Semi-human robot control instruction adaptive allocation method and device based on hierarchical reinforcement learning, equipment, medium, robot and system

Abstract

The application provides a semi-human robot control instruction adaptive allocation method and device based on hierarchical reinforcement learning, an electronic device, a computer readable storage medium, a semi-human robot and a system. The method obtains a current environment perception state and a double-arm operation intention representation of the semi-human robot, generates a quadruped element index and a whole body centroid compensation vector through a top policy network, extracts a target basic control sequence, generates adaptive control instructions and collaborative correction instructions through a bottom execution network, and collaboratively drives the quadruped nodes and the double-arm nodes to perform tasks.

Description

Technical Field

[0001] This application relates to the field of robot intelligent control technology, and more specifically, to a method, apparatus, electronic device, computer-readable storage medium, centaur robot and system for adaptive allocation of control instructions based on hierarchical reinforcement learning. Background Technology

[0002] Centaur robots typically combine quadrupedal locomotion with bi-arm manipulation capabilities, enabling them to perform complex tasks such as movement, obstacle crossing, grasping, carrying, and support operations in complex terrain environments. Because these robots include multiple supporting legs, multi-degree-of-freedom robotic arms, and torso posture adjustment mechanisms, their control process must consider not only the stable movement of the robot within the terrain environment but also the impact of the bi-arms performing operational tasks on the robot's posture and balance. Therefore, this places high demands on the robot control system's real-time computing power, motion coordination capabilities, and disturbance suppression capabilities.

[0003] A common control scheme for centaur robots employs a whole-body control approach to solve multiple degrees of freedom in a unified manner. This scheme first establishes the robot's whole-body dynamics equations, incorporating variables such as foot contact force, joint position, joint velocity, torso posture, and end-effector pose into the same optimization problem. Then, it solves for the control quantities of all joints based on the current task objective and constraints. Finally, the solved control quantities are distributed to the leg and arm drive mechanisms to complete the movement and manipulation tasks.

[0004] However, the aforementioned whole-body control scheme has significant technical shortcomings. Due to the large number of degrees of freedom in the Centaur robot, uniformly solving the optimization problem corresponding to all degrees of freedom would cause the computational load to increase rapidly with the number of degrees of freedom, making it difficult to adapt to the control command generation requirements within a real-time control cycle. Simultaneously, the movement and manipulation tasks are coupled at the control level. When the arms perform operations such as grasping, pushing, pulling, or carrying, the operational load can easily cause the robot's center of gravity to shift and body disturbances. A single whole-body optimization approach cannot simultaneously address foot displacement control, arm manipulation control, and body disturbance suppression within a short control cycle, resulting in low real-time allocation capability and task coordination adaptability for the Centaur robot in complex environments. Summary of the Invention

[0005] This application provides a method, apparatus, electronic device, computer-readable storage medium, centaur robot, and system for adaptive allocation of control instructions based on hierarchical reinforcement learning, so as to at least alleviate the above-mentioned technical problems.

[0006] An adaptive allocation method for control commands of a centaur robot based on hierarchical reinforcement learning includes: Acquire the current environmental perception state and the representation of the operating intentions of the centaur robot's two arms; The current environmental perception state and the dual-arm operation intention representation are input into a pre-constructed top-level policy network, so that the top-level policy network performs action allocation processing on the dual-arm operation intention representation under the current control cycle based on the current environmental perception state, thereby generating the quadruped primitive index and the whole-body centroid compensation vector corresponding to the dual-arm operation intention representation in the current control cycle. Extract the target basic control sequence corresponding to the quadruped primitive index from the preset parameterized primitive library; The target basic control sequence and the whole-body center of mass compensation vector are input into the underlying execution network, so that the underlying execution network performs motion redirection processing on the target basic control sequence based on the whole-body center of mass compensation vector, thereby generating adaptive control instructions for the quadruped nodes and cooperative correction instructions for the bi-arm nodes. The quadruped node is driven to perform displacement tasks based on the adaptive control command, and the dual-arm node is driven to perform target manipulation tasks based on the cooperative correction command, so as to suppress the body disturbance caused by the target manipulation tasks.

[0007] An adaptive allocation device for control commands of a centaur robot based on hierarchical reinforcement learning, comprising: The state acquisition module is used to acquire the current environmental perception state of the Centaur robot and the representation of the operating intentions of its two arms; The top-level strategy processing module is used to input the current environmental perception state and the dual-arm operation intention representation into the pre-constructed top-level strategy network, so that the top-level strategy network performs action allocation processing on the dual-arm operation intention representation under the current control cycle based on the current environmental perception state, thereby generating the quadruped primitive index and the whole-body centroid compensation vector corresponding to the dual-arm operation intention representation in the current control cycle. The basic control sequence extraction module is used to extract the target basic control sequence corresponding to the quadruped primitive index from a preset parameterized primitive library; The underlying execution processing module is used to input the target basic control sequence and the whole-body center of mass compensation vector into the underlying execution network, so that the underlying execution network performs motion redirection processing on the target basic control sequence based on the whole-body center of mass compensation vector, thereby generating adaptive control instructions for the quadruped nodes and cooperative correction instructions for the bi-arm nodes. The task-driven processing module is used to drive the quadruped node to perform a displacement task based on the adaptive control command, and to drive the dual-arm node to perform a target manipulation task based on the cooperative correction command, so as to suppress the body disturbance caused by the target manipulation task.

[0008] An electronic device, including a memory and a processor; The memory is used to store computer programs; The processor is used to execute the computer program to implement the adaptive allocation method for control instructions of the centaur robot based on hierarchical reinforcement learning.

[0009] A computer-readable storage medium storing a computer program; When the computer program is executed by the processor, it implements the adaptive allocation method for control instructions of the centaur robot based on hierarchical reinforcement learning.

[0010] A centaur robot includes a quadruped node, a bi-arm node, an environmental perception component, a body posture acquisition component, and a control processor; The environmental sensing component is used to collect local elevation map data in the current environmental sensing state; The body posture acquisition component is used to acquire six-degree-of-freedom pose data in the current environmental perception state; The control processor is used to acquire the current environmental perception state and the representation of the dual-arm operation intention. The control processor is also used to input the current environmental perception state and the dual-arm operation intention representation into a pre-constructed top-level policy network, so that the top-level policy network performs action allocation processing on the dual-arm operation intention representation under the current control cycle based on the current environmental perception state, thereby generating the quadruped primitive index and the whole-body centroid compensation vector corresponding to the dual-arm operation intention representation in the current control cycle. The control processor is also used to extract the target basic control sequence corresponding to the quadruped primitive index from a preset parameterized primitive library; The control processor is further configured to input the target basic control sequence and the whole-body center of mass compensation vector into the underlying execution network, so that the underlying execution network performs motion redirection processing on the target basic control sequence based on the whole-body center of mass compensation vector, thereby generating adaptive control instructions for the quadruped nodes and cooperative correction instructions for the biarm nodes. The quadruped node is used to perform displacement tasks in response to the adaptive control command; The dual-arm nodes are used to execute target operation tasks in response to the cooperative correction instructions, so as to suppress the machine body disturbances caused by the target operation tasks.

[0011] An adaptive allocation system for control instructions of a centaur robot based on hierarchical reinforcement learning includes a centaur robot, a state acquisition unit, a policy allocation unit, a primitive library management unit, a low-level execution unit, and a task-driven unit. The centaur robot includes a quadruped node and a bi-arm node; The state acquisition unit is used to acquire the current environmental perception state and the representation of the dual-arm operation intention of the centaur robot. The strategy allocation unit is used to input the current environmental perception state and the dual-arm operation intention representation into the pre-constructed top-level strategy network, so that the top-level strategy network performs action allocation processing on the dual-arm operation intention representation under the current control cycle based on the current environmental perception state, thereby generating the quadruped primitive index and the whole-body centroid compensation vector corresponding to the dual-arm operation intention representation in the current control cycle. The primitive library management unit is used to extract the target basic control sequence corresponding to the quadruped primitive index from a preset parameterized primitive library; The underlying execution unit is used to input the target basic control sequence and the whole-body center of mass compensation vector into the underlying execution network, so that the underlying execution network performs motion redirection processing on the target basic control sequence based on the whole-body center of mass compensation vector, thereby generating adaptive control instructions for the quadruped node and cooperative correction instructions for the bi-arm node. The task-driven unit is used to drive the quadruped node to perform a displacement task based on the adaptive control command, and to drive the dual-arm node to perform a target manipulation task based on the cooperative correction command, so as to suppress the body disturbance caused by the target manipulation task.

[0012] Technical advantages of the technical solution provided in this application This application presents a hierarchical reinforcement learning-based adaptive allocation method for centaur robot control commands. Addressing the shortcomings of existing whole-body control schemes that require unified solution of multi-degree-of-freedom optimization problems and suffer from high computational burden during real-time control cycles, this method first acquires the centaur robot's current environmental perception state and the representation of its arm operation intentions. Then, it inputs these representations into a pre-constructed top-level policy network, enabling the network to generate quadruped primitive indices and a whole-body centroid compensation vector within the current control cycle. Compared to traditional methods that directly optimize all degrees of freedom, this method first transforms the current environmental perception state and arm operation intention representations into quadruped primitive indices and a whole-body centroid compensation vector. This allows the control command generation process to complete action allocation at the policy layer, reducing the processing pressure of directly solving all degrees of freedom in the subsequent lower-level execution stages. This improves the real-time adaptability of control command generation in complex control scenarios.

[0013] This application extracts the target basic control sequence corresponding to the quadruped primitive index from a pre-set parameterized primitive library. This allows the quadruped node's displacement task to no longer rely entirely on real-time generation of a complete whole-body control trajectory. Instead, it can use the target basic control sequence based on the quadruped primitive index as the control basis for subsequent motion redirection processing. Compared to traditional whole-body control schemes that reorganize all joint control quantities of the legs and arms in each control cycle, this application utilizes the correspondence between the quadruped primitive index and the target basic control sequence to pre-organize the basic motion content required for quadrupedal movement control into a searchable data foundation. This reduces the repetitive control sequence construction process within the current control cycle, resulting in better control command continuity when the quadruped node performs displacement tasks.

[0014] This application inputs the target basic control sequence and the whole-body center-of-mass compensation vector into the underlying execution network. The underlying execution network then performs motion redirection processing on the target basic control sequence based on the whole-body center-of-mass compensation vector, thereby generating adaptive control commands for the quadruped nodes and cooperative correction commands for the bi-arm nodes. Since the whole-body center-of-mass compensation vector originates from the top-level policy network's action allocation processing of the current environmental perception state and the bi-arm operation intention representation, the underlying execution network can simultaneously incorporate center-of-mass compensation information related to the bi-arm operation task when generating control commands for the quadruped nodes. This ensures that the displacement control of the quadruped nodes and the operation correction of the bi-arm nodes have a correlation within the same control cycle. Compared to traditional solutions where movement and operation tasks are easily mutually constrained within a unified optimization problem, this application outputs adaptive control commands and cooperative correction commands to the quadruped nodes and bi-arm nodes respectively, making the coordination between leg displacement control and arm operation control clearer.

[0015] This application further utilizes adaptive control commands to drive the quadruped nodes to perform displacement tasks and collaborative correction commands to drive the dual-arm nodes to perform target manipulation tasks, thereby suppressing body disturbances caused by target manipulation tasks. Since the adaptive control commands correspond to the displacement execution of the quadruped nodes, and the collaborative correction commands correspond to the target manipulation execution of the dual-arm nodes, both are formed by processing the target-based control sequence and the whole-body centroid compensation vector through a low-level execution network. Therefore, when the dual-arm nodes perform target manipulation tasks, the displacement execution of the quadruped nodes can be linked with the body disturbance suppression requirements. Compared to the technical problem of traditional whole-body control schemes struggling to simultaneously address movement, manipulation, and disturbance suppression within a short control cycle, this application enables a clearer technical connection between quadruped movement tasks, dual-arm manipulation tasks, and body disturbance suppression processing within a hierarchical control link, thereby improving the collaborative control adaptability of the Centaur robot when performing complex tasks in complex environments. Attached Figure Description

[0016] Figure 1This application provides an embodiment of an adaptive allocation scenario for control instructions of a centaur robot based on hierarchical reinforcement learning. Figure 2 This application provides an embodiment of an adaptive allocation method for control instructions of a centaur robot based on hierarchical reinforcement learning. Figure 3 This application provides an embodiment of an adaptive allocation device for control instructions of a centaur robot based on hierarchical reinforcement 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 an adaptive allocation system for control instructions of a centaur robot based on hierarchical reinforcement learning. Detailed Implementation

[0017] like Figure 1 As shown, this is an embodiment of the present application of an adaptive allocation scenario for control commands of a centaur robot based on hierarchical reinforcement learning; Figure 2 As shown in the figure, an adaptive allocation method for control commands of a centaur robot based on hierarchical reinforcement learning is proposed according to an embodiment of this application, including: Acquire the current environmental perception state and the representation of the operating intentions of the centaur robot's two arms; The current environmental perception state and the dual-arm operation intention representation are input into a pre-constructed top-level policy network, so that the top-level policy network performs action allocation processing on the dual-arm operation intention representation under the current control cycle based on the current environmental perception state, thereby generating the quadruped primitive index and the whole-body centroid compensation vector corresponding to the dual-arm operation intention representation in the current control cycle. Extract the target basic control sequence corresponding to the quadruped primitive index from the preset parameterized primitive library; The target basic control sequence and the whole-body center of mass compensation vector are input into the underlying execution network, so that the underlying execution network performs motion redirection processing on the target basic control sequence based on the whole-body center of mass compensation vector, thereby generating adaptive control instructions for the quadruped nodes and cooperative correction instructions for the bi-arm nodes. The quadruped node is driven to perform displacement tasks based on the adaptive control command, and the dual-arm node is driven to perform target manipulation tasks based on the cooperative correction command, so as to suppress the body disturbance caused by the target manipulation tasks.

[0018] Optionally, the pre-set parameterized primitive library is used for the quadruped primitive index to retrieve the corresponding target basic control sequence, and the construction steps of the pre-set parameterized primitive library include: The offline trajectory representation of the robot is collected in a preset physical simulator, which is a simulation processing environment configured during the initialization phase of the robot control system to reproduce the cooperative motion process of the quadruped node and the dual-arm node. The offline trajectory representation is input into a temporal dimensionality reduction network to perform pattern decoupling processing on the offline trajectory representation, thereby separating periodic gait features and non-periodic transition features. The periodic gait features are segmented to generate multiple first motion segments, and the non-periodic transition features are segmented to generate multiple second motion segments. Using template kinematic mapping rules, the first motion segment and the second motion segment are respectively mapped to the corresponding underlying joint drive signal sequences; The underlying joint drive signal sequence is associated and encapsulated with the corresponding primitive identifier, and the associated and encapsulated underlying joint drive signal sequence is written into the database representation structure to generate the preset parameterized primitive library for generating the target basic control sequence.

[0019] Preferably, the specific implementation process of the offline trajectory representation of the acquisition robot in the preset physical simulator is as follows: In the robot control system initialization stage, the robot geometric parameters, joint motion boundaries, quadruped node contact parameters, dual-arm node operation parameters and ground interaction parameters of the preset physical simulator are configured first, so that the preset physical simulator can reproduce the cooperative motion process according to the real connection relationship between the quadruped nodes and dual-arm nodes of the Centaur robot. The essence of the preset physical simulator is not simply a simulation interface displaying robot motion, but rather a means to generate offline trajectory representations at the data level that can be subsequently decomposed, extracted, and mapped. The robot geometric parameters configured in the robot control system initialization phase of the preset physical simulator represent the spatial connection relationships between the torso, leg links, and arm links. The joint motion boundaries define the angle and velocity ranges of each joint during simulation. The quadrupedal node contact parameters record the contact, support, and lift-off states between the feet and the ground. The arm node operation parameters record the end-effector movement intentions and load changes of the arm nodes when performing tasks such as grasping, carrying, or pushing / pulling objects. The ground interaction parameters record the contact height, direction, and source of contact constraints when the feet contact the ground. Subsequently, a simulation task sequence containing quadrupedal node displacement tasks and arm node object manipulation tasks is loaded into the preset physical simulator. The torso spatial pose, quadrupedal node joint states, arm node joint states, foot contact states, and arm end-effector operation states of the Centaur robot are continuously sampled according to the control cycle to form an offline trajectory representation. The simulation task sequence is used to incorporate the quadrupedal node displacement task and the dual-arm node target manipulation task into the same simulation sampling caliber, enabling the subsequently formed offline trajectory representation to simultaneously reflect the movement process of the quadrupedal node and the manipulation process of the dual-arm node. The technical essence of this offline trajectory representation is to transform the multi-degree-of-freedom motion process generated by the centaur robot in the simulation environment into a trajectory data structure arranged according to the control cycle. This trajectory data structure is not a single spatial path, but rather contains multi-channel temporal data simultaneously including movement tasks, manipulation tasks, contact switching, and body posture changes. In subsequent processing, this multi-channel temporal data is arranged according to the control cycle into a trajectory temporal expression, which is then fed into a temporal dimensionality reduction network. Therefore, the trajectory data structure and the subsequent trajectory temporal expression have a data source relationship. The offline trajectory representation is subsequently input into the temporal dimensionality reduction network, which separates periodic gait features and aperiodic transition features from the offline trajectory representation. Therefore, the offline trajectory representation has a direct technical source relationship with the quadrupedal primitive index and target basic control sequence in the protection theme.By first forming an offline trajectory representation in the preset physical simulator, the subsequent preset parameterized primitive library can be retrieved by the quadruped primitive index during the running phase, without having to resolve the complete whole-body motion process in the current control cycle, thus enabling the target basic control sequence to have a traceable data source and a stable motion structure foundation.

[0020] Preferably, the specific implementation process of inputting the offline trajectory representation into a temporal dimensionality reduction network to perform pattern decoupling processing on the offline trajectory representation, thereby separating periodic gait features and non-periodic transition features, is as follows: After the offline trajectory representation is formed, the trunk spatial pose, quadrupedal joint states, biarm joint states, foot contact states, and biarm end-effector operation states in the offline trajectory representation are read according to the control cycle, and the above data are arranged into a trajectory temporal expression according to the same time sampling caliber. Each row in the trajectory temporal expression corresponds to a control cycle, and each column corresponds to a motion parameter channel. The element at the intersection of the row and column represents the sampling value or state marker of the corresponding motion parameter channel under the corresponding control cycle. The motion parameter channel originates from the trunk spatial pose, quadrupedal joint states, biarm joint states, foot contact states, and biarm end-effector operation states in the offline trajectory representation. Therefore, the motion parameter channel will not exist independently of the offline trajectory representation. Subsequently, the trajectory temporal representation is input into the temporal dimensionality reduction network, which includes a trajectory field organization layer, a phase continuous encoding layer, a periodic response separation layer, a transition response separation layer, and a trajectory reconstruction and verification layer. The trajectory field organization layer first unifies the arrangement of different motion parameter channels in the trajectory temporal representation to form a field-aligned trajectory representation. This field-aligned trajectory representation then enters the phase continuous encoding layer, which continuously encodes the support phase, swing phase, and contact switching moment of the quadruped nodes based on the continuity of foot contact states within adjacent control cycles to form a phase-encoded trajectory representation. The periodic response separation layer, based on the phase-encoded trajectory representation, performs homogeneous merging processing on recurring support and swing phases to extract periodic gait features that characterize stable gait cycles. The transition response separation layer, based on the phase-encoded trajectory representation, performs differential extraction processing on the start-up, stop, turning, obstacle avoidance, and posture compensation processes caused by arm node operations that cannot be classified into stable gait cycles to form non-periodic transition features. The trajectory restoration and verification layer then restores and compares the periodic gait features and the aperiodic transition features according to the order of the original control cycle to verify whether the periodic gait features and the aperiodic transition features can cover the main motion changes in the offline trajectory representation on the time axis.Through the continuous cooperation between the above structural layers, the temporal dimensionality reduction network does not simply compress the data dimension of the offline trajectory representation, but separates the gait cycle part that can be repeatedly invoked from the transition response part that is difficult to repeat. This allows the first motion segment to be obtained when the periodic gait features are segmented, and the second motion segment to be obtained when the non-periodic transition features are segmented. This provides a data foundation for the preset parameterized primitive library to store the underlying joint drive signal sequence converted from the first motion segment and the underlying joint drive signal sequence converted from the second motion segment, respectively.

[0021] Preferably, in the specific technical implementation of segmenting the periodic gait features to generate multiple first motion segments, the support phase marker, swing phase marker, foot contact switching time, quadrupedal joint state change sequence, and trunk spatial pose change sequence in the periodic gait features are first read, and the foot contact switching time is used as the candidate position of the segment boundary. The support phase marker, the swing phase marker, and the foot contact switching time are derived from the phase-encoded trajectory expression, and the quadrupedal joint state change sequence and the trunk spatial pose change sequence are derived from the field-aligned trajectory expression. Therefore, the above data in the periodic gait features can be traced back to the offline trajectory representation. The technical essence of the periodic gait features is that the quadrupedal movement data that can repeat in a similar phase order in the offline trajectory representation mainly reflects the regular relationship between foot support, foot swing, and trunk posture changes when the quadrupedal nodes perform displacement tasks. Since the target basic control sequence in the protected subject needs to be quickly extracted within the current control cycle and participate in the motion redirection processing of the underlying execution network, directly saving the complete offline trajectory representation would result in the target basic control sequence containing a large amount of irrelevant control cycle data, which is not conducive to the retrieval of quadruped primitive indexes. Therefore, it is necessary to perform segmentation processing on the periodic gait features to organize the reusable motion intervals in the stable gait cycle into multiple first motion segments. In specific implementation, the start and end control cycles of a gait cycle are first identified based on the support phase marker and swing phase marker. Then, the quadruped node joint state change sequence and trunk spatial pose change sequence within the gait cycle are read. Consistency screening is performed on the phase length, foot contact sequence, and trunk posture change amplitude between adjacent gait cycles to eliminate abnormal sampling intervals caused by initial simulation disturbances or temporary posture corrections. The gait cycle data retained after consistency screening is segmented into first motion segments. Each first motion segment contains the quadruped node joint state change sequence, foot contact state change sequence, and trunk spatial pose change sequence within the corresponding gait cycle. The first motion segment is not a local action segment of a single joint, but rather a motion unit that describes how the quadruped nodes complete support transitions and displacement propulsion within a reusable gait range. The first motion segment is subsequently processed by template kinematics mapping rules, which convert the quadruped node joint state change sequence, foot contact state change sequence, and trunk spatial pose change sequence in the first motion segment into a low-level joint drive signal sequence. Therefore, the first motion segment serves as the data source for stable gait cycles within a pre-set parameterized primitive library.By extracting periodic gait features into multiple first motion segments, the quadruped primitive index can directly retrieve the target basic control sequence that matches the current environmental perception state during the runtime phase, so that the quadruped node does not have to generate a complete joint control sequence from scratch when performing displacement tasks.

[0022] Preferably, the specific implementation process of segmenting the aperiodic transition feature to generate multiple second motion segments is as follows: After the aperiodic transition feature is formed, the starting change interval, stopping change interval, turning change interval, obstacle avoidance change interval, and posture compensation interval caused by the dual-arm node operation are read from the aperiodic transition feature. The starting change interval, stopping change interval, turning change interval, obstacle avoidance change interval, and posture compensation interval caused by the dual-arm node operation are then mapped to the control cycle marker in the offline trajectory representation to determine the entry and exit positions of each transition action on the time axis. The control cycle marker comes from the trajectory data structure arranged according to the control cycle in the offline trajectory representation. The entry position and the exit position are used to define the segmentation boundary of the second motion segment. The technical essence of the aperiodic transition feature is the motion adjustment data generated by the Centaur robot outside the stable gait cycle. It mainly comes from motion processes that do not have a fixed repetition cycle, such as changes in movement direction, foot landing point adjustment, changes in dual-arm node load, or torso posture compensation. Compared to the periodic gait features, the aperiodic transition features are not suitable for segmentation according to a complete gait cycle because the duration and amplitude of the aperiodic transition features change with environmental fluctuations, the representation of the arm operation intentions, and changes in body posture. Therefore, when segmenting the aperiodic transition features, the quadrupedal joint states and trunk spatial pose before the transition action are first read based on the entry position, and then the quadrupedal joint states and trunk spatial pose after the transition action are read based on the exit position. The continuously changing quadrupedal joint states, foot contact states, and arm node posture compensation content between the entry and exit positions are then segmented into a second motion segment. The difference between the second motion segment and the first motion segment is that the first motion segment is used to describe a repeatable stable gait cycle, while the second motion segment is used to describe the transition changes between stable gait cycles or between a stable gait cycle and arm node operations. The boundary of the first motion segment is mainly determined by the foot contact switching moment, while the boundary of the second motion segment is determined by the entry and exit positions of the transition action. To enable the second motion segment to be used in conjunction with the whole-body center-of-mass compensation vector in subsequent motion redirection processing, the second motion segment also retains the torso spatial pose changes before and after the transition action, allowing the underlying execution network to identify the posture compensation direction corresponding to the second motion segment when calling the target basic control sequence. The second motion segment is subsequently entered into the template kinematics mapping rule and converted into the underlying joint drive signal sequence. Therefore, the second motion segment is not additional explanatory data, but rather basic motion data from a pre-set parameterized primitive library used to cover the processes of starting, stopping, turning, obstacle avoidance, and the connection of disturbances in the two-arm node operations.By extracting non-periodic transition features into multiple second motion segments, the quadruped primitive index can retrieve the target basic control sequence corresponding to the transition action when faced with changes in the current environmental perception state or changes in the representation of the dual-arm operation intention, thereby improving the adaptability of the target basic control sequence to complex task scenarios.

[0023] Preferably, the specific implementation process of mapping the first motion segment and the second motion segment to the corresponding underlying joint drive signal sequence using template kinematic mapping rules is as follows: After obtaining multiple first motion segments and multiple second motion segments, the template kinematic mapping rules configured during the initialization phase of the robot control system are read first. The technical essence of the template kinematic mapping rules is to convert the trajectory expression in the first motion segment and the second motion segment into parameter conversion rules for joint drive data that can be executed by the quadruped nodes and the bi-arm nodes. The template kinematic mapping rules include joint channel arrangement, link connection, foot coordinate transformation, joint angle constraint, and joint velocity constraint. The joint channel arrangement aperture is used to specify the arrangement order of each joint in the quadrupedal node and the dual-arm node in the underlying joint drive signal sequence; the link connection aperture is used to record the connection relationship between the torso, leg links and dual-arm links, and to enable the changes in the spatial position of the foot end in the first motion segment and the second motion segment to be converted according to the same link connection relationship; the foot end coordinate transformation aperture is used to convert the changes in the spatial position of the foot end in the first motion segment and the second motion segment into the angle changes of the corresponding joints; the joint angle constraint aperture and the joint velocity constraint aperture are used to limit the converted joint angles and joint velocities to be within the robot's executable range. In specific mapping, the quadrupedal node joint state change sequence in the first motion segment and the quadrupedal node joint state change sequence in the second motion segment are first read according to the joint channel arrangement aperture, and then kinematic inverse solving is performed according to the link connection aperture and the foot end coordinate transformation aperture to form a candidate sequence of joint angles; then, the candidate sequence of joint angles is subjected to boundary correction processing according to the joint angle constraint aperture to form an executable joint angle sequence, and an executable joint velocity sequence is generated according to the change in the executable joint angle sequence in adjacent control cycles. The executable joint angle sequence and the executable joint velocity sequence are further time-registered with the foot contact state in the first motion segment, the foot contact state in the second motion segment, the trunk spatial pose change content in the first motion segment, and the trunk spatial pose change content in the second motion segment to generate a low-level joint drive signal sequence. The technical essence of the low-level joint drive signal sequence is time-seriesd joint control data that can be read by the subsequent low-level execution network. It includes at least the desired joint position data, desired joint velocity data, and feedforward torque data corresponding to the foot contact state in each control cycle. Since the first motion segment corresponds to a stable gait cycle, the low-level joint drive signal sequence mapped from the first motion segment is mainly used to form the target basic control sequence corresponding to the stable gait cycle; since the second motion segment corresponds to a non-periodic transition process, the low-level joint drive signal sequence mapped from the second motion segment is mainly used to form the target basic control sequence corresponding to the non-periodic transition process.By performing mapping processing through the template kinematic mapping rules, the first motion segment and the second motion segment formed in the simulation stage can be converted into a sequence of underlying joint drive signals that can be retrieved and executed in the running stage, so that each data item in the preset parameterized primitive library can establish an executable technical association with the subsequent adaptive control command generation process.

[0024] Preferably, the specific implementation process of associating and encapsulating the underlying joint drive signal sequence with the corresponding primitive identifier, and writing the associated and encapsulated underlying joint drive signal sequence into the database representation structure to generate the preset parameterized primitive library for generating the target basic control sequence is as follows: After the underlying joint drive signal sequence is formed, the corresponding primitive identifier is generated based on the motion segment source, supporting leg identifier, swing leg identifier, action phase marker, transition action marker, and adaptation scene state marker of the underlying joint drive signal sequence. The motion segment source is used to represent that the underlying joint drive signal sequence originates from a first motion segment or a second motion segment. The supporting leg identifier and the swing leg identifier originate from the foot contact state in the first motion segment or the second motion segment. The action phase marker originates from the supporting phase marker and swing phase marker in the periodic gait features. The transition action marker originates from the starting change interval, stopping change interval, turning change interval, obstacle avoidance change interval, and posture compensation interval caused by the dual-arm node operation in the non-periodic transition features. The adaptation scene state marker originates from the terrain undulation state corresponding to the current environmental perception state and the operation load state corresponding to the dual-arm operation intention representation. The technical essence of the primitive identifier is that it serves as an index key value used by the quadruped primitive index to retrieve the target basic control sequence. The primitive identifier is not arbitrarily numbered; rather, it incorporates the motion segment source, quadruped node contact state, and adaptation scene state markers corresponding to the underlying joint drive signal sequence into the same retrieval scope. Subsequently, the primitive identifier, the underlying joint drive signal sequence, the motion segment source corresponding to the underlying joint drive signal sequence, the control cycle arrangement scope corresponding to the underlying joint drive signal sequence, and the joint channel arrangement scope corresponding to the underlying joint drive signal sequence are associated and encapsulated to form a primitive encapsulation record. This primitive encapsulation record is used to map the primitive identifier that can be retrieved by the quadruped primitive index to the underlying joint drive signal sequence that can be read by the underlying execution network, enabling the quadruped primitive index to retrieve the primitive identifier and then read the complete target basic control sequence along the primitive encapsulation record. The database representation structure includes a primitive identifier field, a motion segment source field, a joint channel arrangement field, a control cycle arrangement field, a bottom-level joint drive signal sequence field, and an adaptation scene state marker field. After the primitive encapsulation records are written into the database representation structure, a retrieval index is built based on the primitive identifier field, and an auxiliary retrieval index is built based on the adaptation scene state marker field to form the preset parameterized primitive library. The technical essence of the preset parameterized primitive library is an action retrieval data structure oriented towards the current control cycle. It can read the corresponding primitive encapsulation records according to the four-legged primitive index generated by the top-level policy network and extract the target basic control sequence from the primitive encapsulation records.The target basic control sequence is then input into the underlying execution network along with the whole-body center-of-mass compensation vector. The underlying execution network performs motion redirection processing on the target basic control sequence based on the whole-body center-of-mass compensation vector, thereby generating adaptive control instructions for the quadruped nodes and cooperative correction instructions for the bi-arm nodes. Therefore, the associated encapsulation process and the database representation structure writing process are not simply data storage actions, but rather a chaining of offline trajectory representation, first motion segment, second motion segment, underlying joint drive signal sequence, primitive identifier, quadruped primitive index, and target basic control sequence into a continuous and searchable control data link. This enables the pre-set parameterized primitive library to rapidly generate the target basic control sequence within the current control cycle in this application.

[0025] Preferably, the training sample construction process of the temporal dimensionality reduction network is as follows: After the preset physical simulator has generated multiple sets of offline trajectory representations, the torso spatial pose, quadrupedal joint states, biarm joint states, foot contact states, and biarm end-effector operation states in each set of offline trajectory representations are read according to the control cycle. The torso spatial pose, quadrupedal joint states, biarm joint states, foot contact states, and biarm end-effector operation states are arranged according to the same time sampling caliber to form a training trajectory temporal representation. The training trajectory temporal representation adopts the same row and column meaning as the aforementioned trajectory temporal representation, that is, each row corresponds to a control cycle, each column corresponds to a motion parameter channel, and the element at the intersection of the row and column represents the sampling value or state marker of the corresponding motion parameter channel under the corresponding control cycle. The motion parameter channels are derived from the torso spatial pose, quadrupedal joint states, biarm joint states, foot contact states, and biarm end-effector operation states. Therefore, the training trajectory temporal representation can maintain the same data source relationship as the offline trajectory representation. Subsequently, training support phase markers, training swing phase markers, and training contact switching times are generated based on the foot contact states in the training trajectory temporal representation. These markers are then written into the training sample record, providing a phase continuity reference for the phase continuity coding layer. This phase continuity reference is used to constrain the output phase-coded trajectory representation of the phase continuity coding layer during subsequent training. Next, periodic gait annotations are generated based on the repetitive relationship of foot contact states between adjacent control cycles, the changing patterns of quadrupedal joint states, and the changing patterns of trunk spatial pose in the training trajectory temporal representation. Non-periodic transition annotations are generated based on the posture compensation processes caused by starting, stopping, turning, obstacle avoidance, and arm node operations in the training trajectory temporal representation. The periodic gait annotations indicate the periodic gait features that the periodic response separation layer should extract from the training trajectory temporal representation, and the non-periodic transition annotations indicate the non-periodic transition features that the transition response separation layer should extract from the training trajectory temporal representation. Through the above processing, each training sample record includes a training trajectory temporal representation, a training support phase marker, a training swing phase marker, a training contact switching time, a training periodic gait annotation, and a training non-periodic transition annotation. This enables the temporal dimensionality reduction network to simultaneously learn the field alignment relationship of the trajectory field organization layer, the phase continuation relationship of the phase continuity encoding layer, the periodic extraction relationship of the periodic response separation layer, the transition extraction relationship of the transition response separation layer, and the restoration comparison relationship of the trajectory restoration verification layer during the training process.The field alignment relationship is used to transfer the motion parameter channels in the training trajectory temporal representation to the field-aligned trajectory representation. The phase continuation relationship is used to transfer the training support phase marker, the training swing phase marker, and the training contact switching moment to the phase-encoded trajectory representation. The period extraction relationship is used to transfer the repetitive motion patterns corresponding to the training periodic gait annotations to the periodic gait features. The transition extraction relationship is used to transfer the non-repetitive motion patterns corresponding to the training non-periodic transition annotations to the non-periodic transition features. The restoration comparison relationship is used to restore the periodic gait features and non-periodic transition features together to the restored trajectory representation.

[0026] Preferably, the loss function of the temporal dimensionality reduction network includes a field alignment loss function, a phase continuity loss function, a periodic response separation loss function, a transition response separation loss function, and a trajectory reconstruction verification loss function. The field alignment loss function constrains the difference in motion parameter channel arrangement between the field-aligned trajectory expression output by the trajectory field organization layer and the training trajectory temporal expression, ensuring that the torso spatial pose, quadrupedal joint states, bi-arm joint states, foot contact states, and bi-arm end-effector operation states in the field-aligned trajectory expression follow the channel source and control cycle order of the training trajectory temporal expression. The phase continuity loss function constrains the difference between the phase-encoded trajectory expression output by the phase continuity encoding layer and the training support phase marker, training swing phase marker, and training contact switching time, ensuring that the phase-encoded trajectory expression maintains the continuity between support and swing phases between adjacent control cycles. The periodic response separation loss function constrains the difference between the periodic gait features output by the periodic response separation layer and the periodic gait annotations used for training, ensuring that the periodic gait features retain the sequence of quadrupedal joint state changes, foot contact state changes, and trunk spatial pose changes in a stable gait cycle. The transition response separation loss function constrains the difference between the non-periodic transition features output by the transition response separation layer and the non-periodic transition annotations used for training, ensuring that the non-periodic transition features retain the start-up change interval, stop change interval, turning change interval, obstacle avoidance change interval, and posture compensation interval caused by bi-arm node operations. The trajectory reconstruction verification loss function constrains the difference between the reconstructed trajectory expression generated by the trajectory reconstruction verification layer based on the periodic gait features and non-periodic transition features and the temporal expression of the training trajectory, ensuring that the periodic gait features and non-periodic transition features, after separation, can still jointly cover the main motion changes in the offline trajectory representation. The field alignment loss function, phase continuity loss function, periodic response separation loss function, transition response separation loss function, and trajectory reconstruction verification loss function are jointly calculated during training using the same training sample record. The field alignment loss function first constrains the input order of the motion parameter channels, the phase continuity loss function then constrains the phase continuity relationship corresponding to the foot contact state, the periodic response separation loss function and the transition response separation loss function constrain the separation results of periodic gait features and non-periodic transition features respectively, and the trajectory reconstruction verification loss function finally constrains whether the periodic gait features and non-periodic transition features can be jointly reconstructed to the time axis of the trajectory time series expressed in the training.Therefore, the field alignment loss function corresponds to the field alignment relationship of the trajectory field organization layer, the phase continuity loss function corresponds to the phase continuation relationship of the phase continuity encoding layer, the period response separation loss function corresponds to the period extraction relationship of the period response separation layer, the transition response separation loss function corresponds to the transition extraction relationship of the transition response separation layer, and the trajectory reconstruction verification loss function corresponds to the reconstruction comparison relationship of the trajectory reconstruction verification layer, so that the training objectives of each structural layer are consistent with the construction process of the preset parameterized primitive library.

[0027] Preferably, the training process of the temporal dimensionality reduction network is as follows: First, multiple training sample records are divided into training sample record batches, and the parts of multiple sets of offline trajectory representations that did not participate in the construction of the training sample record batches are divided into verification offline trajectory representations. The verification offline trajectory representations are used to verify whether the segment boundaries corresponding to periodic gait features and the segment boundaries corresponding to non-periodic transition features remain stable during parameter update. At the same time, a preset convergence amplitude is configured before training begins. The preset convergence amplitude is used to limit the decrease threshold of the total training loss value in multiple consecutive training sample record batches. Subsequently, the temporal expression of the training trajectory is input to the trajectory field organizing layer in sequence according to the training sample record batches, so that the trajectory field organizing layer outputs the field-aligned trajectory expression. The field-aligned trajectory expression is input to the phase continuous encoding layer, so that the phase continuous encoding layer learns the continuity relationship of the foot contact state in adjacent control cycles based on the training support phase marker, the training swing phase marker, and the training contact switching time, so as to output the phase-encoded trajectory expression. The phase-encoded trajectory representation is then fed into a periodic response separation layer and a transition response separation layer. The periodic response separation layer separates periodic gait features from the phase-encoded trajectory representation based on the periodic gait annotations used in training, while the transition response separation layer separates aperiodic transition features from the phase-encoded trajectory representation based on the aperiodic transition annotations used in training. Subsequently, a trajectory reconstruction and verification layer generates a reconstructed trajectory representation based on the periodic gait features and aperiodic transition features, and performs a reconstruction comparison with the temporal representation of the training trajectory to obtain the loss value corresponding to the trajectory reconstruction and verification loss function. During training, the field alignment loss function, phase continuity loss function, periodic response separation loss function, transition response separation loss function, and trajectory reconstruction verification loss function each yield their corresponding loss values. These loss values ​​are then weighted and synthesized according to the order of the structural layers to form the total training loss value. Specifically, the field alignment loss function and phase continuity loss function in the total training loss value are used to stabilize the input representation of the temporal dimensionality reduction network; the periodic response separation loss function and transition response separation loss function are used to strengthen the pattern boundary between periodic gait features and non-periodic transition features; and the trajectory reconstruction verification loss function is used to suppress the loss of major motion changes in the offline trajectory representation after separation.The parameters of the trajectory field organization layer, phase continuity encoding layer, periodic response separation layer, transition response separation layer, and trajectory reconstruction verification layer in the temporal dimensionality reduction network are updated based on the total training loss value. This parameter update process is repeated in subsequent training sample batches. When the decrease in the total training loss value in multiple consecutive training sample batches is lower than a preset convergence range, and the segment boundaries corresponding to the periodic gait features and the segment boundaries corresponding to the non-periodic transition features remain stable in the offline trajectory representation used for verification, the parameter update is stopped, thus forming the trained temporal dimensionality reduction network. The trained temporal dimensionality reduction network is then used to perform pattern decoupling processing on new offline trajectory representations and outputs periodic gait features and non-periodic transition features, enabling subsequent segment truncation processing to generate the first motion segment and the second motion segment, respectively.

[0028] Optionally, the step of inputting the current environmental perception state and the dual-arm operation intention representation into a pre-constructed top-level policy network, so that the top-level policy network performs action allocation processing on the dual-arm operation intention representation under the current control cycle based on the current environmental perception state, and then generates the quadruped primitive index and whole-body centroid compensation vector corresponding to the dual-arm operation intention representation in the current control cycle, includes: The environmental feature encoding layer in the top-level policy network is used to perform spatial topology parsing on the current environmental perception state to generate access mask features; Using the operation intent decoding layer in the top-level policy network, the dual-arm operation intent representation is subjected to temporal attention extraction processing to generate the expected centroid offset caused by the arm movements. The access mask features and the expected centroid offset are concatenated to generate a joint state characterization matrix. The joint state representation matrix is ​​value-predicted using the action evaluation layer in the top-level policy network to generate the selection probability of each primitive in the parameterized primitive library. According to the principle of selecting the highest probability, the quadruped primitive index is determined from the parameterized primitive library, and the expected centroid offset is reversed in direction to generate the whole-body centroid compensation vector.

[0029] Preferably, the specific implementation process of concatenating the access mask features and the expected centroid offset to generate a joint state representation matrix is ​​as follows: After the access mask features and the expected centroid offset are formed, the access mask features are first processed to adjust their spatial dimensions, so that they form an access mask arrangement expression according to the spatial arrangement order of the candidate landing regions; simultaneously, the expected centroid offset is processed to adjust its directional dimensions, so that it forms a centroid offset arrangement expression according to the order of the torso's front-back direction, left-right direction, and up-down direction. Each element of the access mask arrangement expression corresponds to a candidate landing region, and each element of the centroid offset arrangement expression corresponds to a centroid offset direction. The two are not the same in terms of data dimensions, so they cannot be directly added or compared. To avoid mixing of dimensions, the access mask arrangement expression is first processed to encode the access state to form a access state encoded expression; then, the centroid offset arrangement expression is processed to normalize the scale to form a centroid offset encoded expression. The passage state encoding is derived from the passage mask arrangement, retaining the passage state of the candidate landing area; the centroid offset encoding is derived from the centroid offset arrangement, retaining the centroid change trend caused by the dual-arm node operation. Subsequently, the passage state encoding and the centroid offset encoding are concatenated according to a channel concatenation method to generate a joint state representation matrix. The technical essence of the joint state representation matrix is ​​to incorporate the spatial passage constraints corresponding to the current environmental perception state and the centroid offset trend corresponding to the dual-arm operation intention representation into the same action evaluation data structure; the rows in the joint state representation matrix correspond to the candidate landing areas, and the columns correspond to different data channels in the passage state encoding and centroid offset encoding. The element at the intersection of rows and columns represents the encoded value of the corresponding candidate landing area under the corresponding data channel. The joint state representation matrix is ​​then input to the action evaluation layer, which performs value prediction processing based on the joint state representation matrix to generate the selection probability of each primitive in a pre-set parameterized primitive library. By organizing the access mask features and expected centroid offsets in the form of a joint state representation matrix, the top-level policy network can simultaneously consider the spatial access constraints corresponding to the current environmental perception state and the centroid offset trend corresponding to the dual-arm operation intention representation in the same value prediction process. This avoids the quadruped primitive index being determined solely based on the current environmental perception state or solely based on the dual-arm operation intention representation, thereby establishing a data connection relationship between the quadruped primitive index and the whole-body centroid compensation vector.

[0030] Preferably, the specific implementation process of using the action evaluation layer in the top-level policy network to perform value prediction processing on the joint state representation matrix to generate the selection probability of each primitive in the preset parameterized primitive library is as follows: After the joint state representation matrix is ​​formed, it is input to the action evaluation layer. The action evaluation layer first reads the data channel corresponding to the access state encoding expression in the joint state representation matrix to obtain the spatial access constraints of each candidate landing area; then it reads the data channel corresponding to the centroid offset encoding expression in the joint state representation matrix to obtain the centroid offset trend of the dual-arm node target operation task on the body centroid. In the same value prediction processing, the action evaluation layer correlates and evaluates the spatial access constraints and the centroid offset trend to form a primitive adaptation score. The technical essence of the primitive adaptation score is to represent the matching degree between each primitive in the preset parameterized primitive library and the current environmental perception state and dual-arm operation intention representation within the current control cycle. The primitive adaptation score is not directly used as a quadruped primitive index, but first undergoes probability normalization processing to convert the primitive adaptation scores corresponding to different primitives into selection probabilities under the same probability caliber. The technical essence of the selection probability is the numerical expression of the relative probability of each primitive in the preset parameterized primitive library being selected within the current control cycle by the action evaluation layer. The selection probability corresponds one-to-one with the primitive identifiers in the preset parameterized primitive library. Specifically, the action evaluation layer first reduces the primitive adaptation scores that conflict with unsuitable landing areas based on the passage status encoding expression, and then increases the primitive adaptation scores that can compensate for the expected centroid offset based on the centroid offset encoding expression. After probability normalization processing, the action evaluation layer generates the selection probability of each primitive in the preset parameterized primitive library. The selection probabilities of each primitive in the pre-set parameterized primitive library are subsequently used to determine the quadruped primitive index. Therefore, the selection probabilities are not the final control commands, but rather intermediate evaluation results connecting the joint state representation matrix and the quadruped primitive index. Through value prediction processing by the action evaluation layer, the top-level policy network can map the current environmental perception state and the dual-arm operation intention representation to the primitive selection range of the pre-set parameterized primitive library without directly solving all joint control variables. This allows the selection probabilities of each primitive in the pre-set parameterized primitive library to simultaneously reflect the spatial passage constraints expressed by the passage mask features and the centroid offset trend expressed by the expected centroid offset, thus providing data for subsequently determining the quadruped primitive index from the pre-set parameterized primitive library.

[0031] Preferably, the specific implementation process of determining the quadruped primitive index from the preset parameterized primitive library according to the principle of the highest selection probability, and inverting the direction of the expected centroid offset to generate the whole-body centroid compensation vector is as follows: After generating the selection probability of each primitive in the preset parameterized primitive library at the action evaluation layer, the selection probability of each primitive is first read in correspondence with the primitive identifier in the preset parameterized primitive library, and the primitive identifiers are arranged in descending order of selection probability to form a primitive candidate sorting expression. The primitive candidate sorting expression is used to represent the adaptation order between different primitives and the joint state representation matrix in the current control cycle. The joint state representation matrix already contains the pass mask feature and the expected centroid offset. Therefore, the primitive candidate sorting expression can bear the combined influence of the current environmental perception state and the dual-arm operation intention representation. Subsequently, the primitive identifier located at the highest selection probability position in the primitive candidate sorting expression is read according to the principle of the highest selection probability, and the primitive identifier located at the highest selection probability position is determined as the quadruped primitive index. The technical essence of the quadruped primitive index is to retrieve the index result of the target basic control sequence from a preset parameterized primitive library within the current control cycle. This quadruped primitive index will subsequently be used as a retrieval condition in the preset parameterized primitive library to extract the target basic control sequence corresponding to the quadruped primitive index. Simultaneously, the expected centroid offset is inverted to generate a whole-body centroid compensation vector. The technical essence of this inversion is to convert the action direction of each directional element in the expected centroid offset to the opposite compensation trend while maintaining the directional dimension and numerical scale of the expected centroid offset. This allows the whole-body centroid compensation vector to express the compensation trend that counteracts the expected centroid offset caused by arm movements. The elements in the whole-body centroid compensation vector still correspond to the torso's forward / backward, left / right, and up / down directions. The whole-body centroid compensation vector maintains the same directional dimension as the expected centroid offset; therefore, the whole-body centroid compensation vector can be input into the underlying execution network along with the target basic control sequence. The underlying execution network then performs motion redirection processing on the target basic control sequence based on the whole-body centroid compensation vector, thereby generating adaptive control instructions for the quadruped nodes and cooperative correction instructions for the bi-arm nodes. Through the above processing, the quadruped primitive index is used to solve the problem of which target basic control sequence should be extracted from the preset parameterized primitive library in the current control cycle, and the whole-body centroid compensation vector is used to solve the problem of how to compensate for the centroid change trend generated by the object operation task of the bi-arm nodes. Together, they constitute the output result of the top-level policy network for action allocation processing in the current control cycle.

[0032] Optionally, the current environmental perception state includes local elevation map data generated by an airborne depth camera and six-degree-of-freedom pose data generated by an inertial measurement unit. The step of utilizing the environmental feature encoding layer in the top-level policy network to perform spatial topology parsing processing on the current environmental perception state to generate access mask features includes: Using the three-dimensional convolutional kernel in the environmental feature encoding layer, spatial depth stripping is performed on the local elevation map data to generate a terrain undulation feature map; Using the temporal memory neurons in the environmental feature encoding layer, the six-degree-of-freedom pose data is processed to track historical poses in order to generate body tilt angle change rate features. The body tilt angle change rate feature is mapped to the channel dimension of the terrain undulation feature map to generate attitude terrain joint features; The posture and terrain joint features are subjected to cross-channel feature extraction processing to generate the access mask features, which are used to characterize the safe landing point area of ​​the robot in the next control cycle.

[0033] Preferably, the current environmental perception state includes local elevation map data generated by the airborne depth camera and six-degree-of-freedom pose data generated by the inertial measurement unit. The specific implementation process of using the three-dimensional convolutional kernel in the environmental feature encoding layer to perform spatial depth stripping processing on the local elevation map data to generate a terrain undulation feature map is as follows: Within the current control cycle, the depth sampling results continuously output by the airborne depth camera are first read, and the depth sampling results are spatially rasterized according to the Centaur robot's forward direction, lateral direction, and height direction to form local elevation map data. The depth sampling results are used to record the spatial distance information obtained by the airborne depth camera within the current control cycle. The spatial distance information, after spatial rasterization, forms the local elevation map data, allowing the local elevation map data to utilize the spatial source of the depth sampling results. The technical essence of the local elevation map data is spatial data used to characterize the ground height distribution around the Centaur robot in the current environmental perception state. Each elevation element in the local elevation map data corresponds to a spatial grid position around the Centaur robot, and the elevation element is used to represent the height difference between the spatial grid position and the coordinate aperture of the Centaur robot. The local elevation map data is not an ordinary two-dimensional image, but rather a terrain height expression that converts the spatial distance information obtained by the airborne depth camera into a terrain height expression that can be processed by the environmental feature encoding layer. The terrain height expression is carried by the local elevation map data and is subsequently processed by spatial depth stripping through a three-dimensional convolutional kernel to generate a terrain undulation feature map. Therefore, the local elevation map data has a priori data relationship with the determination of the quadruped primitive index. Subsequently, the three-dimensional convolutional kernel in the environmental feature encoding layer slides and samples the local elevation map data along the Centaur robot's forward direction, lateral direction, and height direction, and reads the elevation element changes between adjacent spatial grid positions during each slide sampling to form a local spatial depth response. The technical essence of the 3D convolution kernel is a data processing unit that simultaneously reads the changes in terrain (front-to-back, left-to-right, and height) within a 3D neighborhood. Its function is not simply to compress the local elevation map data, but rather to extract ground protrusions, depressions, step edges, and slope changes related to foot stability from the local elevation map data, and incorporate these into the local spatial depth response. This local spatial depth response undergoes further spatial depth stripping processing to distinguish between gentle terrain areas, abrupt terrain areas, and areas of height discontinuity. The gentle terrain areas, abrupt terrain areas, and areas of height discontinuity are then written into the channel content of the terrain undulation feature map, thereby forming the terrain undulation feature map.The technical essence of the terrain undulation feature map is the representation of terrain structure extracted from local elevation map data. Each feature location in the terrain undulation feature map corresponds to a spatial raster neighborhood in the local elevation map data. The channel content of the feature location is used to represent the ground undulation amplitude, ground undulation direction, and height continuity within the spatial raster neighborhood. The terrain undulation feature map subsequently receives attitude channel content formed by the aircraft tilt angle change rate feature mapping to generate joint attitude and terrain features. Therefore, the terrain undulation feature map not only provides terrain undulation information but also provides a channel carrying foundation for the aircraft tilt angle change rate feature to enter spatial topology parsing processing, enabling the pass-through mask feature to simultaneously reflect terrain undulation information and aircraft tilt angle change rate feature.

[0034] Preferably, in the specific technical implementation of using the temporal memory neurons in the environmental feature encoding layer to perform historical attitude tracking processing on the six-DOF pose data to generate the body tilt angle change rate feature, while forming the terrain undulation feature map, the six-DOF pose data generated by the inertial measurement unit of the body in the current control cycle and multiple adjacent historical control cycles is read. The technical essence of the six-DOF pose data is temporal data used to characterize the spatial motion state of the Centaur robot body in the current environmental perception state. The six-DOF pose data includes translational and rotational states along the Centaur robot body coordinate system, where the rotational state reflects the changing trends of the body's pitch, roll, and yaw. The six-DOF pose data is not directly equivalent to the pass-through mask feature, because the pass-through mask feature needs to express the safe landing point area for the next control cycle, while the six-DOF pose data still needs to undergo historical attitude tracking processing to transform the temporal changes in the body's attitude into data that can affect the landing judgment. In specific implementation, the temporal memory neurons in the environmental feature encoding layer read the six-degree-of-freedom pose data sequentially according to the control cycle. They then perform attitude difference processing on the six-degree-of-freedom pose data of the current control cycle and the six-degree-of-freedom pose data of adjacent historical control cycles to obtain an attitude difference representation. The technical essence of the temporal memory neurons is that they are temporal state units used to retain the body's attitude change states in adjacent historical control cycles. They can establish a temporal continuity relationship between the attitude rotation state of the current control cycle and the attitude rotation state of adjacent historical control cycles. This temporal continuity relationship continues to participate in the attitude difference processing, enabling the environmental feature encoding layer to determine whether the body tilt angle is stabilizing, continuously increasing, or undergoing a directional change. The attitude difference representation then enters the tilt angle change extraction processing to extract the change components related to landing stability from the pitch, roll, and yaw changes, thereby forming the body tilt angle change rate feature. The technical essence of the body tilt angle change rate feature is to represent the temporal attitude characteristics of the Centaur robot's body tilt state as it changes with the control cycle. Different elements in the body tilt angle change rate feature correspond to the forward / backward pitch change rate, the left / right roll change rate, and the horizontal yaw change rate, respectively. The forward / backward pitch change rate represents the tilt angle change trend of the Centaur robot's body along the forward / backward direction, the left / right roll change rate represents the tilt angle change trend of the Centaur robot's body along the left / right direction, and the horizontal yaw change rate represents the change trend of the Centaur robot's body in the horizontal turning direction. The body tilt angle change rate feature is subsequently mapped to the channel dimension of the terrain undulation feature map. Therefore, the body tilt angle change rate feature is not an independent result for determining whether the body attitude is stable, but rather it is used to supplement the missing dynamic attitude information of the body in the terrain undulation feature map, enabling the subsequently generated attitude-terrain joint feature to simultaneously include terrain spatial changes and body attitude changes.

[0035] Preferably, the specific implementation process of mapping the aircraft tilt rate change feature to the channel dimension of the terrain undulation feature map to generate attitude-terrain joint features is as follows: After the terrain undulation feature map and the aircraft tilt rate change feature are formed respectively, the spatial grid arrangement relationship and channel arrangement relationship of the terrain undulation feature map are read first, and the pitch rate, roll rate, and yaw rate of the aircraft tilt rate change feature are read. Since the terrain undulation feature map has a spatial grid dimension and a channel dimension, while the aircraft tilt rate change feature mainly expresses aircraft changes along the time dimension and attitude direction dimension, the two cannot be directly numerically superimposed. In order to enable the aircraft tilt rate change feature to enter the spatial analysis process of the terrain undulation feature map, the aircraft tilt rate change feature is first spatially broadcast based on the spatial grid arrangement relationship of the terrain undulation feature map to form an attitude change broadcast expression. The technical essence of the attitude change broadcast expression is to replicate the aircraft's tilt angle change rate feature within the same control cycle to each spatial grid position according to the spatial grid arrangement of the terrain undulation feature map, so that each spatial grid position can simultaneously contain ground undulation content and aircraft attitude change content. Each spatial grid position in the attitude change broadcast expression includes pitch change rate, roll change rate, and yaw change rate. The pitch change rate, roll change rate, and yaw change rate are used as attitude channel content in the attitude change broadcast expression for subsequent channel embedding processing. Subsequently, according to the channel arrangement of the terrain undulation feature map, the attitude change broadcast expression is embedded into the channel dimension of the terrain undulation feature map to form a joint attitude and terrain feature. The technical essence of this mapping is to transform the aircraft tilt rate change characteristic from a time-attitude expression into a channel expression that can be used in calculations together with the terrain undulation feature map, rather than directly adding data of different dimensions. Specifically, the original channels in the terrain undulation feature map continue to represent the continuous state of ground undulation amplitude, direction, and height, while the newly added channels corresponding to the attitude change broadcast expression represent the pitch rate of change, roll rate of change, and yaw rate of change. The technical essence of the attitude-terrain joint feature is to incorporate the terrain undulation information corresponding to the local elevation map data and the aircraft tilt rate change characteristic corresponding to the six-degree-of-freedom pose data into the same spatial channel expression data structure. Each spatial grid position in the attitude-terrain joint feature simultaneously contains terrain undulation information and aircraft attitude change information. The attitude-terrain joint feature will subsequently undergo cross-channel feature extraction processing, which will determine whether the candidate landing area is unsuitable for landing due to the combined effect of terrain undulation information and aircraft attitude change information.Therefore, the attitude and terrain joint feature connects the local elevation map data and the six-degree-of-freedom pose data under the same spatial reference caliber, so that the access mask feature not only reflects the static elevation changes of the terrain, but also reflects the attitude change trend of the Centaur robot in the current control cycle.

[0036] Preferably, the cross-channel feature extraction process is performed on the joint posture and terrain features to generate the access mask features, which characterize the safe landing point area of ​​the robot in the next control cycle. The specific implementation process is as follows: After the joint posture and terrain features are formed, the environmental feature encoding layer first reads the ground undulation amplitude, ground undulation direction, height continuity, pitch change rate, roll change rate, and yaw change rate corresponding to each spatial grid position in the joint posture and terrain features, and performs cross-channel feature extraction processing on the above channel contents. The technical essence of the cross-channel feature extraction processing is to jointly read the terrain channel and the posture channel within the same spatial grid position and determine whether the terrain channel and the posture channel will jointly affect the landing stability of the quadruped node in the next control cycle. The terrain channel includes the ground undulation amplitude, ground undulation direction, and height continuity; the posture channel includes the pitch change rate, roll change rate, and yaw change rate; the terrain channel originates from the terrain undulation feature map, and the posture channel originates from the posture change broadcast expression formed by the body tilt angle change rate feature mapping. For example, when a spatial grid location has a low altitude continuity and a large pitch change rate, even if the location is still within the reach of the quadruped nodes in terms of pure altitude, the landing adaptation degree will be reduced due to changes in the aircraft's attitude. Similarly, when the direction of ground undulation at a spatial grid location overlaps with the direction of its roll change rate, the location will also negatively impact the support stability of the quadruped nodes. The cross-channel feature extraction process first generates a terrain channel response based on the ground undulation amplitude, direction, and altitude continuity. Then, it generates an attitude channel response based on the pitch change rate, roll change rate, and yaw change rate. Subsequently, channel association determination is performed on the terrain channel response and attitude channel response to generate a landing adaptation response. This landing adaptation response indicates the adaptation degree of each spatial grid location as a candidate landing area in the next control cycle. The landing adaptation response further undergoes mask transformation to generate a passage mask feature. The technical essence of the access mask feature is to convert the continuous landing adaptation response in the attitude and terrain joint features into spatial constraint data that can be used by the top-level policy network for action allocation. Each element in the access mask feature corresponds to a spatial grid position, and the element value in the access mask feature is used to indicate whether the corresponding spatial grid position belongs to the safe landing point area of ​​the next control cycle. The safe landing point area is not simply determined by the ground height in the local elevation map data, but is formed by the joint participation of the terrain undulation feature map and the body tilt angle change rate feature in cross-channel feature extraction processing. The access mask feature is then concatenated with the expected centroid offset to generate a joint state representation matrix, and the action evaluation layer generates the selection probability of each primitive in the preset parameterized primitive library based on the joint state representation matrix.Therefore, by generating the access mask features, the local elevation map data and six-degree-of-freedom pose data in the current environmental perception state can enter the determination process of the quadruped primitive index, so that the target basic control sequence extracted from the preset parameterized primitive library can correspond to the safe landing point area of ​​the next control cycle.

[0037] Preferably, the training sample construction process of the top-level policy network is as follows: After the preset physical simulator has formed multiple sets of offline trajectory representations, and the preset parameterized primitive library has stored primitive identifiers and corresponding target basic control sequences, the current environmental perception state and the dual-arm operation intention representation for training are first read from the multiple sets of offline trajectory representations according to the control cycle. The current environmental perception state for training includes local elevation map data generated by the airborne depth camera and six-degree-of-freedom pose data generated by the fuselage inertial measurement unit. The current environmental perception state for training is used to provide the source of terrain spatial structure and body attitude changes under the current control cycle. The dual-arm operation intention representation for training includes the expected trajectory points of the end effector within a preset time window and the expected load mass data corresponding to each expected trajectory point of the end effector. The dual-arm operation intention representation for training is used to provide the source of the expected centroid offset caused by the object operation task of the dual-arm node. Subsequently, a training access mask annotation is generated based on the local elevation map data and six-degree-of-freedom pose data in the current environmental perception state for training. The training access mask annotation is used to indicate the access mask features that the environmental feature encoding layer should parse from the current environmental perception state for training. Based on the end effector expected trajectory point and expected load mass data in the training dual-arm operation intention representation, combined with the torso spatial pose change content of the corresponding control cycle in the offline trajectory representation, a training expected centroid offset annotation is generated. The training expected centroid offset annotation is used to indicate the expected centroid offset that the operation intention decoding layer should extract from the training dual-arm operation intention representation. Then, based on the primitive identifiers that are actually invoked or should be invoked in the corresponding control cycle according to the offline trajectory representation, training quadruped primitive annotations are generated. These training quadruped primitive annotations maintain the same retrieval criterion as the primitive identifiers in the preset parameterized primitive library. The training quadruped primitive annotations are used to constrain the action evaluation layer to increase the selection probability of the corresponding primitive identifier in the preset parameterized primitive library. Simultaneously, the direction of the training expected centroid offset annotation is inverted to generate a training full-body centroid compensation vector annotation. This training full-body centroid compensation vector annotation is used to constrain the full-body centroid compensation vector output by the top-level policy network. Through the above processing, each top-level policy training sample record includes the current environmental perception state for training, the training dual-arm operation intention representation, the training pass-through mask annotation, the training expected centroid offset annotation, the training quadruped primitive annotation, and the training full-body centroid compensation vector annotation. The top-level policy training sample record incorporates the technical relationship between the current environmental perception state, the representation of the dual-arm operation intention, the quadruped primitive index, and the whole-body centroid compensation vector into the same training caliber. This enables the top-level policy network to learn during the training phase how to jointly determine the quadruped primitive index from the expected centroid offset caused by the terrain spatial structure and the representation of the dual-arm operation intention, and to simultaneously generate the whole-body centroid compensation vector corresponding to the expected centroid offset.

[0038] Preferably, the specific process by which the top-level policy training sample records are processed by the top-level policy network is as follows: During training, the current environmental perception state is first input to the environmental feature encoding layer, which performs spatial topology parsing on the current environmental perception state to generate training output access mask features. The three-dimensional convolutional kernel in the environmental feature encoding layer first performs spatial depth stripping on the local elevation map data in the current environmental perception state to generate a training output terrain undulation feature map. The temporal memory neurons in the environmental feature encoding layer then perform historical pose tracking on the six-degree-of-freedom pose data in the current environmental perception state to generate training output body tilt rate of change features. Subsequently, the training output body tilt rate of change features are mapped to the channel dimension of the training output terrain undulation feature map to generate training output pose-terrain joint features, and cross-channel feature extraction is performed on the training output pose-terrain joint features to generate training output access mask features. The training output access mask features are used for comparison with the training access mask annotations and continue to participate in the generation of the subsequent training output joint state representation matrix. Simultaneously, the training dual-arm maneuver intention representation is input into the maneuver intention decoding layer, which performs temporal attention extraction processing on the training dual-arm maneuver intention representation to generate the expected centroid shift of the training output. The maneuver intention decoding layer first arranges the training dual-arm maneuver intention representation into a temporal expression of the training output maneuver intention. Then, based on the amplitude and direction of the arm end-effector movement and the expected load change in the temporal expression of the training output maneuver intention, it forms a weighted expression of the training output maneuver intention, and generates the expected centroid shift of the training output based on this weighted expression. This expected centroid shift of the training output is compared with the labeled expected centroid shift of the training and continues to participate in the generation of the joint state representation matrix of the training output and the whole-body centroid compensation vector of the training output. Subsequently, the spatial dimension of the training output pass mask features is processed to form the training output pass state encoding representation; the direction dimension of the expected centroid offset of the training output is processed and the scale is normalized to form the training output centroid offset encoding representation; then, the training output pass state encoding representation and the training output centroid offset encoding representation are concatenated according to the channel concatenation method to generate the training output joint state representation matrix. The training output joint state representation matrix is ​​input to the action evaluation layer, which performs value prediction processing on the training output joint state representation matrix to generate the training output selection probability. The training output selection probability corresponds one-to-one with the primitive identifier in the preset parameterized primitive library and is used for comparison with the training quadruped primitive annotation; the action evaluation layer also determines the training output quadruped primitive index according to the principle of the highest training output selection probability, and performs direction inversion processing on the expected centroid offset of the training output to generate the training output whole-body centroid compensation vector.The training output quadruped primitive index is used to verify whether the top-level policy network can select a primitive identifier consistent with the training quadruped primitive label under the joint constraints of the current environment perception state and the training dual-arm operation intention representation. The training output whole-body centroid compensation vector is used to verify whether the top-level policy network can convert the training output expected centroid offset into a compensation trend consistent with the training whole-body centroid compensation vector label. The training output quadruped primitive index and the training output whole-body centroid compensation vector will also be jointly included in the calculation process of the action allocation collaborative loss function, so that the target basic control sequence corresponding to the training output quadruped primitive index can maintain the same action allocation caliber as the training output whole-body centroid compensation vector before subsequent motion redirection processing.

[0039] Preferably, the loss function of the top-level policy network includes a passage mask loss function, a centroid offset loss function, a primitive selection loss function, a compensation vector consistency loss function, and an action allocation collaboration loss function. The passage mask loss function constrains the difference between the training output passage mask features and the training passage mask annotations, enabling the environment feature encoding layer to extract spatial constraints related to the safe landing point region for the next control cycle from the current environment perception state used in training. The centroid offset loss function constrains the difference between the training output expected centroid offset and the training expected centroid offset annotations, enabling the operation intention decoding layer to extract the centroid change trend caused by arm movements from the training dual-arm operation intention representations. The primitive selection loss function constrains the difference between the training output selection probability and the training quadruped primitive annotations, enabling the action evaluation layer to select from a pre-set parameterized primitive library. The selection probability of the primitive label corresponding to the training quadruped primitive label is increased. The compensation vector consistency loss function is used to constrain the difference between the training output whole-body centroid compensation vector and the training whole-body centroid compensation vector label, so that the training output whole-body centroid compensation vector obtained by direction inversion can maintain a compensation trend opposite to the training expected centroid offset label. The action allocation collaborative loss function is used to jointly constrain the collaborative relationship between the training output quadruped primitive index and the training output whole-body centroid compensation vector, so that the target basic control sequence corresponding to the training output quadruped primitive index can enter the motion redirection processing of the subsequent low-level execution network together with the training output whole-body centroid compensation vector. During training, the loss values ​​corresponding to the passage mask loss function, the centroid offset loss function, the primitive selection loss function, the compensation vector consistency loss function, and the action allocation collaborative loss function are weighted and synthesized according to the processing order of the structural layers to form the total training loss value of the top-level policy. Based on the total training loss value of the top-level policy, the parameters of the environment feature encoding layer, the operation intention decoding layer, and the action evaluation layer are updated, and the above parameter update process is repeated in the subsequent top-level policy training sample records. Before training begins, the top-level policy convergence magnitude and the top-level policy training sample records for verification are configured. The top-level policy convergence magnitude is used to limit the decrease threshold of the total training loss value of the top-level policy in multiple consecutive top-level policy training sample records. The top-level policy training sample records for verification are derived from the top-level policy training sample records that have not participated in parameter updates. The top-level policy training sample records for verification are used to check whether the training output pass-through mask feature, the training output expected centroid offset, the training output quadruped primitive index, and the training output whole-body centroid compensation vector remain stable on the data that have not participated in parameter updates.When the decrease in the total loss value of the top-level policy training across multiple consecutive top-level policy training sample records is less than the convergence magnitude of the top-level policy, and when the training output quadruped primitive index stably corresponds to the training quadruped primitive label in the validation top-level policy training sample records, and the training output whole-body centroid compensation vector stably corresponds to the training whole-body centroid compensation vector label in the validation top-level policy training sample records, parameter updates are stopped to form the trained top-level policy network. The trained top-level policy network is then used to receive the current environmental perception state and the dual-arm operation intention representation, and to generate the quadruped primitive index and whole-body centroid compensation vector corresponding to the current control cycle.

[0040] Optionally, the dual-arm operation intention representation includes the expected trajectory points of the end effector within a preset time window and the expected load quality data corresponding to each expected trajectory point of the end effector. The preset time window is a time sampling interval configured in the operation task planning stage to limit the forward movement range of the dual-arm nodes. The step of utilizing the operation intent decoding layer in the top-level policy network to perform temporal attention extraction processing on the dual-arm operation intent representation to generate the expected centroid shift caused by the arm movements includes: Using the multi-head attention mechanism in the operation intent decoding layer, spatial dependency analysis is performed on the expected trajectory points of the end effector to generate a range penalty factor. The expected load mass data is processed by torque mapping using the physical dynamics feedback layer in the operation intent decoding layer to generate additional joint torque disturbance features. The additional joint torque perturbation features and the extension range penalty factor are input into a nonlinear multilayer perceptron to perform physical mapping processing on the arm movements corresponding to the dual-arm operation intention representation, thereby generating the expected centroid offset.

[0041] Preferably, the dual-arm operation intention representation includes the expected trajectory points of the end effector within a preset time window and the expected load quality data corresponding to each expected trajectory point of the end effector. The specific implementation process of using the multi-head attention mechanism in the operation intention decoding layer to perform spatial dependency analysis on the expected trajectory points of the end effector to generate the extension range penalty factor is as follows: In the operation task planning stage, a preset time window is first configured based on the duration of the target operation task, the control cycle of the dual-arm nodes, and the range of forward-looking actions that the underlying execution network can receive. The duration of the action is used to limit the time covered by the target operation task from the start to the end of the action. The control cycle of the dual-arm node is used to limit the sampling cycle of the expected trajectory point of the end effector. The range of forward-looking actions that the underlying execution network can receive is used to limit the range of dual-arm node actions that the top-level strategy network can read in advance within the current control cycle. The preset time window is configured by the duration of the action, the control cycle of the dual-arm node, and the range of forward-looking actions that the underlying execution network can receive. It is used to include the dual-arm node actions in several consecutive control cycles after the current control cycle into the same time sampling interval, so that the top-level strategy network can perceive the extension trend and load change trend of the dual-arm node in advance within the current control cycle. Subsequently, within the preset time window, the expected trajectory points of the end effector are read according to the control cycle. The technical essence of these expected trajectory points is the expected spatial position data of the end effectors of both arms relative to the torso reference position. These expected trajectory points include the expected position components of the end effectors in the forward / backward, left / right, and up / down directions. The expected trajectory points of the end effectors from multiple consecutive control cycles collectively express the arm movement path of the two arms within the preset time window. This arm movement path originates from the expected trajectory points of the end effectors from multiple consecutive control cycles and continues to participate in the formation of the operation point timing expression, enabling the operation point timing expression to retain the continuous spatial changes of the two arms within the preset time window. The operation intent decoding layer organizes the expected trajectory points of the end effectors into an operation point timing expression according to the control cycle sequence. This operation point timing expression then enters the multi-head attention mechanism. The technical essence of the multi-head attention mechanism is to read the mutual influence relationship between the expected trajectory points of the end effector from multiple spatially dependent observation apertures. Among them, one spatially dependent observation aperture focuses on reading the extension distance of the expected trajectory point of the end effector relative to the torso reference position, another spatially dependent observation aperture focuses on reading the lateral expansion relationship between the left and right arm nodes, and yet another spatially dependent observation aperture focuses on reading the influence relationship of the expected trajectory point of the end effector on the body attitude when it changes along the height direction.The multi-head attention mechanism performs spatial dependency analysis on the temporal representation of the operation points to form a point-dependency weighted representation. This weighted representation indicates the numerical contribution of the end-effector's desired trajectory point in different control cycles to the arm's range of motion. Subsequently, the operation intent decoding layer, based on the weighted representation, performs a weighted calculation on the extension distance of the end-effector's desired trajectory point relative to the torso reference position to generate an extension range penalty factor. The technical essence of this extension range penalty factor is to characterize whether the arm movements of the two arm nodes within the preset time window are close to the unfavorable region for body posture adjustment. This unfavorable region is jointly defined by the torso reference position, the extension range of the two arm nodes, and the body posture adjustment margin, providing a spatial boundary reference for the formation of the extension range penalty factor. The extension range penalty factor, along with additional joint torque perturbation features, is then input into a nonlinear multilayer perceptron, enabling the nonlinear multilayer perceptron to simultaneously consider the extension range influence from the end-effector's desired trajectory point and the load perturbation influence from the expected load mass data when generating the expected centroid offset.

[0042] Preferably, in the specific technical implementation of using the physical dynamics feedback layer in the operation intention decoding layer to perform torque mapping processing on the expected load mass data to generate additional joint torque disturbance features, after the expected trajectory points of the end effector are organized into the operation point timing expression, the expected load mass data corresponding to each expected trajectory point of the end effector is read synchronously. The technical essence of the expected load mass data is the mass data that the dual-arm nodes are expected to bear by the end effector in the target operation task. The expected load mass data corresponds one-to-one with the expected trajectory points of the end effector, so that each expected trajectory point of the end effector in each control cycle has a corresponding load source. The expected load mass data cannot be directly used as the expected center of mass offset because the expected load mass data only represents the load magnitude, while the expected center of mass offset needs to reflect the trend of the change in the body's center of mass caused by the combined action of the dual-arm nodes and the load. Therefore, the operation intention decoding layer sets up a physical dynamics feedback layer to convert the expected load mass data into additional joint torque disturbance features related to the changes in joint forces. In specific implementation, the physical dynamics feedback layer first reads the spatial distance between the expected trajectory point of each end effector and the torso reference position, and performs lever arm correlation processing with the corresponding expected load mass data to form a load lever arm expression. The spatial distance originates from the positional difference between the expected trajectory point of the end effector and the torso reference position. The spatial distance and the expected load mass data are jointly entered into the load lever arm expression, so that the load lever arm expression retains both the source of the load magnitude and the source of the spatial effect of the load relative to the torso reference position. The technical essence of the load lever arm expression is to convert the expected load mass data from a simple mass value into a spatial load expression that can affect the force on the joints of the two arm nodes. The load lever arm expression continues to participate in torque mapping processing to map the load changes in each control cycle to the torque change aperture of the shoulder joint, elbow joint, and wrist joint of the two arm nodes. The torque variation parameters of the shoulder, elbow, and wrist joints of the dual-arm node are derived from the joint connection relationship of the dual-arm node. This connection relationship is formed during the assembly and calibration phase of the dual-arm node and is used to represent the transmission relationship between the desired trajectory point of the end effector and the shoulder, elbow, and wrist joints of the dual-arm node. The torque mapping process does not directly generate the final control command. Instead, based on the desired trajectory point of the end effector, the expected load mass data, and the joint connection relationship of the dual-arm node, it calculates the disturbance contribution of the expected load mass data at different dual-arm node joints to generate additional joint torque disturbance characteristics.The technical essence of the additional joint torque perturbation feature is that it represents the additional changes in the joint torque requirements of the two arm nodes caused by the expected load mass data in the target operation task. Different elements in the additional joint torque perturbation feature correspond to the perturbation contributions of the shoulder joint, elbow joint, and wrist joint of the two arm nodes, respectively. The additional joint torque perturbation feature is then input together with the extension range penalty factor into a nonlinear multilayer perceptron, enabling the nonlinear multilayer perceptron to incorporate the joint torque perturbation caused by the expected load mass data and the extension range constraint caused by the expected trajectory point of the end effector into the same physical mapping process, thereby providing a load perturbation source for generating the expected centroid offset.

[0043] Preferably, the specific implementation process of inputting the additional joint torque perturbation features and the extension range penalty factor into a nonlinear multilayer perceptron to perform physical mapping processing on the arm movements corresponding to the dual-arm operation intention representation, and thus generate the expected centroid offset, is as follows: After the additional joint torque perturbation features and the extension range penalty factor are formed, the additional joint torque perturbation features are first processed by joint dimension adjustment, so that the additional joint torque perturbation features form a joint perturbation encoding expression according to the arrangement order of the shoulder joint, elbow joint, and wrist joint of the dual-arm nodes; at the same time, the extension range penalty factor is processed by scale adjustment, so that the extension range penalty factor is converted into an extension constraint encoding expression that can enter the nonlinear multilayer perceptron together with the joint perturbation encoding expression. The joint perturbation encoding expression retains the torque perturbation source formed by torque mapping processing of the expected load mass data, and the extension constraint encoding expression retains the extension constraint source formed by spatial dependency analysis processing of the expected trajectory points of the end effector. Subsequently, the joint perturbation encoding expression and the extension constraint encoding expression are concatenated to form a joint arm movement encoding expression, which is then input into a nonlinear multilayer perceptron. The technical essence of this joint arm movement encoding expression is to incorporate the load perturbation source corresponding to the additional joint torque perturbation feature and the extension constraint source corresponding to the extension range penalty factor into the same input expression, enabling the nonlinear multilayer perceptron to read the coupling relationship between the joint perturbation encoding expression and the extension constraint encoding expression during the same physical mapping process. The technical essence of the nonlinear multilayer perceptron is a numerical mapping structure formed by cascading multiple nonlinear mapping layers. The nonlinear mapping layer closest to the input side is used to read the coupling relationship between the joint perturbation encoding expression and the extension constraint encoding expression; the middle nonlinear mapping layer is used to convert the coupling relationship into contributions to the centroid change in the anterior-posterior, lateral, and vertical directions of the torso; and the nonlinear mapping layer closest to the output side is used to organize the centroid change contributions into the expected centroid offset. The essence of the physical mapping process is to convert the arm movements corresponding to the intention of dual-arm operation from the end-effector trajectory and load mass parameters into the body center of gravity offset parameter. The end-effector trajectory is formed by the desired trajectory points of the end effector, the load mass is formed by the expected load mass data, and the body center of gravity offset is carried by the expected center of gravity offset. Therefore, the physical mapping process can jointly convert the desired trajectory points of the end effector and the expected load mass data into the expected center of gravity offset. During this physical mapping process, the nonlinear multilayer perceptron does not process the additional joint torque perturbation features or the extension range penalty factor in isolation, but rather performs a joint mapping between the load perturbation source represented by the additional joint torque perturbation features and the extension constraint source represented by the extension range penalty factor. After processing by the nonlinear multilayer perceptron, the expected center of gravity offset is formed.The technical essence of the expected centroid offset is a vectorized data representing the predicted offset trend of the body's centroid relative to the torso reference position when the dual-arm nodes perform target manipulation tasks. The elements in the expected centroid offset correspond to the centroid change trends in the forward / backward, left / right, and up / down directions of the torso. The expected centroid offset is subsequently concatenated with the passage mask features to generate a joint state representation matrix. Furthermore, the expected centroid offset undergoes direction inversion to generate a whole-body centroid compensation vector. Therefore, by generating the expected centroid offset, the end-effector expected trajectory points and expected load mass data in the dual-arm manipulation intent representation can enter the quadruped primitive index determination process and the whole-body centroid compensation vector generation process, enabling the top-level policy network to simultaneously handle the arm movements of the dual-arm nodes and the body disturbance suppression requirements during the current control cycle's action allocation processing.

[0044] Optionally, the step of inputting the target basic control sequence and the whole-body center of mass compensation vector into the lower-level execution network, so that the lower-level execution network performs motion redirection processing on the target basic control sequence based on the whole-body center of mass compensation vector, thereby generating adaptive control instructions for the quadruped nodes and cooperative correction instructions for the bi-arm nodes, includes: A four-legged impedance branch and a two-arm admittance branch are configured in the underlying execution network; The target basic control sequence is input to the quadruped impedance branch to perform foot support force analysis on the target basic control sequence, thereby generating reference foot force data containing the desired contact force of each foot. The whole-body center of mass compensation vector is input into the dual-arm admittance branch to perform inverse kinematics processing on the whole-body center of mass compensation vector, thereby generating the desired angular acceleration data of each dual-arm node; The adaptive control command for the quadruped nodes is generated based on the reference foot force data, and the cooperative correction command for the biarm nodes is generated based on the desired angular acceleration data.

[0045] Preferably, the specific implementation process of configuring the quadruped impedance branch and the dual-arm admittance branch in the underlying execution network is as follows: After the top-level strategy network has generated the quadruped primitive index and the whole-body centroid compensation vector, and extracted the target basic control sequence corresponding to the quadruped primitive index from the preset parameterized primitive library, a branch configuration record is first established, and the target basic control sequence, the whole-body centroid compensation vector, the quadruped node, the dual-arm node, and the current control cycle marker are written into the branch configuration record. The branch configuration record is used to include the quadruped node displacement control source and the dual-arm node compensation correction source within the same current control cycle into the same execution caliber, so that the data received by the subsequent quadruped impedance branch and dual-arm admittance branch have a common source relationship. Subsequently, a quadruped impedance branch is configured according to the branch configuration record. The technical essence of this quadruped impedance branch is a data processing branch that adjusts impedance based on the support contact state and foot force state of the quadruped nodes. The quadruped impedance branch receives the target basic control sequence and forms reference foot force data based on the foot position data, foot velocity data, and joint drive signal sequence corresponding to the current action phase in the target basic control sequence. The foot position data, foot velocity data, and joint drive signal sequence all originate from the target basic control sequence. The foot position data characterizes the spatial position of the quadruped node in the current action phase; the foot velocity data characterizes the foot motion change of the quadruped node in the current action phase; and the joint drive signal sequence characterizes the basic drive source when the quadruped node performs a displacement task. The foot position data, foot velocity data, and joint drive signal sequence all enter the quadruped impedance branch so that the quadruped impedance branch can form the reference foot force data corresponding to the current action phase. The reference foot force data is subsequently used to generate the adaptive control command for the quadruped nodes. Therefore, the quadruped impedance branch is not an independent control structure, but rather a foot force analysis step before converting the target basic control sequence into executable control quantities for the quadruped nodes. Simultaneously, a dual-arm admittance branch is configured according to the branch configuration record. The technical essence of the dual-arm admittance branch is a data processing branch that adjusts admittance based on the compensation and correction requirements and center of mass compensation requirements of the dual-arm nodes. The dual-arm admittance branch receives the whole-body center of mass compensation vector and generates the desired angular acceleration data for each of the quadruped nodes based on the whole-body center of mass compensation vector.The dual-arm admittance branch and the quadrupedal impedance branch are configured in parallel within the same underlying execution network, but they are not unrelated parallel processing stages. The quadrupedal impedance branch connects to the target basic control sequence to generate the reference foot force data, and the dual-arm admittance branch connects to the whole-body center of mass compensation vector to generate the desired angular acceleration data. The reference foot force data and the desired angular acceleration data then jointly enter the instruction generation process corresponding to the instruction generation record. By configuring the quadrupedal impedance branch and the dual-arm admittance branch in the underlying execution network, the quadrupedal node displacement control source corresponding to the target basic control sequence and the dual-arm node compensation correction source corresponding to the whole-body center of mass compensation vector can enter their respective data processing branches within the same current control cycle, thereby ensuring a clear source relationship and coordination between the subsequently generated adaptive control instructions and the cooperative correction instructions.

[0046] Preferably, in the specific technical implementation of inputting the whole-body center of mass compensation vector into the dual-arm admittance branch to perform inverse kinematics processing on the whole-body center of mass compensation vector, thereby generating the expected angular acceleration data of each of the dual-arm nodes, the whole-body center of mass compensation vector is first read, and then the whole-body center of mass compensation vector is processed according to the torso's forward / backward, left / right, and up / down directions to form a compensation direction encoding expression. The whole-body center of mass compensation vector originates from the inversion of the expected center of mass offset. The compensation direction encoding expression is used to retain the compensation direction and compensation magnitude of each directional element in the whole-body center of mass compensation vector, so that the dual-arm admittance branch can read the dual-arm node compensation correction direction that should be used when the dual-arm nodes perform the target operation task. Subsequently, the dual-arm admittance branch reads the dual-arm node kinematic chain description, which is formed during the robot assembly and calibration phase and is used to represent the spatial transmission relationship between the shoulder, elbow, and wrist joints of the dual-arm nodes and the end effector. The dual-arm node kinematic chain description does not generate control results independently, but rather provides a joint space transformation calibrator for the whole-body center-of-mass compensation vector to enter inverse kinematics processing. This joint space transformation calibrator is used to convert the trunk direction element in the compensation direction encoding expression into joint space data for the shoulder, elbow, and wrist joints of the dual-arm nodes, enabling the compensation direction encoding expression to enter inverse kinematics processing along the dual-arm node kinematic chain description. Next, the dual-arm admittance branch performs inverse kinematics processing based on the compensation direction encoding expression and the dual-arm node kinematic chain description. The technical essence of this inverse kinematics processing is to convert the center-of-mass compensation direction at the trunk reference position into the trend of joint angular acceleration changes that each of the dual-arm nodes should produce in joint space. Specifically, the inverse kinematics processing first determines the compensation correction direction of the dual-arm node end effectors relative to the torso reference position based on the compensation direction encoding expression. Then, based on the dual-arm node kinematic chain description, the compensation correction direction is assigned to the shoulder, elbow, and wrist joints of the dual-arm nodes, thereby forming the desired angular acceleration data for each dual-arm node. The technical essence of the desired angular acceleration data is the joint angular acceleration target data that the dual-arm nodes need to generate in response to the whole-body center of mass compensation vector within the current control cycle. Different elements in the desired angular acceleration data correspond to the joint angular acceleration change trends of the shoulder, elbow, and wrist joints of the dual-arm nodes. The desired angular acceleration data is subsequently used to generate the cooperative correction command for the dual-arm nodes. Therefore, the desired angular acceleration data is not a repetitive description of the target object operation task, but rather intermediate data that converts the whole-body center of mass compensation vector into executable compensation correction actions for the dual-arm nodes.Through the inverse kinematics process described above, the whole-body center of mass compensation vector can be described along the kinematic chain of the two arm nodes and enter the joint space of each of the two arm nodes, so that the cooperative correction command has a joint angular acceleration source corresponding to the body disturbance suppression requirements when it is generated.

[0047] Preferably, the specific implementation process of generating the adaptive control command for the quadruped nodes based on the reference foot force data and the cooperative correction command for the biarm nodes based on the desired angular acceleration data is as follows: After the quadruped impedance branch has generated the reference foot force data according to the target basic control sequence, and the biarm admittance branch has generated the desired angular acceleration data according to the whole-body center of mass compensation vector, an instruction generation load record is first established, and the reference foot force data, the desired angular acceleration data, the target basic control sequence, the whole-body center of mass compensation vector, and the current control cycle marker are written into the instruction generation load record. The instruction generation load record is used to incorporate the foot force analysis results of the quadruped nodes and the joint angular acceleration correction results of the biarm nodes into the instruction generation scope of the same current control cycle, so that the adaptive control command and the cooperative correction command can correspond to the same motion redirection process. Subsequently, a bearing record is generated according to the instructions, reading the reference foot force data. The reference foot force data and the quadrupedal joint drive signal sequence in the target basic control sequence are then processed through instruction field conversion to form the adaptive control instruction for the quadrupedal node. The instruction field conversion process converts the foot support force requirement represented by the reference foot force data and the joint drive source represented by the quadrupedal joint drive signal sequence into instruction fields that the quadrupedal node can execute, enabling the adaptive control instruction to inherit the reference foot force data and the target basic control sequence. The technical essence of the adaptive control instruction is a control instruction that drives the quadrupedal node to perform a displacement task. The adaptive control instruction inherits the foot support force requirement represented by the reference foot force data and the basic displacement action content represented by the target basic control sequence, allowing the quadrupedal node to adjust its control according to the foot support state within the current control cycle when performing the displacement task. The basic displacement action content originates from the target basic control sequence and enters the quadrupedal node's displacement task execution process through the adaptive control instruction. Subsequently, based on the instructions, a bearing record is generated to read the desired angular acceleration data. The correspondence between the desired angular acceleration data and the direction dimension of the whole-body center of mass compensation vector is then processed through instruction field conversion to form the collaborative correction instruction for the two arm nodes. The direction dimension correspondence of the whole-body center of mass compensation vector represents the correspondence between the joint angular acceleration variation trend in the desired angular acceleration data and the compensation direction in the whole-body center of mass compensation vector. After the direction dimension correspondence is processed through instruction field conversion, the collaborative correction instruction can simultaneously accept both the desired angular acceleration data and the whole-body center of mass compensation vector.The technical essence of the cooperative correction instruction is a control instruction that superimposes compensation and correction actions when driving the dual-arm nodes to perform target manipulation tasks. The cooperative correction instruction inherits the joint angular acceleration variation trend represented by the desired angular acceleration data and the compensation direction source represented by the whole-body center-of-mass compensation vector, enabling the dual-arm nodes to adjust arm movements according to the body disturbance suppression requirements when performing target manipulation tasks. The adaptive control instruction and the cooperative correction instruction are directed to the quadruped node and the dual-arm node respectively, but both are generated by the same instruction carrying the reference foot force data and the desired angular acceleration data in the record. Therefore, they have a homogeneous motion redirection processing relationship within the current control cycle. Through the above processing, the reference foot force data can be converted into the adaptive control command, and the desired angular acceleration data can be converted into the cooperative correction command. This allows the underlying execution network to map the quadruped node displacement task and the bi-arm node target operation task into executable control commands without resolving all degrees of freedom control quantities. The adaptive control command and the cooperative correction command are used together to suppress the body disturbance caused by the target operation task.

[0048] Optionally, the step of inputting the target foundation control sequence to the quadruped impedance branch to perform foot support force analysis on the target foundation control sequence, thereby generating reference foot force data containing the desired contact force of each foot, includes: Extract the swing leg identifier and supporting leg identifier of the current action phase from the target basic control sequence; For the first quadruped node corresponding to the swing leg identifier, a zero-position contact force command is generated to release the ground contact constraint of the first quadruped node through the zero-position contact force command; For the second quadruped node corresponding to the support leg identifier, a nonlinear constraint optimization solver is used to perform feasible region projection processing on the expected friction angle variable in the target foundation control sequence to generate support reaction force components that meet the anti-slip conditions. The zero-position contact force command and the support reaction force component are merged and encapsulated to generate the reference foot force data containing the desired contact force of each foot end.

[0049] Preferably, the specific implementation process for extracting the swing leg identifier and supporting leg identifier of the current action phase from the target basic control sequence is as follows: After the target basic control sequence has been extracted from the preset parameterized primitive library by the quadruped primitive index, the current control cycle marker, phase sequence field, foot contact state field, foot position field, foot velocity field, and joint drive signal sequence in the target basic control sequence are read first, and a phase resolution record is established. The phase resolution record is used to include the action timing source, foot contact source, and joint drive source corresponding to the current control cycle in the target basic control sequence into the same resolution caliber, so that the subsequently extracted current action phase, swing leg identifier, and supporting leg identifier can all correspond to the same current control cycle. The technical essence of the current action phase is the quadrupedal node motion stage data corresponding to the target basic control sequence within the current control cycle. The current action phase is not a simple time label, but a foot motion state jointly defined by the phase sequence field and the foot contact state field, used to distinguish whether each foot in the quadrupedal node is in the off-ground swinging process or the ground-touching support process. In specific implementation, firstly, the current control cycle sequence segment is located from the target basic control sequence based on the current control cycle marker. Then, the phase order field in the current control cycle sequence segment is read to determine the current action phase. Subsequently, the foot contact state field in the current control cycle sequence segment is read based on the current action phase. The quadruped nodes representing the off-ground state in the foot contact state field are marked as the swing leg identifier, and the quadruped nodes representing the ground contact state in the foot contact state field are marked as the support leg identifier. The technical essence of the swing leg identifier is to point to the identification data of the quadruped nodes that need to release the ground contact constraint within the current control cycle. The swing leg identifier is subsequently used to determine the first quadruped node and to generate the zero-position contact force command. The technical essence of the support leg identifier is to point to the identification data of the quadruped nodes that need to bear the foot support force within the current control cycle. The support leg identifier is subsequently used to determine the second quadruped node and to participate in the feasible domain projection processing of the expected friction angle variable. Through the above processing, the current action phase associates the current control cycle sequence segment in the target basic control sequence with the contact state of the four-legged node. The swing leg identifier and the supporting leg identifier further divide the four-legged node in the current control cycle into different foot end force processing objects. This allows the four-legged impedance branch to distinguish between the first four-legged node that needs to generate the zero-position contact force command and the second four-legged node that needs to generate the support reaction force component before generating the reference foot end force data. This avoids using the same foot end force analysis caliber for the foot end corresponding to the first four-legged node and the foot end corresponding to the second four-legged node in the same current control cycle.

[0050] Preferably, the specific implementation process of generating a zero-position contact force command for the first quadruped node corresponding to the swing leg identifier, so as to release the ground contact constraint of the first quadruped node through the zero-position contact force command, is as follows: After the swing leg identifier has been extracted from the foot contact state field in the target basic control sequence, the foot position field, foot velocity field, and joint drive signal sequence corresponding to the swing leg identifier in the target basic control sequence are read first, and the swing leg identifier, the current action phase, the foot position field, the foot velocity field, and the joint drive signal sequence are written into the swing leg contact release record. The swing leg contact release record is used to correspond the source of the first quadruped node's off-ground swing process with the source of the subsequent zero-position contact force command, so that the zero-position contact force command can use the source of the first quadruped node's off-ground swing process in the target basic control sequence, instead of applying contact force zeroing encoding to the first quadruped node in isolation. The technical essence of the first quadruped node is that the swing leg identifier points to the quadruped node in the current control cycle. The first quadruped node needs to perform the ground-lifting swing process according to the foot position field and foot velocity field in the target base control sequence. The first quadruped node does not bear the ground support reaction force during the ground-lifting swing process, therefore the first quadruped node should not enter the solution process of the support reaction force component. Subsequently, the quadruped impedance branch reads the foot position field and foot velocity field corresponding to the first quadruped node according to the swing leg contact release record, and performs contact force zeroing encoding processing on the foot contact channel of the first quadruped node to generate the zero-position contact force command. The technical essence of the zero-position contact force command is to set the expected contact force of each foot end of the corresponding foot contact channel of the first quadruped node to zero-position contact force control data. The zero-position contact force command does not cancel the ground-lifting swing process of the first quadruped node, but rather ensures that the first quadruped node, while maintaining the source of the ground-lifting swing process given by the target base control sequence, is not misjudged by the foot support force analysis processing as a quadruped node that needs to bear the support reaction force component. The technical essence of the ground contact constraint is the contact condition between the foot tip of the quadruped node and the ground contact surface, used to limit the foot tip's slippage, penetration, and lift-off states. For the first quadruped node, the ground contact constraint needs to be released within the current action phase, allowing the first quadruped node to complete the lift-off swing process according to the foot tip position field and the foot tip velocity field. The zero-position contact force command will subsequently be merged and encapsulated with the support reaction force component to generate reference foot tip force data containing the desired contact force of each foot tip. Therefore, the zero-position contact force command is not an independent control output result, but rather the desired contact force field of each foot tip corresponding to the swing leg identifier in the reference foot tip force data.By generating the zero-position contact force command for the first quadruped node, the quadruped impedance branch can retain the ground-free state of the first quadruped node in the same reference foot force data, so that when the adaptive control command is generated based on the reference foot force data, the first quadruped node will not be configured as a quadruped node that bears the support reaction force component.

[0051] Preferably, for the second quadruped node corresponding to the supporting leg identifier, a nonlinear constraint optimization solver is used to perform feasible region projection processing on the expected friction angle variable in the target base control sequence to generate a supporting reaction force component that satisfies the anti-slip condition. The zero-position contact force command and the supporting reaction force component are merged and encapsulated to generate the reference foot force data containing the expected contact force of each foot. The specific implementation process is as follows: After the supporting leg identifier has been extracted from the foot contact state field in the target base control sequence, the foot position field, foot velocity field, expected friction angle variable, and joint drive signal sequence corresponding to the supporting leg identifier in the target base control sequence are read according to the supporting leg identifier. The supporting leg identifier, the current action phase, the foot position field, the foot velocity field, the expected friction angle variable, and the joint drive signal sequence are written into the supporting leg force analysis record. The technical essence of the second quadruped node is that the supporting leg identifier points to the quadruped node in the current control cycle. The difference between the second and first quadruped nodes is that the second quadruped node needs to maintain ground contact and bear the foot support force in the current action phase, while the first quadruped node needs to release the ground contact constraint and perform the off-ground swing process. The technical essence of the expected friction angle variable is the friction constraint parameter in the target basic control sequence used to characterize the allowable deviation between the tangential and normal forces on the foot contact surface. The expected friction angle variable, together with the foot position field and foot velocity field of the second quadruped node, determines the candidate range of the support reaction force. Subsequently, the nonlinear constraint optimization solver reads the force analysis record of the supporting leg and constructs the candidate range of support reaction force based on the expected friction angle variable, the foot position field of the second quadruped node, and the foot velocity field of the second quadruped node. The technical essence of the nonlinear constraint optimization solver is a data solution structure that iteratively filters the candidate range of support reaction force under the conditions of friction constraints, foot force direction constraints, and joint drive signal sequence constraints. Its function is not to regenerate complete whole-body joint control quantities, but rather, given that the target basic control sequence has already provided the source of the basic movement, to determine the support reaction force component that meets the anti-slip condition for the second quadruped node. The technical essence of the feasible region projection processing is to adjust the candidate support reaction forces that exceed the friction constraint boundary defined by the expected friction angle variable within the candidate range of support reaction force to within the friction constraint feasible region, and to retain the candidate support reaction forces that match the foot position field and foot velocity field of the second quadruped node, thereby forming the support reaction force component that meets the anti-slip condition.The technical essence of the anti-slip condition is a force constraint that limits the tangential force on the foot tip of the second quadruped node to within the friction range that the foot contact surface can withstand. This anti-slip condition is determined by the expected friction angle variable and enters the generation process of the support reaction force component through the feasible domain projection process. The technical essence of the support reaction force component is the expected contact force component of each foot tip that the second quadruped node needs to form within the current action phase. This support reaction force component is subsequently merged and encapsulated with the zero-position contact force command to generate the reference foot tip force data containing the expected contact forces of each foot tip. In the specific encapsulation process, firstly, a foot-end force encapsulation record is established according to the foot-end arrangement order of the four-legged nodes. This record carries the zero-position contact force command corresponding to the first four-legged node and the support reaction force component corresponding to the second four-legged node, according to the foot-end arrangement order. Then, the zero-position contact force command corresponding to the swing leg identifier is written into the swing foot-end field of the foot-end force encapsulation record, and the support reaction force component corresponding to the support leg identifier is written into the support foot-end field. The swing foot-end field carries the expected contact force fields for each foot-end corresponding to the first four-legged node, and the support foot-end field carries the expected contact force fields for each foot-end corresponding to the second four-legged node. Subsequently, the foot-end force encapsulation records are merged and encapsulated to form the reference foot-end force data. The technical essence of the expected contact force of each foot-end is the expected force data corresponding to the foot-end of each four-legged node in the reference foot-end force data. The expected contact force of each foot-end includes both the zero-position contact force command corresponding to the first four-legged node and the support reaction force component corresponding to the second four-legged node. The technical essence of the reference foot force data is that the quadruped impedance branch forms a foot force expression based on the current action phase, the swing leg identifier, the supporting leg identifier, the zero-position contact force command, and the supporting reaction force component. The reference foot force data is subsequently used to generate the adaptive control command for the quadruped node, enabling the quadruped node to distinguish the force requirements of the first quadruped node and the second quadruped node when performing displacement tasks.

[0052] Optionally, driving the quadruped node to perform a displacement task based on the adaptive control command includes: Extract the desired joint position data, desired joint velocity data, and feedforward torque data contained in the adaptive control command; The actual joint position data of the four-legged nodes is captured using the underlying motor driver, and the difference between the actual joint position data and the expected joint position data is calculated to generate the position deviation. The actual joint velocity data of the four-legged nodes is captured using the underlying motor driver, and the difference between the actual joint velocity data and the expected joint velocity data is calculated to generate a velocity deviation. The position deviation and the velocity deviation are input to the proportional-derivative controller for closed-loop regulation, thereby generating feedback compensation torque data. Based on the feedforward torque data and the feedback compensation torque data, the final motor control current is generated; The final motor control current is injected into the corresponding joint motor to drive the quadruped nodes to perform the displacement task.

[0053] Preferably, the specific implementation process for extracting the desired joint position data, desired joint velocity data, and feedforward torque data contained in the adaptive control command is as follows: After the adaptive control command has been generated from the reference foot force data and the target basic control sequence, the current control cycle marker, quadruped node identifier, joint arrangement field, foot force source field, and joint control field in the adaptive control command are read first, and an adaptive control command parsing record is established. The adaptive control command parsing record is used to associate the adaptive control command with the reference foot force data, the target basic control sequence, and the current control cycle marker on which the adaptive control command is based, so that the subsequently extracted desired joint position data, desired joint velocity data, and feedforward torque data can maintain the same current control cycle and the same quadruped node arrangement. The foot force source field is used to record the source of the reference foot force data in the adaptive control command, so that the feedforward torque data can be traced back to the foot force analysis result corresponding to the reference foot force data. The joint control field is used to carry the desired joint position data, the desired joint velocity data, and the feedforward torque data, so that the adaptive control command parsing record can distinguish and read the control data of different quadruped nodes according to the joint arrangement field. In specific implementation, firstly, based on the quadruped node identifier and the joint arrangement field, the joint control sub-field corresponding to each quadruped node is located from the joint control field; then, the position target field is read from the joint control sub-field to form the desired joint position data, the velocity target field is read from the joint control sub-field to form the desired joint velocity data, and the torque feedforward field is read from the joint control sub-field to form the feedforward torque data. The position target field is used to provide the field source of the desired joint position data, the velocity target field is used to provide the field source of the desired joint velocity data, and the torque feedforward field is used to provide the field source of the feedforward torque data. After completing the reading of the desired joint position data, the desired joint velocity data, and the feedforward torque data, the joint control subfield continues to maintain the correspondence between the record and the current control cycle marker through the adaptive control instruction parsing.The technical essence of the expected joint position data is the target angular position data that each joint of the quadruped node should reach within the current control cycle. The expected joint position data originates from the basic displacement action content in the target basic control sequence and has been adapted through the foot force analysis results corresponding to the reference foot force data. The technical essence of the expected joint velocity data is the target angular velocity data that each joint of the quadruped node should form within the current control cycle. The expected joint velocity data is used to calculate the difference with the subsequently captured actual joint velocity data to determine the velocity following state of the quadruped node when performing displacement tasks. The technical essence of the feedforward torque data is the joint torque compensation data pre-given by the quadruped node within the current control cycle based on the reference foot force data and the target basic control sequence. The feedforward torque data is used to provide a basic driving source before the feedback compensation torque data is formed. The desired joint position data, the desired joint velocity data, and the feedforward torque data are subsequently incorporated into the position deviation generation, velocity deviation generation, and final motor control current generation processes, respectively. Therefore, the desired joint position data, the desired joint velocity data, and the feedforward torque data are not isolated instruction fields, but rather three continuous control sources that drive the quadruped node to perform displacement tasks, as described by the adaptive control instructions.

[0054] Preferably, the specific implementation process of capturing the actual joint position data of the quadruped nodes using a low-level motor driver and calculating the difference between the actual joint position data and the desired joint position data to generate the position deviation is as follows: After the desired joint position data has been extracted from the adaptive control command, the quadruped node identifier and the joint arrangement field in the adaptive control command parsing record are used to determine the joint reading order of the quadruped nodes that need to perform position following, and the low-level motor driver captures the actual joint position data of the quadruped nodes according to the joint reading order. The joint reading order comes from the joint arrangement field, and the joint reading order is used to correspond the sampling order of the actual joint position data with the field order of the desired joint position data, so that the actual joint position data and the desired joint position data can be used to perform difference calculation under the same joint diameter. The technical essence of the low-level motor driver is a data driving unit that performs drive interaction and state sampling with the corresponding joint motor of the quadruped node. The low-level motor driver reads the coded position feedback value of the corresponding joint motor in the current control cycle and organizes the coded position feedback value into the actual joint position data according to the joint arrangement field. The coded position feedback value provides the sampling source for the actual joint position data. After processing, the actual joint position data retains the quadruped node identifier and joint arrangement field corresponding to the coded position feedback value. The technical essence of the actual joint position data is the angular position state data of each joint of the quadruped node within the current control cycle. The actual joint position data and the desired joint position data have the same joint arrangement diameter and angular position dimension, thus allowing for difference calculation. To avoid field misalignment between different quadruped nodes or joints, the actual joint position data is first written into a joint position feedback record, which includes the current control cycle marker, the quadruped node identifier, and the joint arrangement field. Subsequently, the actual joint position data and the desired joint position data are aligned joint by joint according to the joint position feedback record, ensuring that each actual joint position element corresponds to the desired joint position element of the same joint. After joint-by-joint alignment, the difference between the actual joint position data and the desired joint position data is calculated to generate the position deviation. The technical essence of the position deviation is a data result used to represent the degree of deviation of the actual angular position of each joint of the quadruped node from the expected angular position in the current control cycle. Each element in the position deviation corresponds to the angular position difference of a joint.The position deviation will subsequently be input to the proportional-derivative controller along with the velocity deviation, and participate in the generation process of feedback compensation torque data. Therefore, the position deviation is not simply a state monitoring result, but a feedback source used to correct joint position following errors when the adaptive control command drives the quadruped node to perform displacement tasks.

[0055] Preferably, the specific implementation process of using the underlying motor driver to capture the actual joint speed data of the quadruped node and calculating the difference between the actual joint speed data and the desired joint speed data to generate the speed deviation is as follows: After the desired joint speed data has been extracted from the adaptive control command and the actual joint position data has been captured according to the joint arrangement field, the underlying motor driver continues to capture the actual joint speed data of the quadruped node according to the same quadruped node identifier and the same joint arrangement field. The actual joint speed data can be formed by the speed feedback value of the corresponding joint motor, or by time difference of the actual joint position data of adjacent control cycles; regardless of which formation method is used, the actual joint speed data is arranged according to the joint arrangement field so that the actual joint speed data and the desired joint speed data maintain the same joint sequence. The speed feedback value is used to provide a direct sampling source for the actual joint speed data, and the actual joint position data of adjacent control cycles is used to provide a source for differential calculation of the actual joint speed data. After being formed, the actual joint speed data continues to maintain a correspondence with the current control cycle marker, the quadruped node identifier, and the joint arrangement field. The technical essence of the actual joint velocity data is the angular velocity state data of each joint of the quadruped node within the current control cycle. This actual joint velocity data reflects the rate of motion change of the quadruped node when performing a displacement task. The technical essence of the desired joint velocity data is the angular velocity target data given in the adaptive control command. This desired joint velocity data constrains the motion change trend of each joint of the quadruped node. Specifically, the actual joint velocity data is first written into a joint velocity feedback record, which includes the current control cycle marker, the quadruped node identifier, the joint arrangement field, and the velocity sampling aperture. The velocity sampling aperture records whether the actual joint velocity data originates from the rotational speed feedback value or from the actual joint position data of the adjacent control cycle, ensuring that the formation method of the actual joint velocity data can be reused by subsequent feedback adjustment records. Then, based on the joint velocity feedback record, the actual joint velocity data and the desired joint velocity data are aligned joint by joint, and the difference is calculated after alignment to generate the velocity deviation. The technical essence of the velocity deviation is a data result used to represent the degree of deviation of the actual angular velocity of each joint of the quadruped node from the expected angular velocity in the current control cycle. Each element in the velocity deviation corresponds to the angular velocity difference of a joint.The velocity deviation and the position deviation play different data roles in the subsequent proportional-derivative (PD) controller. The position deviation mainly reflects the deviation of the joint angular position target, while the velocity deviation mainly reflects the change in the joint angular velocity target. By transmitting both the velocity deviation and the position deviation to the PDC controller, the quadruped nodes can simultaneously generate feedback compensation torque data based on both joint position and joint velocity states when performing displacement tasks. This avoids the loss of a source of joint velocity change due to relying solely on joint position states to correct rapid motion.

[0056] Preferably, the specific implementation process of inputting the position deviation and the velocity deviation to the proportional-derivative controller for closed-loop regulation to generate feedback compensation torque data is as follows: After the position deviation and the velocity deviation are formed, a feedback regulation load record is first established, and the position deviation, the velocity deviation, the current control cycle marker, the quadruped node identifier, and the joint arrangement field are written into the feedback regulation load record. The feedback regulation load record is used to correspond the source of position deviation and the source of velocity deviation under the same joint arrangement within the same current control cycle, so that the proportional-derivative controller can read the position deviation and the velocity deviation in the same joint sequence. The feedback regulation load record also reads the joint position feedback record and the joint velocity feedback record, so that the position deviation can be traced back to the difference calculation process between the actual joint position data and the desired joint position data, and the velocity deviation can be traced back to the difference calculation process between the actual joint velocity data and the desired joint velocity data. The technical essence of the proportional-derivative controller is a data adjustment unit that converts position deviation contributions and velocity deviation contributions into joint torque compensation amounts. The proportional adjustment parameters and derivative adjustment parameters in the proportional-derivative controller are configured during the quadruped node assembly and calibration stage or the control parameter initialization stage. The proportional adjustment parameters are used to limit the contribution of the position deviation amount to the joint torque compensation amount, and the derivative adjustment parameters are used to limit the contribution of the velocity deviation amount to the joint torque compensation amount. Both the proportional adjustment parameters and the derivative adjustment parameters are configured according to the quadruped node identifier and the joint arrangement field to avoid using the same adjustment parameters for different joints. In specific implementation, the proportional-derivative controller first reads the position deviation based on the feedback adjustment load record, and performs proportional adjustment processing on the position deviation according to the proportional adjustment parameters to form a position feedback torque component. Then, the proportional-derivative controller reads the velocity deviation based on the feedback adjustment load record, and performs differential adjustment processing on the velocity deviation according to the derivative adjustment parameters to form a velocity feedback torque component. Next, the position feedback torque component and the velocity feedback torque component are synthesized according to the same joint arrangement field to generate the feedback compensation torque data. The technical essence of this closed-loop adjustment processing is to form a difference feedback between the desired joint position data, the desired joint velocity data, and the actual joint position data and the actual joint velocity data, and to convert the difference feedback into torque compensation data that can participate in joint actuation.The technical essence of the feedback compensation torque data is that it is the joint torque correction data formed based on the position deviation and the speed deviation. The feedback compensation torque data will subsequently participate in the final motor control current generation process together with the feedforward torque data. Therefore, the feedback compensation torque data is not an independent replacement of the control quantity of the feedforward torque data, but is used to compensate for the joint state deviation in the current control cycle based on the feedforward torque data.

[0057] Preferably, the specific implementation process for generating the final motor control current based on the feedforward torque data and the feedback compensation torque data is as follows: After the feedforward torque data has been extracted by the adaptive control command parsing record and the feedback compensation torque data has been generated by the proportional-derivative controller, a motor control current generation record is first established, and the feedforward torque data, the feedback compensation torque data, the current control cycle marker, the quadruped node identifier, and the joint arrangement field are written into the motor control current generation record. The motor control current generation record is used to incorporate the basic torque source represented by the feedforward torque data and the deviation correction source represented by the feedback compensation torque data into the same motor control current generation caliber, so that the final motor control current can simultaneously bear the basic displacement action content corresponding to the target basic control sequence and the joint state deviation of the quadruped node in the current control cycle. In specific implementation, the feedforward torque data and the feedback compensation torque data are first aligned joint by joint according to the joint arrangement field, so that each feedforward torque element corresponds to the feedback compensation torque element of the same joint. Then, the feedforward torque data and the feedback compensation torque data that have completed joint-by-joint alignment are subjected to torque cascading configuration processing to form target joint torque data. This torque cascading configuration processing is not a simple superposition of data from different sources, but rather first retains the basic drive contribution from the reference foot force data and the target basic control sequence in the feedforward torque data, and then writes the correction contribution from the position deviation and the velocity deviation in the feedback compensation torque data into the same joint torque caliber, thereby forming the target joint torque data corresponding to each joint. The target joint torque data is then further processed by current mapping according to the torque-current conversion relationship of the corresponding joint motor to generate the final motor control current. The torque-current conversion relationship is configured during the assembly and calibration phase of the corresponding joint motor. This relationship represents the conversion between the target joint torque data and the input current of the corresponding joint motor, and is stored in correspondence with the quadruped identifier and the joint arrangement field. The input current of the corresponding joint motor provides the current mapping for the final motor control current. After generation, the final motor control current continues to maintain a correspondence with the corresponding joint motor according to the quadruped identifier and the joint arrangement field. The technical essence of the final motor control current is to drive the corresponding joint motor to output motor control data that matches the target joint torque data. Each current element in the final motor control current corresponds to one corresponding joint motor. The final motor control current is subsequently injected into the corresponding joint motor to drive the quadruped to perform the displacement task. Therefore, the final motor control current is not simply calculated from the feedforward torque data, but rather is motor control data formed by the feedforward torque data and the feedback compensation torque data.

[0058] Preferably, the specific implementation process of injecting the final motor control current into the corresponding joint motor to drive the quadruped node to perform the displacement task is as follows: After the final motor control current is formed, the corresponding joint motor corresponding to each current element in the final motor control current is determined based on the quadruped node identifier and the joint arrangement field in the motor control current generation record, and a current injection execution record is established. The current injection execution record is used to map the final motor control current, the corresponding joint motor, the current control cycle marker, and the execution relationship between the quadruped nodes, so that the final motor control current can enter the correct corresponding joint motor according to the joint arrangement field. The technical essence of the corresponding joint motor is an execution component that is mechanically connected to each joint in the quadruped node and is used to output joint driving torque. After receiving the final motor control current, the corresponding joint motor converts the final motor control current into the joint driving torque of the corresponding joint. In specific implementation, the final motor control current is first processed by joint channel allocation according to the current injection execution record to form a joint current allocation expression. This expression establishes a one-to-one correspondence between each current element in the final motor control current and its corresponding joint motor, preventing joint channel mismatch during injection. Subsequently, each current element in the joint current allocation expression is injected into its corresponding joint motor, causing the corresponding joint motor to generate the joint driving torque based on the final motor control current. Next, the joint driving torque drives the quadruped nodes to perform joint movements according to the desired joint position data and the desired joint velocity data, and adjusts the foot displacement state of the quadruped nodes through a driving source formed by the feedforward torque data and the feedback compensation torque data. The foot displacement state represents the spatial movement result of the quadruped nodes under the joint driving torque. This foot displacement state continues to participate in the execution of the displacement task, allowing the quadruped nodes to utilize the displacement control source in the adaptive control command. The technical essence of the displacement task is that the quadruped node completes the foot movement, foot support switching, or torso posture coordination movement process according to the adaptive control command within the current control cycle. The displacement task is not an action result independent of the adaptive control command, but is the quadruped node execution result formed by the continuous participation of the expected joint position data, the expected joint velocity data, the feedforward torque data, the actual joint position data, the actual joint velocity data, the position deviation, the velocity deviation, the feedback compensation torque data, and the final motor control current.After the quadruped node performs the displacement task, it can participate in the body disturbance suppression process together with the cooperative correction instructions formed when the dual-arm node performs the target operation task. This allows the displacement execution of the quadruped node to follow the motion redirection processing result of the underlying execution network and maintain control connection with the target operation task of the dual-arm node under the same current control cycle.

[0059] Optionally, the step of driving the dual-arm nodes to perform target manipulation tasks based on the cooperative correction instructions to suppress machine disturbances caused by the target manipulation tasks includes: Read the physical property label data configured on the surface of the target object, and parse the physical property label data to obtain the target stiffness level parameter of the target object; The safe torque threshold range is determined based on the target stiffness level parameters, and it is determined whether the preset clamping torque data contained in the collaborative correction instruction exceeds the safe torque threshold range. If the judgment result indicates that the preset clamping torque data exceeds the safe torque threshold range, then the preset clamping torque data is subjected to amplitude limiting and truncation processing to generate safe correction torque data, and the dual arm nodes are driven to perform compliant contact actions based on the safe correction torque data. If the judgment result indicates that the preset clamping torque data does not exceed the safe torque threshold range, then the dual-arm node is driven to perform a rigid clamping action based on the preset clamping torque data; The force feedback generated by the compliant contact action or the rigid clamping action is used to suppress the machine body disturbance caused by the target object operation task.

[0060] Preferably, the specific implementation process of reading the physical attribute label data configured on the surface of the target object and parsing the physical attribute label data to obtain the target stiffness level parameters of the target object is as follows: After the cooperative correction command has been generated by the desired angular acceleration data and the dual-arm node has entered the contact preparation stage corresponding to the target object operation task, the current control cycle flag, dual-arm node identifier, end effector identifier, clamping posture field, and preset clamping torque data in the cooperative correction command are read first, and a target object attribute reading load record is established. The target object attribute reading load record is used to include the dual-arm node execution source, end effector contact source, and preset clamping torque data source in the current target object operation task into the same reading caliber, so that the subsequently read physical attribute label data can correspond to the same execution process of the cooperative correction command. Subsequently, according to the contact direction corresponding to the end effector identifier, the physical attribute label data configured on the target surface is read as the dual-arm node approaches the target surface. This physical attribute label data can come from a machine-readable label attached to the target surface or from a graphical attribute mark printed on the target surface. Both the machine-readable label and the graphical attribute mark are used to carry the physical attribute label data. The physical attribute label data is configured before the target enters the operation preparation stage corresponding to the target operation task and includes the target material code, target contact area code, stiffness grade code, deformation tolerance grade code, and label verification field. The technical essence of this physical attribute label data is to convert the contact attributes of the target under clamping force into readable data fields, enabling the dual-arm node to establish a correspondence between the target contact area and the preset clamping torque data source in the cooperative correction instruction before executing the target operation task. After reading the physical attribute label data, the physical attribute label data is first written into the target object attribute reading and bearing record, and the field integrity of the physical attribute label data is checked according to the label verification field. If the field integrity check indicates that the physical attribute label data can be parsed, the stiffness level code and the deformation tolerance level code are read from the physical attribute label data, and the stiffness level code and the deformation tolerance level code are mapped to obtain the target stiffness level parameter. The technical essence of the target stiffness level parameter is a data-driven expression of the resistance of the target object to changes in clamping torque within the contact area of ​​the target object. The target stiffness level parameter is not simply a material name, nor is it a subjective description of the hardness or softness of the target object, but rather a source of torque constraint obtained by parsing the stiffness level code and the deformation tolerance level code from the physical attribute label data.The target stiffness level parameter will subsequently serve as the source of torque constraints for determining the safe torque threshold range. Therefore, the target stiffness level parameter connects the process of reading the contact attributes of the target surface with the process of judging the preset clamping torque data in the collaborative correction instruction, so that when the dual-arm node performs the target operation task, it can select a compliant contact action or a rigid clamping action based on the contact attributes of the target.

[0061] Preferably, in the specific technical implementation of determining the safe torque threshold range based on the target stiffness level parameter and determining whether the preset clamping torque data included in the collaborative correction instruction exceeds the safe torque threshold range, after the target stiffness level parameter has been parsed from the physical attribute label data, a clamping torque determination bearing record is first established, and the target stiffness level parameter, the physical attribute label data, the collaborative correction instruction, the preset clamping torque data, the current control cycle marker, the dual-arm node identifier, and the end effector identifier are written into the clamping torque determination bearing record. The clamping torque determination bearing record is used to include the contact attribute source of the target object and the source of the preset clamping torque data of the collaborative correction instruction in the same judgment criterion, avoiding the preset clamping torque data from being directly sent to the dual-arm node without deviating from the target stiffness level parameter. The technical essence of the safe torque threshold range is the allowable clamping torque range determined based on the target stiffness level parameter. The safe torque threshold range includes a lower torque boundary and an upper torque boundary. The lower torque boundary limits the minimum clamping torque required for the dual-arm node to stably contact the target object, while the upper torque boundary limits the maximum clamping torque that the dual-arm node can use without exceeding the allowable deformation range of the target object. Specifically, the torque threshold configuration representation corresponding to the contact area of ​​the target object is first read based on the target stiffness level parameter. This torque threshold configuration representation is configured during the operation task planning stage based on the target object material code, target object contact area code, stiffness level code, and deformation tolerance level code, and is used to write the allowable torque range under different stiffness levels into the same data caliber. Subsequently, the lower torque boundary and the upper torque boundary corresponding to the target stiffness level parameter are read based on the torque threshold configuration representation to form the safe torque threshold range. The preset clamping torque data is the initial target clamping torque data configured during the operation task planning phase based on the expected trajectory points of the end effector, expected load mass data, clamping posture fields, and the joint reachability range of the dual-arm nodes. It enters the current control cycle along with the collaborative correction command. Technically, the preset clamping torque data is the initial target clamping torque data formed by the dual-arm nodes before they are combined with the physical attribute label data on the target object surface. Subsequently, the preset clamping torque data and the safe torque threshold range are subjected to boundary judgment processing under the same torque dimension. When the preset clamping torque data is higher than the upper torque boundary or lower than the lower torque boundary, a judgment result is generated indicating that the preset clamping torque data exceeds the safe torque threshold range; when the preset clamping torque data is between the lower torque boundary and the upper torque boundary, a judgment result is generated indicating that the preset clamping torque data does not exceed the safe torque threshold range.The judgment result is subsequently processed into either the amplitude limiting and truncation process or the rigid clamping action generation process. Therefore, the judgment result is not an independent discrimination information, but rather the control basis for the clamping torque diversion based on the target stiffness level parameter before the execution of the collaborative correction instruction.

[0062] Preferably, if the judgment result indicates that the preset clamping torque data exceeds the safe torque threshold range, the preset clamping torque data is subjected to amplitude limiting and truncation processing to generate safe correction torque data. The specific implementation process of driving the dual-arm nodes to perform compliant contact actions based on the safe correction torque data is as follows: After the judgment result indicates that the preset clamping torque data exceeds the safe torque threshold range, the preset clamping torque data, the safe torque threshold range, the target stiffness level parameter, the clamping posture field, the dual-arm node identifier, and the end effector identifier are read from the clamping torque judgment load record, and a compliant contact control load record is established. The compliant contact control load record is used to match the source of the exceeding of the preset clamping torque data, the boundary source of the safe torque threshold range, and the execution source of the dual-arm nodes, so that the subsequently formed safe correction torque data can correspond to the cooperative correction command in the current control cycle. Subsequently, the preset clamping torque data is subjected to amplitude limiting and truncation processing. The essence of this amplitude limiting and truncation processing is to adjust the preset clamping torque data that exceeds the safe torque threshold range to the torque boundary diameter within the safe torque threshold range. When the preset clamping torque data is higher than the upper limit torque boundary, the preset clamping torque data is adjusted to torque data that is not higher than the upper limit torque boundary. When the preset clamping torque data is lower than the lower limit torque boundary, the preset clamping torque data is adjusted to torque data that is not lower than the lower limit torque boundary. After the amplitude limiting and truncation process, the safety correction torque data is formed. The technical essence of this safety correction torque data is that, while retaining the clamping direction and clamping posture source in the collaborative correction command, the preset clamping torque data is constrained by boundary conditions based on the target stiffness level parameters. This safety correction torque data is not a regenerated object operation task, nor does it replace all control content of the collaborative correction command; rather, it is used to replace the preset clamping torque data in the collaborative correction command that exceeds the safety torque threshold range. Subsequently, joint drive correction data corresponding to the compliant contact action is generated based on the safety correction torque data and the clamping posture field, and this joint drive correction data is written into the compliant contact control bearing record. The joint drive correction data is used to convert the safety correction torque data into the joint execution aperture of the dual-arm nodes, enabling the compliant contact action to continue to inherit the clamping posture field in the collaborative correction command. The technical essence of the compliant contact action is that when the dual-arm nodes contact the target object, they gradually establish the clamping contact force according to the safety correction torque data. The compliant contact action does not change the contact position of the end effector of the target object operation task, but reduces the sudden change of clamping torque when the end effector of the dual-arm nodes contacts the surface of the target object.The safety correction torque data drives the dual-arm nodes to perform the compliant contact action. The cooperative correction command can continue to be executed under the constraint of the target contact attribute, and the force feedback generated by the compliant contact action is fed into the subsequent body disturbance suppression process.

[0063] Preferably, if the judgment result indicates that the preset clamping torque data does not exceed the safe torque threshold range, the specific implementation process of driving the dual-arm node to perform a rigid clamping action based on the preset clamping torque data is as follows: After the judgment result indicates that the preset clamping torque data does not exceed the safe torque threshold range, the preset clamping torque data, the safe torque threshold range, the target stiffness level parameter, the clamping posture field, the dual-arm node identifier, and the end effector identifier are read from the clamping torque judgment load record, and a rigid clamping control load record is established. The rigid clamping control load record is used to associate the judgment result that the preset clamping torque data is within the safe torque threshold range with the clamping execution source of the dual-arm node, so that the preset clamping torque data can enter the rigid clamping action generation process without undergoing the amplitude limiting truncation process. The configuration method for the preset clamping torque data is as follows: During the operation task planning phase, the end effector contact position of the dual-arm node is determined based on the expected trajectory point of the end effector; the load source in the target operation task is determined based on the expected load mass data; the contact direction corresponding to the end effector identifier is determined based on the clamping posture field; and then the end effector contact position, the load source, and the contact direction are written into the initial clamping torque configuration record to form the preset clamping torque data that can be issued with the cooperative correction command. The initial clamping torque configuration record is used to illustrate that the preset clamping torque data is not an arbitrarily set torque value, but rather the initial target clamping torque data obtained by the constraints of the dual-arm operation intention representation, the expected angular acceleration data, and the clamping posture field. The technical essence of the preset clamping torque data is that it is the target torque data used to drive the end effector of the dual-arm node to form clamping contact with the target object within the current control cycle. When the preset clamping torque data does not exceed the safe torque threshold range, the preset clamping torque data matches the force boundary corresponding to the target stiffness level parameter. Therefore, the preset clamping torque data can serve as the torque source for rigid clamping actions. Subsequently, joint-driven clamping data corresponding to the rigid clamping action is generated based on the preset clamping torque data and the clamping posture field, and the joint-driven clamping data is written into the rigid clamping control bearing record. The joint-driven clamping data is used to convert the preset clamping torque data into the joint execution aperture of the dual-arm node, enabling the rigid clamping action to continue to inherit the clamping posture field in the collaborative correction instruction.The technical essence of the rigid clamping action is that the dual-arm nodes form a stable clamping contact with the target object within the torque range defined by the preset clamping torque data. This rigid clamping action continues to receive the attitude correction source from the collaborative correction command, ensuring that the dual-arm nodes maintain clamping contact while retaining the attitude correction relationship derived from the whole-body center-of-mass compensation vector when performing the target object manipulation task. The force feedback generated by the rigid clamping action subsequently enters the body disturbance suppression process. Therefore, the rigid clamping action is not an end-point clamping action independent of disturbance suppression, but rather a source of contact feedback formed when the collaborative correction command is executed within the safe torque threshold range.

[0064] Preferably, in the specific technical implementation of the force feedback generated by the compliant contact action or the rigid clamping action, after the compliant contact action or the rigid clamping action begins to be executed, the current control cycle marker, dual-arm node identifier, end effector identifier, safety correction torque data, or preset clamping torque data are first read according to the compliant contact control load record or the rigid clamping control load record, and a force feedback acquisition load record is established. The force feedback acquisition load record is used to write the torque source, end effector contact source, and current control cycle source corresponding to the compliant contact action or the rigid clamping action into the same acquisition caliber, so that the acquired force feedback can correspond to the actual contact process when the dual-arm node performs the target object operation task. The technical essence of the force feedback is that after the end effector of the dual-arm node comes into contact with the target object surface, the feedback data is formed by the change in the end effector contact state, the change in joint force, and the clamping torque response. The force feedback is used to represent the actual contact force change formed by the compliant contact action or the rigid clamping action on the target object surface. In specific implementation, firstly, end-effector contact feedback data is read according to the end-effector identifier, and joint torque feedback data is read according to the dual-arm node identifier. Then, the end-effector contact feedback data and the joint torque feedback data are written into the force feedback acquisition and bearing record according to the current control cycle marker to form the force feedback. The end-effector contact feedback data represents the change in contact intensity between the end-effector and the target surface, and the joint torque feedback data represents the change in joint force on the dual-arm nodes when performing the compliant contact action or the rigid clamping action. The end-effector contact feedback data and the joint torque feedback data form the same feedback caliber through the force feedback acquisition and bearing record, enabling the force feedback to simultaneously reflect the end-effector contact result and the joint force result. Subsequently, the force feedback undergoes time alignment processing to ensure consistency with the cooperative correction command, the adaptive control command, and the current control cycle marker. The time-aligned force feedback continues into the body disturbance suppression processing process and forms a correspondence with the whole-body center of mass compensation vector, the desired angular acceleration data, and the reference foot force data. By reading the force feedback, the contact force changes generated when the dual-arm nodes perform the target operation task can be returned from the end-effector process to the disturbance suppression processing link, so that the subsequent suppression processing of the body disturbance can not only be based on the pre-generated cooperative correction instructions, but also combined with the actual contact feedback source of the compliant contact action or the rigid clamping action.

[0065] Preferably, the specific implementation process for suppressing the machine disturbance caused by the target operation task using the force feedback generated by the compliant contact action or the rigid clamping action is as follows: After the force feedback has completed time alignment processing according to the current control cycle mark, a machine disturbance suppression load record is first established, and the force feedback, the cooperative correction command, the adaptive control command, the whole-body center of mass compensation vector, the desired angular acceleration data, the reference foot force data, and the current control cycle mark are written into the machine disturbance suppression load record. The machine disturbance suppression load record is used to include the force feedback source of the bi-arm node and the displacement control source of the quadruped node in the same current control cycle, so that the machine disturbance suppression processing can simultaneously read the force feedback when the bi-arm node performs the target operation task and the foot force source when the quadruped node performs the displacement task. The technical essence of the body disturbance is the tendency of the torso posture deviation caused by the contact reaction force of the target object, the joint force changes of the two arm nodes, and the change of the body's center of mass when the dual-arm nodes perform the target object operation task. The body disturbance is not a separate posture deviation data, but rather the disturbance result after the target object operation task is transmitted to the Centaur robot's body posture and quadrupedal support state through the force feedback. In specific implementation, firstly, the direction and magnitude of the contact force change are extracted based on the force feedback, and the direction and magnitude of the contact force change are read in correspondence with the direction dimension of the whole-body center of mass compensation vector to form a disturbance direction correspondence expression. The disturbance direction correspondence expression is used to indicate whether the actual disturbance direction caused by the force feedback corresponds to the compensation direction given by the whole-body center of mass compensation vector. Subsequently, the expected angular acceleration data and the reference foot force data are read according to the disturbance direction correspondence expression, and disturbance suppression configuration processing is performed on the expected angular acceleration data and the reference foot force data to form disturbance suppression correction data. The disturbance suppression configuration process is used to write the actual disturbance direction corresponding to the force feedback into the attitude correction aperture of the dual-arm node and the force aperture of the foot end of the quadruped node, respectively. The technical essence of the disturbance suppression correction data is that, based on the already generated collaborative correction command and adaptive control command, the data results of synchronously adjusting the attitude correction trend of the dual-arm node and the force aperture of the foot end of the quadruped node according to the force feedback. Subsequently, the disturbance suppression correction data is written into the next control cycle command generation aperture corresponding to the collaborative correction command and the adaptive control command, respectively, so that when the dual-arm node continues to perform the target operation task, it can continue to use the contact correction source corresponding to the force feedback, and when the quadruped node continues to perform the displacement task, it can continue to use the support correction source corresponding to the force feedback.Through the above processing, the force feedback generated by the compliant contact action or the rigid clamping action can enter the body disturbance suppression processing, so that the body disturbance caused by the target operation task is gradually transformed into the dual-arm node correction source and quadruped node support source in the next control cycle along the force feedback, the corresponding expression of the disturbance direction and the disturbance suppression correction data.

[0066] like Figure 3 As shown in the figure, an adaptive allocation device for control commands of a centaur robot based on hierarchical reinforcement learning is provided in this application embodiment, which includes: The state acquisition module is used to acquire the current environmental perception state of the Centaur robot and the representation of the operating intentions of its two arms; The top-level strategy processing module is used to input the current environmental perception state and the dual-arm operation intention representation into the pre-constructed top-level strategy network, so that the top-level strategy network performs action allocation processing on the dual-arm operation intention representation under the current control cycle based on the current environmental perception state, thereby generating the quadruped primitive index and the whole-body centroid compensation vector corresponding to the dual-arm operation intention representation in the current control cycle. The basic control sequence extraction module is used to extract the target basic control sequence corresponding to the quadruped primitive index from a preset parameterized primitive library; The underlying execution processing module is used to input the target basic control sequence and the whole-body center of mass compensation vector into the underlying execution network, so that the underlying execution network performs motion redirection processing on the target basic control sequence based on the whole-body center of mass compensation vector, thereby generating adaptive control instructions for the quadruped nodes and cooperative correction instructions for the bi-arm nodes. The task-driven processing module is used to drive the quadruped node to perform a displacement task based on the adaptive control command, and to drive the dual-arm node to perform a target manipulation task based on the cooperative correction command, so as to suppress the body disturbance caused by the target manipulation task.

[0067] 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; The processor is used to execute the computer program to implement the adaptive allocation method for control instructions of a centaur robot based on hierarchical reinforcement learning as described in any one of the claims of this application.

[0068] like Figure 5 As shown, this is a computer-readable storage medium according to an embodiment of the present application, wherein the computer-readable storage medium stores a computer program; When the computer program is executed by the processor, it implements the adaptive allocation method for control instructions of a centaur robot based on hierarchical reinforcement learning as described in any one of the claims of this application.

[0069] like Figure 6 As shown, this is an embodiment of a centaur robot according to this application, which includes a quadruped node, a dual-arm node, an environmental perception component, a body posture acquisition component, and a control processor. The environmental sensing component is used to collect local elevation map data in the current environmental sensing state; The body posture acquisition component is used to acquire six-degree-of-freedom pose data in the current environmental perception state; The control processor is used to acquire the current environmental perception state and the representation of the dual-arm operation intention. The control processor is also used to input the current environmental perception state and the dual-arm operation intention representation into a pre-constructed top-level policy network, so that the top-level policy network performs action allocation processing on the dual-arm operation intention representation under the current control cycle based on the current environmental perception state, thereby generating the quadruped primitive index and the whole-body centroid compensation vector corresponding to the dual-arm operation intention representation in the current control cycle. The control processor is also used to extract the target basic control sequence corresponding to the quadruped primitive index from a preset parameterized primitive library; The control processor is further configured to input the target basic control sequence and the whole-body center of mass compensation vector into the underlying execution network, so that the underlying execution network performs motion redirection processing on the target basic control sequence based on the whole-body center of mass compensation vector, thereby generating adaptive control instructions for the quadruped nodes and cooperative correction instructions for the biarm nodes. The quadruped node is used to perform displacement tasks in response to the adaptive control command; The dual-arm nodes are used to execute target operation tasks in response to the cooperative correction instructions, so as to suppress the machine body disturbances caused by the target operation tasks.

[0070] like Figure 7 As shown in the figure, this application provides an embodiment of a centaur robot control instruction adaptive allocation system based on hierarchical reinforcement learning, which includes a centaur robot, a state acquisition unit, a strategy allocation unit, a primitive library management unit, a low-level execution unit, and a task-driven unit. The centaur robot includes a quadruped node and a bi-arm node; The state acquisition unit is used to acquire the current environmental perception state and the representation of the dual-arm operation intention of the centaur robot. The strategy allocation unit is used to input the current environmental perception state and the dual-arm operation intention representation into the pre-constructed top-level strategy network, so that the top-level strategy network performs action allocation processing on the dual-arm operation intention representation under the current control cycle based on the current environmental perception state, thereby generating the quadruped primitive index and the whole-body centroid compensation vector corresponding to the dual-arm operation intention representation in the current control cycle. The primitive library management unit is used to extract the target basic control sequence corresponding to the quadruped primitive index from a preset parameterized primitive library; The underlying execution unit is used to input the target basic control sequence and the whole-body center of mass compensation vector into the underlying execution network, so that the underlying execution network performs motion redirection processing on the target basic control sequence based on the whole-body center of mass compensation vector, thereby generating adaptive control instructions for the quadruped node and cooperative correction instructions for the bi-arm node. The task-driven unit is used to drive the quadruped node to perform a displacement task based on the adaptive control command, and to drive the dual-arm node to perform a target manipulation task based on the cooperative correction command, so as to suppress the body disturbance caused by the target manipulation task.

[0071] Figures 2-7 For an exemplary description, please refer to the above. Figure 1 This will not be elaborated upon here.

Claims

1. An adaptive allocation method for control commands of a centaur robot based on hierarchical reinforcement learning, characterized in that, include: Acquire the current environmental perception state and the representation of the operating intentions of the centaur robot's two arms; The current environmental perception state and the dual-arm operation intention representation are input into a pre-constructed top-level policy network, so that the top-level policy network performs action allocation processing on the dual-arm operation intention representation under the current control cycle based on the current environmental perception state, thereby generating the quadruped primitive index and the whole-body centroid compensation vector corresponding to the dual-arm operation intention representation in the current control cycle. Extract the target basic control sequence corresponding to the quadruped primitive index from the preset parameterized primitive library; The target basic control sequence and the whole-body center of mass compensation vector are input into the underlying execution network, so that the underlying execution network performs motion redirection processing on the target basic control sequence based on the whole-body center of mass compensation vector, thereby generating adaptive control instructions for the quadruped nodes and cooperative correction instructions for the bi-arm nodes. The quadruped node is driven to perform displacement tasks based on the adaptive control command, and the dual-arm node is driven to perform target manipulation tasks based on the cooperative correction command, so as to suppress the body disturbance caused by the target manipulation tasks.

2. The adaptive allocation method for control commands of a centaur robot based on hierarchical reinforcement learning according to claim 1, characterized in that, The pre-set parameterized primitive library is used for the quadruped primitive index to retrieve the corresponding target basic control sequence. The construction steps of the pre-set parameterized primitive library include: The offline trajectory representation of the robot is collected in a preset physical simulator, which is a simulation processing environment configured during the initialization phase of the robot control system to reproduce the cooperative motion process of the quadruped node and the dual-arm node. The offline trajectory representation is input into a temporal dimensionality reduction network to perform pattern decoupling processing on the offline trajectory representation, thereby separating periodic gait features and non-periodic transition features. The periodic gait features are segmented to generate multiple first motion segments, and the non-periodic transition features are segmented to generate multiple second motion segments. Using template kinematic mapping rules, the first motion segment and the second motion segment are respectively mapped to the corresponding underlying joint drive signal sequences; The underlying joint drive signal sequence is associated and encapsulated with the corresponding primitive identifier, and the associated and encapsulated underlying joint drive signal sequence is written into the database representation structure to generate the preset parameterized primitive library for generating the target basic control sequence.

3. The adaptive allocation method for control commands of a centaur robot based on hierarchical reinforcement learning according to claim 1, characterized in that, The step of inputting the current environmental perception state and the dual-arm operation intention representation into a pre-constructed top-level policy network, so that the top-level policy network performs action allocation processing on the dual-arm operation intention representation under the current control cycle based on the current environmental perception state, and then generates the quadruped primitive index and whole-body centroid compensation vector corresponding to the dual-arm operation intention representation in the current control cycle, including: The environmental feature encoding layer in the top-level policy network is used to perform spatial topology parsing on the current environmental perception state to generate access mask features; Using the operation intent decoding layer in the top-level policy network, the dual-arm operation intent representation is subjected to temporal attention extraction processing to generate the expected centroid offset caused by the arm movements. The access mask features and the expected centroid offset are concatenated to generate a joint state characterization matrix. The joint state representation matrix is ​​value-predicted using the action evaluation layer in the top-level policy network to generate the selection probability of each primitive in the parameterized primitive library. According to the principle of selecting the highest probability, the quadruped primitive index is determined from the parameterized primitive library, and the expected centroid offset is reversed in direction to generate the whole-body centroid compensation vector.

4. The adaptive allocation method for control commands of a centaur robot based on hierarchical reinforcement learning according to claim 3, characterized in that, The current environmental perception state includes local elevation map data generated by the airborne depth camera and six-degree-of-freedom pose data generated by the fuselage inertial measurement unit; The step of utilizing the environmental feature encoding layer in the top-level policy network to perform spatial topology parsing processing on the current environmental perception state to generate access mask features includes: Using the three-dimensional convolutional kernel in the environmental feature encoding layer, spatial depth stripping is performed on the local elevation map data to generate a terrain undulation feature map; Using the temporal memory neurons in the environmental feature encoding layer, the six-degree-of-freedom pose data is processed to track historical poses in order to generate body tilt angle change rate features. The body tilt angle change rate feature is mapped to the channel dimension of the terrain undulation feature map to generate attitude terrain joint features; The posture and terrain joint features are subjected to cross-channel feature extraction processing to generate the access mask features, which are used to characterize the safe landing point area of ​​the robot in the next control cycle.

5. The adaptive allocation method for control commands of a centaur robot based on hierarchical reinforcement learning according to claim 3, characterized in that, The dual-arm operation intention representation includes the expected trajectory points of the end effector within a preset time window and the expected load quality data corresponding to each expected trajectory point of the end effector. The preset time window is a time sampling interval configured in the operation task planning stage to limit the forward movement range of the dual-arm nodes. The step of utilizing the operation intent decoding layer in the top-level policy network to perform temporal attention extraction processing on the dual-arm operation intent representation to generate the expected centroid shift caused by the arm movements includes: Using the multi-head attention mechanism in the operation intent decoding layer, spatial dependency analysis is performed on the expected trajectory points of the end effector to generate a range penalty factor. The expected load mass data is processed by torque mapping using the physical dynamics feedback layer in the operation intent decoding layer to generate additional joint torque disturbance features. The additional joint torque perturbation features and the extension range penalty factor are input into a nonlinear multilayer perceptron to perform physical mapping processing on the arm movements corresponding to the dual-arm operation intention representation, thereby generating the expected centroid offset.

6. An adaptive allocation device for control commands of a centaur robot based on hierarchical reinforcement learning, characterized in that, include: The state acquisition module is used to acquire the current environmental perception state of the Centaur robot and the representation of the operating intentions of its two arms; The top-level strategy processing module is used to input the current environmental perception state and the dual-arm operation intention representation into the pre-constructed top-level strategy network, so that the top-level strategy network performs action allocation processing on the dual-arm operation intention representation under the current control cycle based on the current environmental perception state, thereby generating the quadruped primitive index and the whole-body centroid compensation vector corresponding to the dual-arm operation intention representation in the current control cycle. The basic control sequence extraction module is used to extract the target basic control sequence corresponding to the quadruped primitive index from a preset parameterized primitive library; The underlying execution processing module is used to input the target basic control sequence and the whole-body center of mass compensation vector into the underlying execution network, so that the underlying execution network performs motion redirection processing on the target basic control sequence based on the whole-body center of mass compensation vector, thereby generating adaptive control instructions for the quadruped nodes and cooperative correction instructions for the bi-arm nodes. The task-driven processing module is used to drive the quadruped node to perform a displacement task based on the adaptive control command, and to drive the dual-arm node to perform a target manipulation task based on the cooperative correction command, so as to suppress the body disturbance caused by the target manipulation task.

7. An electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is used to execute the computer program to implement the adaptive allocation method for control instructions of a centaur robot based on hierarchical reinforcement learning as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program; When the computer program is executed by the processor, it implements the adaptive allocation method for control instructions of a centaur robot based on hierarchical reinforcement learning as described in any one of claims 1 to 5.

9. A centaur robot, characterized in that, It includes a quadruped node, a dual-arm node, an environmental perception component, a body posture acquisition component, and a control processor; The environmental sensing component is used to collect local elevation map data in the current environmental sensing state; The body posture acquisition component is used to acquire six-degree-of-freedom pose data in the current environmental perception state; The control processor is used to acquire the current environmental perception state and the representation of the dual-arm operation intention. The control processor is also used to input the current environmental perception state and the dual-arm operation intention representation into a pre-constructed top-level policy network, so that the top-level policy network performs action allocation processing on the dual-arm operation intention representation under the current control cycle based on the current environmental perception state, thereby generating the quadruped primitive index and the whole-body centroid compensation vector corresponding to the dual-arm operation intention representation in the current control cycle. The control processor is also used to extract the target basic control sequence corresponding to the quadruped primitive index from a preset parameterized primitive library; The control processor is further configured to input the target basic control sequence and the whole-body center of mass compensation vector into the underlying execution network, so that the underlying execution network performs motion redirection processing on the target basic control sequence based on the whole-body center of mass compensation vector, thereby generating adaptive control instructions for the quadruped nodes and cooperative correction instructions for the biarm nodes. The quadruped node is used to perform displacement tasks in response to the adaptive control command; The dual-arm nodes are used to execute target operation tasks in response to the cooperative correction instructions, so as to suppress the machine body disturbances caused by the target operation tasks.

10. An adaptive allocation system for control commands of a centaur robot based on hierarchical reinforcement learning, characterized in that, It includes a centaur robot, a state acquisition unit, a strategy allocation unit, a primitive library management unit, a low-level execution unit, and a task-driven unit; The centaur robot includes a quadruped node and a bi-arm node; The state acquisition unit is used to acquire the current environmental perception state and the representation of the dual-arm operation intention of the centaur robot. The strategy allocation unit is used to input the current environmental perception state and the dual-arm operation intention representation into the pre-constructed top-level strategy network, so that the top-level strategy network performs action allocation processing on the dual-arm operation intention representation under the current control cycle based on the current environmental perception state, thereby generating the quadruped primitive index and the whole-body centroid compensation vector corresponding to the dual-arm operation intention representation in the current control cycle. The primitive library management unit is used to extract the target basic control sequence corresponding to the quadruped primitive index from a preset parameterized primitive library; The underlying execution unit is used to input the target basic control sequence and the whole-body center of mass compensation vector into the underlying execution network, so that the underlying execution network performs motion redirection processing on the target basic control sequence based on the whole-body center of mass compensation vector, thereby generating adaptive control instructions for the quadruped node and cooperative correction instructions for the bi-arm node. The task-driven unit is used to drive the quadruped node to perform a displacement task based on the adaptive control command, and to drive the dual-arm node to perform a target manipulation task based on the cooperative correction command, so as to suppress the body disturbance caused by the target manipulation task.

Images