determination, the source of support stability range, and the source of influence of center of mass height state into the same determination caliber, so that the dynamic stability criterion can constrain the vertical axial position of the centaur robot's center of mass. Subsequently, the geometric distance between the predicted zero-moment point position and the zero-moment point boundary is calculated within the ground support reference plane. This geometric distance includes the distance from the inner edge of the zero-moment point boundary when the predicted zero-moment point position is inside the boundary, and the distance from the outer edge of the zero-moment point boundary when the predicted zero-moment point position is outside the boundary. The inner edge distance expresses the spatial margin of the current gait control process relative to the support stability range, while the outer edge distance expresses the spatial offset of the current gait control process from the support stability range. Then, based on the correspondence between the inner edge distance, the outer edge distance, the vertical axial position, and the linkage characterization, the dynamic stability standard is constructed. The technical essence of the dynamic stability standard is to combine the horizontal inertial change state caused by the centaur robot's center of mass height state, the planar distribution of the predicted zero-moment point position, and the support stability range defined by the zero-moment point boundary into a single stability state determination rule. When the predicted zero-moment point is located inside the boundary of the zero-moment point, the Centaur robot is determined to be in a stable state based on the distance inside the boundary, and a stable state determination result is generated. When the predicted zero-moment point is located outside the boundary of the zero-moment point, the Centaur robot is determined to be in a unstable state based on the distance outside the boundary, and a unstable state determination result is generated. Subsequently, the distance outside the boundary, the vertical axis position, and the gait category identifier corresponding to the unstable state determination result are associated and recorded as stability deviation data. The stability deviation data continues to participate in the subsequent iterative solution of the standard gait feature data, so that the subsequently generated multi-band decoupled target control strategy can refer to the source of support stability deviation in the current gait control process.Through the above processing, the dynamic stability criterion is not only statically judged based on the geometric range of the foot support area, but also incorporates the vertical axial position into the calculation of the predicted zero moment point position through the linkage characterization, and then forms a stability state judgment result through the geometric distance between the predicted zero moment point position and the zero moment point boundary, so that the dynamic stability criterion can serve the stability constraint in the subsequent multimodal gait switching process.

[0035] Optionally, the introduction of a partially observable Markov mechanism defines the gait control process as an interaction hierarchy encompassing the environmental state space and the action space, including: The gait control process of the centaur robot is modeled as a partially observable Markov decision process, defining the environmental state space, partial observation information, and action space. The environmental state space is configured to include the centaur robot's body joint angles, body base linear velocity, body base angular velocity, and historical interaction actions. The motion space is configured to include the desired angle increments of the target joints of the centaur robot; The partial observation information is obtained by using inertial measurement units and joint encoders mounted on the centaur robot; The observation information is input into a policy processing network, which generates target interaction actions corresponding to the action space to participate in the subsequent iterative solution of the standard gait feature data.

[0036] Preferably, the specific implementation process of the step "modeling the gait control process of the centaur robot as a partially observable Markov decision process, defining the environmental state space, partial observation information, and action space" is as follows: After the dynamic stability standard has been formed, the standard gait feature data, the dynamic stability standard, the unified time axis index, and the gait category identifier corresponding to the current gait control process are read first, and the standard gait feature data, the dynamic stability standard, the unified time axis index, and the gait category identifier are written into the interaction hierarchy construction record. The interaction hierarchy construction record is used to incorporate the previously generated standard gait feature data and the dynamic stability standard into the same interaction description scope, so that the subsequently formed environmental state space, partial observation information, and action space can all correspond to the same current gait control process. Subsequently, the gait control process of the centaur robot is modeled as a partially observable Markov decision process. The essence of this partially observable Markov decision process is that, when the current gait control process cannot directly and completely read all external support conditions and all internal motion variables, it uses state fields in the environmental state space to express gait changes and uses the target joint desired angle increment in the action space to express the control selection for the next control cycle. Based on this partially observable Markov decision process, the environmental state space is first defined. This environmental state space is used to contain state fields that can describe the current gait control process. The environmental state space includes subsequently configured body joint angles, body base linear velocity, body base angular velocity, and historical interaction actions. Then, the partially observed information is defined. This partially observed information is used to express the observable fields that can be read by inertial measurement elements and joint encoders in the current control cycle. This partially observed information is not equivalent to the entire content of the environmental state space, but rather selects observable fields that can be directly obtained within the current control cycle or calculated from adjacent control cycles. The action space is then defined to limit the range of target joint desired angle increments that the Centaur robot can output in the next control cycle. The action space includes subsequently configured target joint desired angle increments. Thus, the interaction layer is formed by the environmental state space, the partial observation information, and the action space. This interaction layer continues to participate in the generation of subsequent target interaction actions, ensuring that the target interaction actions are no longer isolated joint changes detached from the gait state, but rather action generation results based on the current gait control process, the standard gait feature data, and the dynamic stability standard.

[0037] Preferably, in the specific technical implementation of the step "configuring the environmental state space to include the body joint angles, body base linear velocity, body base angular velocity, and historical interaction actions of the Centaur robot", after the environmental state space is defined, an environmental state space configuration record is first established, and the unified time axis index, gait category identifier, standard gait feature data, and dynamic stability standard corresponding to the current gait control process are written into the environmental state space configuration record. Subsequently, the target joint angle data corresponding to the current control cycle is read from the standard gait feature data, and the target joint angle data is converted into the body joint angle; the body joint angle is used to express the angular state of each joint motor of the Centaur robot relative to the joint zero position in the current control cycle. The body joint angle continues to participate in the generation of the expected angle increment of the subsequent target joint, so that the expected angle increment of the subsequent target joint can be calculated based on the body joint angle of the current control cycle. Then, the linear velocity of the body base is extracted from the changes in the torso's spatial position in the current control cycle and adjacent control cycles. These changes in torso's spatial position are derived from the spatial position readings of the torso structure relative to the ground support reference plane within the current and adjacent control cycles. The linear velocity of the body base is used to express the displacement rate of the centaur robot's torso structure within the ground support reference plane and in the vertical direction. This linear velocity of the body base continues to participate in the formation of subsequent observation information, enabling the policy processing network to read the base displacement change state in the current control cycle. Next, the angular velocity of the body base is extracted from the changes in torso posture in the current and adjacent control cycles. These changes in torso posture are derived from the posture readings of the torso structure relative to the ground support reference plane within the current and adjacent control cycles. The angular velocity of the body base is used to express the rotational change rate of the centaur robot's torso structure around different posture directions. This angular velocity of the body base continues to participate in the formation of subsequent observation information, enabling the policy processing network to identify the impact of torso posture changes on the gait control process. Furthermore, the target interactive action formed in the previous control cycle is written into the environment state space configuration record to form the historical interactive action; the historical interactive action is used to express the target joint expected angle increment that has been output to the action space in the previous control cycle. The historical interactive action continues to participate in the state description of the current control cycle, so that the generation of the target interactive action in the current control cycle can take into account the influence of the historical interactive action of the previous control cycle on the current body joint angle, body base linear velocity and body base angular velocity.With the above configuration, the environmental state space is not a simple parallel arrangement of several motion data, but rather organizes the body joint angles, the body base linear velocity, the body base angular velocity, and the historical interaction actions according to a unified time axis index into a state expression of the current gait control process. The acquisition of subsequent observation information and the generation of target interaction actions all take the environmental state space as the state source.

[0038] Preferably, the specific implementation process of the step "configuring the motion space to include the target joint expected angle increment of the Centaur robot" is as follows: After the environmental state space has been configured, the body joint angles, historical interaction actions, and dynamic stability criteria in the environmental state space configuration record are read first, and the physical limits of the Centaur robot's joint rotation angles are read simultaneously to establish a motion space configuration record. The motion space configuration record is used to incorporate the body joint angles of the current control cycle, the historical interaction actions of the previous control cycle, the dynamic stability criteria, and the physical limits of the joint rotation angles into the same motion constraint caliber, so that the target joint expected angle increment in the motion space can correspond to the angle range that the joint motor can execute. Subsequently, the target joint expected angle increment is used as the motion field of the motion space. The target joint expected angle increment does not directly replace the absolute angle value of the body joint angle, but is the expected change amount of the next control cycle formed on the basis of the body joint angle of the current control cycle. Using the target joint expected angle increment to express the motion space can make the joint changes between adjacent control cycles continue in an incremental form, avoiding abrupt changes in joint drive response caused by directly outputting a large span of absolute joint angles between different gait categories. Then, based on the physical limits of joint angle, an incremental boundary is configured for the target joint's desired angle increment. This incremental boundary is used to ensure that the sum of the body joint angle and the target joint's desired angle increment in the current control cycle remains within the range that the joint motor can execute. Simultaneously, based on the historical interaction actions, an action continuity relationship is configured for the target joint's desired angle increment. This action continuity relationship is used to ensure that the incremental change between the target joint's desired angle increment in the current control cycle and the historical interaction actions in the previous control cycle remains continuously readable. Afterwards, the target joint's desired angle increment, the incremental boundary, the action continuity relationship, and the corresponding joint identifier are written into the action space configuration record to form the action space. The action space continues to participate in the generation of subsequent target interaction actions, ensuring that the target interaction actions output by the strategy processing network fall within the target joint's desired angle increment range defined by the action space. Through the above configuration, the action space transforms the body joint angle and historical interaction actions in the environmental state space into a selectable target joint's desired angle increment range for the next control cycle. This ensures that subsequent target interaction actions reflect the motion requirements of the current gait control process while also being constrained by the physical limits of joint angle and the dynamic stability criteria.

[0039] Preferably, in the specific technical implementation of the step "acquiring the partial observation information through the inertial measurement unit and joint encoder mounted on the centaur robot", after the environmental state space and the motion space have been configured, the trunk motion observation data output by the inertial measurement unit and the joint angle observation data output by the joint encoder are first read, and the trunk motion observation data and the joint angle observation data are written into the partial observation information acquisition record. The inertial measurement unit is installed on the trunk structure of the centaur robot. The inertial measurement unit is used to output trunk motion observation data related to trunk posture changes, trunk linear motion changes, and trunk angular motion changes within the current control cycle. The joint encoder is set corresponding to the joint motors respectively. The joint encoder is used to output the joint angle observation data corresponding to each joint motor within the current control cycle. Subsequently, the linear velocity and angular velocity of the body base are calculated based on the trunk motion observation data under adjacent unified time axis indices, and the joint angles of the body are read based on the joint angle observation data, so that the partial observation information can cover the state fields in the environmental state space that can be directly read by the current control cycle. Then, the target interaction action formed in the previous control cycle is read, and this target interaction action is written into the partial observation information acquisition record as the historical interaction action of the current control cycle. This allows the partial observation information to express the influence of the historical interaction action of the previous control cycle on the body joint angle, body base linear velocity, and body base angular velocity of the current control cycle. Afterwards, the body joint angle, body base linear velocity, body base angular velocity, historical interaction action, unified time axis index, and gait category identifier are encapsulated as observation fields to form the partial observation information. The technical essence of this partial observation information is to extract observable fields from the environmental state space that can be directly read in the current control cycle or calculated from adjacent control cycles. This partial observation information differs from the complete environmental state space because the external support conditions, support contact details, and future gait changes during the current gait control process cannot all be directly read within the same control cycle. The observation information is further input into the strategy processing network, enabling the network to generate target interactive actions based on the body joint angles, body base linear velocity, body base angular velocity, and historical interactive actions of the current control cycle, and then participate in the subsequent iterative solution of the standard gait feature data.

[0040] Preferably, the specific implementation process of the step "inputting the partial observation information into the strategy processing network, and having the strategy processing network generate the target interactive action corresponding to the action space to participate in the subsequent iterative solution of the standard gait feature data" is as follows: After the partial observation information is formed, the partial observation information is first input into the strategy processing network according to a unified time axis index. The strategy processing network sequentially performs observation field organization processing, temporal state retention processing, and action increment generation processing in the data processing flow. The observation field organization processing is used to read the body joint angle, body base linear velocity, body base angular velocity, historical interactive action, and gait category identifier from the partial observation information, and form an observation field sequence according to a pre-configured field arrangement caliber. The field arrangement caliber is configured when the strategy processing network is initialized. The configuration method is to write the reading order, numerical normalization caliber, and missing field filling caliber of the body joint angle, body base linear velocity, body base angular velocity, historical interactive action, and gait category identifier into the observation field organization record, so that partial observation information from different control cycles can enter the subsequent processing in a consistent field order. The observation field sequence continues to serve as input data for temporal state retention processing. Subsequently, the temporal state retention process reads the observation field sequence of the current control cycle and combines it with the temporal state data retained from the previous control cycle to generate the temporal state data for the current control cycle. This temporal state data expresses the temporal continuity between partial observation information of the current control cycle and historical partial observation information, enabling the policy processing network to retain the influence of previous motion changes on the current action generation even when some observation information is incomplete. Afterward, the action increment generation process reads the temporal state data of the current control cycle and combines it with the target joint expected angle increment, increment boundary, and action continuity relationship in the action space configuration record to generate the target interactive action. The technical essence of the target interactive action is the joint increment control result corresponding to the action space. The target interactive action includes the target joint expected angle increment of each joint motor in the next control cycle, and the target interactive action is constrained by the increment boundary and action continuity relationship in the action space. Subsequently, the target interactive action is associated with the standard gait feature data, the dynamic stability standard, and the unified time axis index, so that the target interactive action can serve as the action input for subsequent iterative solutions to the standard gait feature data. Through the above processing, the strategy processing network does not output isolated actions based solely on the current joint angle. Instead, it incorporates the body joint angle, body base linear velocity, and body base angular velocity from the partial observation information, the source of the preceding actions in the historical interactive actions, and the target joint expected angle increment constraint in the action space into the action generation process. This ensures that the target interactive action maintains the same time source and the same action caliber as the standard gait feature data in the subsequent iterative solution process.

[0041] Optionally, within the interaction layer, based on the proximal policy optimization mechanism and the dynamic stability criterion, the standard gait feature data is iteratively solved to generate a multi-band decoupled target control strategy, including: Within the interaction layer, the partial observation information, the target interaction action, and the standard gait feature data are input into the policy processing network to generate the current generated gait data for the current iteration cycle; The ratio of the probability of generating a new action in the current iteration cycle to the probability of generating an old action in the previous iteration cycle is calculated. This ratio represents the magnitude of the action update. Based on the determination result of the dynamic stability criterion and the similarity between the currently generated gait data and the standard gait feature data, the motion evaluation deviation value in the current state is calculated; The fluctuation range of the ratio is limited by a pruning mechanism, and the weight parameters of the policy processing network are updated by gradient according to the action evaluation deviation value until the action evaluation deviation value converges. Based on the converged policy processing network, the target control policy for multi-band decoupling is determined. The target control policy integrates the interaction level, the dynamic stability criterion, and the standard gait feature data.

[0042] Preferably, the specific implementation process of step "within the interaction level, inputting the partial observation information, the target interaction action, and the standard gait feature data into the policy processing network to generate the current generated gait data for the current iteration cycle" is as follows: After the interaction level has been formed by the environment state space, the partial observation information, and the action space, the partial observation information, the target interaction action, the standard gait feature data, the unified time axis index, and the gait category identifier corresponding to the current iteration cycle are read first, and the partial observation information, the target interaction action, the standard gait feature data, the unified time axis index, and the gait category identifier are written into the current iteration input record. The current iteration input record is used to include the body joint angles, body base linear velocity, body base angular velocity, historical interaction actions, and the target interaction actions corresponding to the action space that can be observed in the current control cycle corresponding to the current iteration cycle into the same iterative solution caliber, so that the subsequently generated gait data can be traced back to the partial observation information, the target interaction action, and the standard gait feature data. Subsequently, the strategy processing network reads the body joint angles, body base linear velocity, body base angular velocity, and historical interaction actions from the partial observation information according to the unified time axis index. It also reads the target joint expected angle increment from the target interaction action and performs incremental superposition processing with the body joint angles to form candidate joint angle data for the current iteration cycle. The candidate joint angle data is further temporally correlated with the body base linear velocity and body base angular velocity to form candidate motion state data. This candidate motion state data expresses the correspondence between the Centaur robot's joint angle state, base linear motion state, and base angular motion state after adopting the target interaction action within the current iteration cycle. Afterward, the strategy processing network reads the standard gait feature record corresponding to the same unified time axis index from the standard gait feature data and performs same-cycle mapping processing between the candidate motion state data and the standard gait feature record to form the currently generated gait data. The currently generated gait data includes candidate joint angle data for the current iteration cycle, candidate mechanical spatial position data corresponding to the candidate joint angle data, a unified time axis index, and a gait category identifier. The technical essence of this currently generated gait data is that, based on the partial observation information providing the current state source, the target interaction action providing the action increment source, and the standard gait feature data providing the gait temporal change relationship source, the gait data is generated by the policy processing network and can participate in the calculation of subsequent action generation probability ratios and action evaluation deviation values.Through the above processing, the currently generated gait data is not an arbitrary output that is detached from the standard gait feature data. Instead, it is formed within the same interaction level by the constraints of the partial observation information, the target interaction action, and the standard gait feature data. This provides a data source for subsequent calculation of the ratio of the probability of generating a new action in the current iteration cycle to the probability of generating an old action in the historical iteration cycle.

[0043] Preferably, in the specific implementation of the step "calculating the ratio of the probability of generating a new action in the current iteration cycle to the probability of generating an old action in the historical iteration cycle, whereby the ratio characterizes the magnitude of the action update", after the current generated gait data has been formed, the partial observation information corresponding to the current iteration cycle, the target interaction action, the current generated gait data, and the action probability record formed by the strategy processing network in the current iteration cycle are first read. The action probability record is used to record the probability distribution state when the strategy processing network generates the target interaction action for the expected angle increment of the target joint in the action space. The probability distribution state is used to express the probability caliber of the strategy processing network selecting the corresponding target interaction action under the partial observation information corresponding to the current iteration cycle. Subsequently, the probability of generating a new action in the current iteration cycle is extracted from the action probability record. The probability of generating a new action in the current iteration cycle corresponds to the target interaction action and is used to express the probability of the strategy processing network generating the target interaction action under the partial observation information corresponding to the current iteration cycle. Afterward, the historical action probability record saved in the historical iteration cycle is read, and the probability of generating an old action in the historical iteration cycle is extracted from the historical action probability record. The probability of generating old actions in the historical iteration cycle corresponds to the same target interaction action and the same part of the observation information, and is used to express the probability that the policy processing network generates the same target interaction action in the historical iteration cycle. To ensure that the probability of generating new actions in the current iteration cycle and the probability of generating old actions in the historical iteration cycle have comparable data, before calculating the ratio, it is first verified whether the corresponding unified time axis index, gait category identifier, action space configuration record, and target joint expected angle increment are consistent. When the above data are consistent, the ratio of the probability of generating new actions in the current iteration cycle to the probability of generating old actions in the historical iteration cycle is calculated to form the action generation probability ratio. The action generation probability ratio is used to characterize the action update magnitude. When the action generation probability ratio is close to the allowable update range adopted by the subsequent pruning mechanism, it indicates that the update magnitude of the target interaction action by the policy processing network in the current iteration cycle is relatively small; when the action generation probability ratio deviates from the allowable update range adopted by the subsequent pruning mechanism, it indicates that the update magnitude of the target interaction action by the policy processing network in the current iteration cycle is relatively large. The action generation probability ratio continues to participate in the subsequent pruning mechanism, so that the weight parameter update of the policy processing network will not produce an excessively large action update magnitude based solely on a single currently generated gait data, thereby providing a source of probabilistic constraints for subsequent action evaluation deviation calculation and gradient update.

[0044] Preferably, the specific implementation process of the step "calculating the action evaluation deviation value in the current state based on the determination result of the dynamic stability criterion and the similarity between the currently generated gait data and the standard gait feature data" is as follows: After the action generation probability ratio is formed, the determination result of the dynamic stability criterion, the currently generated gait data, the standard gait feature data, and the unified time axis index under the same current iteration cycle are read first, and the determination result of the dynamic stability criterion, the currently generated gait data, the standard gait feature data, and the unified time axis index are written into the action evaluation deviation value calculation record. The determination result of the dynamic stability criterion comes from the geometric distance between the predicted zero moment point position and the zero moment point boundary, and is used to express whether the gait state corresponding to the currently generated gait data meets the support stability range requirements; the currently generated gait data comes from the gait data formed by the policy processing network processing the partial observation information, the target interaction action, and the standard gait feature data under the current iteration cycle; the standard gait feature data is used to provide the standard gait input basis after denoising and temporal smoothing processing. Subsequently, according to the unified timeline index, the candidate joint angle data in the currently generated gait data and the target joint angle data in the standard gait feature data are compared using similar data difference calculations to form joint angle similarity data. Simultaneously, the candidate mechanical spatial position data in the currently generated gait data and the mechanical spatial coordinates within the physical shape range in the standard gait feature data are compared using similar data difference calculations to form mechanical spatial position similarity data. The joint angle similarity data is calculated only between angle data, and the mechanical spatial position similarity data is calculated only between spatial position data; the dimensions of angle data and spatial position data are not mixed for comparison. Afterwards, a similarity score is formed between the currently generated gait data and the standard gait feature data based on the joint angle similarity data and the mechanical spatial position similarity data. This similarity score expresses the degree to which the currently generated gait data preserves the gait temporal change relationship in the standard gait feature data. Furthermore, the determination result of the dynamic stability criterion is correlated with the similarity for evaluation: when the determination result of the dynamic stability criterion indicates that a stable state is met and the similarity is at a high level, the motion evaluation deviation value in the current state is relatively low; when the determination result of the dynamic stability criterion indicates that a stable state is not met, or the similarity is at a low level, the motion evaluation deviation value in the current state is relatively high. The resulting motion evaluation deviation value does not simply measure the degree of similarity between the currently generated gait data and the standard gait feature data, but rather incorporates both the determination result of the dynamic stability criterion and the similarity, enabling the motion evaluation deviation value to express the degree of deviation of the current target interaction action in terms of both gait reproduction and support stability.The motion evaluation deviation value continues to participate in subsequent gradient updates, enabling the policy processing network to iteratively solve for the standard gait feature data and the dynamic stability criterion.

[0045] Preferably, in the specific implementation of the step "using a pruning mechanism to limit the fluctuation range of the ratio, and performing gradient updates on the weight parameters of the policy processing network according to the action evaluation deviation value until the action evaluation deviation value converges", after both the action generation probability ratio and the action evaluation deviation value have been formed, a gradient update carrying record is first established, and the action generation probability ratio, the action evaluation deviation value, the currently generated gait data, the target interaction action, and the weight parameters of the policy processing network are written into the gradient update carrying record. The gradient update carrying record is used to include the sources of the action generation probability ratio, the action evaluation deviation value, the currently generated gait data, the target interaction action, and the weight parameters of the policy processing network into the same update caliber, so that subsequent gradient updates can correspond to the specific target interaction action within the current iteration cycle. Subsequently, a pruning mechanism is used to limit the fluctuation range of the action generation probability ratio. The essence of this pruning mechanism is to limit the action generation probability ratio to within the allowable update range when it exceeds it, and to retain its original value when it is within the allowable update range. The allowable update range is configured based on the target joint's expected angle increment range in the action space, the preceding action changes expressed by the historical interaction actions, and the determination result of the dynamic stability criterion. This allowable update range is used to limit excessively large action update amplitudes of the target interaction actions in the current iteration cycle relative to historical iteration cycles. Afterwards, the action generation probability ratio processed by the pruning mechanism is jointly calculated with the action evaluation deviation value to form the weight parameter update amount. The weight parameter update amount is used to express the direction and magnitude of adjustment of the weight parameters of the policy processing network in the current iteration cycle. The higher the action evaluation deviation value, the more the weight parameter update amount tends to reduce the probability of generating target interaction actions that do not meet the stable state determination result or have low similarity; the lower the action evaluation deviation value, the more the weight parameter update amount tends to retain the probability of generating target interaction actions that match the stable state determination result and standard gait feature data. Subsequently, the weight parameters of the policy processing network are updated using gradients based on the weight parameter update amount, and the currently generated gait data and action evaluation deviation value for the next iteration cycle are regenerated after the gradient update is completed. This process is repeated for multiple iteration cycles. When the change magnitude of the action evaluation deviation value is lower than a preset convergence boundary within multiple consecutive iteration cycles, the action evaluation deviation value is considered to have converged. The preset convergence boundary is configured during the training configuration phase based on the control cycle density of the standard gait feature data and the stability determination granularity of the dynamic stability standard. The preset convergence boundary is used to determine whether the action evaluation deviation value has entered a stable change range.Therefore, the pruning mechanism, the action evaluation bias, and the gradient update together form an iterative constraint on the policy processing network, so that the weight parameter update process of the policy processing network can be simultaneously constrained by the action generation probability ratio, similarity, and dynamic stability criteria.

[0046] Preferably, the specific implementation process of the step "determining the multi-band decoupled target control strategy based on the converged strategy processing network, wherein the target control strategy integrates the interaction level, the dynamic stability criterion, and the standard gait feature data" is as follows: After the action evaluation deviation value reaches the convergence condition, the weight parameters of the converged strategy processing network, the environmental state space, partial observation information and action space in the interaction level, the dynamic stability criterion, and the standard gait feature data are read first, and the weight parameters of the converged strategy processing network, the interaction level, the dynamic stability criterion, and the standard gait feature data are written into the target control strategy generation record. The target control strategy generation record is used to incorporate the sources of the weight parameters, interaction levels, dynamic stability criterion, and standard gait feature data of the converged strategy processing network into the same strategy encapsulation scope, so that the subsequently determined target control strategy can be traced back to the various technical processing results in the iterative solution process. Subsequently, based on the converged strategy processing network, the mapping relationship between the partial observation information and the action space is solidified into an action generation mapping relationship. This action generation mapping relationship expresses how the strategy processing network generates corresponding target interactive actions under different body joint angles, body base linear velocities, body base angular velocities, and historical interactive actions. Then, the dynamic stability criteria are written into the stability constraint caliber of the action generation mapping relationship, enabling the target control strategy to use the stability state determination result between the predicted zero-moment point position and the zero-moment point boundary when generating target interactive actions. Further, the standard gait feature data is written into the gait reference caliber of the action generation mapping relationship, allowing the target interactive actions generated by the target control strategy to correspond to the gait temporal change relationship in the standard gait feature data. The gait reference caliber continues to participate in the subsequent division of the temporal frequency band execution part, ensuring that the target interactive action generation process under different gait categories can use the standard gait feature data. Subsequently, based on the gait category identifier and unified time axis index in the standard gait feature data, the action generation mapping relationship is divided into multiple time-series frequency band execution parts corresponding to walking gait, running gait, and climbing gait. Category continuity relationships are maintained between different time-series frequency band execution parts, forming the multi-frequency band decoupled target control strategy. The category continuity relationship is used to record the unified time axis index connection position and gait category identifier switching position between adjacent time-series frequency band execution parts. This category continuity relationship allows the multi-frequency band decoupled target control strategy to continue using the action generation mapping relationship when switching between different gait categories.The technical essence of the multi-band decoupled target control strategy is to organize the execution parts of different time-series frequency bands corresponding to different time-state categories into a unified action generation mapping relationship based on the same converged strategy processing network. This allows walking gait, running gait, and climbing gait to be executed according to their respective control cycle densities, while using the same interaction level, the same dynamic stability standard, and the same standard gait feature data source. The target control strategy determined in this way integrates the interaction level, the dynamic stability standard, and the standard gait feature data. The target control strategy continues to participate in the generation of subsequent joint drive commands, ensuring that the joint drive commands originate from the target control strategy after stability constraints and iterative convergence processing, rather than from independent fixed control templates.

[0047] Preferably, the strategy processing network structurally includes an observation field organization layer, a temporal state retention layer, an action probability generation layer, an action boundary verification layer, and an action increment generation layer. The observation field organization layer corresponds to the aforementioned observation field organization process. This layer reads the body joint angles, body base linear velocity, body base angular velocity, historical interaction actions, unified time axis index, and gait category identifier from the partial observation information, and organizes these observation fields sequentially according to the field arrangement criteria to form an observation field sequence. The temporal state retention layer corresponds to the aforementioned temporal state retention process. This layer reads the observation field sequence and combines it with the temporal state data retained from the previous control cycle to form the temporal state data for the current control cycle. The action probability generation layer reads the temporal state data for the current control cycle and generates action probability records for the expected angle increments of target joints in the action space, enabling the strategy processing network to express the generation tendency of different expected angle increments of target joints under the same set of observation information. The action boundary verification layer is used to read the incremental boundaries and action continuity relationships in the action probability record and the action space configuration record, and to perform action boundary verification on the target joint expected angle increment corresponding to the action probability record, so that the target joint expected angle increment that does not conform to the incremental boundary or the action continuity relationship will not enter the subsequent action increment generation process. The action increment generation layer is used to read the target joint expected angle increment after being verified by the action boundary verification layer, and to generate the target interaction action in combination with the temporal state data of the current control cycle. Thus, the various structural layers in the policy processing network are not independent data processing units, but are passed step by step along the partial observation information, the observation field sequence, the temporal state data of the current control cycle, the action probability record, and the target interaction action, so that the target interaction action can simultaneously originate from the current observable motion state, the historical interaction actions of the previous control cycle, and the target joint expected angle increment constraints in the action space.

[0048] Preferably, the specific processing method of the observation field sorting layer is as follows: First, the field arrangement criteria are read, and the body joint angles, body base linear velocity, body base angular velocity, historical interaction actions, and gait category identifiers are sequentially read from the partial observation information according to the field arrangement criteria. Subsequently, the body joint angles are arranged according to the joint identifiers, the body base linear velocity is arranged according to the front-back direction, left-right direction, and vertical direction, the body base angular velocity is arranged according to the posture direction of the torso structure, and the historical interaction actions are arranged in the same reading position as the corresponding joint identifiers to form the observation field sequence with fixed field positions. The technical essence of the observation field sequence is to transform partial observation information from different sources into a state field arrangement result that the policy processing network can continuously read. The observation field sequence not only records the motion state of the current control cycle, but also retains the action source of the previous control cycle through the historical interaction actions. To ensure consistent data caliber across the observation field sequences of different control cycles, the observation field processing layer performs numerical normalization and missing field imputation based on the field arrangement caliber. The numerical normalization process is performed only within the same data type, ensuring that the body joint angles, body base linear velocities, and body base angular velocities enter subsequent processing according to their respective data ranges, without mixing the dimensions of angle, linear, and angular velocity data. The missing field imputation process fills in missing joint angle observation data continuously based on the body joint angles corresponding to the same joint identifier under adjacent unified time axis indices when a joint encoder fails to return joint angle observation data for a short period. The imputed body joint angles are then continuously written into the observation field sequence. After the observation field sequence is formed, it is input into the timing state retention layer, enabling the timing state retention layer to read the motion state of the current control cycle at fixed field positions, thereby reducing state reading deviations caused by inconsistent field arrangements between different control cycles.

[0049] Preferably, the temporal state retention layer includes observation write gating, history retention gating, and state fusion processing. The observation write gating is used to determine which fields in the observation field sequence need to be written into the temporal state data of the current control cycle; for example, when the body base angular velocity changes continuously between adjacent unified time axis indices, the observation write gating writes the body base angular velocity into the temporal state data of the current control cycle, so that the temporal state data of the current control cycle can reflect the torso posture change. The history retention gating is used to determine which fields in the temporal state data retained in the previous control cycle need to be retained; for example, when the historical interaction action still affects the change in body joint angle in the current control cycle, the history retention gating continues to retain the state fields corresponding to the historical interaction action in the previous control cycle, so that the temporal state data of the current control cycle can reflect the continued influence of the previous target joint's expected angle increment on the current body joint angle. The state fusion processing is used to fuse the current observation state field obtained from the gating processing and the historical state field obtained from the historical retention gating processing at the same field position to form the temporal state data of the current control cycle. The technical essence of the temporal state data of the current control cycle is that, when some observation information cannot cover all gait states, it organizes the observation fields that can be read in the current control cycle and the motion change trends retained in the previous control cycle into the same temporal state expression. The temporal state data of the current control cycle is then input into the action probability generation layer, enabling the action probability generation layer to generate action probability records based on the motion state of the current control cycle and the action influence of the previous control cycle, rather than generating target interaction actions solely based on the body joint angles of a single control cycle.

[0050] Preferably, the specific processing method of the action probability generation layer is as follows: First, read the temporal state data of the current control cycle, and read the target joint expected angle increment, the increment boundary, the action continuity relationship, and the corresponding joint identifier from the action space configuration record; then, the action probability generation layer calculates the generation tendency of the target joint expected angle increment corresponding to each joint identifier in the action space according to the temporal state data of the current control cycle, and writes the above generation tendency into the action probability record. The action probability record includes the new action generation probability under the current iteration cycle and the data caliber corresponding to the old action generation probability under the historical iteration cycle. The action probability record is used to form the action generation probability ratio later. After reading the action probability record, the action boundary verification layer first excludes the target joint expected angle increment that would cause the body joint angle to exceed the physical limit of the joint rotation angle according to the increment boundary, and then excludes the target joint expected angle increment that has too large a change span with the historical interaction action according to the action continuity relationship, so as to form the action probability record after boundary verification. The boundary-verified motion probability record is then input into the motion increment generation layer. The motion increment generation layer selects the target joint expected angle increment from the boundary-verified motion probability record that matches the temporal state data of the current control cycle. It then associates and encapsulates the selected target joint expected angle increment with the corresponding joint identifier, unified time axis index, and gait category identifier to form the target interactive motion. This target interactive motion continues to participate in the generation of the currently generated gait data, enabling the currently generated gait data to correspond to the motion probability record, the increment boundary, and the motion continuity relationship in terms of the motion source of the target interactive motion.

[0051] Preferably, the weight parameters of the strategy processing network include field encoding weight parameters, temporal state weight parameters, action probability weight parameters, and action increment weight parameters. The field encoding weight parameters are configured within the observation field processing layer, and are used to determine the numerical contribution relationship of body joint angles, body base linear velocity, body base angular velocity, historical interaction actions, and gait category identifiers in subsequent state representations within the observation field sequence. The temporal state weight parameters are configured within the temporal state retention layer, and are used to determine the state retention ratio between the observation writing gating process, the history retention gating process, and the state fusion process, so that the temporal state data of the current control cycle can simultaneously include current observation fields and historical state fields. The action probability weight parameters are configured within the action probability generation layer, and are used to determine the mapping relationship between the temporal state data of the current control cycle and the action probability records, so that different body joint angles, body base linear velocities, body base angular velocities, and historical interaction actions can correspond to the generation probability of different target joint desired angle increments. The action increment weight parameters are configured within the action increment generation layer. These parameters determine the mapping between the boundary-checked action probability record and the target interaction action, ensuring that the target interaction action corresponds to the expected angle increment range of the target joint in the action space. All weight parameters are written into the policy processing network parameter record. This record establishes a correspondence between the weight parameters in each structural layer and the corresponding observation field processing, temporal state preservation processing, action probability generation processing, action boundary check processing, and action increment generation processing. Therefore, the weight parameters of the policy processing network are not abstract parameter names, but rather correspond to the data reading, state preservation, probability generation, and action generation processes of different structural layers within the network. Subsequent gradient updates adjust the weight parameters of the policy processing network according to the aforementioned correspondence between structural layers.

[0052] Preferably, the specific implementation process of the gradient update is as follows: After the action generation probability ratio and the action evaluation deviation value are formed, the field encoding weight parameter, temporal state weight parameter, action probability weight parameter, and action increment weight parameter in the policy processing network parameter record are read first, and the field encoding weight parameter, the temporal state weight parameter, the action probability weight parameter, the action increment weight parameter, the action generation probability ratio, the action evaluation deviation value, the currently generated gait data, and the standard gait feature data are written into the weight parameter update record. The weight parameter update record is used to include the weight parameter source, the action update amplitude source, the action evaluation deviation source, and the gait reference source into the same update caliber. Subsequently, based on the source of deviation of the motion evaluation deviation value in the currently generated gait data, the deviation contribution relationships corresponding to the field encoding weight parameter, the temporal state weight parameter, the motion probability weight parameter, and the motion increment weight parameter are determined. Specifically, when the deviation between the candidate joint angle data in the currently generated gait data and the target joint angle data in the standard gait feature data is large, the deviation contribution relationship corresponding to the motion increment weight parameter and the motion probability weight parameter is increased. When the deviation between the candidate mechanical spatial position data in the currently generated gait data and the mechanical spatial coordinates within the physical shape range in the standard gait feature data is large, the deviation contribution relationship corresponding to the field encoding weight parameter and the temporal state weight parameter is increased. When the determination result of the dynamic stability criterion indicates that a stable state is not met, the deviation contribution relationships corresponding to the temporal state weight parameter, the motion probability weight parameter, and the motion increment weight parameter all participate in subsequent updates. Subsequently, the update magnitude corresponding to the deviation contribution relationship is limited based on the action generation probability ratio processed by the pruning mechanism, and the update magnitudes of the field encoding weight parameter, temporal state weight parameter, action probability weight parameter, and action increment weight parameter are formed according to the limited update magnitude. Then, the field encoding weight parameter, temporal state weight parameter, action probability weight parameter, and action increment weight parameter are updated using the update magnitudes of these parameters respectively, so that the target interaction action generated by the policy processing network in the next iteration cycle can reduce the deviation from the standard gait feature data and reduce the tendency to generate actions that do not meet the dynamic stability criteria.

[0053] Preferably, after completing one gradient update, the policy processing network re-executes the observation field sorting process based on the updated field encoding weight parameters, re-executes the temporal state retention process based on the updated temporal state weight parameters, re-executes the action probability generation process based on the updated action probability weight parameters, and re-executes the action increment generation process based on the updated action increment weight parameters, so as to form new target interaction actions and new currently generated gait data in the next iteration cycle. Subsequently, the action generation probability ratio and action evaluation deviation value are recalculated based on the new currently generated gait data, and the weight parameters of the policy processing network are continued to be updated with gradients based on the new action evaluation deviation value. After repeating the above process for multiple iteration cycles, when the change amplitude of the action evaluation deviation value is lower than the preset convergence boundary in multiple consecutive iteration cycles, the action evaluation deviation value is determined to have converged. The preset convergence boundary is configured during the training configuration phase based on the control cycle density of the standard gait feature data, the target joint expected angle increment range in the action space, and the stability determination granularity of the dynamic stability standard. The preset convergence boundary is used to determine whether the weight parameters of the policy processing network have entered a stable update state. After convergence, the policy processing network retains the updated field encoding weight parameters, temporal state weight parameters, action probability weight parameters, and action increment weight parameters, and solidifies the action generation mapping relationship based on the updated field encoding weight parameters, temporal state weight parameters, action probability weight parameters, and action increment weight parameters. This enables the subsequently determined multi-band decoupled target control strategy to inherit the iterative constraint relationship between the partial observation information, the action space, the standard gait feature data, and the dynamic stability standard.

[0054] Preferably, the training process of the strategy processing network is executed after the standard gait feature data, the dynamic stability standard, the motion space configuration record, and the field sources of the partial observation information have all been formed. Before training begins, the standard gait feature record, unified time axis index, gait category identifier, skeleton acquisition point identifier, target joint angle data, and mechanical space coordinates within the physical body range are read from the standard gait feature data. Simultaneously, the target joint expected angle increment, increment boundary, motion continuity relationship, and joint identifier in the motion space configuration record are read to establish a training sample construction record. The training sample construction record is used to include the gait reference source, time axis source, motion boundary source, and joint identifier source into the same training data caliber, so that the subsequently formed training samples can be consistent with the partial observation information, motion space, and standard gait feature data used in the aforementioned iterative solution process. Subsequently, the target joint angle data corresponding to the same joint identifier in adjacent control cycles is read according to the unified time axis index, and the change in the target joint angle data of the next control cycle relative to the target joint angle data of the current control cycle is used to form the reference target joint expected angle increment. The technical essence of the reference target joint expected angle increment is that it is a joint increment reference extracted from the standard gait feature data. This reference is used to express the joint angle change that should occur in the next control cycle relative to the current control cycle under the standard gait temporal change relationship. Subsequently, the body joint angles, body base linear velocity, body base angular velocity, historical interaction actions, gait category identifier, and reference target joint expected angle increment of the current control cycle are written into the same training sample. This training sample is not a separate joint angle record, but rather organizes the current observable motion state, the action source of the previous control cycle, and the action reference of the next control cycle into a single trainable data unit. This allows the policy processing network to learn the action generation mapping relationship from the partial observation information to the target interaction action.

[0055] Preferably, the body joint angles in the training samples are converted from the target joint angle data corresponding to the current control cycle. The body base linear velocity in the training samples is formed based on the changes in trunk spatial position under adjacent unified time axis indices. The body base angular velocity in the training samples is formed based on the changes in trunk posture under adjacent unified time axis indices. The historical interaction actions in the training samples originate from the expected angle increment of the reference target joint corresponding to the previous control cycle. To enable the training samples to reflect the gait control process under partially observable conditions, fields that can be directly read in the current control cycle or deduced from adjacent control cycles are selected from the standard gait feature data to form partial observation information for training. The partial observation information for training has the same field structure as the aforementioned partial observation information, including body joint angles, body base linear velocity, body base angular velocity, historical interaction actions, unified time axis index, and gait category identifier. Subsequently, the partial observation information for training is used as the training input field, the expected angle increment of the reference target joint is used as the training target field, and the standard gait feature records under the same unified time axis index are used as the gait reference field. The gait reference field is used to constrain the currently generated gait data generated by the policy processing network during training, ensuring that the currently generated gait data and the standard gait feature data share the same time source and gait category source. Furthermore, the determination result of the dynamic stability criterion is written into the training samples as a stability determination field, so that each training sample not only contains motion increment references but also the source of the stable state between the predicted zero-moment point position and the zero-moment point boundary. Through the above training sample construction method, the training samples can simultaneously express gait reproduction requirements, motion continuity requirements, and support stability constraint requirements, providing a directly readable data source for subsequent training of the policy processing network with the above-mentioned layered structure.

[0056] Preferably, when training the policy processing network with the above-described layered structure, the training observation information is first input into the observation field organization layer, which forms a training observation field sequence according to the field arrangement criteria. The training observation field sequence is then input into the temporal state preservation layer, which combines this sequence with the training temporal state data preserved from the previous training sample to form the training temporal state data corresponding to the current training sample. The training temporal state data is then input into the action probability generation layer, which generates training action probability records based on the expected angle increment of the target joint in the action space. The training action probability records are then input into the action boundary verification layer, which performs action boundary verification based on the incremental boundary and the action continuity relationship to obtain a training boundary-verified action probability record. Finally, the training boundary-verified action probability record is input into the action increment generation layer, which generates training target interaction actions. Subsequently, the expected angle increment of the target joint in the training target interaction action is compared with the expected angle increment of the reference target joint in the training sample using the same joint identifier to form action increment difference data. The difference between the training current generated gait data generated based on the training target interaction action and the standard gait feature record in the gait reference field is calculated under the same unified time axis index to form gait reproduction difference data. The difference between the stability determination result corresponding to the training current generated gait data and the stability determination field in the training sample is calculated to form stability constraint difference data. The action increment difference data, the gait reproduction difference data, and the stability constraint difference data continue to participate in subsequent loss function calculations, ensuring that the training of the policy processing network is not only focused on action increment values ​​but also on gait temporal changes and dynamic stability criteria.

[0057] Preferably, the training of the policy processing network uses a combined loss function, which includes a motion increment loss function, a gait reproduction loss function, a stability constraint loss function, a motion continuation loss function, a boundary constraint loss function, and a probability update loss function. The motion increment loss function is formed based on the motion increment difference data and is used to constrain the expected angle increment of the target joint in the training target interaction motion to be close to the expected angle increment of the reference target joint. The gait reproduction loss function is formed based on the gait reproduction difference data, and it calculates the difference in similar data between candidate joint angle data and target joint angle data, and the difference in similar data between candidate mechanical spatial position data and mechanical spatial coordinates within the physical shape range; the gait reproduction loss function does not perform dimensional mixing comparisons of angle data and spatial position data. The stability constraint loss function is formed based on the stability constraint difference data and is used to increase the training penalty when the predicted zero moment point position corresponding to the currently generated gait data is outside the zero moment point boundary, and decrease the training penalty when the predicted zero moment point position corresponding to the currently generated gait data is inside the zero moment point boundary. The action continuity loss function is formed based on the incremental change between the target interaction action used in training and the historical interaction actions, and is used to constrain the target joint expected angle increment between adjacent control cycles to maintain a continuously readable change relationship. The boundary constraint loss function is formed based on whether the target interaction action used in training falls within the incremental boundary, and is used to limit the target joint expected angle increment that exceeds the physical limit of the joint rotation angle during training. The probability update loss function is formed based on the change relationship between the action generation probability in the current training round and the action generation probability in the previous training round, and is used to constrain the action probability generation layer to avoid excessive probability update amplitude during training. The above loss functions together form the combined loss function, so that the training objective of the policy processing network simultaneously covers action increment generation, gait temporal reproduction, support stability determination, action continuity, angle boundary constraint, and probability update amplitude constraint.

[0058] Preferably, the combined loss function is calculated using weighted summations based on training weight records during training. These training weight records are configured during the training configuration phase according to the gait category identifier, control cycle density, and stability determination granularity of the standard gait feature data. For training samples corresponding to walking gait, the training weight records ensure that the action continuity loss function and the gait reproduction loss function contribute significantly to the combined loss function, enabling the policy processing network to focus more on maintaining the continuity of joint changes and the temporal relationship of gait changes in constant-speed walking gait during training. For training samples corresponding to running gait, the training weight records ensure that the probability update loss function, the action continuity loss function, and the stability constraint loss function all contribute significantly, reducing the tendency for the policy processing network to perform excessively large action update amplitudes under denser control cycles. For training samples corresponding to climbing gait, the training weight records ensure that the stability constraint loss function and the gait reproduction loss function contribute significantly to the combined loss function, allowing the policy processing network to continue referencing the dynamic stability standard and the standard gait feature data when vertical support switching is involved. The training weight records do not alter the individual technical meanings of the action increment loss function, gait reproduction loss function, stability constraint loss function, action continuity loss function, boundary constraint loss function, and probability update loss function. Instead, they adjust the numerical contribution relationships of these loss functions in the combined loss function based on the differences in control cycle density, support switching methods, and joint change speeds among different gait categories. Therefore, the combined loss function can correspond to the different control requirements of walking gait, running gait, and climbing gait, rather than using the same training penalty caliber to handle all gait categories.

[0059] Preferably, when training the policy processing network based on the combined loss function, the field encoding weight parameters, temporal state weight parameters, action probability weight parameters, and action increment weight parameters in the policy processing network parameter record are first read, and the combined loss function, the training target interaction action, the training currently generated gait data, the reference target joint expected angle increment, the gait reference field, and the stability determination field are written into the training gradient update record. Subsequently, the training update amounts corresponding to the field encoding weight parameters, temporal state weight parameters, action probability weight parameters, and action increment weight parameters are calculated according to the combined loss function. The training update amount corresponding to the field encoding weight parameters mainly comes from the gait reproduction loss function and the action increment loss function, and is used to adjust the numerical contribution relationship of the observation field processing layer to the body joint angle, body base linear velocity, body base angular velocity, historical interaction action, and gait category identifier. The training update amount corresponding to the temporal state weight parameters mainly comes from the action continuation loss function, the gait reproduction loss function, and the stability constraint loss function, and is used to adjust the retention ratio of the temporal state retention layer for the current observation field and the historical state field. The training update amount corresponding to the action probability weight parameter mainly comes from the action increment loss function, the probability update loss function, and the stability constraint loss function, and is used to adjust the probability distribution of the action probability generation layer in generating the expected angle increment of the target joint under different training time-series state data. The training update amount corresponding to the action increment weight parameter mainly comes from the action increment loss function, the boundary constraint loss function, and the action continuation loss function, and is used to adjust the generation method of the action increment generation layer for the target interaction action. Then, the field encoding weight parameter, the time-series state weight parameter, the action probability weight parameter, and the action increment weight parameter are updated according to the above training update amount, and the updated weight parameters are rewritten into the policy processing network parameter record, so that the next training round can continue to perform observation field sorting processing, time-series state preservation processing, action probability generation processing, action boundary verification processing, and action increment generation processing using the updated weight parameters.

[0060] Preferably, the following steps are repeated across multiple training rounds: training sample reading, training target interaction action generation, training current gait data generation, combined loss function calculation, and weight parameter update processing, until the change in the combined loss function is lower than the training convergence boundary. The training convergence boundary is configured during the training configuration phase based on the control cycle density of the standard gait feature data, the expected angle increment range of the target joint in the action space, the stability determination granularity of the dynamic stability standard, and the number of training samples. The training convergence boundary is used to determine whether the training state of the policy processing network has entered a stable change range. After training, the updated field encoding weight parameters, temporal state weight parameters, action probability weight parameters, and action increment weight parameters are read from the policy processing network parameter record, and written into the training completion parameter record. The training completion parameter record continues to be used in subsequent iterative solutions, enabling the policy processing network to generate target interaction actions according to the field organization method, temporal state retention method, action probability generation method, and action increment generation method formed during the actual iterative solution. Therefore, the training process of the policy processing network maintains the same technical source relationship with the aforementioned interaction level, action space, standard gait feature data and dynamic stability standard, so that the policy processing network after training can directly participate in the subsequent generation of current generated gait data, calculation of action generation probability ratio, calculation of action evaluation deviation value, and determination of multi-band decoupled target control strategy.

[0061] Optionally, the target control strategy for multi-band decoupling possesses multi-band decoupling execution capability through the following means: Obtain the gait temporal mapping relationship corresponding to the standard gait feature data; A time scaling adjustment parameter is configured within the target control strategy. The time scaling adjustment parameter represents the scaling ratio of the execution cycle when the gait timing mapping relationship is mapped to the gait control sequence. Based on the time scaling adjustment parameter, the execution cycle in the gait timing mapping relationship is scaled to generate gait control sequences with different timing cycles; When the time scaling adjustment parameter is adjusted, the target control strategy maps the same standard gait feature data to gait control sequences with different time periods, so that the centaur robot can generate corresponding gaits at different time execution frequencies without re-collecting the multimodal motion time trajectory of the target heterogeneous organism.

[0062] Preferably, the specific implementation process of the step "obtaining the gait temporal mapping relationship corresponding to the standard gait feature data" is as follows: After the multi-band decoupled target control strategy has been determined by the converged strategy processing network, the standard gait feature record, unified time axis index, gait category identifier, skeleton acquisition point identifier, target joint angle data, and mechanical space coordinates within the physical body range in the standard gait feature data are read first. The standard gait feature record, the unified time axis index, the gait category identifier, the skeleton acquisition point identifier, the target joint angle data, and the mechanical space coordinates within the physical body range are then written into the gait temporal mapping relationship generation record. The gait temporal mapping relationship generation record is used to incorporate the time source, gait category source, joint angle source, and mechanical space position source in the standard gait feature data into the same reading caliber, so that the subsequently formed gait temporal mapping relationship can be traced back to the standard gait input basis that has already undergone denoising and temporal smoothing processing. Subsequently, according to the order of the unified timeline index, multiple standard gait feature records under the same gait category identifier are temporally arranged to form intra-category gait time sequence segments corresponding to walking gait, running gait, and climbing gait, respectively. These intra-category gait time sequence segments are not arbitrary segments of the standard gait feature data, but rather temporally correlated with changes in target joint angle data, mechanical spatial coordinate changes within the physical body area, and the correspondence of skeletal acquisition point identifiers within adjacent control cycles under the same gait category identifier. This ensures that each intra-category gait time sequence segment can express the local movement rhythm within the corresponding gait category. Then, based on the connection position of the unified timeline index between adjacent intra-category gait time sequence segments, the category continuity relationships between walking gait and running gait, and between running gait and climbing gait, are extracted and written into the gait time sequence mapping relationship generation record. These category continuity relationships are used to express the connection boundaries between different gait categories on the timeline, preventing the subsequent scaling of execution cycles from misinterpreting different gait categories as the same continuous gait cycle. Furthermore, the gait time sequence segments within the category and the category continuation relationship are jointly encapsulated into the gait time sequence mapping relationship. The technical essence of this gait time sequence mapping relationship is to reorganize the standard gait feature records in the standard gait feature data according to a unified time axis index and gait category identifier into a time sequence correspondence that can be read by the multi-band decoupled target control strategy. This gait time sequence mapping relationship not only preserves the gait time sequence change relationship in the standard gait feature data but also provides a reference source for subsequent configuration of time scaling adjustment parameters, enabling the time scaling adjustment parameters to be configured based on the gait time sequence segments within the category and the category continuation relationship in the gait time sequence mapping relationship.

[0063] Preferably, in the specific implementation of the step "configuring time scaling adjustment parameters within the target control strategy, wherein the time scaling adjustment parameters characterize the execution cycle scaling ratio when mapping the gait temporal mapping relationship to a gait control sequence", after the gait temporal mapping relationship has been formed, the action generation mapping relationship, the gait temporal mapping relationship, the dynamic stability standard, and the target joint expected angle increment in the action space of the multi-band decoupled target control strategy are read first, and a time scaling configuration record is established. The time scaling configuration record is used to include the action generation source of the multi-band decoupled target control strategy, the time source of the gait temporal mapping relationship, the stability constraint source of the dynamic stability standard, and the action boundary source of the target joint expected angle increment into the same configuration caliber, so that the subsequently configured time scaling adjustment parameters do not deviate from the action generation mapping relationship already formed by the multi-band decoupled target control strategy. Subsequently, based on the interval relationship between adjacent unified time axis indices in the gait temporal mapping relationship, the basic execution cycle corresponding to each gait category is extracted; the basic execution cycle is used to express the original control cycle rhythm of the standard gait feature data when no time scaling processing is performed. Subsequently, based on the determination results of the dynamic stability criterion, a scalable range is configured for the basic execution cycle. This scalable range is used to limit the basic execution cycle to maintain the same control caliber as the target joint desired angle increment range, the zero torque point boundary, and the motion generation mapping relationship, even when shortened or extended. The scalable range is not an arbitrarily set time range, but rather a range of execution cycle variations jointly defined by the basic execution cycle, the target joint desired angle increment, and the dynamic stability criterion. For example, when the control cycle density corresponding to the running gait category is high, the scalable range ensures that shortening the execution cycle will not exceed the range within which the target joint desired angle increment can be continuously connected. When there are many vertical support switches corresponding to the climbing gait category, the scalable range ensures that extending the execution cycle retains the time interval required for foot support changes. Subsequently, the time scaling adjustment parameter is configured within the multi-band decoupled target control strategy, and a correspondence is established between the time scaling adjustment parameter and the basic execution cycle, the scalable range, the gait category identifier, and the unified time axis index. The technical essence of the time scaling adjustment parameter is that it is a data parameter used to describe the scaling ratio of the execution cycle when the gait time sequence mapping relationship is mapped to the gait control sequence. The time scaling adjustment parameter continues to participate in the subsequent execution cycle scaling process, so that the same standard gait feature data can form corresponding gait control sequences according to different execution cycles within the multi-band decoupled target control strategy, and the gait control sequence can continue to use the action generation mapping relationship and the dynamic stability standard.

[0064] Preferably, the specific implementation process of the step "scaling the execution cycle in the gait timing mapping relationship according to the time scaling adjustment parameter to generate gait control sequences with different timing cycles" is as follows: After the time scaling adjustment parameter is configured, the gait timing mapping relationship, the time scaling adjustment parameter, the basic execution cycle, the scalable range, the gait category identifier, and the unified time axis index are read from the time scaling configuration record, and a cycle scaling processing record is established. The cycle scaling processing record is used to include the time source to be scaled, the scaling ratio source, the scaling boundary source, and the gait category source into the same processing caliber, so that the subsequently generated gait control sequence can continue to trace back to the standard gait feature data. Subsequently, the basic execution cycle in the gait timing mapping relationship is shortened or lengthened according to the time scaling adjustment parameter to form a scaled execution cycle. The scaled execution cycle is used to express the control cycle interval of gait time sequence segments within the same category at different execution frequencies. The shortening process allows the target joint angle data and mechanical space coordinates within the physical body range corresponding to adjacent unified time axis indices to be read at a more intensive control rhythm, while the lengthening process allows the target joint angle data and mechanical space coordinates within the physical body range corresponding to adjacent unified time axis indices to be read at a more moderate control rhythm. Then, the standard gait feature records in the gait time sequence segments within the category are rearranged according to the scaled execution cycle to form scaled gait time sequence segments. The scaled gait time sequence segments still retain the corresponding gait category identifier, skeleton acquisition point identifier, target joint angle data, and mechanical space coordinates within the physical body range. Furthermore, the scaled gait time sequence segments do not change the sequence of actions in the standard gait feature data; they only change the execution cycle interval when adjacent actions are read. Subsequently, the scaled gait time sequence fragment is input into the action generation mapping relationship in the multi-band decoupled target control strategy. The action generation mapping relationship generates a corresponding target interactive action sequence based on the scaled gait time sequence fragment, and the target interactive action sequence is associated and encapsulated with the scaled execution cycle to form gait control sequences with different time cycles. The gait control sequence includes the scaled execution cycle, gait category identifier, target interactive action sequence, and target joint expected angle increment. The technical essence of the gait control sequence is to convert the gait time sequence mapping relationship into a time-series action sequence that can be read by subsequent joint drive command generation, without changing the gait time sequence change relationship of the standard gait feature data. Through the above scaling process, the multi-band decoupled target control strategy can form gait control sequences with different time cycles between walking gait, running gait, and climbing gait, and the gait control sequences with different time cycles continue to be constrained by the scalable range and the dynamic stability standard.

[0065] Preferably, the specific implementation process of step "when adjusting the time scaling adjustment parameter, the target control strategy maps the same standard gait feature data to gait control sequences with different time periods" is as follows: When it is necessary to change the gait execution frequency, first read the time scaling adjustment parameter, the gait time mapping relationship, the standard gait feature data, and the period scaling processing record already configured within the multi-band decoupled target control strategy, and establish an execution frequency switching record. The execution frequency switching record is used to write the time scaling adjustment parameter before adjustment, the time scaling adjustment parameter after adjustment, the gait time mapping relationship, and the standard gait feature data into the same data caliber, so that the gait control sequences with different time periods formed subsequently can correspond to the same standard gait input basis. Subsequently, the time scaling adjustment parameter is updated, and the adjusted time scaling adjustment parameter is checked against the scalable range. When the adjusted time scaling adjustment parameter is within the scalable range, the basic execution cycle in the gait timing mapping relationship is scaled according to the adjusted time scaling adjustment parameter. When the adjusted time scaling adjustment parameter exceeds the scalable range, the adjusted time scaling adjustment parameter is boundary-limited according to the scalable range, ensuring that the adjusted time scaling adjustment parameter maintains a consistent data caliber with the target joint expected angle increment, the motion generation mapping relationship, and the dynamic stability standard. Afterward, the multi-band decoupled target control strategy reads the same standard gait feature data and uses the same intra-category gait timing segment and category continuation relationship in the same gait timing mapping relationship, changing the scaled execution cycle only according to the adjusted time scaling adjustment parameter. Since the standard gait feature data was neither re-acquired nor replaced with new motion samples, the multi-band decoupled target control strategy maintains consistency in the sources of target joint angle data, mechanical spatial coordinates within the physical body range, gait category identifiers, and skeletal acquisition point identifiers when adjusting the time scaling adjustment parameter, only changing the time interval between the reading and mapping of adjacent standard gait feature records. Subsequently, the gait control sequences with different time periods are regenerated based on the adjusted time scaling adjustment parameter, and the gait control sequences with different time periods are written into the execution frequency switching record. The gait control sequences with different time periods can correspond to different time state execution frequencies such as normal walking, accelerated running, or slow climbing; the normal walking, accelerated running, or slow climbing are all formed by changing the scaled execution period through the time scaling adjustment parameter, and the gait control sequences with different time periods all originate from the same standard gait feature data and the same motion generation mapping relationship.Through this process, the multi-band decoupled target control strategy can avoid re-establishing independent control templates for different time-state execution frequencies, and also avoid the problem of inconsistent parameter calibers introduced by switching different control templates. The multi-band decoupled execution capability is reflected in the separable processing between the adjustment of the time scaling parameter on the scaled execution cycle and the generation of the gait control sequence of different time-sequence periods by the action generation mapping relationship.

[0066] Preferably, in the specific technical implementation of the step "enabling the centaur robot to generate corresponding gaits at different time-mode execution frequencies without re-collecting the multimodal motion temporal trajectory of the target heterogeneous organism," after the multi-band decoupled target control strategy has formed gait control sequences with different time-mode periods according to the adjusted time-scaling adjustment parameters, the gait control sequences with different time-mode periods, the gait category identifier, the category succession relationship, the dynamic stability standard, and the standard gait feature data are first read, and a corresponding gait generation record is established. The corresponding gait generation record is used to include the source of the gait control sequence, the source of the gait category, the source of the category succession, the source of the stability constraint, and the source of the standard gait feature data into the same generation caliber, so that when generating corresponding gaits at different time-mode execution frequencies, the preceding data source of the same multimodal motion temporal trajectory can be traced back. Subsequently, the multi-band decoupled target control strategy reads gait control sequences with different time periods according to the gait category identifier, and determines the switching position between adjacent gait categories based on the category succession relationship. When reading the gait control sequence corresponding to walking gait, the multi-band decoupled target control strategy generates a target interaction action sequence corresponding to walking gait according to a relatively smooth scaling execution cycle. When reading the gait control sequence corresponding to running gait, the multi-band decoupled target control strategy generates a target interaction action sequence corresponding to running gait according to a relatively dense scaling execution cycle. When reading the gait control sequence corresponding to climbing gait, the multi-band decoupled target control strategy generates a target interaction action sequence corresponding to climbing gait according to a scaling execution cycle adapted to the vertical support switching. Afterward, the multi-band decoupled target control strategy correlates and verifies the target interaction action sequence under the corresponding gait category with the dynamic stability standard, so that the stability state determination result between the predicted zero moment point position and the zero moment point boundary continues to participate in the action generation process under different time execution frequencies. The target interaction sequence, after correlation verification, serves as the source for generating the corresponding gait and continues to participate in the generation of subsequent joint drive commands. This allows the joint drive commands to drive the Centaur robot to execute the corresponding gait at different time-mode execution frequencies. Thus, without re-collecting the multimodal motion temporal trajectory of the target heterogeneous organism, the Centaur robot can generate gait control sequences with different time-series periods based on the same standard gait feature data, the same gait temporal mapping relationship, and different time-scaling adjustment parameters, and generate corresponding gaits at different time-mode execution frequencies. This processing ensures that the execution frequency change of multimodal gait switching is mainly adjusted by the time-scaling adjustment parameter on the scaled execution period, rather than by switching between multiple independent motion control templates. This allows the multi-band decoupled target control strategy to maintain a unified standard gait input basis and a unified source of stable constraints in complex work scenarios.

[0067] Optionally, generating joint drive commands according to the target control strategy to drive the joint motors of the centaur robot and control the centaur robot to perform multimodal gait switching includes: When the centaur robot responds to the first rated load task, it obtains the normal speed walking gait control sequence corresponding to the standard gait feature data from the target control strategy, generates a normal speed walking gait drive command as the joint drive command, regulates the joint motor to generate drive with the first task power, and executes the normal speed walking gait. When the centaur robot responds to the second rated load task, the data processing structure of the target control strategy remains unchanged, and the time scaling adjustment parameter is adjusted. Based on the adjusted time scaling parameters, the execution cycle in the gait timing mapping relationship is scaled to generate an accelerated running gait control sequence. The accelerated running gait control sequence generates an accelerated running gait drive command as the joint drive command, and the joint motor is controlled to generate drive at the second task power to execute the accelerated running gait.

[0068] Preferably, the specific implementation process of step "obtaining the constant-speed walking gait control sequence corresponding to the standard gait feature data from the target control strategy when the Centaur robot responds to the first rated load task" is as follows: After the multi-band decoupled target control strategy has been formed based on the standard gait feature data, the gait timing mapping relationship, and the time scaling adjustment parameters, the current task trigger information is first read, and the first rated load task identifier, first task power, target walking execution frequency, and current unified time axis index corresponding to the first rated load task are identified from the current task trigger information. The current task trigger information is used to express the task trigger source when the Centaur robot enters the first rated load task, so that the first rated load task identifier, the first task power, the target walking execution frequency, and the current unified time axis index can enter the same task reading caliber. The first rated load task identifier originates from the rated load task configuration record generated before the centaur robot executes the task. This record, during the task configuration phase, includes the task load level, gait category requirement, task power reading caliber, and gait execution frequency reading caliber. Specifically, the task load level distinguishes the load requirement source between the first rated load task and the subsequent second rated load task; the gait category requirement maps the first rated load task to a normal-speed walking gait; the task power reading caliber reads the first task power; and the gait execution frequency reading caliber reads the target walking execution frequency, ensuring that the first rated load task identifier corresponds to the required task load level, first task power, and target walking execution frequency for a normal-speed walking gait. Subsequently, the first rated load task identifier, the first task power, the target walking execution frequency, the standard gait characteristic data, the gait timing mapping relationship, and the multi-band decoupling target control strategy are written into the normal-speed walking gait control sequence acquisition record. The constant-speed walking gait control sequence acquisition record is used to incorporate the task source, power source, timing source, and strategy source into the same reading caliber. This ensures that the subsequently acquired constant-speed walking gait control sequences are not isolated, fixed templates, but rather traceable to the motion generation mapping relationship already formed within the multi-band decoupled target control strategy. Then, based on the first rated load task identifier, a gait timing segment corresponding to the walking gait is selected from the gait timing mapping relationship. Using the unified time axis index, gait category identifier, target joint angle data, and mechanical space coordinates within the physical body range of the gait timing segment, a constant-speed walking gait timing reading segment is formed.The constant-speed walking gait timing read segment continues to input the action generation mapping relationship in the multi-band decoupled target control strategy. The action generation mapping relationship reads the corresponding scaled execution cycle according to the target walking execution frequency, and the constant-speed walking gait timing read segment is sequentially mapped according to the scaled execution cycle to form a constant-speed walking gait target interactive action sequence. This sequence is associated and encapsulated with the first task power, the scaled execution cycle, and the gait category identifier to form the constant-speed walking gait control sequence. The constant-speed walking gait control sequence includes the constant-speed walking gait target interactive action sequence, the scaled execution cycle, the first task power, the gait category identifier, a unified time axis index, and the target joint expected angle increment. The technical essence of this constant-speed walking gait control sequence is to map the temporal change relationship corresponding to the walking gait in the standard gait feature data into constant-speed walking control data that can be continuously read by the joint motors. This allows the Centaur robot to use the same standard gait input basis when responding to the first rated load task, without needing to switch to another independent motion control template.

[0069] Preferably, in the specific technical implementation of the step "generating constant speed walking gait drive instructions as joint drive instructions, regulating the joint motors to generate drive at the first task power, and executing constant speed walking gait", after the constant speed walking gait control sequence is formed, the constant speed walking gait target interaction action sequence, target joint expected angle increment, scaling execution cycle, first task power, unified time axis index, and gait category identifier in the constant speed walking gait control sequence are read first, and the joint identifier, joint angle physical limit, and joint motor drive aperture of the Centaur robot are read simultaneously to establish a constant speed walking gait drive instruction generation record. The joint motor drive specifications are written when the Centaur robot completes joint calibration and power calibration. These specifications include the correspondence between joint identifiers and joint motors, the target angle reading format for the joint motors, the power reading format for the joint motors, and the execution cycle reading format for the joint motors. Specifically, the correspondence between joint identifiers and joint motors ensures that the desired angle increment of the target joint in the normal-speed walking gait target interaction sequence matches the corresponding joint motor. The target angle reading format for the joint motors carries the subsequently formed target joint angle for the normal-speed walking gait. The power reading format for the joint motors carries the first task power, and the execution cycle reading format for the joint motors carries the scaled execution cycle. The joint motor drive specifications enable the normal-speed walking gait control sequence to be converted into drive data readable by the joint motors, and ensure that the drive data maintains the same instruction reading relationship with the joint identifiers, the first task power, and the scaled execution cycle. Subsequently, based on the unified timeline index, the expected angle increments of the target joints in the normal-speed walking gait target interaction action sequence are read sequentially. Each expected angle increment is then incrementally superimposed with the corresponding joint motor's body joint angle in the current control cycle to form the normal-speed walking gait target joint angle. This normal-speed walking gait target joint angle undergoes angle boundary verification by the joint rotation physical limit to ensure it remains within the executable angle range of the corresponding joint motor. Afterward, the normal-speed walking gait target joint angle that has passed the angle boundary verification is encapsulated with the scaled execution cycle, the first task power, and the joint identifier in an instruction field to form the normal-speed walking gait drive instruction. When the normal-speed walking gait drive instruction serves as the joint drive instruction, it does not only contain the target angle value but also includes the normal-speed walking gait target joint angle, the scaled execution cycle, the first task power, the joint identifier, and the gait category identifier, enabling each joint motor to read the corresponding angle drive content and power drive content according to the same timeline.Subsequently, the joint motors read the corresponding target joint angles for the normal-speed walking gait based on the joint identifiers in the joint drive commands, and execute drive outputs on the target joint angles for the normal-speed walking gait according to the first task power. This causes the centaur robot's torso structure and quadrupedal support process to execute the normal-speed walking gait in the sequential order of the target interactive action sequence. Through the above processing, the normal-speed walking gait drive commands can connect the first rated load task, the normal-speed walking gait control sequence, the first task power, and the execution actions of the joint motors, allowing the execution source of the normal-speed walking gait to be traced from the standard gait feature data all the way to the final joint drive command.

[0070] Preferably, the specific implementation process of step "keeping the data processing structure of the target control strategy unchanged and adjusting the time scaling adjustment parameter when the centaur robot responds to the second rated load task" is as follows: When the centaur robot switches from the first rated load task to the second rated load task, first read the second rated load task identifier, second task power, target running execution frequency and current unified time axis index corresponding to the second rated load task, and write the second rated load task identifier, second task power, target running execution frequency, multi-band decoupled target control strategy, action generation mapping relationship, gait timing mapping relationship and time scaling adjustment parameter into the time scaling adjustment parameter adjustment record. The second rated load task identifier also originates from the rated load task configuration record. The task load level in the rated load task configuration record is used to distinguish the load requirement difference between the second rated load task and the first rated load task. The gait category requirement in the rated load task configuration record is used to map the second rated load task to a running gait category. The task power reading caliber in the rated load task configuration record is used to read the power of the second task. The gait execution frequency reading caliber in the rated load task configuration record is used to read the target running execution frequency. The second rated load task identifier is used to correspond to a running gait category with a higher gait execution frequency requirement than the first rated load task. The second task power is used to correspond to the power reading caliber used when the joint motor is driven under the second rated load task. Subsequently, the data processing structure of the multi-band decoupled target control strategy remains unchanged. This means that the action generation mapping relationship is not replaced, the interaction hierarchy is not reconstructed, the standard gait feature data is not regenerated, and the multimodal motion temporal trajectory of the target heterologous organism is not re-acquired. Instead, the action generation mapping relationship, the gait temporal mapping relationship, and the dynamic stability standard already formed within the multi-band decoupled target control strategy are continued to be used. Then, based on the target running execution frequency, the basic execution cycle corresponding to the running gait in the gait temporal mapping relationship is read, and the target running execution frequency and the basic execution cycle corresponding to the running gait are subjected to period scaling matching processing to form candidate time scaling adjustment parameters. The candidate time scaling adjustment parameters are further boundary-checked with the scalable range, which is derived from the execution cycle variation range configured previously based on the basic execution cycle corresponding to the running gait, the target joint expected angle increment, and the dynamic stability standard.When the aforementioned candidate time-scaling adjustment parameter is within the scalable range, it is determined as the adjusted time-scaling adjustment parameter. When the aforementioned candidate time-scaling adjustment parameter exceeds the scalable range, the candidate time-scaling adjustment parameter is boundary-limited according to the scalable range to form the adjusted time-scaling adjustment parameter. The adjusted time-scaling adjustment parameter is then written into the time-scaling adjustment parameter adjustment record and a correspondence is established with the second rated load task identifier, the second task power, the target running execution frequency, and the basic execution cycle corresponding to the running gait. Thus, the second rated load task causes the adjustment of the time-scaling adjustment parameter, rather than the replacement of the data processing structure of the multi-band decoupled target control strategy, ensuring that the subsequent generation of the accelerated running gait control sequence still uses the same standard gait input basis and action generation mapping relationship as the normal speed walking gait control sequence.

[0071] Preferably, in the specific technical implementation of the step "scaling the execution cycle in the gait timing mapping relationship according to the adjusted time scaling adjustment parameters to generate an accelerated running gait control sequence", after the adjusted time scaling adjustment parameters have been formed, the adjusted time scaling adjustment parameters, the gait timing mapping relationship, the second rated load task identifier, the second task power, the target running execution frequency, the basic execution cycle corresponding to the running gait, and the scalable range are first read from the time scaling adjustment parameter adjustment record, and an accelerated running gait control sequence generation record is established. The accelerated running gait control sequence generation record is used to include the second rated load task source, time scaling source, running gait timing source, power source, and scaling boundary source into the same generation caliber, so that the subsequently generated accelerated running gait control sequence can maintain a direct correspondence with the second rated load task and the multi-band decoupled target control strategy. Subsequently, based on the second rated load task identifier, a gait time sequence segment corresponding to the running gait is selected from the gait time sequence mapping relationship. The standard gait feature records, unified time axis indexes, target joint angle data, mechanical space coordinates within the physical body area, and skeletal acquisition point identifiers are read from the gait time sequence segment to form a running gait time sequence reading segment. Then, the basic execution cycle corresponding to the running gait is shortened according to the adjusted time scaling parameter to form a scaled execution cycle for the running gait. The scaled execution cycle for the running gait is used to express the control cycle interval when adjacent standard gait feature records are read and mapped under the second rated load task. The scaled execution cycle for the running gait does not change the order of the target joint angle data and mechanical space coordinates within the physical body area in the running gait time sequence reading segment, but changes the reading rhythm between adjacent unified time axis indices. Subsequently, the running gait time sequence reading segment is time-rearranged according to the scaled execution cycle for the running gait, forming an accelerated running gait time sequence segment. The accelerated running gait timing segment continues to input the action generation mapping relationship in the multi-band decoupled target control strategy. Based on the accelerated running gait timing segment, the action generation mapping relationship generates an accelerated running gait target interaction action sequence. This accelerated running gait target interaction action sequence is then associated and encapsulated with the scaled execution cycle of the running gait, the second task power, and the gait category identifier to form the accelerated running gait control sequence. The accelerated running gait control sequence includes the accelerated running gait target interaction action sequence, the scaled execution cycle of the running gait, the second task power, the gait category identifier, a unified time axis index, and the target joint desired angle increment.The accelerated running gait control sequence continues to participate in the generation of subsequent accelerated running gait drive commands, enabling the accelerated running gait target interactive action sequence, the execution cycle of the running gait after scaling, the second task power, and the target joint expected angle increment to be converted into command fields readable by the joint motor. Through the above processing, the accelerated running gait control sequence is not generated from new motion samples or new control templates, but rather generated under the same gait timing mapping relationship and the same action generation mapping relationship, by changing the execution cycle corresponding to the running gait through the adjusted time scaling parameter, thus maintaining a consistent data source relationship between the accelerated running gait control sequence and the normal speed walking gait control sequence.

[0072] Preferably, the specific implementation process of the step "generating an accelerated running gait drive command as the joint drive command based on the accelerated running gait control sequence, regulating the joint motor to generate drive with the second task power, and executing the accelerated running gait" is as follows: After the accelerated running gait control sequence is formed, the accelerated running gait target interactive action sequence, target joint desired angle increment, running gait scaling execution cycle, second task power, unified time axis index, and gait category identifier in the accelerated running gait control sequence are read first, and the joint identifier, the joint angle physical limit, and the joint motor drive aperture are read simultaneously to establish an accelerated running gait drive command generation record. The accelerated running gait drive command generation record is used to incorporate the accelerated running gait control sequence source, the second task power source, the joint angle boundary source, and the joint motor reading aperture into the same command generation aperture, so that the subsequently formed accelerated running gait drive command can correspond to the second rated load task. Subsequently, the expected angle increments of the target joints in the target interactive action sequence of the accelerated running gait are read sequentially according to the unified time axis index. These expected angle increments are then incrementally superimposed with the body joint angle of the corresponding joint motor in the current control cycle to form the target joint angle for the accelerated running gait. The target joint angle for the accelerated running gait is further checked against the physical limit of the joint angle. After passing the angle boundary check, it is encapsulated with the execution cycle after scaling the running gait, the second task power, and the joint identifier in an instruction field to form the accelerated running gait drive instruction. When the accelerated running gait drive instruction serves as the joint drive instruction, it includes the target joint angle for the accelerated running gait, the execution cycle after scaling the running gait, the second task power, the joint identifier, and the gait category identifier. This allows the joint motor to read the corresponding angle drive content according to a more frequent execution cycle after scaling the running gait and generate the drive according to the second task power. Subsequently, the joint motors read the corresponding target joint angles for the accelerated running gait based on the joint identifiers in the joint drive commands, and drive the corresponding joint motors according to the second task power, causing the centaur robot to execute the accelerated running gait according to the sequence of the target interactive actions. During the execution of the accelerated running gait, the accelerated running gait drive commands continue to be correlated and verified with the dynamic stability standard, so that the stability determination result between the predicted zero-moment point position and the zero-moment point boundary can participate in the action adjustment of subsequent control cycles. The action adjustment of subsequent control cycles continues to act in reverse on the generation process of the target interactive actions, so that the accelerated running gait drive commands can still use the action generation mapping relationship in the multi-band decoupled target control strategy and the dynamic stability standard in subsequent control cycles.Therefore, both the normal walking gait drive command and the accelerated running gait drive command are generated by the multi-band decoupled target control strategy. The two are consistent in terms of standard gait input basis, action generation mapping relationship and source of stability constraints. They only form drive differences under different gait execution frequencies through the time scaling adjustment parameter, the scaled execution cycle and the corresponding task power. This enables the Centaur robot to reduce the sudden changes in joint drive response caused by the switching of independent control templates when switching between multimodal gait modes.

[0073] like Figure 3 As shown in the figure, a multimodal gait generation and control device for a centaur robot based on imitation learning is provided in an embodiment of this application, comprising: The trajectory acquisition and mapping module is used to acquire the multimodal motion temporal trajectory of the target heterogeneous organism, and to redirect and map the multimodal motion temporal trajectory according to the physical kinematic constraints of the Centaur robot to generate the initial control trajectory of the Centaur robot. The gait feature generation module is used to input the initial control trajectory into the gait generation network, and to first denoise the initial control trajectory in the gait generation network, and then perform temporal smoothing on the denoising result to generate standard gait feature data. A stable interaction construction module is used to construct a dynamic stability standard for constraining the vertical axis position of the centaur robot's center of mass based on the variable axis inverted pendulum mechanism, and introduces a partially observable Markov mechanism to define the gait control process as an interaction hierarchy that includes the environmental state space and the action space. The control strategy generation module is used to generate a multi-band decoupled target control strategy by iteratively solving the standard gait feature data based on the near-end strategy optimization mechanism and the dynamic stability criterion within the interaction level. The joint drive control module is used to generate joint drive commands according to the target control strategy to drive the joint motors of the centaur robot and control the centaur robot to perform multimodal gait switching.

[0074] like Figure 4 As shown, an electronic device according to an embodiment of this application includes a memory and a processor; The memory is used to store computer programs; When the processor is used to execute the computer program, it implements the method described in any one of the claims of this application.

[0075] like Figure 5 As shown, a computer-readable storage medium according to an embodiment of this application stores a computer program, which, when executed by a processor, implements the method described in any one of the claims of this application.

[0076] like Figure 6 As shown, this is an embodiment of a centaur robot according to this application, which includes a torso structure, joint motors, sensor acquisition components and a gait controller; The sensing and acquisition component is used to collect motion state data of the centaur robot; The gait controller is used to acquire the multimodal motion time-series trajectory of the target heterogeneous organism, and to redirect and map the multimodal motion time-series trajectory according to the physical kinematic constraints of the centaur robot to generate the initial control trajectory of the centaur robot. The gait controller is also used to input the initial control trajectory into the gait generation network, and to first denoise the initial control trajectory in the gait generation network, and then to perform temporal smoothing on the denoising result to generate standard gait feature data. The gait controller is also used to construct a dynamic stability standard for constraining the vertical axial position of the centaur robot's center of mass based on the variable axis inverted pendulum mechanism, and introduces a partially observable Markov mechanism to define the gait control process as an interactive hierarchy that includes the environmental state space and the action space. The gait controller is also used to iteratively solve the standard gait feature data based on the proximal strategy optimization mechanism and the dynamic stability criterion within the interaction level to generate a multi-band decoupled target control strategy. The gait controller is further configured to generate joint drive commands according to the target control strategy, and send the joint drive commands to the joint motor; The joint motor is used to drive the trunk structure to perform multimodal gait switching according to the joint drive command.

[0077] like Figure 7 As shown in the figure, this is an embodiment of a multimodal gait generation and control system for a centaur robot based on imitation learning, which includes a motion trajectory acquisition end, a gait control end, and a centaur robot execution end. The motion trajectory acquisition terminal is used to acquire the multimodal motion temporal trajectory of the target heterologous organism and send the multimodal motion temporal trajectory to the gait control terminal; The gait control terminal is used to redirect and map the multimodal motion temporal trajectory according to the physical kinematic constraints of the Centaur robot, thereby generating the initial control trajectory of the Centaur robot. The gait control terminal is also used to input the initial control trajectory into the gait generation network, and to first denoise the initial control trajectory in the gait generation network, and then to perform temporal smoothing on the denoising result to generate standard gait feature data. The gait control terminal is also used to construct a dynamic stability standard for constraining the vertical axial position of the centaur robot's center of mass based on the variable axis inverted pendulum mechanism, and to introduce a partially observable Markov mechanism to define the gait control process as an interactive hierarchy that includes the environmental state space and the action space. The gait control terminal is also used to iteratively solve the standard gait feature data based on the proximal strategy optimization mechanism and the dynamic stability criterion within the interaction layer, thereby generating a multi-band decoupled target control strategy. The gait control terminal is also used to generate joint drive commands according to the target control strategy, and send the joint drive commands to the centaur robot execution terminal; The centaur robot actuator is used to drive the joint motors according to the joint drive commands, so as to control the centaur robot to perform multimodal gait switching.

[0078] 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 generating and controlling multimodal gait of a centaur robot based on imitation learning, characterized in that, Includes the following steps: The multimodal motion time-series trajectory of the target heterogeneous organism is acquired, and the multimodal motion time-series trajectory is redirected and mapped according to the physical kinematic constraints of the centaur robot to generate the initial control trajectory of the centaur robot. The initial control trajectory is input into the gait generation network. The initial control trajectory is first denoised within the gait generation network, and then the denoised result is temporally smoothed to generate standard gait feature data. Based on the variable axis inverted pendulum mechanism, a dynamic stability criterion is constructed to constrain the vertical axis position of the centaur robot's center of mass. A partially observable Markov mechanism is introduced to define the gait control process as an interactive hierarchy that includes the environmental state space and the action space. Within the interaction layer, based on the near-end policy optimization mechanism and the dynamic stability criterion, the standard gait feature data is iteratively solved to generate a multi-band decoupled target control strategy. The target control strategy generates joint drive commands to drive the joint motors of the centaur robot and control the centaur robot to perform multimodal gait switching.

2. The method for generating and controlling multimodal gait of a centaur robot based on imitation learning according to claim 1, characterized in that, Before acquiring the multimodal motion temporal trajectory of the target heterologous organism, the following steps are included: Using the target heterologous organism as the motion acquisition object, walking gait data, running gait data and climbing gait data of the target heterologous organism are collected by optical capture device; Based on the acquisition time markers corresponding to each gait data, the walking gait data, the running gait data, and the climbing gait data are time-aligned to generate gait time-series data corresponding to different gait categories; The key bone contact point temporal change features are extracted from the gait temporal data. The key bone contact point temporal change features characterize the bone position change state of the target heterologous organism at different times when it comes into contact with the supporting environment. According to the gait category, the temporal change features of the corresponding key bone contact points are sequentially spliced ​​to generate the multimodal motion temporal trajectory of the target heterologous organism.

3. The method for generating and controlling multimodal gait of a centaur robot based on imitation learning according to claim 1, characterized in that, The step of retargeting and mapping the multimodal motion temporal trajectory based on the physical kinematic constraints of the centaur robot to generate the initial control trajectory of the centaur robot includes: Obtain the physical limits of the link length and joint angle of the centaur robot to construct the physical kinematic constraints; Based on the physical kinematic constraints, the mechanical space coordinate range of the centaur robot is determined as the target mechanical motion range for the multimodal motion temporal trajectory mapping. Based on the target's mechanical motion range, a spatial geometric transformation matrix is ​​established to transform the spatial coordinates of the target heterogeneous organism into the mechanical spatial coordinates of the centaur robot. The spatial geometric transformation matrix represents the physical limitations of the link length and the physical limits of the joint angle. The spatial geometric transformation matrix is ​​used to project the multimodal motion temporal trajectory into coordinate space to scale the multimodal motion temporal trajectory to the physical shape range of the centaur robot, thereby generating the initial control trajectory.

4. The method for generating and controlling multimodal gait of a centaur robot based on imitation learning according to claim 1, characterized in that, The initial control trajectory is input into the gait generation network. Within the gait generation network, the initial control trajectory is first denoised, and then the denoising result is temporally smoothed to generate standard gait feature data, including: The sliding window length used to input the gait generation network is determined based on the number of consecutive control cycles corresponding to the initial control trajectory. According to the sliding window length, the initial control trajectory is divided into temporal sliding window segment data and input into the gait generation network; The temporal sliding window segmented data is sequentially input into the gait generation network. The gait generation network uses a recurrent layer with a memory gating structure to remove jump noise from the temporal sliding window segmented data, thereby generating a denoised trajectory. A smoothing window is configured at the generation end of the loop layer, and a first-order difference constraint is applied to the denoised trajectory through the smoothing window to obtain the time-smoothed trajectory. The standard gait feature data is generated based on the trajectory after time-series smoothing, and the standard gait feature data maintains the gait time-series change relationship in the initial control trajectory.

5. The method for generating and controlling multimodal gait of a centaur robot based on imitation learning according to claim 1, characterized in that, The dynamic stability criterion for constraining the vertical axial position of the centaur robot's center of mass, based on the variable-axis inverted pendulum mechanism, includes: The foot support area of ​​the centaur robot is collected in real time, and the zero moment point boundary is determined based on the foot support area. The zero moment point boundary limits the support stability range of the centaur robot during the current gait control process. The real-time three-dimensional position of the centaur robot's center of mass is obtained to extract the vertical axial position of the center of mass in the vertical direction. Based on the variable axial inverted pendulum mechanism, a linkage characterization between the vertical axial position and the acceleration of the center of mass in the horizontal direction is established, so that the vertical axial position participates in the determination of the stable state. The predicted zero-moment point location is calculated based on the aforementioned linkage characterization; Based on the geometric distance between the predicted zero-moment point position and the zero-moment point boundary, a dynamic stability criterion is constructed to constrain the vertical axial position of the centaur robot's center of mass. When the predicted zero-moment point position is located inside the zero-moment point boundary, the centaur robot is determined to be in a stable state.

6. A multimodal gait generation and control device for centaur robots based on imitation learning, characterized in that, include: The trajectory acquisition and mapping module is used to acquire the multimodal motion temporal trajectory of the target heterogeneous organism, and to redirect and map the multimodal motion temporal trajectory according to the physical kinematic constraints of the Centaur robot to generate the initial control trajectory of the Centaur robot. The gait feature generation module is used to input the initial control trajectory into the gait generation network, and to first denoise the initial control trajectory in the gait generation network, and then perform temporal smoothing on the denoising result to generate standard gait feature data. A stable interaction construction module is used to construct a dynamic stability standard for constraining the vertical axis position of the centaur robot's center of mass based on the variable axis inverted pendulum mechanism, and introduces a partially observable Markov mechanism to define the gait control process as an interaction hierarchy that includes the environmental state space and the action space. The control strategy generation module is used to generate a multi-band decoupled target control strategy by iteratively solving the standard gait feature data based on the near-end strategy optimization mechanism and the dynamic stability criterion within the interaction level. The joint drive control module is used to generate joint drive commands according to the target control strategy to drive the joint motors of the centaur robot and control the centaur robot to perform multimodal gait switching.

7. An electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; When the processor is used to execute the computer program, it implements the method according to any one of claims 1-9.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method described in any one of claims 1-9.

9. A centaur robot, characterized in that, This includes the trunk structure, joint motors, sensor acquisition components, and gait controller; The sensing and acquisition component is used to collect motion state data of the centaur robot; The gait controller is used to acquire the multimodal motion time-series trajectory of the target heterogeneous organism, and to redirect and map the multimodal motion time-series trajectory according to the physical kinematic constraints of the centaur robot to generate the initial control trajectory of the centaur robot. The gait controller is also used to input the initial control trajectory into the gait generation network, and to first denoise the initial control trajectory in the gait generation network, and then to perform temporal smoothing on the denoising result to generate standard gait feature data. The gait controller is also used to construct a dynamic stability standard for constraining the vertical axial position of the centaur robot's center of mass based on the variable axis inverted pendulum mechanism, and introduces a partially observable Markov mechanism to define the gait control process as an interactive hierarchy that includes the environmental state space and the action space. The gait controller is also used to iteratively solve the standard gait feature data based on the proximal strategy optimization mechanism and the dynamic stability criterion within the interaction level to generate a multi-band decoupled target control strategy. The gait controller is further configured to generate joint drive commands according to the target control strategy, and send the joint drive commands to the joint motor; The joint motor is used to drive the trunk structure to perform multimodal gait switching according to the joint drive command.

10. A multimodal gait generation and control system for a centaur robot based on imitation learning, characterized in that, This includes a motion trajectory acquisition unit, a gait control unit, and a Centaur robot execution unit; The motion trajectory acquisition terminal is used to acquire the multimodal motion temporal trajectory of the target heterologous organism and send the multimodal motion temporal trajectory to the gait control terminal; The gait control terminal is used to redirect and map the multimodal motion temporal trajectory according to the physical kinematic constraints of the Centaur robot, thereby generating the initial control trajectory of the Centaur robot. The gait control terminal is also used to input the initial control trajectory into the gait generation network, and to first denoise the initial control trajectory in the gait generation network, and then to perform temporal smoothing on the denoising result to generate standard gait feature data. The gait control terminal is also used to construct a dynamic stability standard for constraining the vertical axial position of the centaur robot's center of mass based on the variable axis inverted pendulum mechanism, and to introduce a partially observable Markov mechanism to define the gait control process as an interactive hierarchy that includes the environmental state space and the action space. The gait control terminal is also used to iteratively solve the standard gait feature data based on the proximal strategy optimization mechanism and the dynamic stability criterion within the interaction layer, thereby generating a multi-band decoupled target control strategy. The gait control terminal is also used to generate joint drive commands according to the target control strategy, and send the joint drive commands to the centaur robot execution terminal; The centaur robot actuator is used to drive the joint motors according to the joint drive commands, so as to control the centaur robot to perform multimodal gait switching.

Images