Multi-robot cooperative control method and system

Abstract

This invention relates to the field of automation control, and more specifically, to a method and system for collaborative control of multiple robotic arms. The method first acquires motion data of multiple robotic arms in an overlapping space and maps it to three-dimensional poses. Then, based on the residuals between the poses and the ideal model, it calculates the information entropy decay coefficient, corrects the state transition logic to obtain a spatiotemporal occupancy probability distribution predicting the future probability of the spatial grid being occupied. Based on this distribution and combined with process flow dependencies, it sets an asymmetric task bias factor, constructs and solves a strategy utility mapping table, and obtains an asymmetric action command combination that satisfies Nash equilibrium. By integrating the rate of change of the probability distribution, it obtains the cumulative value of the belief state gradient, generates a state trigger signal through a double-threshold hysteresis comparison, and smoothly issues the execution command. This invention effectively improves the prediction reliability, overall decision-making optimality, and command execution smoothness of multi-arm collaborative operations.

Description

Technical Field

[0001] This invention relates to the field of automation control technology. More specifically, this invention relates to a method and system for the collaborative control of multiple robotic arms. Background Technology

[0002] In the field of industrial automation, especially in scenarios such as precision assembly and material sorting, the collaborative operation of multiple robotic arms in overlapping or shared workspaces has become a key technology for improving production efficiency.

[0003] Typically, sensors are used to obtain the joint angles or end-effector positions of the robotic arm, and forward kinematics is used to calculate its definite pose in three-dimensional space. Based on this, static or dynamic interference checks are performed, and then the path is replanned through optimization algorithms to avoid conflicts.

[0004] However, with the increasing complexity of tasks and the dynamic nature of the environment, the reliance on deterministic geometric models and instantaneous observations makes them lack robustness to sensor noise, model errors, and unstructured environmental disturbances. Minor execution deviations or observation noises can easily be misjudged as impending collisions, leading to unnecessary emergency braking or avoidance maneuvers, frequent interruptions to the work cycle, reduced overall efficiency, and increased equipment wear.

[0005] Secondly, most collaborative algorithms assume that all robotic arms are on equal footing when making decisions and use symmetric utility functions for optimization. This can easily lead to decision deadlock in assembly tasks involving process flow dependencies, where multiple robotic arms may simultaneously choose to avoid or compete for the same space due to similar evaluations, causing the system to stagnate or oscillate.

[0006] Finally, the transition from decision-making to execution is often based on fixed cycles or simple threshold triggers, without fully considering the continuity of system state changes. This may lead to non-smooth issuance of control commands and trigger mechanical shocks.

[0007] In summary, existing multi-manipulator collaborative control technologies have three major drawbacks: 1) It relies on deterministic pose, has poor robustness to noise and model errors, and is prone to false braking; 2) Using symmetric decision-making can easily lead to deadlock, oscillations, and passage conflicts; 3) The command triggering and issuance lacks a smooth hysteresis mechanism, which can easily cause mechanical shock. Summary of the Invention

[0008] To address the problems of perception uncertainty, symmetric decision-making prone to deadlock, and non-smooth command issuance in existing technologies, this invention provides a multi-manipulator collaborative control method and system that achieves robust prediction, asymmetric collaborative decision-making, and smooth execution.

[0009] In a first aspect, the present invention provides a multi-manipulator collaborative control method, the method comprising: Acquire motion data of multiple robotic arms located in an overlapping workspace, and map the motion data to the pose of each robotic arm in three-dimensional space; Based on the pose and the observation residuals of the actual motion data of the manipulator and the preset ideal model, the information entropy decay coefficient is calculated, and the preset state transition logic is corrected using the information entropy decay coefficient to generate a spatiotemporal occupancy probability distribution that characterizes the probability of spatial grid cells being occupied in the future. Based on the spatiotemporal occupancy probability distribution, discrete motion strategies are assigned to the robotic arms that are about to engage in spatial interaction, and an asymmetric task bias factor is set for the robotic arms according to the dependency of the tasks performed by each robotic arm in the overall process flow. By integrating the spatiotemporal occupancy probability distribution in the interaction region with the asymmetric task bias factor, a policy utility mapping table is constructed, and the policy utility mapping table is solved to obtain a policy combination that optimizes the overall system utility and satisfies Nash equilibrium, and asymmetric action instructions are output. The cumulative value of the belief state gradient is obtained by integrating the derivative of the spatiotemporal occupancy probability distribution over time. A state trigger signal is generated by hysteresis comparison of the accumulated value of the belief state gradient with preset activation and release thresholds; when the state trigger signal is valid, the execution command corresponding to the asymmetric action command is smoothly issued.

[0010] Preferably, acquiring motion data of multiple manipulators within an overlapping workspace includes: collecting static obstacle geometry information of the workspace through a global visual perception network, and collecting joint angle increments and end-effector load factors in real time through the built-in encoders and end-effector torque sensors of each manipulator; mapping the pose of each manipulator in three-dimensional space includes: calculating the absolute pose matrix of each manipulator component based on the joint angle increments and the rigid link length parameters of the manipulator using a robot forward kinematics mapping algorithm.

[0011] Preferably, the calculation of the information entropy decay coefficient includes: calculating the multimodal observation residual vector at the current sampling time; calculating its dispersion and disorder based on the information entropy of the multimodal observation residual vector; and normalizing the information entropy before substituting it into an exponential decay function to obtain the information entropy decay coefficient.

[0012] Preferably, the step of using the information entropy decay coefficient to correct the preset state transition logic and generate a spatiotemporal occupancy probability distribution to characterize the probability of spatial grid cells being occupied in the future includes: weighting and fusing a preset basic state transition matrix with a unit equal-dimensional matrix representing a uniform spatial distribution using the information entropy decay coefficient as the weight; when the observation noise is extremely high, causing the information entropy to increase, the confidence of the robot motion model decreases. At this time, it is assumed that it may appear in any reachable space at the next moment, so a uniform distribution is used as an informationless prior for fusion to obtain a corrected set of state transition beliefs; iterative calculation is performed based on the set of state transition beliefs to output a set of probability distributions.

[0013] Preferably, the discrete motion allocation strategy includes: setting an acceleration follow-up strategy, maintaining the original constant speed operation strategy, and deceleration retreat and avoidance strategy; the setting of asymmetric task bias factors includes: assigning a positive asymmetric task bias factor to the robot on the core process path and assigning a zero or negative asymmetric task bias factor to the robot on the auxiliary process path based on the topological dependency node position of the task performed by each robot in the total process flow.

[0014] Preferably, the step of constructing a strategy utility mapping table and solving the strategy utility mapping table to obtain a strategy combination that optimizes the overall system utility and satisfies Nash equilibrium includes: using the integral result of the spatiotemporal occupancy probability distribution in the interaction region as a strategy overlap penalty factor; integrating the strategy overlap penalty factor with the asymmetric task bias factor to construct a utility table that comprehensively evaluates safety costs and task advancement utility; and searching for a strategy combination that optimizes the overall system utility and where no single robot can obtain higher returns by changing its strategy alone by traversing the utility table.

[0015] Preferably, obtaining the cumulative value of the belief state gradient includes: calculating the partial derivative sequence of the spatiotemporal occupancy probability distribution with respect to time within a set sliding time observation window; and performing continuous time integration on all positive change components in the partial derivative sequence to obtain the cumulative value of the belief state gradient characterizing the trend of escalating physical intrusion intent.

[0016] Preferably, the generation of the state trigger signal includes: generating a valid state trigger signal when the accumulated value of the belief state gradient exceeds the activation threshold; invalidating the state trigger signal only when the accumulated value of the belief state gradient falls below the release threshold; and maintaining the state trigger signal from the previous moment when the accumulated value of the belief state gradient is between the activation threshold and the release threshold.

[0017] In a second aspect, the present invention provides a multi-manipulator collaborative control system, comprising: a global visual perception module, a multi-manipulator data acquisition module, a collaborative decision processor, a command smoothing module, and a memory; the memory is used to store a computer program, which, when executed by the collaborative decision processor, implements the steps of the multi-manipulator collaborative control method as described above.

[0018] The beneficial effects of this invention are: 1. By introducing dynamic residual compensation logic, the observation residuals of multimodal motion data and ideal model are transformed into information entropy decay coefficients that quantify the confidence level. This is used to correct the state transition logic and finally output the spatiotemporal occupancy probability distribution. This enables the logical distinction between real collision threats and background noise, avoiding unnecessary emergency braking triggered by misjudging small deviations as collisions. This ensures the continuity of the work process and the efficiency of equipment operation.

[0019] 2. An asymmetric task bias factor reflecting the dependence of the final assembly process is embedded in the game theory solution framework. This factor is assigned a differentiated value according to the topological position of the component held by the robot in the process chain. When constructing the strategy utility mapping table, the asymmetric task bias factor is weighted and fused with the probability distribution integral term representing safety risk. By solving the Nash equilibrium of this asymmetric game, a set of pure strategies with optimal overall utility and unique stability can be obtained. This allows the robot holding a high asymmetric task bias factor to obtain priority passage, while other robots follow a retreat strategy. This eliminates decision oscillations and stagnation caused by the isomorphism of multi-machine behavior, ensuring the smoothness of the assembly process and the correctness of the timing.

[0020] 3. By integrating the positive component of the time partial derivative of the spatiotemporal occupancy probability distribution, the cumulative value of the belief state gradient is obtained. This value accumulates the trend of escalating threat and filters out instantaneous pullback noise. Subsequently, a hysteresis comparison logic with Schmitt triggering characteristics is used to set activation and release thresholds. This ensures that the command is triggered only when the cumulative value of the belief state gradient strongly exceeds the activation threshold, and is released only when it falls back below the activation threshold. This constructs a stable trigger-hold region, ensuring that the state transition decision has the necessary inertia and stability, realizing the smooth and reliable issuance of control commands, protecting the servo system, and ensuring the stability of bus communication. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall process of the multi-manipulator collaborative control method provided in an embodiment of the present invention; Wherein: S1 is the multi-manipulator pose acquisition and 3D mapping step, S2 is the information entropy decay and spatiotemporal occupancy probability distribution generation step, S3 is the discrete action strategy allocation and asymmetric task bias factor setting step, S4 is the strategy utility table construction and Nash equilibrium solution step, S5 is the belief state gradient accumulation value calculation step, S6 is the double threshold hysteresis comparison and state trigger signal generation step, and S7 is the execution instruction smoothing step. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] S1: Acquire motion data of multiple robotic arms in an overlapping workspace and map the motion data to the pose of each robotic arm in three-dimensional space.

[0024] In complex working conditions, the raw data obtained from heterogeneous sensors such as global vision and joint encoders have different formats and physical meanings, and only reflect some aspects of the robot's movement, such as joint angles or end-effector loads.

[0025] If these discrete and heterogeneous data fragments are used directly for collaborative decision-making, it will lead to serious biases and lags in the understanding of the overall space occupation of the robotic arm, which in turn will cause incorrect obstacle avoidance commands or task deadlock.

[0026] To fuse and map multi-source, asynchronous motion sensing data onto the same three-dimensional reference frame, heterogeneous motion data is synchronously acquired and encapsulated. Readings from different sensors are aligned in time and space to form a multi-dimensional vector describing the instantaneous state of a single robot. Then, through a deterministic kinematic model, the internal motion data corresponding to the multi-dimensional vector describing the instantaneous state of a single robot is converted into a geometric matrix of the absolute position and attitude of each component of the robot in a shared three-dimensional workspace. In this way, the configuration of the joint space is uniquely mapped to the position in Cartesian space.

[0027] In practice, point cloud data is collected at a fixed frequency by a global 3D depth vision system located above the work area. The global 3D depth vision system may be an Intel RealSense D435i or a similar model. The implementer may adjust the frequency according to the specific implementation scenario. The fixed frequency is taken as an empirical value of 30Hz, or it may be adjusted by the implementer according to the specific implementation scenario.

[0028] Then, after voxel mesh filtering and downsampling preprocessing, the geometric contours of the static workbench and known large obstacles are extracted using region growing and segmentation to form an initial set of static obstacle boundary geometry, denoted as . Among them, voxel mesh filtering, downsampling, and region growing segmentation are all well-known techniques, and will not be described in detail in this invention.

[0029] The angle values ​​of each joint are read in real time by high-precision absolute encoders installed at each joint of each robotic arm. subscript Indicates the robot arm number. Indicates joint number K represents the total number of joints. The absolute encoder can be a multi-turn absolute encoder, which can be adjusted by the implementer according to the specific implementation scenario.

[0030] Simultaneously, by reading the force and torque components of the end effector in three coordinate axes using a six-dimensional force and torque sensor rigidly connected to the end effector, the load factor scalar of the external load state is obtained. .

[0031] The collected data is then encapsulated into a structured multimodal motion data packet. Used to represent robotic arms At any moment The instantaneous state.

[0032] This multimodal motion data packet contains the following data: timestamps. Joint angle vector Joint angle increment ,in The sampling period is determined by the fixed frequency of data acquisition and the end load factor. .

[0033] All robotic arms It can be synchronized to the scheduling master gateway via Ethernet, so that all data has a unified time base.

[0034] For each robotic arm It calls the corresponding robot forward kinematics model.

[0035] The model is based on the Denavit-Hartenberg (DH) parameters of the manipulator and describes the chain transformation relationship from the base coordinate system to the end effector coordinate system.

[0036] The input is the joint angle vector at the current moment. and the inherent link length of the robotic arm and connecting rod torsion angle By obtaining the DH parameters, the angle values ​​of each joint can be obtained. The corresponding homogeneous transformation matrix The method of using DH parameters to obtain the homogeneous transformation matrix is ​​a well-known technique and will not be elaborated further.

[0037] For example, for a six-DOF manipulator, the pose of its end effector relative to the base coordinate system can be determined by a 4x4 homogeneous transformation matrix. The precise representation is as follows: `end` represents the position of the actuator corresponding to the end effector of the robot. A 4×4 homogeneous transformation matrix is ​​used to characterize the pose because its dimension is strictly bound only to the inherent geometric properties of the three-dimensional Cartesian workspace, and is completely independent of the number of degrees of freedom of the robot itself. The number of degrees of freedom of the robot only determines the number of product terms of the local transformation matrix during the forward kinematic mapping process, and does not change the mathematical dimension of the final output. Therefore, regardless of the different degrees of freedom of the robot within the collaborative workspace, its absolute spatial state can converge to a standard 4×4 matrix.

[0038] in, Indicates based on the first The homogeneous transformation matrix between adjacent link coordinate systems calculated from the joint angles and corresponding DH parameters. This refers to a 6-DOF (degrees of freedom) robotic arm.

[0039] Through matrix multiplication, we finally obtain Its top-left 3x3 submatrix is ​​a rotation matrix. The 3x1 submatrix in the upper right corner is the position vector. Together, they define the absolute pose of the end effector in three-dimensional space, where It is a tool bias matrix. Since the gripper is a fixed rigid body with no relative motion degrees of freedom, the relative positional relationship from the mechanical wrist coordinate system to the fingertip is constant, which is expressed as a constant matrix and can be obtained through the production calibration parameters of the robot.

[0040] However, in order to calculate the space occupied by the manipulator links themselves for interference checks, it is also necessary to calculate the pose of each key point of the manipulator links, such as the joint axis and the midpoint of the link. The pose of each key point of the link can be achieved by calculating the intermediate transformation matrix.

[0041] For example, from the position of the robot arm base to the first pose matrix of each joint It can be accessed through the corresponding front Multiplying the transformation matrices together yields: .

[0042] Output a set of pose descriptions for each robotic arm i. }, where base represents the starting point corresponding to the robot arm base, and the corresponding set of key point positions { }

[0043] The process involves real-time acquisition of point cloud data of the scene using a global 3D depth vision system located above the work area. The point cloud data is then preprocessed with voxel mesh filtering and downsampling. A plane fitting algorithm is used to remove known background point clouds, such as those on the fixed workbench surface, thereby extracting the geometric contour information of other obstacles in the space. This forms an initial static obstacle boundary geometry set, denoted as... The plane fitting algorithm is a well-known technique and will not be elaborated further.

[0044] Where T represents the homogeneous transformation matrix. for The 3x3 rotation matrix extracted from it represents the spatial attitude of the robot's end effector; for The extracted 3x1 position vector represents the spatial position of the robotic arm's end effector. (Subscript) The specific robot arm number is identified, with superscripts base, end, and k representing the base, end effector, and first robot arm, respectively. Each joint.

[0045] S2: Based on the pose and the observation residuals of the actual motion data of the robot and the preset ideal model, calculate the information entropy decay coefficient, and use the information entropy decay coefficient to correct the preset state transition logic to obtain the spatiotemporal occupancy probability distribution used to characterize the probability of spatial grid cells being occupied in the future.

[0046] In the face of control errors, sensor noise, and unstructured environmental disturbances during continuous motion, data can be fragmented into independent data, making it impossible to characterize the uncertainties in the motion process. At the same time, small execution deviations or load fluctuations can easily be misjudged as impending collisions, causing the robot to frequently perform unnecessary emergency braking or avoidance maneuvers, which seriously disrupts the production cycle.

[0047] Furthermore, by comparing the actual motion data of the robotic arm with a preset ideal motion model, the observation residual, which reflects the current control accuracy and stability, is calculated. The larger the residual, the greater the degree to which the actual motion deviates from the ideal trajectory, and the less reliable the prediction of the future position.

[0048] Furthermore, the uncertainty is calculated using the information entropy theory to form a dynamic information entropy decay coefficient, which is used to correct the pre-set state transition, and finally obtains a probability distribution field that can represent the possibility of different spatial grid units being occupied at different times in the future.

[0049] Obtain the time of each robotic arm Key point location set and motion data packets back.

[0050] A linear predictor based on the assumption of uniform motion, the ideal model is based on the previous moment... Ideal location and speed The initial state of the ideal model is obtained based on the filtered state estimate from the previous sampling time, and the ideal prediction position is predicted at the current time. And calculate the multimodal observation residual vector. This vector is not a single scalar, but a collection that comprehensively reflects deviations in different dimensions, mainly containing two core components: positional residuals. and load dynamic residual .

[0051] Location residual , representing the Euclidean distance between the actual end position and the model-predicted position, is used to improve the numerical stability of multimodal residual fusion, specifically for the position residual. Divide by the preset maximum allowable position error Perform normalization to map its range to The interval, in which The empirical value can be set according to the repeatability and working conditions of the specific robot model. It can be taken as 0.01m and can be adjusted by the implementer according to the specific implementation scenario.

[0052] Load dynamic residual ,in The actual end-point load factor at the current moment. This is the rated load factor for this process. This is the upper limit of the sensor's range; this component is used to sense anomalies caused by external disturbances or unstable gripping.

[0053] final, , For positional residuals, For a normalized to The scalar value of the interval, thus allowing the multimodal observation residual vector to be used as a basis. Calculate the information entropy decay coefficient .

[0054] Due to information entropy It is an indicator that measures the uncertainty of a probability distribution, and it combines the multimodal observation residual vectors from the current and several recent historical moments. Treating it as a sample set, its empirical distribution is fitted using methods such as kernel density estimation, and the information entropy of this distribution is calculated. The higher the information entropy value, the more drastic and unpredictable the residual fluctuations, indicating higher uncertainty. The information entropy is mapped to an exponentially decaying function. Information entropy decay coefficient within the interval: .

[0055] in, The normalization constant is a hyperparameter whose empirical value is usually set based on the fluctuation range of information entropy in historical data, and can be taken as... , The number of state discretizations is adjusted by the implementer to adapt to the decay characteristics of the exponential function, depending on the specific implementation scenario.

[0056] The closer to 1, the higher the confidence level and the lower the uncertainty of the current motion; the closer to 0, the lower the confidence level and the higher the uncertainty.

[0057] After obtaining the information entropy decay coefficient, the preset state transition logic is modified to generate the spatiotemporal occupancy probability distribution.

[0058] Pre-set a basic state transition matrix This matrix is ​​built based on the kinematic constraints and default motion strategy of the manipulator, and represents the probability of moving from the current spatial grid cell to the adjacent cell at the next moment. For example, for a manipulator moving in a certain direction, the probability of moving to the grid cell in front of it is relatively high.

[0059] To introduce uncertainty, the information entropy decay coefficient is used as a weight for... With a uniform distribution matrix representing completely randomness and no prior information Perform convex combination: ; In the formula, Represents scalar multiplication, matrix This is the corrected set of state transition beliefs, using the current time's determined robot arm grid occupancy as the initial probability distribution, and utilizing the corrected transition matrix. Perform multi-step iterative forward prediction.

[0060] Let the initial time be A robotic arm occupies the grid. The probability is 1, and the future number can be predicted. Step occupancy probability distribution It can be accessed through Iterative calculation, where , To predict the upper limit of the number of steps; This represents matrix multiplication. Stacking the probability distributions of all prediction steps along the time dimension yields the spatiotemporal occupancy probability distribution. ,in, for The row vector, Given the total number of spatial grids, the transition matrix during a complete forward prediction process is... The information entropy decay coefficient is calculated based on the current time t and remains static during the iteration process.

[0061] Each value in this distribution is in The space between these two points represents the probability that a specific spatial location will be occupied by a robotic arm at some point in the future, thus expanding the deterministic geometric trajectory into probabilistic data that includes uncertainty.

[0062] S3: Based on the spatiotemporal occupancy probability distribution, discrete motion strategies are assigned to the robotic arms that are about to engage in spatial interaction, and an asymmetric task bias factor is set for the robotic arms according to the dependency relationship of the tasks performed by each robotic arm in the overall process flow.

[0063] When making collaborative decisions for multiple robotic arms based on the probability distribution of space occupancy, assigning equal avoidance strategies to all robotic arms solely based on symmetrical space occupancy threat information can easily lead to decision deadlock.

[0064] In highly coordinated assembly tasks, the sensory inputs and decision-making objectives of multiple robotic arms are highly similar, often resulting in the simultaneous calculation of the exact same retreat or advance trajectory.

[0065] This isomorphism in behavior can cause robotic arms to linger in space, yielding to each other, or to confront each other while simultaneously trying to occupy the same safe space, thus disrupting the continuity of the entire production process.

[0066] To address the collaborative deadlock problem caused by decision symmetry, the topological dependency of the final assembly process flow is quantified as an asymmetric task bias factor, which is then introduced as a top-level constraint into the game decision-making process.

[0067] The robots that are about to interact in the shared workspace are not in a completely equal competitive relationship; according to the process design, some of the components carried by the robots are the benchmark or prerequisite for the subsequent process, and their tasks have a higher time urgency.

[0068] By assigning a positive asymmetric task bias factor to robotic arms on such critical paths, an additional reward term is essentially added to their policy utility function, thereby breaking the symmetry in utility calculation among multiple decision-makers.

[0069] This naturally leads to a preference for prioritizing the manipulator with a high asymmetric task bias factor when solving for collaborative strategies, thus avoiding decision loops.

[0070] In practice, the collaborative scheduling master control core gateway first identifies all robotic arm pairs whose predicted occupancy probability distribution overlaps in space in the future short time domain, based on the preset spatial interaction early warning rules. The future short time domain refers to the next 2 seconds, which can be adjusted by the implementer according to the specific implementation scenario.

[0071] For each identified interactive robotic arm pair Each of them is assigned a discrete, finite set of action strategies. .

[0072] This set typically contains three basic strategies: This indicates an accelerated follow-up. Indicates maintaining a constant speed. It indicates slowing down and yielding.

[0073] S4: Integrate the spatiotemporal occupancy probability distribution in the interaction region with the asymmetric task bias factor, construct a policy utility mapping table, and solve the policy utility mapping table to obtain the Nash equilibrium policy combination that optimizes the overall system utility and prevents each robot from obtaining higher returns by changing its policy alone, and output asymmetric action commands.

[0074] In overlapping workspaces, if multiple robotic arms independently choose evasive actions based solely on symmetrical spatial threat information, they are highly susceptible to stalemates or standoffs due to isomorphic decision-making objectives.

[0075] Specifically, when faced with potential conflicts, multiple robotic arms may simultaneously calculate and execute the same deceleration and yielding strategy, resulting in a sharp drop in overall efficiency; or they may simultaneously determine that they can pass and compete for the same space, causing a collision risk.

[0076] The root cause lies in the lack of high-level constraints that can break the symmetry of decision-making, resulting in each decision-making entity being completely equal in utility calculation, and thus unable to form a collaborative order with a distinction between primary and secondary roles.

[0077] To address the behavioral conflicts and efficiency bottlenecks caused by decision symmetry, abstract process logic dependencies are transformed into computable decision bias weights.

[0078] The process flow diagram defines the sequence and dependencies of component assembly. The tasks performed by the robot are prerequisites for the subsequent processes to be carried out, and have higher time urgency and system importance.

[0079] By assigning positive asymmetric task bias factors to these robotic arms on the critical path, it is essentially to inject an additional reward term into their decision utility function, giving them priority when facing competition for space resources.

[0080] In practice, the collaborative scheduling master control core gateway first determines the spatiotemporal occupancy probability distribution. It can detect robotic arm pairs that may interact spatially in the near future in a short time domain, where the near future time domain is set to 2 seconds, which can be adjusted by the implementer according to the specific implementation method.

[0081] The judgment rule is: for any two robotic arms and If in the short term in the future, in the spatial region Points on Exceeding the preset interaction warning threshold If the condition is met, it is determined that the pair of robotic arms is about to enter an interactive state, where the interaction warning threshold is set. The empirical value is usually set to 0.4, but can be adjusted by the implementer according to the specific implementation method.

[0082] For each identified pair of interactions Each of them is predefined with a discrete, finite set of primitive action strategies. .

[0083] in, This represents an accelerated follow-up strategy. This represents a strategy of maintaining a constant speed. This represents a slowdown and yielding strategy.

[0084] Subsequently, asymmetric task bias factors are assigned to manipulators i and j in the interaction pair, respectively. and The asymmetric task bias factor is assigned based on the process flow data obtained from the upper-level manufacturing execution, the current assembly task sequence is parsed, and a task dependency directed graph is constructed.

[0085] By analyzing the node position of the component currently held by the robot in the graph, its critical path weight is calculated. For example, if the component of the robot is the only input for multiple subsequent assembly operations, its critical path weight is high.

[0086] Mapping function Transform this weight into a specific bias value: A weight mapping table is predefined. For example, nodes with 2 or more dependent processes are defined as absolute core paths with the highest weight; nodes with 1 or more subsequent dependent processes are defined as ordinary paths with medium weight; and nodes with no subsequent dependencies are defined as auxiliary paths with the lowest weight.

[0087] For robotic arms on the absolute core path It can be set to +Kk, where Kk is an empirical value of 50, which can be adjusted by the implementer according to the specific implementation method; for auxiliary path robotic arms, The value can be set to 0, and can be adjusted by the implementer according to the specific implementation method. In some scenarios where explicit yielding instructions are required, a small negative bias, such as -5, can even be set for the secondary robot arm, which can be adjusted by the implementer according to the specific implementation method.

[0088] This process ensures This introduces asymmetry at the decision input level. Next, we will construct a system targeting the interaction pairs. Strategy utility mapping table The table is a two-dimensional matrix, with each row corresponding to a robotic arm. strategy The column corresponds to the robotic arm strategy .

[0089] Each cell in the table stores a utility pair. robotic arm The utility Calculated by the following formula: ; In the formula, This is a global task progress weighting coefficient used to adjust the overall scale of progress gains. It is a hyperparameter, and its empirical value can be set according to production cycle requirements. It can be adjusted by the implementer according to the specific implementation method.

[0090] It is a strategy The basic driving benefits are defined as follows: It can be adjusted by the implementer according to the specific implementation method.

[0091] This is the aforementioned asymmetric task bias factor.

[0092] The risk sensitivity coefficient is used to amplify or reduce the impact of collision penalties. It is another hyperparameter with an empirical value of 50 to match the order of magnitude of the terms in the utility function. It can be adjusted by the implementer according to the specific implementation method.

[0093] It is a strategy pair The inherent conflict cost coefficient, exemplarily, The combination with the highest conflict cost can be set to 1.0; The lowest value for the combination can be set to 0.1; other combinations should take the middle value of 0.6, which can be adjusted by the implementer according to the specific implementation method.

[0094] Integral term Calculated in the overlapping region The joint probability of conflict occurring within the term is used as a modulation factor for the penalty term.

[0095] After constructing the utility mapping table, solve for the Nash equilibrium strategy combination. The Nash equilibrium is defined as follows: under this combination, no single robot arm can obtain higher utility by unilaterally changing its strategy, which is expressed as requiring the simultaneous satisfaction of: 1) for all ,have ;2) For all ,have .

[0096] Because of the introduction of asymmetric Δi and Δj, the utility matrix is ​​usually asymmetric, which guarantees the existence and uniqueness of pure policy Nash equilibrium solutions.

[0097] By traversal By considering several possible strategy combinations and verifying the above conditions, an equilibrium solution can be found. This is a stable strategy that achieves optimal overall utility and where no robotic arm is willing to deviate unilaterally.

[0098] Finally, this strategy combination is interpreted into specific asymmetric motion instructions, for example, for a robotic arm. Increased output speed The instructions for the robotic arm Output offset along preset path The instructions are given to complete collaborative decision-making.

[0099] S5: Integrate the derivative of the spatiotemporal occupancy probability distribution over time to obtain the cumulative value of the belief state gradient.

[0100] After making collaborative decisions based on the spatiotemporal occupancy probability distribution and generating asymmetric action instructions, the key to ensuring stable operation is how to safely and smoothly convert discrete decision instructions into continuous control signals of the underlying driver.

[0101] At the end of a high-dynamic, high-precision manufacturing production line, even minor environmental disturbances can affect the spatiotemporal occupancy probability distribution. It itself fluctuates on a microscopic timescale.

[0102] If the underlying network directly uses the decision logic of polling at fixed time intervals, or relies solely on a single truncation point... Whether the amplitude exceeds a certain absolute threshold to trigger the issuance of action commands will pose significant security and stability risks.

[0103] Because the probability amplitude is prone to fluctuating repeatedly due to noise near the trigger threshold, this decision mechanism will directly cause high-frequency jumps in the underlying control signal.

[0104] For heavy-duty, high-inertia robotic arms, this vibration reaching the drive end will cause the servo motor to repeatedly perform sudden starts and stops, exacerbating reducer wear and potentially causing the fieldbus to become blocked and out of control due to instantaneous overload.

[0105] Therefore, a smoothing driving mechanism based on time cumulative trend analysis and dual-threshold hysteresis comparison is introduced. Specifically, the probability distribution fluctuation at a single moment may just be noise, while the real risk accumulation manifests as an irreversible and continuously aggravating trend. Therefore, judgments should not be made based solely on instantaneous values, but rather on the cumulative effect of the acceleration of the probability distribution change over time.

[0106] By observing within a sliding observation time window, Integrating the forward derivative with respect to time yields a scalar of the cumulative gradient of the belief state. .

[0107] This value filters out high-frequency random fluctuations, indicating an irreversible trend of the spatial intrusion risk becoming clearer and more aggravated.

[0108] The specific implementation process is as follows: After obtaining the Nash equilibrium strategy combination and its corresponding asymmetric action instruction set Subsequently, instead of immediately sending the data to the driver, it continuously monitors the spatiotemporal occupancy probability distribution related to the interaction area of ​​the target robotic arm. The evolution trend is observed, and a stable, jitter-free state trigger signal is generated accordingly. .

[0109] First, within a pre-defined sliding time observation window of a fixed length... ,in The window width is a crucial hyperparameter, and its empirical value needs to be matched with the electromechanical inertia time constant of the field servo motor. It is set to... The time in seconds can be adjusted by the implementer according to the specific implementation scenario to ensure that noise at the control cycle level is smoothed out, while not lagging excessively behind actual risk changes.

[0110] Within the aforementioned sliding window, historical moments are sampled at a frequency higher than the control period. Snapshot of the spatiotemporal occupancy probability distribution The frequency higher than the control cycle is 1kHz, which can be adjusted by the implementer according to the specific implementation scenario.

[0111] Subsequently, the sequence of partial derivatives of this distribution with respect to the time dimension is calculated. Reflecting the moment The instantaneous rate at which the threat of space occupation intensifies or diminishes.

[0112] To focus on risk accumulation, through a The function performs half-wave rectification on the derivative sequence, removing all negative change components representing a decrease in threat and retaining only the positive change components. .

[0113] Next, these positive change components are processed in a sliding window. By performing continuous-time integration, the cumulative value of the belief state gradient is obtained. The calculation formula is as follows: ; In the formula, It is the cumulative value of the belief state gradient, representing the value in the past. The cumulative total of the increasing trend of spatial intrusion risk over a period of time, expressed as probability multiplied by time.

[0114] This is the width of the sliding observation window, a hyperparameter, with an empirical value of [value missing]. Second.

[0115] It is a historic moment The spatiotemporal occupancy probability distribution.

[0116] The integration operation itself acts as a low-pass filter, effectively smoothing out high-frequency spikes caused by momentary sensor obstruction or electrical noise, thus... It becomes a reliable indicator that smoothly and monotonously reflects macroeconomic risk trends.

[0117] S6: Generate a state trigger signal by hysteresis comparison of the accumulated value of the belief state gradient with preset activation and release thresholds.

[0118] Obtain the cumulative value of the belief state gradient Then, based on the accumulated value of the state gradient of this belief, a stable and jitter-free state trigger signal is obtained to ensure that the collaborative decision-making instructions can be issued and executed safely and reliably.

[0119] In practical industrial control, sensor noise, computational delay, and minor environmental disturbances can lead to… High-frequency, small-amplitude oscillations occur near the critical point characterizing risk accumulation.

[0120] If a single threshold decision mechanism is used, when Exceeding a certain fixed threshold It is triggered immediately and is extremely easy to cause exist The repeated up-and-down movement causes frequent changes in the state trigger signal.

[0121] This critical oscillation phenomenon, when transmitted to the underlying driver, will directly cause the robotic arm actuator to switch between start and stop states at high frequency. This not only severely wears down the hardware, but may also cause instability or even safety accidents due to command conflicts.

[0122] To address the inherent critical oscillation problem of the single-threshold decision mechanism, two different thresholds are set: a higher activation threshold and a lower activation threshold. and a lower release threshold It also provides state memory functionality, thereby constructing a stable trigger-hold-release state machine.

[0123] Specifically, when monitoring is in an untriggered state, only when the cumulative value of the belief state gradient is... A strong upward trend and a breakthrough to higher levels Only when the risk accumulation trend is confirmed to have clearly formed will the state trigger signal be set to valid.

[0124] Once triggered, it enters a hold state; in this state, even The noise level briefly dropped, but as long as it doesn't fall to a lower level... The trigger signal will remain valid from now on.

[0125] This characteristic of being easy to enter but difficult to exit, and A stable hysteresis range was created between them, effectively shielding... Random fluctuations within the critical region interfere with the triggering state.

[0126] The specific implementation process is as follows.

[0127] The cumulative value of the belief state gradient is obtained in real time. Then, it is input into a digital hysteresis comparator, with two key hyperparameters preset: activation threshold. and release threshold .

[0128] At the same time, a status register is maintained to store the current status trigger signal. Its initial value is usually set to 0, representing an untriggered state.

[0129] It executes cyclically in discrete control cycles, and within each cycle, it is based on the current time... The state trigger signal from the previous moment Update according to the following defined rules : If the state in the previous moment was not triggered, that is If it is in the monitoring waiting stage, then only focus on... Whether a sufficiently high level of risk confirmation has been achieved.

[0130] Only when Only then is it considered that the risk accumulation trend is undeniable, and the state trigger signal is set to valid, that is... .

[0131] like Regardless of how its value fluctuates, it will remain in a non-triggered state. Keep it at 0.

[0132] This ensures that triggering actions require clear, high-confidence risk signals, avoiding false triggering caused by transient noise spikes.

[0133] If the state in the previous moment was already triggered, that is If it is in the trigger-hold phase, it is relatively lenient in this phase in order to maintain the continuity of instruction execution.

[0134] Only when It continued to decline and eventually fell completely below the lower release threshold. Only when the risk accumulation trend has largely subsided is the state trigger signal removed, i.e., the state is set to [a state trigger signal]. .

[0135] like Still greater than or equal to Even if its value is and Fluctuations between We will also firmly maintain the value of 1.

[0136] Ensure that the trigger status signal is not affected by exist The immediate cancellation of minor dips nearby provides a stable time window for the smooth execution of instructions.

[0137] when It falls exactly between the two thresholds, i.e. If the previous state is a stable state within the interval, the state of the previous cycle remains unchanged.

[0138] The calculation formula corresponding to the above operation is: ; In the formula, It is the cumulative value of the belief state gradient calculated in real time, representing the integral of the risk escalation trend.

[0139] This is the activation threshold, a hyperparameter. Its empirical value needs to be set according to the system's stringency in risk assessment. It can be adjusted by the implementer according to the specific implementation scenario.

[0140] It is the release threshold, also a hyperparameter, and its empirical value must be less than 1 / 3. To form a hysteresis interval, it is usually taken as It can be adjusted by the implementer according to the specific implementation scenario.

[0141] This is the final generated state trigger signal, with a value of or .

[0142] If and only if the state trigger signal from jump to The daemon process is activated only when the instruction is issued at the rising edge of the signal.

[0143] S7: When the state trigger signal is valid, smoothly issue the execution instruction corresponding to the asymmetric action instruction.

[0144] A stable, jitter-free trigger signal is obtained. Next, the temporarily stored, discrete asymmetric action instructions need to be processed. It is transformed into a continuous flow of control commands that can be safely and smoothly executed by the driver.

[0145] If in Effective, meaning the instant it jumps to 1, will Parameters defined in the code, such as target speed and offset position, are sent directly and in a step-like manner to the servo driver, which will cause a sudden change in the motion state of the robot.

[0146] For high-inertia industrial robots, step commands to speed or position can cause huge acceleration changes. At best, this can lead to mechanical vibration, trajectory overshoot, and end effector jitter, seriously affecting assembly accuracy. At worst, it may trigger the overload protection of the servo drive or cause mechanical structural impact, damaging the life of the equipment.

[0147] When the state trigger signal When it is effective, it is not issued directly. Instead of the target value, it uses the actual motion state of the current robotic arm, such as the actual speed of the end effector. Actual location As the starting point, the instruction The target value extracted from the data, such as the target velocity. Target location As the endpoint, a trajectory with continuous acceleration, i.e. finite acceleration, is planned between the two states.

[0148] An S-shaped velocity curve can be used to ensure a smooth velocity curve and no sudden changes in acceleration, thus eliminating the impact of step commands.

[0149] The specific implementation process is as follows: When the collaborative scheduling master control core gateway detects a status trigger signal When a rising edge from 0 to 1 occurs, a high-priority instruction smoothing interrupt service is immediately triggered.

[0150] First, calculate and temporarily store the asymmetric action instruction set. , It is a structure that typically contains the identifier of the target robotic arm. Target action type and the corresponding specific parameter values. .

[0151] At the same time, query the target robotic arm The current actual motion state, including the real-time linear velocity of its end effector in Cartesian space. and real-time location .

[0152] Subsequently, according to and Call the corresponding S-shaped velocity planning algorithm, and set... The desired target speed .

[0153] Given current velocity Target speed And preset the maximum allowable acceleration during the motion process. and maximum permissible jerk That is, the derivative of acceleration.

[0154] and These are hyperparameters, and their empirical values ​​are determined by the robot's dynamic performance and safety regulations. For example, for a medium-load robot, Possible choice , Pick It can be adjusted by the implementer according to the specific implementation scenario.

[0155] Therefore, the calculation speed is reduced from Change to The required S-curve typically consists of seven stages: acceleration, uniform acceleration, deceleration, constant speed, acceleration-deceleration, uniform deceleration, and deceleration-deceleration.

[0156] First, determine the change in velocity. Whether it is sufficient to achieve a uniform acceleration segment, thereby determining the actual number of curve segments to be used, and calculating the total time required for the entire transition process. And every moment Planning speed Accelerated Planning .

[0157] Next, the S-shaped velocity curve Discretized into a series of high-frequency setpoint sequences At the same time, through integration And combined with real-time location Generate the corresponding position setpoint sequence .

[0158] Finally, these setpoint sequences Along with the robotic arm logo Encapsulate it into a timestamped final execution instruction package. .

[0159] Command Packet The actuators are periodically streamed to the target robot via a real-time motion control bus.

[0160] The driver strictly follows the received instructions. The sequence is used for tracking control. Since the sequence itself is generated by an S-shaped curve, the continuity of speed and acceleration is guaranteed, allowing the robot to smoothly and without impact transition from the current state to the new state required by collaborative decision-making.

[0161] The entire distribution process continues until the total time is reached. End, after which the robotic arm will remain in the position of... It operates under the defined new motion state, thus safely and stably realizing smooth dispatching operations in multi-manipulator collaborative control.

[0162] The present invention also provides a multi-manipulator collaborative control system. The system includes: a global visual perception module, a multi-manipulator data acquisition module, a collaborative decision processor, a command smoothing module, and a memory; the memory is used to store a computer program, which, when executed by the collaborative decision processor, implements the steps of the multi-manipulator collaborative control method described above.

[0163] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, specific embodiments have been described above. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims can be performed in a different order than that shown in the embodiments and still achieve the desired result. Additionally, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0164] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0165] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for collaborative control of multiple robotic arms, characterized in that, Includes the following steps: Acquire motion data of multiple robotic arms located in an overlapping workspace, and map the motion data to the pose of each robotic arm in three-dimensional space; Based on the pose and the observation residuals of the actual motion data of the manipulator and the preset ideal model, the information entropy decay coefficient is calculated, and the preset state transition logic is corrected using the information entropy decay coefficient to generate a spatiotemporal occupancy probability distribution that characterizes the probability of spatial grid cells being occupied in the future. Based on the spatiotemporal occupancy probability distribution, discrete motion strategies are assigned to the robotic arms that are about to engage in spatial interaction, and an asymmetric task bias factor is set for the robotic arms according to the dependency of the tasks performed by each robotic arm in the overall process flow. By integrating the spatiotemporal occupancy probability distribution in the interaction region with the asymmetric task bias factor, a policy utility mapping table is constructed, and the policy utility mapping table is solved to obtain a policy combination that optimizes the overall system utility and satisfies Nash equilibrium, and asymmetric action instructions are output. The cumulative value of the belief state gradient is obtained by integrating the derivative of the spatiotemporal occupancy probability distribution over time. A state trigger signal is generated by hysteresis comparison of the accumulated value of the belief state gradient with preset activation and release thresholds; when the state trigger signal is valid, the execution command corresponding to the asymmetric action command is smoothly issued.

2. The multi-manipulator collaborative control method according to claim 1, characterized in that, The acquisition of motion data from multiple robotic arms within an overlapping workspace includes: The system collects static obstacle geometry information in the workspace through a global visual perception network, and collects joint angle increments and end-effector load factors in real time through the built-in encoders and end-effector torque sensors of each robot arm. The mapping of the pose of each manipulator in three-dimensional space includes: calculating the absolute pose matrix of each manipulator component based on the joint angle increment and the rigid link length parameters of the manipulator using a robot forward kinematics mapping algorithm.

3. The multi-manipulator collaborative control method according to claim 1, characterized in that, The calculation of the information entropy attenuation coefficient includes: Calculate the multimodal observation residual vector at the current sampling time; The discreteness and disorder of the multimodal observation residual vector are calculated based on its information entropy. After normalizing the information entropy, the exponential decay function is input to obtain the information entropy decay coefficient.

4. The multi-manipulator collaborative control method according to claim 1, characterized in that, The step of using the information entropy decay coefficient to correct the preset state transition logic and generate a spatiotemporal occupancy probability distribution to characterize the probability of spatial grid cells being occupied in the future includes: The preset basic state transition matrix and the unit equal-dimensional matrix representing the uniform distribution of the space are weighted and fused with the information entropy decay coefficient as the weight. When the observation noise is extremely large, which leads to an increase in information entropy, the confidence of the manipulator motion model decreases. At this time, it is assumed that the next moment may appear in any reachable space. Therefore, uniform distribution is used as an informationless prior for fusion to obtain the corrected state transition belief set. Iterative calculations are performed based on the state transition belief set to output a probability distribution set.

5. The multi-manipulator collaborative control method according to claim 1, characterized in that, The discrete allocation action strategy includes: Set up an acceleration follow-up strategy, a constant speed operation strategy, and a deceleration retreat and avoidance strategy; The setting of asymmetric task bias factors includes: assigning positive asymmetric task bias factors to robots on the core process path and assigning zero or negative asymmetric task bias factors to robots on the auxiliary process path, based on the topological dependency node position of the task performed by each robot in the total process flow.

6. The multi-manipulator collaborative control method according to claim 1, characterized in that, The process of constructing a policy utility mapping table and solving the policy utility mapping table to obtain a policy combination that optimizes the overall system utility and satisfies Nash equilibrium includes: The integral result of the spatiotemporal occupancy probability distribution in the interaction area is used as the policy overlap penalty factor; A utility table is constructed by integrating the strategy overlap penalty factor and the asymmetric task bias factor to comprehensively evaluate the safety cost and task advancement utility. The strategy combination that optimizes the overall system utility and ensures that no single robot can achieve higher returns by traversing the utility table is found.

7. The multi-manipulator collaborative control method according to claim 1, characterized in that, The process of obtaining the cumulative value of the belief state gradient includes: Within the set sliding time observation window, calculate the sequence of partial derivatives of the spatiotemporal occupancy probability distribution with respect to time; By performing continuous-time integration on all positively changing components in the partial derivative sequence, the cumulative value of the belief state gradient, which characterizes the trend of escalating physical intrusion intent, is obtained.

8. The multi-manipulator collaborative control method according to claim 1, characterized in that, The generation state trigger signal includes: When the accumulated value of the belief state gradient exceeds the activation threshold, a valid state trigger signal is generated; The state trigger signal is only invalidated when the accumulated value of the belief state gradient falls below the release threshold; When the cumulative value of the belief state gradient is between the activation threshold and the release threshold, the state trigger signal of the previous moment is maintained.

9. A multi-robot collaborative control system, characterized in that, The system includes: The system includes a global visual perception module, a multi-robotic arm data acquisition module, a collaborative decision processor, a smooth instruction delivery module, and a memory. The memory is used to store a computer program, which, when executed by the collaborative decision processor, implements the steps of the multi-manipulator collaborative control method as described in any one of claims 1 to 8.

Images