A mobile robot cluster distributed safety learning control method

Abstract

This invention discloses a distributed safety learning control method for mobile robot swarms, relating to the field of robot cooperative control technology. The method includes: calculating the future short-term motion position based on a nominal control quantity, and calculating the predicted neighborhood distance by combining it with a neighborhood state set to form multiple sets of safety constraints; combining these multiple sets of safety constraints with environmental boundary information to form obstacle avoidance constraints, and establishing constraint priority rules with actuator limits; obtaining a hierarchical safety constraint set by adding constraint level identifiers; using the nominal control quantity as the optimization objective and the hierarchical safety constraint set as the constraint conditions to construct a local optimization subproblem, and performing a distributed solution using the alternating direction multiplier method to generate a safe and feasible control quantity. This invention achieves the output of safe and feasible control quantities without centralized computation, thereby improving the real-time performance and reliability of swarm cooperative control.

Description

Technical Field

[0001] This invention relates to the field of robot collaborative control technology, and in particular to a distributed safety learning control method for mobile robot clusters. Background Technology

[0002] With the development of multi-robot systems and swarm intelligence theory, mobile robot swarms have received widespread attention in scenarios such as warehousing and logistics, environmental monitoring, disaster relief, and intelligent manufacturing. In existing technologies, control methods for multi-mobile robot collaborative tasks typically achieve overall behavioral coordination through information interaction and collaborative decision-making. Common methods include consensus control, distributed optimization, and reinforcement learning. In this technological framework, each mobile robot acquires the state information of surrounding robots through neighborhood communication and updates its decisions within a certain control cycle, thus forming a distributed collaborative control strategy. Simultaneously, to ensure the stability and safety of system operation, related research is increasingly combining learning mechanisms with safety constraints. By constructing state feedback and control strategy update mechanisms within a process control framework, robot swarms can achieve autonomous collaboration and task execution in dynamic environments.

[0003] In existing technologies, most mobile robot swarm control methods, when performing distributed learning and collaborative decision-making, typically focus on the convergence or coordination of control strategies, with relatively limited consideration for dynamic safety constraints between robots and the environment. When robots update their strategies based on local information, without a unified mechanism for constructing and hierarchically processing safety constraints, it may be difficult to simultaneously achieve both effective collaborative control and operational safety under complex environments or high-density swarm operation conditions, thus affecting the overall reliability of the system control. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a distributed safety learning control method for mobile robot clusters, which solves the problem of safety constraint coordination in the collaborative control process of mobile robot clusters.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a distributed safety learning control method for mobile robot clusters, comprising: each mobile robot collecting its own real-time state within a fixed control period and receiving neighborhood state information from neighboring mobile robots through neighborhood communication to construct a local state vector and a neighborhood state set; performing differential calculation on the local state vector and the neighborhood state set to obtain a cooperative control error, and constructing an initial nominal control strategy based on the local state vector; updating the parameters of the initial nominal control strategy online using the cooperative control error to generate a nominal control quantity; calculating the future short-term motion position based on the nominal control quantity, and calculating the neighborhood prediction distance in conjunction with the neighborhood state set to form a multi-class safety constraint set; combining the multi-class safety constraint set with environmental boundary information to form obstacle avoidance constraints, and establishing constraint priority rules in conjunction with actuator limits, obtaining a hierarchical safety constraint set by adding constraint level identifiers; using the nominal control quantity as the optimization objective and the hierarchical safety constraint set as the constraint condition to construct a local optimization sub-problem, and performing a distributed solution using the alternating direction multiplier method to generate a safe and feasible control quantity.

[0008] As a preferred embodiment of the distributed secure learning control method for mobile robot clusters described in this invention, each mobile robot collects its own real-time state within a fixed control period and receives neighborhood state information from neighboring mobile robots through neighborhood communication to construct a local state vector and a neighborhood state set. The specific steps are as follows:

[0009] At the beginning of each fixed control cycle, each mobile robot acquires the cycle timestamp and latches the positioning, odometry and inertial measurement data in a fixed order to form its own real-time state. Based on its own real-time state, it performs local state updates and generates a local state vector.

[0010] The neighborhood broadcast message is constructed based on the local state vector, and the neighborhood state information is received and the validity verification, deduplication and disorder processing, time freshness filtering and quality flag filtering are performed in sequence, and the updated neighbor table is output.

[0011] Perform a freshness scan on the updated neighbor table and assemble it into a neighborhood state set by the neighbor identifier index.

[0012] As a preferred embodiment of the distributed safety learning control method for mobile robot clusters described in this invention, the specific steps for obtaining the cooperative control error by performing difference calculation between the local state vector and the neighborhood state set are as follows:

[0013] Within each fixed control cycle, the motion state quantities in the local state vector are read, and the quality flag and timestamp freshness are determined for each item in the neighborhood state set to obtain the set of valid neighborhood entries.

[0014] The motion state information of each neighbor in the set of valid entries in the neighborhood is classified and organized by field, and the central representative value of each field is extracted and combined to form the neighborhood reference state.

[0015] Based on the neighborhood reference state and the local state vector, differential calculation is performed in the order of preset fields, and each difference is encapsulated and output in a fixed order to generate the collaborative control error.

[0016] As a preferred embodiment of the distributed safety learning control method for mobile robot clusters described in this invention, the step of constructing an initial nominal control strategy based on the local state vector refers to reading the local linear velocity and local angular velocity from the local state vector as the initial linear velocity command and the initial angular velocity command, respectively, and executing the actuator to limit the amplitude to obtain the initial nominal control quantity. At the same time, the linear velocity parameter group and the angular velocity parameter group are initialized to form the initial nominal control strategy.

[0017] As a preferred embodiment of the distributed safety learning control method for mobile robot clusters described in this invention, the specific steps for updating the initial nominal control strategy online using cooperative control errors to generate nominal control quantities are as follows:

[0018] Linear velocity difference and heading difference are extracted from the cooperative control error, which drive the update of the linear velocity parameter set and angular velocity parameter set in the initial nominal control strategy. The two parameter sets are then updated once according to the preset update direction and fixed learning rate to obtain the updated two parameter sets.

[0019] Parameter limiting is applied to the two updated parameter sets, and the updated nominal control strategy is output. The initial nominal control quantity is corrected online using the updated nominal control strategy, and actuator limiting is applied again to output the nominal control quantity.

[0020] As a preferred embodiment of the distributed safety learning control method for mobile robot clusters described in this invention, the specific steps for calculating the future short-term motion position based on the nominal control quantity are as follows:

[0021] Read the nominal control variables and the current position and current heading from the local state vector. At the same time, determine the control cycle duration as the future short-term prediction time window and write it into the prediction buffer. Output the short-term prediction state set.

[0022] Short-term motion propulsion calculations are performed on the local state vector based on the short-term predicted state set to obtain the future short-term motion position.

[0023] As a preferred embodiment of the distributed safety learning and control method for mobile robot clusters described in this invention, the specific steps for forming multiple sets of safety constraints are as follows:

[0024] Perform short-term motion propagation calculations on each neighbor state in the neighborhood state set to obtain the neighbor's future short-term position, and calculate the neighborhood prediction distance between the future short-term motion position and the neighbor's future short-term position for each state.

[0025] The predicted distance of each neighborhood is compared with the safety distance threshold. When the predicted distance of the neighborhood meets the safety distance threshold, a predicted collision constraint is generated. When the predicted distance does not meet the safety distance threshold, a safety maintenance constraint is generated. The two types of constraints are classified and encapsulated to form a set of multiple safety constraints.

[0026] As a preferred embodiment of the distributed safety learning and control method for mobile robot clusters described in this invention, the specific steps for forming obstacle avoidance constraints by combining multiple sets of safety constraints with environmental boundary information are as follows:

[0027] Read multiple sets of safety constraints and environmental boundary information, and organize the environmental boundary information into environmental boundary constraint terms;

[0028] The boundary determination of the future short-term movement position generates environmental boundary obstacle avoidance constraints, which are then merged with various constraint items in multiple safety constraint sets to form obstacle avoidance constraints.

[0029] As a preferred embodiment of the distributed safety learning control method for mobile robot clusters described in this invention, the specific steps of establishing constraint priority rules by combining actuator limits and obtaining a hierarchical safety constraint set by adding constraint level identifiers are as follows:

[0030] Read the actuator limits and determine the limit proximity of the nominal control quantity to form a constraint priority rule;

[0031] Based on obstacle avoidance constraints and constraint priority rules, constraint level identifiers are sequentially added to the predicted collision constraints and safety maintenance constraints, and then encapsulated into high-level constraint subsets and low-level constraint subsets according to the constraint level to generate a hierarchical safety constraint set.

[0032] As a preferred embodiment of the distributed safety learning control method for mobile robot clusters described in this invention, the method involves using the nominal control quantity as the optimization objective and the hierarchical safety constraint set as the constraint condition to construct a local optimization sub-problem. The alternating direction multiplier method is then used for distributed solution to generate a safe and feasible control quantity. The specific steps are as follows:

[0033] Using the nominal control quantity as the optimization target reference value, and taking the high-level constraint subset and low-level constraint subset of the hierarchical safety constraint set as constraints, a local optimization subproblem is constructed;

[0034] The local optimization subproblem is decomposed into multiple distributed subproblems. The distributed subproblems are solved iteratively using the alternating direction multiplier method. In each iteration, local variable updates, neighborhood consistency variable updates, and multiplier variable updates are performed sequentially, and the updated control variables are output.

[0035] Perform hierarchical safety constraint verification and actuator limit verification on the updated control variables, and output safe and feasible control quantities.

[0036] The beneficial effects of this invention are as follows: by classifying multi-source safety constraints and setting priority rules, the key safety requirements under constraint conflicts are prioritized and satisfied, enabling the mobile robot to maintain safe and stable operation in process control; and by using the nominal control quantity as the target and combining the hierarchical safety constraints with the distributed solution of the alternating direction multiplier method, safe and feasible control quantities can be output without centralized calculation, thereby improving the real-time performance and reliability of cluster collaborative control. Attached Figure Description

[0037] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 A flowchart for a distributed safety learning control method for mobile robot clusters.

[0039] Figure 2 A flowchart for generating multiple sets of safety constraints.

[0040] Figure 3 Flowchart for distributed solution of control quantities for safety and feasibility.

[0041] Figure 4 Solve the residual convergence plot for ADMM.

[0042] Figure 5 This is a comparison chart of the real-time performance of cluster control. Detailed Implementation

[0043] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0044] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0045] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0046] Reference Figures 1-5 As one embodiment of the present invention, this embodiment provides a distributed safety learning control method for mobile robot clusters, comprising the following steps:

[0047] S1. Each mobile robot collects its own real-time status within a fixed control cycle and receives the neighborhood status information of neighboring mobile robots through neighborhood communication to construct a local state vector and a neighborhood state set.

[0048] S1.1 At the beginning of each fixed control cycle, each mobile robot obtains the cycle timestamp and latches the positioning, odometer and inertial measurement data in a fixed order to form its own real-time state. Based on its own real-time state, it performs local state updates and generates a local state vector.

[0049] It should be noted that at the beginning of each fixed control cycle, each mobile robot reads the clock and writes the cycle timestamp as the time reference for this cycle; it reads the current position and heading information output by the positioning system in a fixed order and latches it as positioning data; it reads the wheel speed or displacement increment information output by the odometer and latches it as odometer data; it reads the angular velocity and acceleration information output by the inertial measurement system and latches it as inertial measurement data, so that the positioning data, odometer data, and inertial measurement data form a consistent data frame within the same fixed control cycle; it uses the consistent data frame as input to perform local state update calculation, and calculates motion state variables such as position, heading, linear velocity, and angular velocity through state fusion to generate a local state vector containing motion state variables and cycle timestamps.

[0050] It should also be noted that the fixed order is the sensor reading order, and the example order is the positioning data reading order, the odometer data reading order, and the inertial measurement data reading order.

[0051] S1.2 Construct neighborhood broadcast messages based on local state vectors, and receive neighborhood state information to perform legality verification, deduplication and out-of-order processing, time freshness filtering and quality flag filtering in sequence, and output the updated neighbor table.

[0052] It should be noted that the process involves encapsulating neighborhood broadcast messages in field order based on the local state vector and sending them via neighborhood communication, while simultaneously receiving neighborhood state information. The process includes checking the integrity and value range of each field in the neighborhood state information for validity; removing duplicate neighborhood state information by neighbor identifier and periodic timestamp and rearranging it by periodic timestamp to complete deduplication and disordering; calculating the time difference between the periodic timestamp of the neighborhood state information and the local periodic timestamp and removing neighborhood state information exceeding the freshness threshold to complete time freshness filtering; comparing the quality flag with preset quality flag conditions and removing neighborhood state information that does not meet the preset quality flag conditions to complete quality flag filtering; writing the filtered and retained neighborhood state information into the neighbor table by neighbor identifier index; and outputting the updated neighbor table.

[0053] It should also be noted that the freshness threshold is set as the upper limit of the allowed communication delay range according to the fixed control period duration. The example value is the fixed control period duration or an integer multiple of the fixed control period duration.

[0054] The quality flag condition is set according to the quality flag written when the neighborhood state information is generated. The quality flag condition is that the quality flag is in a valid state.

[0055] The field order is set by the fixed order in which motion state variables are written into the local state vector during the local state vector generation phase;

[0056] Neighborhood broadcast messages are state data frames that a mobile robot encapsulates based on its local state vector and sends to other mobile robots in its neighborhood during each fixed control cycle. These frames are used to share information such as the robot's current position, heading, linear velocity, angular velocity, and cycle timestamp, enabling neighboring mobile robots to obtain the robot's latest state.

[0057] S1.3 Perform a freshness scan on the updated neighbor table and assemble it into a neighborhood state set according to the neighbor identifier index.

[0058] It should be noted that, for each neighbor status information in the updated neighbor table, the neighbor identifier and periodic timestamp are read one by one. The time difference between the periodic timestamp and the local periodic timestamp is calculated, and the neighbor status information exceeding the freshness threshold is deleted to complete the freshness scan. The neighbor status information retained through the freshness scan is indexed and organized according to the neighbor identifier. The neighbor status information corresponding to the same neighbor identifier is recorded and assembled according to the preset field order to form a neighbor status set.

[0059] S2. Perform differential calculation on the local state vector and the neighborhood state set to obtain the cooperative control error, and construct an initial nominal control strategy based on the local state vector. Update the parameters of the initial nominal control strategy online through the cooperative control error to generate the nominal control quantity.

[0060] S2.1. Read the motion state quantities in the local state vector within each fixed control cycle, and perform quality flag determination and timestamp freshness determination on each item in the neighborhood state set to obtain the set of valid entries in the neighborhood.

[0061] It should be noted that within each fixed control cycle, position, heading, linear velocity, and angular velocity are read from the local state vector as motion state variables, and neighborhood state information in the neighborhood state set is read one by one; for each piece of neighborhood state information, a quality flag is read and compared with preset quality flag conditions, and neighborhood state information that does not meet the preset quality flag conditions is deleted to complete the quality flag determination; for neighborhood state information that passes the quality flag determination, the cycle timestamp is read and the time difference between the cycle timestamp and the local cycle timestamp is calculated, and neighborhood state information whose time difference exceeds the freshness threshold is deleted to complete the timestamp freshness determination; the neighborhood state information that passes the quality flag determination and timestamp freshness determination is aggregated to form a set of valid neighborhood entries.

[0062] S2.2. Classify and organize the motion state information of each neighbor in the set of valid entries in the neighborhood by field, and extract the central representative value of each field to form the neighborhood reference state.

[0063] It should be noted that the motion state information of each neighbor is read one by one from the set of valid entries in the neighborhood, and the position, heading, linear velocity, and angular velocity are summarized into position field set, heading field set, linear velocity field set, and angular velocity field set, respectively. The center representative value is calculated for each of the position field set, heading field set, linear velocity field set, and angular velocity field set. The example value is the mean or median. The center representative values ​​of position, heading, linear velocity, and angular velocity are combined in the order of preset fields to form the neighborhood reference state.

[0064] It should also be noted that the neighbor motion state information refers to the state data generated by the neighbor mobile robot within its fixed control cycle based on its own local state vector and sent through the neighborhood broadcast message. It usually includes fields such as the neighbor robot's position, heading, linear velocity, angular velocity, and period timestamp. The neighbor motion state information is obtained by this robot after receiving the neighborhood broadcast message through neighborhood communication and after legality verification and freshness filtering.

[0065] S2.3 Based on the neighborhood reference state and the local state vector, perform differential calculations in the order of preset fields and encapsulate and output the collaborative control error in a fixed order.

[0066] It should be noted that motion state quantities are read based on neighborhood reference states and local state vectors, and the neighborhood reference state field sequence and local state vector field sequence are aligned according to a preset field order; differential calculation is performed on the aligned motion state quantities to obtain each motion state difference; each motion state difference is encapsulated in a fixed order to form a cooperative control error.

[0067] S2.4 Read the local linear velocity and local angular velocity from the local state vector as the initial linear velocity command and the initial angular velocity command respectively, and execute the actuator to limit the amplitude to obtain the initial nominal control quantity. At the same time, initialize the linear velocity parameter group and the angular velocity parameter group to form the initial nominal control strategy.

[0068] It should be noted that the local linear velocity is read from the local state vector and assigned as the initial linear velocity command, and the local angular velocity is read from the local state vector and assigned as the initial angular velocity command. The initial linear velocity command is subjected to range constraint processing with the upper and lower limits of linear velocity, and the initial angular velocity command is subjected to range constraint processing with the upper and lower limits of angular velocity to obtain the initial nominal control quantity. The initial linear velocity command is written into the linear velocity parameter group as the initial value of the linear velocity parameter group, and the initial angular velocity command is written into the angular velocity parameter group as the initial value of the angular velocity parameter group. The initial nominal control quantity, the linear velocity parameter group, and the angular velocity parameter group are combined to form the initial nominal control strategy.

[0069] It should also be noted that the upper limit of linear speed is set based on the maximum permissible linear speed parameter of the mobile robot actuator; the lower limit of linear speed is set based on the minimum permissible linear speed parameter of the mobile robot actuator.

[0070] The initial nominal control strategy refers to the basic control strategy generated based on the current motion state of the machine without considering neighborhood cooperative errors and safety constraints. It is used to provide initial control references for linear velocity and angular velocity, and to provide a benchmark control framework for subsequent parameter updates and online correction of control quantities based on cooperative control errors.

[0071] S2.5 Extract the linear velocity difference and heading difference from the cooperative control error, drive the update of the linear velocity parameter set and angular velocity parameter set in the initial nominal control strategy, and update the two parameter sets once according to the preset update direction and fixed learning rate to obtain the updated two parameter sets.

[0072] It should be noted that in the cooperative control error, the linear velocity differential field and the heading differential field are located according to a fixed field order. Based on the field position, the linear velocity differential field value is written to the linear velocity parameter group update input position, and the heading differential field value is written to the angular velocity parameter group update input position. The direction of increase or decrease of the linear velocity parameter group parameter value is determined according to the preset update direction, and the linear velocity parameter group is updated once according to a fixed learning rate. The direction of increase or decrease of the angular velocity parameter group parameter value is determined according to the preset update direction, and the angular velocity parameter group is updated once according to a fixed learning rate, resulting in the updated linear velocity parameter group and angular velocity parameter group.

[0073] It should also be noted that the update direction is set according to the sign of the corresponding differential component in the collaborative control error. When the collaborative control error is positive, it is set to the direction of parameter increase, and when the collaborative control error is negative, it is set to the direction of parameter decrease.

[0074] S2.6. Perform parameter limiting on the two updated parameter groups, output the updated nominal control strategy, use the updated nominal control strategy to correct the initial nominal control quantity online, and perform actuator limiting again to output the nominal control quantity.

[0075] It should be noted that parameter limiting processing is performed on the updated linear velocity parameter set and angular velocity parameter set. This is achieved by applying interval constraints to the linear velocity parameter set values ​​against its upper and lower limits, and vice versa. This results in parameter-limited linear velocity parameter sets and parameter-limited angular velocity parameter sets, forming the updated nominal control strategy. The updated nominal control strategy is then used to calculate the online correction to the initial nominal control quantity, resulting in the corrected nominal control quantity. Finally, actuator limiting processing is performed on the corrected nominal control quantity, and the nominal control quantity is obtained through interval constraints on both linear velocity and angular velocity commands.

[0076] It should also be noted that the upper limit values ​​of the linear velocity parameter group and the angular velocity parameter group are set synchronously when initializing the linear velocity parameter group and the angular velocity parameter group. The upper limit values ​​are determined according to the maximum control command range allowed by the mobile robot actuator.

[0077] The lower limits of the linear velocity parameter group and the angular velocity parameter group are set synchronously when initializing the linear velocity parameter group and the angular velocity parameter group. The lower limits of the parameters are determined according to the minimum control command range allowed by the mobile robot actuator.

[0078] S3. Calculate the future short-term motion position based on the nominal control quantity, and combine it with the neighborhood state set to calculate the neighborhood prediction distance, forming a set of multiple safety constraints.

[0079] S3.1 Read the nominal control quantity and the current position and current heading in the local state vector. At the same time, determine the control cycle duration as the future short-term prediction time window and write it into the prediction buffer. Output the short-term prediction state set.

[0080] It should be noted that within each fixed control cycle, the linear velocity command and angular velocity command in the nominal control quantity are located and the nominal control quantity is retrieved. The current position and current heading in the local state vector are located and the current position and current heading are retrieved. The control cycle duration corresponding to the fixed control cycle is used as the future short-term prediction time window and written into the prediction buffer. The nominal control quantity, current position, current heading and future short-term prediction time window are combined and encapsulated according to the preset field order, and the short-term prediction state set is output.

[0081] S3.2. Perform short-term motion propulsion calculation on the local state vector based on the short-term predicted state set to obtain the future short-term motion position.

[0082] It should be noted that, based on the linear velocity and angular velocity commands of the nominal control quantities in the short-term prediction state set, the future short-term prediction time window in the short-term prediction state set, and the current position and current heading in the local state vector, short-term motion propulsion calculation is performed, and the displacement increment and heading increment within the future short-term prediction time window are applied to the current position and current heading to obtain the future short-term motion position.

[0083] S3.3 Perform short-term motion propagation calculations on each neighbor state in the neighborhood state set to obtain the future short-term position of the neighbor, and calculate the neighborhood prediction distance between the future short-term motion position and the future short-term position of the neighbor for each state.

[0084] It should be explained that the neighboring states are traversed one by one in the neighborhood state set to locate the position, heading, linear velocity, and angular velocity of each neighboring state. The linear velocity of the neighboring state is combined with the future short-term prediction time window to obtain the neighboring displacement increment. The angular velocity of the neighboring state is combined with the future short-term prediction time window to obtain the neighboring heading increment. The neighboring displacement increment is applied to the position of the neighboring state according to the neighboring heading direction to obtain the neighboring short-term position. The coordinate difference between the future short-term position and the neighboring future short-term position is calculated to obtain the position difference component. The Euclidean distance is calculated on the position difference component to obtain the neighborhood prediction distance. The neighborhood prediction distance is then associated with the neighboring identifier and recorded.

[0085] The neighborhood prediction distance is obtained by performing Euclidean distance calculation on the position difference components, and the expression is:

[0086] ;

[0087] in, It is the neighborhood prediction distance, representing the distance between the future short-term movement position and the future short-term positions of its neighbors, in meters; It is the future short-term motion position in a planar coordinate system Coordinates, in meters; The future short-term position of the neighbor in a planar coordinate system Coordinates, in meters; It is the future short-term motion position in a planar coordinate system Coordinates, in meters; The future short-term position of the neighbor in a planar coordinate system Coordinates, in meters; It is a local mobile robot identifier used to indicate the mobile robot to which a short-term future location of movement belongs; It is a neighbor mobile robot identifier, used to indicate the neighbor mobile robot to which the neighbor's future short-term location belongs.

[0088] S3.4. Compare the predicted distance of each neighborhood with the safety distance threshold. When the predicted distance of the neighborhood meets the safety distance threshold, a predicted collision constraint is generated. When the safety distance threshold is not met, a safety maintenance constraint is generated. The two types of constraints are classified and encapsulated to form a set of multiple safety constraints.

[0089] It should be noted that the neighborhood predicted distance sequence is compared with the safety distance threshold one by one. The comparison and judgment determines the collision risk level by comparing the neighborhood predicted distance with the safety distance threshold. When the neighborhood predicted distance is less than or equal to the safety distance threshold, a predicted collision constraint is generated. When the neighborhood predicted distance is greater than the safety distance threshold, a safety holding constraint is generated. The set of predicted collision constraints and the set of safety holding constraints are encapsulated and combined according to the preset field order to form multiple sets of safety constraints.

[0090] It should also be noted that the safe distance threshold is set based on the outer envelope size of the mobile robot and the safe interval requirements for the operation of the mobile robot. It is calculated by combining the outer envelope radius of the mobile robot with the maximum possible movement distance of the mobile robot within a fixed control cycle.

[0091] Based on the mobile robot's outer envelope size, maximum operating speed, and effective and safe distance thresholds for neighborhood communication, the range of values ​​is greater than the mobile robot's outer envelope diameter and less than the mobile robot's communication neighborhood radius.

[0092] Predictive collision constraints are safety restrictions generated when the predicted distance between neighboring locations in the short term is less than or equal to a safe distance threshold. They are used to constrain the control variables of a mobile robot to avoid potential collisions with neighboring robots in the short term.

[0093] Safety maintenance constraints refer to safety interval constraints generated when the predicted distance to the neighborhood is greater than the safety distance threshold. They are used to maintain a reasonable distance between mobile robots while ensuring a safe distance, thereby maintaining a stable cooperative operation.

[0094] S4. Combine multiple safety constraint sets with environmental boundary information to form obstacle avoidance constraints, and establish constraint priority rules with actuator limits. Obtain hierarchical safety constraint sets by adding constraint level identifiers.

[0095] S4.1 Read multiple sets of safety constraints and environmental boundary information, and organize the environmental boundary information into environmental boundary constraint terms.

[0096] It should be noted that the process involves traversing multiple sets of safety constraints and retrieving them, while simultaneously identifying the boundary type and extracting boundary parameters from the environmental boundary information. This environmental boundary information is then converted into a standard field format that can be used for constraint verification, including boundary position parameters, boundary direction parameters, and boundary effective range parameters. The standard field format of the boundary position parameters, boundary direction parameters, and boundary effective range parameters is then encapsulated to form environmental boundary constraint items.

[0097] It should also be noted that the environmental boundary information is obtained from map data or environmental perception data of the mobile robot's operating environment. It can be derived from environmental maps, positioning map databases, or real-time detection results of environmental boundaries by LiDAR, visual sensors, etc. It is used to describe the boundary position, boundary direction and effective range of the operating area, and provides a basis for generating environmental boundary constraint terms.

[0098] S4.2. Perform boundary discrimination on the future short-term movement position to generate environmental boundary obstacle avoidance constraints, and merge them with the constraint entries in the multi-class safety constraint set to form obstacle avoidance constraints.

[0099] It should be explained that the future short-term movement position is compared with the environmental boundary constraints to determine the boundary. The relationship between the future short-term movement position and the boundary within the boundary defined by the environmental boundary constraints is calculated and the direction of crossing the boundary is identified. The cases in which the future short-term movement position falls into the prohibited area of ​​the environmental boundary constraints or crosses the boundary line of the environmental boundary constraints are encapsulated as environmental boundary obstacle avoidance constraints. The environmental boundary obstacle avoidance constraints are merged with the predicted collision constraints and safety maintenance constraints of the multiple types of safety constraints to form obstacle avoidance constraints.

[0100] It should also be noted that obstacle avoidance constraints refer to safety restrictions generated based on the predicted distance between the mobile robot's future short-term movement position and neighboring robots, as well as the positional relationship of the environmental boundary. These constraints are used to constrain control variables and prevent the mobile robot from colliding with neighboring robots or crossing environmental boundaries during its movement, thereby ensuring the safety of the mobile robot's movement.

[0101] S4.3 Read the actuator limits and make a limit proximity judgment on the nominal control quantity to form a constraint priority rule.

[0102] It should be noted that actuator limits are extracted from the mobile robot actuator parameter configuration. These limits include upper and lower limits for linear velocity, and upper and lower limits for angular velocity. The linear velocity command in the nominal control quantity is compared with the upper and lower limits to calculate the linear velocity approximation. Similarly, the angular velocity command in the nominal control quantity is compared with the upper and lower limits to calculate the angular velocity approximation. The limit approximation is determined by comparing the linear velocity approximation with the angular velocity approximation to determine the limit approximation level of the nominal control quantity. Based on the limit approximation level, the predicted collision constraint in obstacle avoidance constraints is set as a high-priority constraint, and the safety holding constraint in obstacle avoidance constraints is set as a low-priority constraint, forming a constraint priority rule. This rule prioritizes satisfying the predicted collision constraint during control optimization, and only satisfies the safety holding constraint when conditions permit, thus ensuring that critical safety constraints are executed first.

[0103] S4.4 Based on obstacle avoidance constraints and constraint priority rules, constraint level identifiers are sequentially added to the predicted collision constraints and safety maintenance constraints, and they are encapsulated into high-level constraint subsets and low-level constraint subsets according to the constraint level to generate a hierarchical safety constraint set.

[0104] It should be noted that, based on obstacle avoidance constraints and constraint priority rules, the obstacle avoidance constraint entries are traversed one by one. For each predicted collision constraint entry and each safety maintenance constraint entry, a constraint level identifier is written. The process of writing the constraint level identifier is completed by querying the priority identifiers corresponding to the predicted collision constraints and the safety maintenance constraints in the constraint priority rules. The predicted collision constraint entries with high priority identifiers are encapsulated and aggregated into a high-level constraint subset according to a preset field order. The safety maintenance constraint entries with low priority identifiers are encapsulated and aggregated into a low-level constraint subset according to a preset field order. The high-level constraint subset and the low-level constraint subset are merged and encapsulated to generate a hierarchical safety constraint set.

[0105] S5. Using the nominal control quantity as the optimization objective and the hierarchical safety constraint set as the constraint condition, construct a local optimization subproblem. Use the alternating direction multiplier method for distributed solution to generate a safe and feasible control quantity.

[0106] S5.1. Using the nominal control quantity as the optimization target reference value, and taking the high-level constraint subset and low-level constraint subset of the hierarchical safety constraint set as constraints, construct a local optimization subproblem.

[0107] It should be noted that the linear velocity and angular velocity commands in the nominal control quantities are used as reference values ​​for optimization variables to establish the optimization objective. The deviation minimization objective is formed by calculating the deviation between the optimization variables and the nominal control quantities. The high-level constraint subset and the low-level constraint subset in the hierarchical safety constraint set are converted into distance constraint expressions one by one. The high-level constraint subset corresponds to the distance limit of the predicted collision constraint, and the low-level constraint subset corresponds to the distance limit of the safety maintenance constraint. The deviation minimization objective and the distance constraint expression are written together into the optimization problem description to form a local optimization subproblem.

[0108] S5.2 Perform variable decomposition on the local optimization subproblem to form multiple distributed subproblems. Use the alternating direction multiplier method to solve the distributed subproblems iteratively. In each iteration, perform local variable update, neighborhood consistency variable update and multiplier variable update in sequence, and output the update control variable.

[0109] It should be noted that when performing variable decomposition on the local optimization subproblem, the optimization variables in the local optimization subproblem are split into local variables, neighborhood consistency variables, and multiplier variables, and the local optimization subproblem is divided into multiple distributed subproblems according to the variable affiliation. The alternating direction multiplier method is used to solve the multiple distributed subproblems iteratively. In each iteration, local variable updates are performed. Under the condition of fixed neighborhood consistency variables and multiplier variables, the local objective and local constraints are solved to obtain the updated local variables. Neighborhood consistency variable updates are performed. By aggregating the updated local variables in the neighborhood and calculating the consistency objective, the updated neighborhood consistency variables are obtained. Multiplier variable updates are performed. By calculating the consistency residual between the local variables and the neighborhood consistency variables and incrementally updating the multiplier variables, the updated multiplier variables are obtained. After the iterative cycle ends, the control variable corresponding to the neighborhood consistency variable is output as the updated control variable.

[0110] The updated neighborhood consistency variable is obtained by aggregating the updated local variables within the neighborhood and calculating the consistency target. The expression is as follows:

[0111] ;

[0112] in, It is the first The mobile robot in the first The updated neighborhood consistency variable obtained in the next iteration is used to represent the consistent values ​​of the control variable within the neighborhood. It is the first The mobile robot in the first The updated local variables obtained in the next iteration correspond to the control variables obtained by solving the local optimization subproblem. It is a neighborhood set The Middle The neighbor mobile robot in the first The updated local variables obtained in the next iteration; It is the first The neighborhood set of the i-th mobile robot, representing the set of neighbors with the i-th mobile robot. A set of neighboring mobile robots that have neighborhood communication relationships; It is the iteration count index of the alternating direction multiplier method; It is a local mobile robot identifier used to represent a mobile robot that performs the solution of local optimization subproblems; It is a neighbor mobile robot identifier used to represent a neighborhood set. The neighbor mobile robot.

[0113] like Figure 4 As shown, the method of the present invention demonstrates the convergence of the original and dual residuals with the number of iterations when using the Alternating Direction Multiplier Method (ADMM) for distributed solution. The horizontal axis represents the ADMM iteration number k, and the vertical axis represents the residual value. Both the original and dual residual curves gradually decrease and stabilize with increasing iteration count, approaching the preset convergence threshold within a finite number of iterations. This indicates that the present invention, through distributed iterative computation using the Alternating Direction Multiplier Method in solving local optimization subproblems, can achieve stable convergence of neighborhood consistency variables, thereby effectively outputting safe and feasible control quantities that satisfy the hierarchical safety constraint set. This verifies the stability and feasibility of the distributed solution process of the present invention.

[0114] S5.3 Perform hierarchical safety constraint verification and actuator limit verification on the updated control variables, and output safe and feasible control quantities.

[0115] It should be noted that the updated control variables are substituted one by one into the hierarchical safety constraint set for hierarchical safety constraint verification. The hierarchical safety constraint verification is completed by calculating the future short-term motion position corresponding to the updated control variable and checking the satisfaction relationship between the future short-term motion position and the predicted collision constraint and the safety maintenance constraint. If the hierarchical safety constraint verification is not satisfied, the updated control variable is determined to be an invalid control variable and the update control variable is rolled back to the previous round of alternating direction multiplier method. The updated control variables that pass the hierarchical safety constraint verification are substituted into the actuator limits for actuator limit verification. The actuator limit verification is completed by performing range constraint processing on the linear velocity command in the updated control variable with the upper and lower limits of linear velocity, and on the angular velocity command in the updated control variable with the upper and lower limits of angular velocity, thus forming a safe and feasible control quantity.

[0116] like Figure 5As shown, the method of the present invention and the control group method vary in terms of solution time per control cycle under different mobile robot scales. The horizontal axis represents the number of mobile robots N, and the vertical axis represents the 95th percentile (P95) of solution time per control cycle. Figure 5 The curve "Invention" represents a control method that uses nominal control quantities as the optimization objective and combines them with a hierarchical safety constraint set, employing the Alternating Direction Multiplier Method (ADMM) for distributed solution to obtain safe and feasible control quantities; the curve "Control Group" represents a control method using a centralized, unified solution approach. From Figure 5 As can be seen, with the increase in the number of mobile robots, the solution time of the control group method increases rapidly, while the solution time of the method of the present invention increases significantly less. It can still maintain a low computation time in the case of a large-scale mobile robot cluster, indicating that the present invention can effectively output safe and feasible control quantities without centralized computation, thereby significantly improving the real-time performance and scalability of collaborative control of mobile robot clusters.

[0117] exist Figure 5 In the comparative experiment shown, the control group method uses a centralized control optimization approach to achieve mobile robot cluster control. Specifically, within each fixed control cycle, each mobile robot uploads its own state information and the state information of its neighborhood to the central computing node. The central computing node then constructs a global control optimization problem and solves it using the control variables of all mobile robots as the overall optimization variables. This optimization problem uses the nominal control variables of each mobile robot as the reference target, minimizing the deviation between the control variables and the nominal control variables while simultaneously satisfying the safety distance constraints between mobile robots and the environmental boundary constraints, thereby obtaining the global control variable set. Subsequently, the central node sends the solved control commands to each mobile robot for execution.

[0118] In summary, this invention achieves the priority satisfaction of key safety requirements under constraint conflicts by classifying multi-source safety constraints and setting priority rules, thus enabling the mobile robot to maintain safe and stable operation in process control; and by using the nominal control quantity as the target and combining the graded safety constraints with the distributed solution of the alternating direction multiplier method, it achieves the output of safe and feasible control quantities without centralized calculation, thereby improving the real-time performance and reliability of cluster collaborative control.

[0119] It should be noted that 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A distributed safety learning control method for mobile robot clusters, characterized in that, include: Each mobile robot collects its own real-time status within a fixed control cycle and receives neighborhood status information from neighboring mobile robots through neighborhood communication, thus constructing a local state vector and a neighborhood state set. The cooperative control error is obtained by performing differential calculation between the local state vector and the neighborhood state set. An initial nominal control strategy is constructed based on the local state vector. The parameters of the initial nominal control strategy are updated online through the cooperative control error to generate the nominal control quantity. The future short-term motion position is calculated based on the nominal control quantity, and the neighborhood prediction distance is calculated by combining the neighborhood state set to form a multi-class safety constraint set; Multiple safety constraint sets are combined with environmental boundary information to form obstacle avoidance constraints, and constraint priority rules are established with actuator limits. Hierarchical safety constraint sets are obtained by adding constraint level identifiers. The nominal control quantity is used as the optimization objective, and the hierarchical safety constraint set is used as the constraint condition to construct a local optimization subproblem. The alternating direction multiplier method is used for distributed solution to generate a safe and feasible control quantity.

2. The distributed safety learning control method for mobile robot clusters as described in claim 1, characterized in that, Each mobile robot collects its own real-time state within a fixed control cycle and receives neighborhood state information from neighboring mobile robots through neighborhood communication to construct a local state vector and a neighborhood state set. The specific steps are as follows: At the beginning of each fixed control cycle, each mobile robot acquires the cycle timestamp and latches the positioning, odometry and inertial measurement data in a fixed order to form its own real-time state. Based on its own real-time state, it performs local state updates and generates a local state vector. The neighborhood broadcast message is constructed based on the local state vector, and the neighborhood state information is received and the validity verification, deduplication and disorder processing, time freshness filtering and quality flag filtering are performed in sequence, and the updated neighbor table is output. Perform a freshness scan on the updated neighbor table and assemble it into a neighborhood state set by neighbor identifier index.

3. The distributed safety learning control method for mobile robot clusters as described in claim 2, characterized in that, The specific steps for obtaining the cooperative control error by performing difference calculation between the local state vector and the neighborhood state set are as follows: Within each fixed control cycle, the motion state quantities in the local state vector are read, and the quality flag and timestamp freshness are determined for each item in the neighborhood state set to obtain the set of valid entries in the neighborhood. The motion state information of each neighbor in the set of valid entries in the neighborhood is classified and organized by field, and the central representative value of each field is extracted and combined to form the neighborhood reference state. Based on the neighborhood reference state and the local state vector, differential calculation is performed in the order of preset fields, and each difference is encapsulated and output in a fixed order to produce the collaborative control error.

4. The distributed safety learning control method for mobile robot clusters as described in claim 1, characterized in that, The construction of the initial nominal control strategy based on the local state vector refers to reading the local linear velocity and local angular velocity from the local state vector as the initial linear velocity command and the initial angular velocity command, respectively, and executing the actuator to limit the amplitude to obtain the initial nominal control quantity. At the same time, the linear velocity parameter group and the angular velocity parameter group are initialized to form the initial nominal control strategy.

5. The distributed safety learning control method for mobile robot clusters as described in claim 4, characterized in that, The steps for updating the initial nominal control strategy online using the cooperative control error to generate a nominal control quantity are as follows: Linear velocity difference and heading difference are extracted from the cooperative control error, which drive the update of the linear velocity parameter set and angular velocity parameter set in the initial nominal control strategy. The two parameter sets are then updated once according to the preset update direction and fixed learning rate to obtain the updated two parameter sets. Parameter limiting is applied to the two updated parameter sets, and the updated nominal control strategy is output. The initial nominal control quantity is corrected online using the updated nominal control strategy, and actuator limiting is applied again to output the nominal control quantity.

6. The distributed safety learning control method for mobile robot clusters as described in claim 5, characterized in that, The specific steps for calculating the future short-term motion position based on the nominal control quantity are as follows: Read the nominal control variables and the current position and current heading from the local state vector. At the same time, determine the control cycle duration as the future short-term prediction time window and write it into the prediction buffer. Output the short-term prediction state set. Short-term motion propulsion calculations are performed on the local state vector based on the short-term predicted state set to obtain the future short-term motion position.

7. The distributed safety learning control method for mobile robot clusters as described in claim 1, characterized in that, The specific steps for forming multiple sets of security constraints are as follows: Perform short-term motion propagation calculations on each neighbor state in the neighborhood state set to obtain the neighbor's future short-term position, and calculate the neighborhood prediction distance between the future short-term motion position and the neighbor's future short-term position for each state. The predicted distance of each neighborhood is compared with the safety distance threshold. When the predicted distance of the neighborhood meets the safety distance threshold, a predicted collision constraint is generated. When the predicted distance does not meet the safety distance threshold, a safety maintenance constraint is generated. The two types of constraints are classified and encapsulated to form a set of multiple safety constraints.

8. The distributed safety learning control method for mobile robot clusters as described in claim 7, characterized in that, The specific steps for combining multiple sets of safety constraints with environmental boundary information to form obstacle avoidance constraints are as follows: Read multiple sets of safety constraints and environmental boundary information, and organize the environmental boundary information into environmental boundary constraint terms; The boundary determination of the future short-term movement position generates environmental boundary obstacle avoidance constraints, which are then merged with various constraint items in multiple safety constraint sets to form obstacle avoidance constraints.

9. The distributed safety learning control method for mobile robot clusters as described in claim 8, characterized in that, The steps for establishing constraint priority rules by combining actuator limits and obtaining a hierarchical safety constraint set by adding constraint level identifiers are as follows: Read the actuator limits and determine the limit proximity of the nominal control quantity to form a constraint priority rule; Based on obstacle avoidance constraints and constraint priority rules, constraint level identifiers are sequentially added to the predicted collision constraints and safety maintenance constraints, and then encapsulated into high-level constraint subsets and low-level constraint subsets according to the constraint level to generate a hierarchical safety constraint set.

10. The distributed safety learning control method for mobile robot clusters as described in claim 9, characterized in that, The process involves using the nominal control quantity as the optimization objective and a hierarchical set of safety constraints as the constraints to construct a local optimization subproblem. A distributed solution using the alternating direction multiplier method is then employed to generate a safe and feasible control quantity. The specific steps are as follows: Using the nominal control quantity as the optimization target reference value, and taking the high-level constraint subset and low-level constraint subset of the hierarchical safety constraint set as constraints, a local optimization subproblem is constructed; The local optimization subproblem is decomposed into multiple distributed subproblems. The distributed subproblems are solved iteratively using the alternating direction multiplier method. In each iteration, local variable updates, neighborhood consistency variable updates, and multiplier variable updates are performed sequentially, and the updated control variables are output. Perform hierarchical safety constraint verification and actuator limit verification on the updated control variables, and output safe and feasible control quantities.

Images