Digital twinning control method and system for catastrophe rescue robot

Abstract

The invention discloses a catastrophe rescue robot digital twin control method and system, and belongs to the technical field of robot intelligent control. Comprising a virtual-real mapping module which processes the body state of a robot and catastrophe scene point cloud data in real time, and constructs and synchronously updates a digital twinborn body containing a terrain elevation map at a cloud end; the hierarchical decision-making module is used for making a decision in the digital twinborn body by utilizing a hierarchical reinforcement learning model, and outputting a wheel-leg movement mode and an excavation arm cooperation instruction by a high-level strategy network according to the terrain complexity; and the safety deduction module introduces a safety reinforcement learning mechanism, deduces and verifies the generated action instruction in the virtual environment, and restrains the dangerous action by controlling a barrier function. According to the method, the decision risk of the heavy robot in the unstructured catastrophic terrain can be reduced, the stability of motion control is improved, and a high-safety control strategy is provided for various complex obstacle crossing actions based on simulation prediction.

Description

Technical Field

[0001] This invention belongs to the field of robot intelligent control technology, specifically relating to a digital twin control method and system for disaster relief robots. Background Technology

[0002] Heavy-duty wheeled-legged robots combine the efficiency of wheeled mobility with the obstacle-crossing ability of legged movement, making them highly promising for unstructured disaster scenarios such as earthquake rescue and geological disaster relief. However, in practical applications, these robots face significant technical challenges.

[0003] First, decision-making in complex terrain carries high risks. Disaster sites typically include soft soil, collapsed ruins, and other unknown terrain features. Traditional model-based control methods rely on precise physical parameters, making them difficult to adapt to drastic changes in environmental parameters. While end-to-end deep reinforcement learning methods are highly adaptable, they are black-box characteristics, and their output actions often lack safety constraints, which can easily lead to irreversible rollover accidents for the robot.

[0004] Secondly, the collaborative utilization rate of the excavator arm is low. Most existing control strategies treat the excavator arm as a working tool, locking it during movement or using it only as a passive load. In fact, for heavy robots, the excavator arm accounts for a large proportion of the overall mass. Ignoring the dynamic characteristics of the excavator arm not only wastes a huge degree of adjustment freedom, but the large inertia it generates can also become a detrimental factor to the robot's balance.

[0005] Furthermore, there is the issue of model parameter mismatch in dynamic environments. Existing simulation-based predictive control relies on preset physical parameters. However, the geological characteristics of disaster sites are time-varying and non-uniform. This dynamic mismatch of physical parameters causes digital twins to fail to accurately reflect the dynamic response of real robots, leading to problems such as slippage, large tracking errors, and even control divergence when simulation-based planning instructions are executed on the real machine. Summary of the Invention

[0006] The technical problem this invention aims to solve is: addressing the high decision-making risks in complex terrain, low utilization rate of excavator arm dynamics, and parameter mismatch between virtual and real models in dynamic environments faced by existing heavy-duty wheeled-legged robots in disaster scenarios. This invention provides a digital twin control method and system for disaster relief robots. The method encompasses a virtual-real mapping module, a hierarchical decision-making module, and a safety simulation module. It can reduce the decision-making risks of heavy-duty robots in complex disaster terrain, improve the stability of motion control, and provide a highly safe control strategy for various complex obstacle-crossing actions based on simulation prediction. The specific technical solution is as follows:

[0007] A digital twin control method for disaster relief robots, employing a cloud-edge collaborative architecture, includes:

[0008] S110, in the virtual-real mapping module, the cloud server receives the body state data stream and disaster scenario point cloud data uploaded by the robot edge side. After data fusion and physical attribute inversion processing, a high-fidelity digital twin is constructed and updated in real time. The digital twin includes the robot's multi-rigid-body dynamics model and the terrain elevation map of the current environment. The robot has a wheel-legged mobile mechanism and an excavating arm.

[0009] S120, in the hierarchical decision module, the current state of the digital twin is input into the pre-trained hierarchical reinforcement learning model, and the high-level policy network outputs motion mode instructions and excavator arm coordination instructions according to the complexity of the terrain elevation map; the excavator arm coordination instructions include motion parameters for using the excavator arm as a balancing counterweight or auxiliary support limb.

[0010] S130, in the safety simulation module, a safety reinforcement learning mechanism is used to perform a several-time-step advance simulation of the instructions output in step S120 in the digital twin; a safety constraint set is constructed using the control obstacle function, and it is determined whether the simulated robot state is within the safety constraint set;

[0011] S140, Virtual-Real Migration and Local Security Control: The cloud sends the instructions generated by the hierarchical reinforcement learning model to the edge side; the edge controller receives the instructions and uses the security inference module deployed locally to perform control obstacle function constraint detection in combination with the current high-frequency ontology state; if the inference result satisfies the set of security constraints, the joint is driven to execute; if the inference result violates the set of security constraints, the instruction correction or replanning strategy is triggered.

[0012] A digital twin control system for disaster relief robots includes:

[0013] Perception and edge computing unit: Deployed on the robot body, used to collect environmental and state data and perform high-frequency dynamic control;

[0014] Cloud computing platform: used to build digital twins, run hierarchical reinforcement learning models, and perform security filtering corrections;

[0015] Communication module: Used for high-speed data transmission between sensing and edge computing units and cloud computing platforms;

[0016] Memory and processor: The memory stores a computer program, and the processor executes the computer program to implement the digital twin control method for the disaster relief robot.

[0017] The advantages of this invention compared to the prior art are as follows:

[0018] (1) To address the high decision-making risk in complex terrain, this invention introduces a safety extrapolation based on a control obstacle function and an instruction correction mechanism based on quadratic programming. Even if the neural network outputs a high-risk action in unknown terrain, the digital twin can identify the risk in advance through advanced extrapolation and forcibly correct the instruction through safety filtering, thereby fundamentally eliminating the overturning accident caused by algorithm trial and error.

[0019] (2) To address the problem of low collaborative efficiency of the excavator arm, this invention incorporates the excavator arm into the whole-body dynamics planning, constructing a dual collaborative mode of "active counterweight balance" and "assisted support for obstacle crossing". In the suspended state, the excavator arm acts as a dynamic balance bar, providing inertial counter-torque to counteract the tendency of the fuselage to overturn; in the contact state, the excavator arm acts as a fifth limb, providing additional traction and support points. This not only eliminates the negative inertial effects brought by the large mass excavator arm, but also transforms it into a favorable factor for expanding the stable support domain, enabling the robot to conquer steep slopes and cliffs that exceed the kinematic limits of the legs.

[0020] (3) To address the problem of model parameter mismatch in dynamic environments, this invention introduces a physical property inference and mapping mechanism in the construction of digital twins. By integrating visual texture and contact force feedback, the friction coefficient and contact stiffness of the ground are inverted in real time and dynamically updated to the cloud simulation environment. Combined with the "fast and slow dual-loop" architecture of cloud-edge collaboration, the control divergence problem caused by slippage and large tracking error when planning instructions are executed on the real machine is effectively reduced. Attached Figure Description

[0021] Figure 1 The above is an overall flowchart of a digital twin control method for a disaster relief robot provided in an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of the structure of a three-level cascaded hierarchical reinforcement learning model containing prior constraints in an embodiment of the present invention. Detailed Implementation

[0023] The present invention will now be described in further detail with reference to the accompanying drawings.

[0024] This invention relates to a digital twin control method for a disaster relief robot. The method is implemented based on a cloud-edge collaborative architecture, which includes an edge computing unit deployed on the robot itself and a cloud computing platform deployed remotely. To address the issues of computing power allocation and real-time performance, the system employs a "fast-slow dual-loop" strategy: the slow loop on the cloud operates at a frequency of 10Hz-50Hz, responsible for environment reconstruction and strategy deduction under high computing loads; the fast loop on the edge operates at a frequency of 500Hz-1kHz, responsible for high-frequency inertial measurement unit (IMU) data processing and whole-body dynamics control. Simultaneously, if communication latency exceeds a safety threshold, the edge side will automatically switch to a local degradation mode to lock the gait and maintain static balance.

[0025] The structural block diagram of the present invention is as follows: Figure 1 As shown. Figure 1 As shown, this method includes a virtual-real mapping module 1, a hierarchical decision-making module 2, and a safety simulation module 3. It can reduce the decision-making risk of heavy robots in complex and catastrophic terrain, improve the stability of motion control, and provide a highly safe control strategy for various complex obstacle-crossing actions based on simulation prediction.

[0026] like Figure 1 As shown, the specific implementation method is as follows:

[0027] S110, in the virtual-real mapping module 1, the cloud server receives the body state data stream and point cloud data and visual images of the disaster scene uploaded by the robot's edge side. After data fusion and physical attribute inversion processing, a high-fidelity digital twin is constructed and updated in real time. The digital twin includes the robot's multi-rigid-body dynamics model and the terrain elevation map of the current environment, etc.; the robot has a wheel-legged mobility mechanism and a digging arm. The specific implementation is as follows:

[0028] S110-1, Multi-dimensional Information Fusion Mapping: Spatiotemporal registration of point cloud data uploaded from the edge side with visual images to construct a local terrain elevation map containing geometric information; this process may specifically include:

[0029] The cloud server receives robot body status data streams and disaster scenario point cloud data uploaded from the edge. First, it performs spatiotemporal registration of the heterogeneous data based on a unified timestamp. Then, it uses a rasterization method to transform the unstructured point cloud data into a local terrain elevation map. This elevation map Each grid in It not only includes geometric height information It also maps the corresponding visual texture features. The subscript These represent the row and column indices of the terrain grid in the world coordinate system, respectively.

[0030] S110-2, Physical Attribute Inference and Mapping: Based on visual texture features and historical contact force data, geological attributes are identified for each region of the topographic elevation map; the equivalent friction coefficient, contact stiffness and maximum bearing pressure of the ground are estimated, and the above physical parameters are mapped to the corresponding virtual ground of the digital twin.

[0031] To address the problem of inaccurate terrain parameter presets leading to simulation distortion in traditional simulations, this invention designs a "visual-tactile" fusion physical property inversion mechanism. The system targets each grid cell... Define the physical property parameter vector ,in Characterizing the equivalent friction coefficient, To characterize ground contact stiffness, the system dynamically updates this parameter vector using the following weighted fusion model:

[0032] ,

[0033] in, For a parameter estimation operator based on visual priors, the input variables are... Representing the texture feature information corresponding to the mesh, the system uses a semantic segmentation network to identify the terrain texture category and obtains the initial physical parameter estimate corresponding to the material by looking up a table. , These are the ground friction coefficient and ground contact stiffness, respectively, estimated based on visual priors. For the ontology-based posterior correction operator, the input variables include foot contact force. With slip velocity The system is based on the inverse dynamics algorithm to calculate the current physical properties in real time and output a physical property parameter vector based on ontology perception. As a correction to visual priors, These are the ground friction coefficient and ground contact stiffness estimated based on body perception, respectively. This is a time-varying adaptive confidence weight matrix. To achieve decoupled control of the fusion weights for friction coefficient and stiffness, this matrix is ​​designed as a diagonal matrix. ,in, These represent the weighting coefficients for the friction coefficient and contact stiffness, respectively. For attributes that are difficult to ascertain visually, the system sets a lower initial value. The value makes it more dependent on haptic feedback items. For attributes with significant visual features, a higher level is maintained. Value, with contact time With the increase of matrix elements It shows a monotonically decreasing trend This means that as the interaction deepens, the system gradually transitions from visual prediction to tactile measurement, and finally converges to the actual physical contact state.

[0034] Loading the robot's multi-rigid-body dynamics model: To reproduce the robot's dynamic characteristics in the cloud, the system constructs a multi-rigid-body model containing a complete kinematic chain and inertia distribution. This model not only maps the robot's geometric dimensions and mass properties but also integrates the torque constraints and response delay characteristics of the joint actuators. The system defines the virtual robot's state vector as follows: ,in Includes the six-DOF floating base pose and all joint angles of the base. This corresponds to the generalized velocity.

[0035] Real-time synchronization and integration of virtual and real states: In this stage, the system integrates the aforementioned "virtual terrain with physical attributes" and "robot dynamics model" into the physics engine. To ensure the effectiveness of subsequent safety simulations, the system must eliminate the deviation between virtual and real states. The system subscribes to the sensor data of the physical robot at high frequency and uses a state observer to synchronize the current real state of the physical robot. Real-time mapping and forced synchronization to virtual state This state reset mechanism ensures that the initial state of the digital twin is completely consistent with that of the physical robot before each advanced simulation, thus guaranteeing the physical reliability of the prediction results.

[0036] S120, in the hierarchical decision-making module, the current state of the digital twin is input into a pre-trained hierarchical reinforcement learning model. The high-level policy network outputs motion mode instructions and excavator arm coordination instructions based on the complexity of the terrain elevation map. The excavator arm coordination instructions include motion parameters for using the excavator arm as a balancing weight or auxiliary support limb. The specific implementation is as follows:

[0037] like Figure 2 As shown, the hierarchical reinforcement learning model in S120 adopts a three-level cascaded architecture with prior constraints, including:

[0038] (1) High-level policy network: The local terrain elevation map generated by the digital twin, the robot's centroid state, and the target task are used as the input state space. Discrete motion mode switching commands are output at a low control frequency. The motion mode switching commands include at least wheeled movement mode, legged movement mode, and excavator arm-assisted mode. This layer is constructed using a deep reinforcement learning algorithm framework based on policy gradient, specifically using an Actor-Critic network architecture and trained based on the PPO algorithm. The Actor network is responsible for outputting the probability distribution of each motion mode according to the joint state vector, and the Critic network is responsible for evaluating the value function of the current state. During training, a truncated surrogate objective function is used to limit the magnitude of policy updates to ensure convergence stability, and the Adam optimizer is used for parameter updates.

[0039] In the reasoning and decision-making phase, the network first encodes the input local terrain elevation map using a feature extraction network, extracting implicit terrain geometric features and passability features. A multimodal attention fusion mechanism is then introduced to map the robot's centroid state to a query vector and the terrain features to key-value pairs. Through an attention weight matrix, key local terrain features highly correlated with the current motion state are extracted, constructing a context-aware joint state vector. This vector is then input into the policy network, outputting discrete motion mode switching commands. At least including:

[0040] ①Wheel mode: Suitable for flat terrain, locking leg freedom to improve energy efficiency;

[0041] ② Leg-based movement mode: suitable for rugged terrain, activating foot planning to cross obstacles;

[0042] ③ Excavator Arm Assist Mode: Suitable for extreme and catastrophic terrain, activating the collaborative functions of the excavator arm. Specifically, this invention designs two specific excavator arm collaborative logics to fully utilize the large inertia and operational capacity of the excavator arm:

[0043] a) Active counterweight balancing mode: When the digital twin predicts that the robot will become unstable if it relies solely on its legs for support, it generates a center of mass adjustment trajectory or a reverse swing trajectory for the excavator arm; it uses the gravitational potential energy of the excavator arm to adjust the position of the overall composite center of mass to a safe area, or uses the inertial reaction torque generated by the swing of the excavator arm to counteract the overturning torque of the machine body.

[0044] b) Assisted support obstacle crossing mode: When the high-level policy network determines that the span or slope of the terrain ahead exceeds the kinematic limit of the legs, it generates the end contact trajectory of the excavator arm and the support force distribution command; after the end of the excavator arm establishes physical contact with the ground, it estimates the contact force at the end of the excavator arm based on the driving force / torque feedback of each joint of the excavator arm; the excavator arm is included as the fifth supporting limb in the whole-body dynamics optimization solution, and the equivalent stiffness and damping parameters of the end of the excavator arm are actively adjusted through the variable impedance control strategy to provide auxiliary traction force or expand the stable support domain of the robot.

[0045] (2) Mid-level behavior generation network: Receives instructions from the high-level layer and the current body state of the robot, and generates the robot's continuous expected behavior trajectory based on the preset kinematic model and dynamic model; the continuous expected behavior trajectory includes the expected body pose, the expected center of mass velocity and the expected foot landing point that change smoothly over time.

[0046] For the wheel / leg pattern, plan the desired foot landing sequence. and the trajectory of the fuselage center of gravity For the excavator arm-assisted mode, the focus is on planning the contact position and support force distribution at the excavator arm tip, enabling it to participate in overall balance control as a "fifth limb." The final output is a continuous desired state sequence that changes smoothly over time. , among which, among which The desired six-DOF floating base pose and all joint angles are included. This corresponds to the expected generalized velocity.

[0047] (3) Hybrid control network at the bottom layer: The control strategy of “prior model + residual learning” is adopted. The tracking error between the continuous expected behavior trajectory output by the middle layer behavior generation network and the actual fuselage state fed back by the sensor is used as the input state space. The feedforward torque is calculated by using the rigid body dynamics prior model, and the nonlinear compensation torque is calculated by using the residual neural network. The two are superimposed as the final control command to drive the joint to execute.

[0048] This layer operates at a high frequency and uses a rigid body dynamics prior model as the main channel. Utilizing an inverse dynamics algorithm, it calculates the theoretical feedforward torque required to drive the robot's motion based on the desired acceleration input from the middle layer and the current state. The calculation formula is as follows:

[0049] ,

[0050] in, This represents the theoretical feedforward torque vector. This represents the number of degrees of freedom of the joints. This represents the desired joint angle acceleration vector; The joint space inertia matrix represents the robot. Represents the matrix of Coriolis force and centrifugal force; Represents the gravity compensation vector; It includes the six-DOF floating base pose and all joint angles of the base. It corresponds to the expected generalized velocity.

[0051] Furthermore, this layer uses a residual learning network as a compensation channel, and designs a lightweight fully connected residual neural network to address the nonlinear factors that are difficult to model in catastrophic environments. This network tracks the error... Input is nonlinear compensation torque. .in, The desired configuration includes the six-DOF floating base pose and all joint angles of the base. It corresponds to the expected generalized velocity.

[0052] Finally, the outputs of the two channels are superimposed to obtain the final control command sent to the joint actuator. :

[0053] .

[0054] The pre-training process of the hierarchical reinforcement learning model adopts a "terrain complexity increasing" course learning approach, including:

[0055] (1) Proprioception training: In flat terrain, the elevation map input is blocked, and only the ability of the bottom hybrid control network to track the upper command and the balance and coordination ability of the excavator arm are trained.

[0056] (2) Terrain adaptation training: Introduce static wavy road surface and stepped terrain, unlock the middle layer behavior generation network, and train the robot to adaptively adjust gait and center of mass according to the terrain geometry.

[0057] (3) Disaster response training: Introduce dynamic collapse terrain and random external force disturbances, activate the safety simulation module, introduce strong penalty terms for violating safety constraints in the training, and train the model's self-recovery and risk avoidance capabilities under extreme working conditions.

[0058] In S130, within the safety simulation module 3, a safety reinforcement learning mechanism is used to perform a several-time-step advance simulation of the instructions output in step S120 within the digital twin; a safety constraint set is constructed using control obstacle functions, and it is determined whether the simulated robot state falls within the safety constraint set. The specific implementation is as follows:

[0059] S130-1, Constructing a Composite Control Obstacle Function: Based on the robot's multi-rigid-body dynamics model, construct a composite control obstacle function that includes multi-dimensional safety sub-terms. Define a safe set. That is, when At that time, the robot is in a safe state. In this embodiment... It includes the following three core security sub-items:

[0060] ① Dynamically stable sub-item : Stability constraint characterizing whether the zero-moment point lies within the supporting polygon. The system calculates the position of the robot's zero-moment point. It is constrained that it must lie within the convex polygonal support domain formed by all the support feet. The calculation formula is as follows:

[0061] ,

[0062] in, and These are the inward pointing normal vector and the vertex that support the polygon boundary, respectively.

[0063] ②Mechanical Limit Sub-item This characterizes whether the position, speed, and hydraulic cylinder driving force of the excavator arm and wheel leg joints exceed physical threshold constraints. The calculation formula is as follows:

[0064] ,

[0065] in, These are the physical upper and lower threshold values ​​for the joint angle; This represents the maximum permissible speed threshold for the joint motor. This is the joint driving torque vector; This is the maximum output torque saturation threshold of the hydraulic actuator.

[0066] ③Environmental Interaction Sub-items : Characterizes the minimum safe distance constraint between the robot body and excavating arm and obstacles in catastrophic terrain; the system calculates the minimum Euclidean distance between the robot body and excavating arm and the obstacle point cloud in the digital twin. It is constrained that it must be greater than the preset safety collision avoidance threshold. The calculation formula is as follows:

[0067] ,

[0068] The final composite control barrier function is designed as a weighted combination of the above sub-items, serving as the core constraint for subsequent controller optimization:

[0069] ,

[0070] in, These are the weighting coefficients for each safety sub-item, used to adjust the degree of emphasis on stability, self-hardware protection, and environmental obstacle avoidance under different operating conditions.

[0071] S130-2, Advanced Temporal Security Verification: In the digital twin, verify the future outputs of the high-level policy network and the mid-level behavior generation network. Rolling detection is performed by predicting the state sequence step by step.

[0072] Utilizing the high-speed simulation capabilities of digital twins, the system extrapolates the instructions output in step S120. The system uses the current time... Synchronization status Starting from this point, the instruction sequences output by the high-level policy network and the mid-level behavior generation network are input into the digital twin to extrapolate to the future. Each time step.

[0073] S130-3, Safety Criterion: If each step in the predicted state sequence satisfies... And it satisfies the forward invariance condition of the barrier function. (in For class function, It is a composite control barrier function If the time derivative along the system state trajectory is obtained, the instruction sequence is determined to be safe; otherwise, it is determined to have violated the safety constraints, and the time of the violation and the corresponding obstacle function gradient information are transmitted to step four.

[0074] S140, Virtual-Real Migration and Local Security Control: The cloud sends parameter commands generated by the hierarchical reinforcement learning model to the edge side; the edge controller receives the commands and uses a locally deployed security inference module, combined with the current high-frequency ontology state, to perform control obstacle function constraint detection; if the inference result satisfies the security constraint set, the joint is driven to execute; if the inference result violates the security constraint set, an instruction correction or replanning strategy is triggered. The specific implementation is as follows:

[0075] When step S130 determines the current time If the predicted instruction sequence satisfies the forward invariance conditions of all control obstacle functions, it indicates that the output action of the hierarchical reinforcement learning model conforms to the physical constraints. The system directly sends this instruction sequence to the edge controller to drive the robot's hydraulic joints and hub motors, achieving efficient motion control.

[0076] When step S130 determines the current time When the predicted command sequence poses a risk of violating the safety boundary, the command correction mechanism of the rigid body dynamics prior model will be triggered:

[0077] (1) Constraint construction based on prior model: Digital twin calls robot rigid body dynamics prior model (in It is the time derivative of the robot's current state vector. For the system's natural dynamics, For the control matrix, Given the control input vector to be solved, calculate the feasible control space that satisfies the forward invariance of the control barrier function in the current state. The system calculates the control barrier function. Lie differential along system dynamics and Construct the following linear inequality constraints:

[0078] ,

[0079] in, This represents the robot's current state vector; For the control barrier function Along the system's natural dynamics Lie differential; For the control barrier function Along the control matrix Lie differential; For extended classes function.

[0080] This constraint defines a half-plane within the control space; all control inputs falling within this region... All of these measures ensure that the robot does not move out of the safe set in the next moment.

[0081] (2) Security Filter Correction: Construct a security filter based on quadratic programming to filter the instructions sent from the cloud. Using this as a reference input, components that violate physical constraints are filtered out. The optimal correction control command is then solved under the aforementioned linear inequality constraints, with the objective function being the minimization of the correction amount. :

[0082] ,

[0083] ,

[0084] in, This indicates the corrected optimal control command; This indicates that the control commands output by the hierarchical reinforcement learning model in step S120 are sent to the joint actuators, and ; These represent the upper and lower limits of the physical output of the robot's joint actuator, respectively.

[0085] (3) Deadlock handling and replanning: If the safety filter has no solution in the current time domain, the current motion mode is determined to be deadlocked. An abnormal state is reported to the cloud to trigger high-level policy replanning, and a forced switch to the self-recovery control program is performed to execute a center of gravity reduction or shutdown locking action; specifically including:

[0086] ① Policy Update Feedback: The digital twin immediately sends a large negative reward signal to the high-level policy network in the cloud. This signal is used to punish the preceding decision sequence that led to deadlock, prompting the reinforcement learning policy to avoid such states in subsequent training.

[0087] ② Forced Self-Recovery: The edge-side controller forcibly interrupts the current motion task and switches to a preset self-recovery control program. Depending on the current state, it performs actions such as lowering the center of gravity to crouch, actively grounding the excavator arm for support, or braking and locking, until the robot returns to a controllable and safe area. The self-recovery control program is as follows: When the robot is detected to be overturned or fallen, the excavator arm end actively seeks a ground support point; by adjusting the angles of each joint of the excavator arm, a temporary triangular support structure is constructed between the excavator arm end and the body contact point; the extension motion of the excavator arm lifts the body's center of gravity to a recoverable area within the support polygon; subsequently, in conjunction with the contraction and reset of the leg joints, the robot autonomously recovers from a fallen posture to a standing posture.

[0088] In summary, this invention discloses a digital twin control method for disaster relief robots, including a virtual-real mapping module, a hierarchical decision-making module, and a safety simulation module. This method can reduce the decision-making risk of heavy robots in unstructured disaster terrain, improve the stability of motion control, and provide a highly safe control strategy for various complex obstacle-crossing actions based on simulation prediction.

[0089] Furthermore, the present invention also provides a digital twin control system for a disaster relief robot, comprising:

[0090] Perception and edge computing unit: Deployed on the robot body, used to collect environmental and state data and perform high-frequency dynamic control;

[0091] Cloud computing platform: used to build digital twins, run hierarchical reinforcement learning models, and perform security filtering corrections;

[0092] Communication module: Used for high-speed data transmission between sensing and edge computing units and cloud computing platforms;

[0093] Memory and processor: The memory stores a computer program, and the processor executes the computer program to implement the above control method.

[0094] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

[0095] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions. The above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. It should be pointed out that those skilled in the art can make several improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered to fall within the protection scope of the present invention.

Claims

1. A digital twin control method for a disaster relief robot, characterized in that, The cloud-edge collaborative architecture is adopted, including: S110, in the virtual-real mapping module, the cloud server receives the body state data stream and point cloud data and visual images of the disaster scene uploaded by the robot edge side. After data fusion and physical attribute inversion processing, a high-fidelity digital twin is constructed and updated in real time. The digital twin includes the robot's multi-rigid-body dynamics model and the terrain elevation map of the current environment. The robot has a wheel-legged mobile mechanism and an excavating arm. S120, in the hierarchical decision module, the current state of the digital twin is input into the pre-trained hierarchical reinforcement learning model, and the high-level policy network outputs motion mode instructions and excavator arm coordination instructions according to the complexity of the terrain elevation map; the excavator arm coordination instructions include motion parameters for using the excavator arm as a balancing counterweight or auxiliary support limb. S130, in the safety simulation module, a safety reinforcement learning mechanism is used to perform a several-time-step advance simulation of the instructions output in step S120 in the digital twin; a safety constraint set is constructed using the control obstacle function, and it is determined whether the simulated robot state is within the safety constraint set; S140, Virtual-Real Migration and Local Security Control: The cloud sends the instructions generated by the hierarchical reinforcement learning model to the edge side; the edge controller receives the instructions and uses the security inference module deployed locally to perform control obstacle function constraint detection in combination with the current high-frequency ontology state; if the inference result satisfies the set of security constraints, the joint is driven to execute; if the inference result violates the set of security constraints, the instruction correction or replanning strategy is triggered.

2. The digital twin control method for a disaster relief robot according to claim 1, characterized in that, The process of constructing and updating the high-fidelity digital twin in S110 includes: S110-1, Multi-dimensional information fusion mapping: Based on the robot pose information provided by the ontology state data stream, spatiotemporal registration is performed on the point cloud data uploaded from the edge side and the visual image to construct a local terrain elevation map containing geometric information. ; S110-2, Physical Attribute Inference and Mapping: Based on visual texture features and historical contact force data, geological attributes are identified for each region of the topographic elevation map; the equivalent friction coefficient, contact stiffness and maximum bearing pressure of the ground are estimated, and the above physical parameters are mapped to the corresponding virtual ground of the digital twin.

3. The digital twin control method for a disaster relief robot according to claim 1, characterized in that, In S120, the hierarchical reinforcement learning model adopts a three-level cascaded architecture with prior constraints, including: High-level policy network: The input state space consists of a local terrain elevation map generated by the digital twin, the robot's center of mass state, and the target task. The output is discrete motion mode switching commands, which include wheeled movement mode, legged movement mode, and excavator arm assisted mode. The mid-level behavior generation network receives instructions from the high-level layer and the current state of the robot. Based on the preset kinematic and dynamic models, it generates the robot's continuous expected behavior trajectory. The continuous expected behavior trajectory includes the expected body pose, the expected center of mass velocity, and the expected foot landing point that change smoothly over time. The underlying hybrid control network adopts a control strategy of prior model + residual learning. The tracking error between the continuous expected behavior trajectory output by the middle behavior generation network and the actual fuselage state fed back by the sensor is used as the input state space. The feedforward torque is calculated using the rigid body dynamics prior model, and the nonlinear compensation torque is calculated using the residual neural network. The two are superimposed as the final control command to drive the joint execution.

4. The digital twin control method for a disaster relief robot according to claim 3, characterized in that, The high-level policy network is constructed using a deep reinforcement learning algorithm framework based on policy gradients. First, a feature extraction network encodes the input local terrain elevation map, extracting implicit terrain geometric and passability features. A multimodal attention fusion mechanism is then introduced to map the robot's centroid state to a query vector and the terrain geometric features to key-value pairs. Through an attention weight matrix, key local terrain features highly correlated with the current motion state are extracted, constructing a context-aware joint state vector. This vector is then input into the high-level policy network, outputting discrete motion mode switching commands. It includes: Wheeled movement mode: suitable for flat terrain, locking leg freedom to improve energy efficiency; Legged movement mode: suitable for rugged terrain, activating foot planning to cross obstacles; Excavator arm assist mode: suitable for extreme catastrophic terrain, activating the collaborative function of the excavator arm.

5. The digital twin control method for a disaster relief robot according to claim 4, characterized in that, The mid-level behavior generation network plans the desired foot landing sequence for both wheeled and legged locomotion modes. and the trajectory of the fuselage center of gravity For the excavator arm-assisted mode, the focus is on planning the contact position and support force distribution at the end of the excavator arm, enabling it to participate in the overall balance control as a "fifth limb," ultimately outputting a continuous desired state sequence that changes smoothly over time. , among which, among which The desired six-DOF floating base pose and all joint angles are included. This corresponds to the expected generalized velocity.

6. The digital twin control method for a disaster relief robot according to claim 5, characterized in that, The underlying hybrid control network calculates the theoretical feedforward torque required to drive the robot's motion, using the following formula: ， in, This represents the theoretical feedforward torque vector. This refers to the number of degrees of freedom of the joints. This represents the desired joint angle acceleration vector; The joint space inertia matrix represents the robot. Represents the matrix of Coriolis force and centrifugal force; Represents the gravity compensation vector; It includes the six-DOF floating base pose and all joint angles of the base. This corresponds to the generalized velocity; Using a residual learning network as a compensation channel, a lightweight fully connected residual neural network is designed to address the nonlinear factors that are difficult to model in catastrophic environments. This network tracks the error. Input is nonlinear compensation torque. ,in, The desired configuration includes the six-DOF floating base pose and all joint angles of the base. This corresponds to the expected generalized velocity; Finally, the outputs of the two channels are superimposed to obtain the final control command sent to the joint actuator. : 。 7. The digital twin control method for a disaster relief robot according to claim 1, characterized in that, S130 includes: S130-1, Constructing a Composite Control Obstacle Function: Based on the robot's multi-rigid-body dynamics model, construct a composite control obstacle function that includes multi-dimensional safety sub-terms. ;in, The weighting coefficients for each security sub-item are provided, and the security sub-items include at least a dynamically stable sub-item. Mechanical Limits Sub-item , Environment Interaction Sub-items Among them, the dynamic stability sub-item represents the stability constraint of whether the zero torque point is located within the supporting polygon; the mechanical limit sub-item represents the constraint of whether the position, speed and hydraulic cylinder driving force of the excavator arm and wheel leg joints exceed the physical threshold; the environmental interaction sub-item represents the minimum safe distance constraint between the robot body and excavator arm and the catastrophic terrain obstacle. S130-2, Advanced Temporal Security Verification: In the digital twin, verify the future outputs of the high-level policy network and the mid-level behavior generation network. Rolling detection is performed on the predicted state sequence. S130-3, Safety Criterion: If each step in the predicted state sequence satisfies... And it satisfies the forward invariance condition of the barrier function. ,in For class function, It is a composite control barrier function If the time derivative along the system state trajectory is used, the instruction sequence is determined to be safe; otherwise, it is determined to have violated safety constraints.

8. The digital twin control method for a disaster relief robot according to claim 1, characterized in that, The process of triggering instruction correction or replanning strategies in S140 includes: Constraint Construction Based on Prior Model: The digital twin invokes the robot's rigid body dynamics prior model to calculate the feasible control space that satisfies the forward invariance of the control obstacle function in the current state. ;in Let be the robot's current state vector. Let be the control input vector to be solved. For the control barrier function Along the system's natural dynamics Lie Weifen, For the control barrier function Along the control matrix Lie Weifen, For extended classes function; Security Filtering Correction: Construct a security filter based on quadratic programming to filter the instructions sent from the cloud. As input, the optimal modified control command is solved while satisfying the feasible control space constraints. ,in, This indicates the corrected optimal control command; Deadlock handling and replanning: If the safety filter has no solution in the current time domain, the current motion mode is determined to be deadlocked. The abnormal state is reported to the cloud to trigger the high-level policy replanning and to force a switch to the self-recovery control program to perform a center of gravity reduction or shutdown locking action.

9. A digital twin control method for a disaster relief robot according to claim 8, characterized in that, The self-recovery control procedure includes: Policy update feedback: The digital twin sends a negative reward signal to the high-level policy network in the cloud. This signal is used to punish the preceding decision sequence that led to deadlock, prompting the reinforcement learning policy to avoid such states in subsequent training. Forced self-recovery: The edge controller forcibly interrupts the current motion task and switches to a preset self-recovery control program until the robot's state returns to a controllable and safe area.

10. A digital twin control system for a disaster relief robot, characterized in that, include: Sensing and edge computing unit: Deployed on the robot body, used to collect environmental and state data and perform high-frequency dynamic control; Cloud computing platform: used to build digital twins, run hierarchical reinforcement learning models, and perform security filtering corrections; Communication module: Used for high-speed data transmission between sensing and edge computing units and cloud computing platforms; Memory and processor: The memory stores a computer program, and the processor executes the computer program to implement the digital twin control method for disaster relief robots according to any one of claims 1-9.

Images