A robot control strategy selection and maintenance system based on dynamic capability card

Abstract

The application discloses a kind of robot control strategy selection and maintenance system based on dynamic capability card, belong to robot control, control strategy management and continuous learning technical field, by configuring dynamic capability card for each control strategy, with structured way record its applicable condition and safety constraint;When new strategy is added, the applicable condition overlap with existing strategy is identified by overlap detection module, and only relevant capability card is updated by local update module, irrelevant strategy remains unchanged, so as to maintain the stability of the calling relationship of existing strategy in the strategy library expansion process.Control strategy selection module matches each dynamic capability card to determine the target strategy according to task instruction, robot state and environment perception information, and triggers safety processing procedure when there is no applicable strategy or strategy is uncertain;Control output and execution module sends to execution mechanism after safety constraint check on control output, forms full-link protection, improves the accuracy of strategy selection and the safety of robot operation.

Description

Technical Field

[0001] This invention belongs to the fields of robot control, control strategy management and continuous learning, and particularly relates to a robot control strategy selection and maintenance system based on dynamic capability cards. Background Technology

[0002] As robot applications become increasingly complex, modern robot systems typically require multiple control strategies to handle different tasks and working conditions. Examples include strategies for legged robots to walk on flat ground, overcome obstacles, and climb stairs, and strategies for robotic arms to grasp, transport, and place objects. The system needs to select an appropriate strategy from a control strategy library based on the task objective, its own state, and environmental information.

[0003] Different control strategies vary significantly in terms of input requirements, applicable scope, control objectives, output format, and safety constraints. The correctness of strategy selection directly affects the safety and success rate of mission execution. Incorrect selection may lead to problems such as foot landing point deviation, machine instability, trajectory mismatch with terrain, or joint command exceeding limits, which may result in mission failure or even equipment collision accidents.

[0004] In existing technologies, learning-based methods (such as neural selectors or gating networks) implicitly embed policy selection logic within model parameters, lacking interpretability and making it difficult to intuitively verify the input conditions, applicable scope, and safety boundaries of a policy. More importantly, adding a new policy usually requires retraining or updating the selector, which may alter the selection boundaries of existing policies, affecting system stability and maintainability. While methods based on static labels or fixed descriptions facilitate retrieval, their expressive power is limited, making it difficult to accurately characterize multi-dimensional states and complex operating conditions. They also cannot automatically adjust the applicable boundaries of adjacent policies after adding a new policy, easily leading to multi-policy conflicts or matching confusion. Furthermore, existing solutions generally lack judgment and handling mechanisms for ambiguous situations such as the absence of applicable policies or policy uncertainty. Forcibly invoking a policy when information is insufficient or operating conditions exceed boundaries may output control quantities that do not conform to safety constraints, creating safety hazards.

[0005] Therefore, there is an urgent need for a control strategy selection and maintenance system that can achieve transparent matching of strategies, support localized maintenance of new strategies, and have a safe handling mechanism when decision-making conditions are insufficient. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a robot control strategy selection and maintenance system based on dynamic capability cards, comprising: The dynamic capability card registry is used to store the dynamic capability cards corresponding to each control strategy. The dynamic capability cards at least record the applicable conditions and security constraints of the corresponding control strategy. A control strategy library is used to store various robot control strategies; The control strategy application condition overlap detection module is used to compare the application conditions of the new control strategy with those of the existing control strategies when adding a new control strategy, and to identify whether there is any overlap. The Dynamic Capability Card Partial Update Module is used to update only the dynamic capability cards corresponding to newly added control strategies and control strategies whose applicable conditions overlap with them, while other dynamic capability cards remain unchanged. The control strategy selection module is used to determine the target control strategy from the control strategy library based on the matching results of the task instructions, robot status information, and environmental perception information with each dynamic capability card. When there is no applicable strategy or multiple applicable strategies cannot be uniquely determined, the corresponding inapplicable state or uncertain state is output, and the safety processing procedure is triggered. The control output and execution module is used to call the target control strategy to generate control output, and after performing a safety constraint check on the control output, send it to the robot driver or actuator control interface.

[0007] Optionally, the dynamic capability card also records at least one of the following fields: control strategy identifier, control task type, capability description, input state variable, boundary value, control output type, non-boundary factor, strategy initialization or switching condition, and strategy selection reason.

[0008] Optionally, the conditions under which the control strategy application condition overlap detection module determines that overlap exists are: When the same task simultaneously satisfies the boundary conditions of the newly added dynamic capability card and the existing dynamic capability card, or when there is an intersection in the numerical range, category set, inclusion relationship or logical condition of the semantically equivalent boundary factors that can be satisfied simultaneously, it is determined that there is an overlap; the control strategy application condition overlap detection module is also used to treat the situation where the boundary factors and boundary values ​​within the same dynamic capability card do not correspond, or multiple boundary conditions are contradictory and cannot be satisfied simultaneously as a boundary consistency error.

[0009] Optionally, when the dynamic capability card partial update module makes partial modifications, it only modifies the fields required to eliminate applicable condition conflicts or maintain the internal consistency of the capability card, while fields not involved in conflicts remain unchanged; the dynamic capability card partial update module only modifies the dynamic capability card record, without modifying the existing control policy body and its parameters, rules, and calling interfaces; when the modified dynamic capability card fails the integrity or consistency check, the dynamic capability card partial update module does not submit the modification and continues to use the dynamic capability card registry before the update.

[0010] Optionally, the control strategy selection module includes: The hard condition screening unit is used to exclude control strategies that do not meet the boundary hard conditions based on the applicable conditions and safety constraints in each dynamic capability card. The semantic matching unit is used to determine the semantic applicability of each control strategy based on the degree of matching between the task instructions and the capability descriptions and strategy selection reasons in the remaining dynamic capability cards. The decision output unit is used to form a candidate set of control strategies that simultaneously meet the boundary hard conditions and the preset applicability threshold. When the candidate set is empty, the output state is "no applicable strategy". When there are two or more strategies in the candidate set and the difference between the highest applicability and the second highest applicability is less than or equal to the preset uncertainty threshold, the output state is "strategy uncertainty". Otherwise, the strategy with the highest applicability is output as the target control strategy.

[0011] Optionally, when the current control strategy differs from the target control strategy, the control output and execution module is also used to check whether the robot's current state meets the initial state requirements of the target control strategy; if it does, the target control strategy is initialized with the robot's current state, and the strategy switching is completed through transition trajectory, control quantity smoothing, or intermediate stable state; if it does not meet the requirements, deceleration, attitude adjustment, motion stop, or return to a safe state is executed first, and then the strategy switching is executed. In the control output and execution module, when the target control strategy outputs a task execution plan, the trajectory generator or lower-level control strategy converts the task execution plan into a reference trajectory or control target. The robot controller calculates the underlying control instructions based on the reference trajectory or control target and the robot's real-time state. When the target control strategy directly outputs the underlying control instructions, the robot controller no longer performs control quantity conversion. The underlying control instructions are sent to the robot driver or actuator control interface after safety constraint checks.

[0012] Optionally, it also includes a task input module, a robot state acquisition module, and an environment perception module; the task input module is used to receive robot task instructions, the robot state acquisition module is used to acquire the robot's current state information, and the environment perception module is used to acquire environmental perception information.

[0013] Optionally, it also includes a task and state parsing module, used to convert task instructions, robot state information and environmental perception information into structured query conditions required for control strategy selection, and to generate the structured information required for generating an initial dynamic capability card based on the description of the new control strategy, representative tasks and verification tasks when registering a new control strategy.

[0014] Optionally, it also includes an execution feedback module, used to obtain robot status and task progress feedback after execution, and provide the feedback to the next control cycle; when the feedback information exceeds the applicable boundary of the current dynamic capability card, the execution feedback module triggers the control strategy selection module to reselect the control strategy.

[0015] Optionally, the safety processing procedure includes at least one of supplemental sensing, replanning, deceleration, or safe stopping.

[0016] Compared with the prior art, the present invention has the following advantages and technical effects: This system configures dynamic capability cards for each control strategy and explicitly records the applicable conditions and safety constraints of the strategy in the capability cards. This makes the scope of application of the strategy no longer implicit in the model parameters, but clearly visible in the form of structured data. Based on this, the control strategy selection module matches each capability card with the task instructions, robot status, and environmental perception information. It can accurately locate the applicable control strategy under the current working condition, effectively avoiding problems such as foot landing point deviation, robot instability, or joint command exceeding limits caused by incorrect strategy selection, thus reducing the risk of issuing erroneous control commands from the source.

[0017] Secondly, this system, by setting up a control strategy application condition overlap detection module and a dynamic capability card local update module, can automatically identify the overlap between the new and existing strategies in terms of application conditions when adding a new control strategy. It only performs local modifications on capability cards with overlap, while dynamic capability cards unrelated to the new strategy remain unchanged. This mechanism ensures that the expansion of the control strategy library does not trigger global changes in selection relationships, and the application scope of existing strategies remains stable. It avoids the problem of drifting selection boundaries of old task strategies due to retraining the selector in traditional solutions, significantly improving the maintainability and stability of the system during the continuous evolution of the strategy library.

[0018] Furthermore, during the matching process, when no applicable strategy exists or multiple applicable strategies cannot be uniquely determined, the control strategy selection module will not forcibly invoke a candidate strategy, but will instead trigger a safety processing procedure. This mechanism avoids the safety risks caused by the system outputting control quantities without sufficient evidence in scenarios where information is insufficient or the operating conditions exceed the boundaries of all strategies, thus providing safety assurance for the robot's operation in complex open environments.

[0019] Finally, after generating control output by invoking the target control strategy, the control output and execution module further performs a safety constraint check on the control output. Only after confirming that the constraint conditions are met is the output sent to the robot driver or actuator control interface. This process adds a safety barrier between strategy selection and underlying execution. Even if the strategy selection result or the strategy itself has potential risks, there is a chance to intercept them at the execution level, forming a complete safety closed loop from strategy matching to control output, effectively ensuring the safety of robot operation. Attached Figure Description

[0020] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a system overall structure diagram according to an embodiment of the present invention; Figure 2 This is a flowchart of the registration of new robot control strategies and partial update of dynamic capability cards in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the robot control strategy selection and control execution process according to an embodiment of the present invention. Detailed Implementation

[0021] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0023] Example 1 like Figure 1 As shown, this embodiment proposes a robot control strategy selection and maintenance system based on dynamic capability cards, forming a closed-loop control system consisting of task, state, and environment inputs, control strategy selection, control output, safety checks, robot execution, and feedback. It also includes a dynamic capability card maintenance branch for adding new control strategies. The system mainly includes a task input module, a robot state acquisition module, an environment perception module, a task and state parsing module, a dynamic capability card registry, a control strategy library, a control strategy application condition overlap detection module, a dynamic capability card partial update module, a control strategy selection module, a control output and execution module, and an execution feedback module.

[0024] The task input module is used to receive robot task instructions. The task instructions include one or more of the following: natural language task description, target pose, target region, upper-level task planning results, or operation instructions.

[0025] The robot state acquisition module is used to acquire the robot's current state information. Specifically, the robot state acquisition module acquires one or more of the following from sensors or state estimators: joint angles, joint velocities, body position and posture, foot or end effector contact state, force or torque information, and the current motion state.

[0026] The environmental perception module is used to acquire environmental perception information. The environmental perception information includes one or more of the following: obstacle height and width, number and continuity of steps, terrain slope and type, target object location and category, object stiffness or deformation characteristics, and workspace constraints.

[0027] The task and state parsing module is used to convert task instructions, robot state information, and environmental perception information into structured query conditions required for control strategy selection. Furthermore, when registering a new control strategy, the task and state parsing module is also used to generate the structured information required for the initial dynamic capability card.

[0028] The dynamic capability card registry is used to store the dynamic capability cards corresponding to each control strategy and to provide the current capability card to the control strategy selection and applicable condition overlap detection process. The dynamic capability card registry also associates and stores representative tasks, verification tasks, or corresponding record identifiers for each control strategy; these associated records are used for dynamic capability card coverage checks.

[0029] The control strategy library stores various robot control strategies. These control strategies include one or more of the following: motion controllers, reinforcement learning control strategies, task planners, trajectory generators, or robot skill modules.

[0030] Furthermore, the robot control strategy referred to in this embodiment refers to a callable control unit capable of generating reference trajectories, control objectives, low-level control instructions, or task execution plans based on robot state, task objectives, and environmental information. The robot control strategy includes one or more of the following: motion controller, reinforcement learning control strategy, task planner, trajectory generator, or robot skill module. Although the aforementioned control units may be at different control levels, they are uniformly considered as a selectable and callable robot control strategy.

[0031] The control strategy application condition overlap detection module is used to compare the application conditions of the new control strategy with those of existing control strategies when adding a new control strategy, and to identify whether there is any overlap. Specifically, the control strategy application condition overlap detection module compares the task type, state input, environmental conditions, and application boundaries of the new control strategy with those of existing strategies, identifies the intersection, inclusion, or conflict of application conditions, and determines the relevant dynamic capability cards that need to be locally maintained.

[0032] The dynamic capability card partial update module is used to update only newly added control strategies and control strategy capability cards that overlap with their applicable conditions; other control strategy capability cards remain unchanged. This module only modifies the dynamic capability card records and does not modify the existing control strategy body, its parameters, rules, or calling interfaces.

[0033] Furthermore, the control strategy selection based on dynamic capability cards mentioned in this embodiment refers to the process of matching the robot's task, current state, and environmental perception information with the dynamic capability cards corresponding to each control strategy, and determining the applicable control strategy from the control strategy library. This process belongs to control strategy matching and selection, and is not part of robot spatial path planning.

[0034] In this embodiment, the dynamic capability card is an updatable data record corresponding to the robot control strategy, used to record the input requirements, applicable working conditions, control objectives, control output forms, safety constraints, applicable boundaries, and initialization or switching conditions of the control strategy. "Dynamic" refers to the partial update of the relevant capability card when a new control strategy is registered, the applicable scope of the control strategy is re-verified, or an applicable boundary conflict occurs, rather than continuously changing the capability card in every underlying real-time control cycle.

[0035] As a specific implementation method, the control strategy selection module is used to exclude strategies that do not meet the input requirements, applicable conditions, or safety constraints based on the task objective, the robot's current state, environmental conditions, and dynamic capability card, and output the target control strategy, the state of no applicable control strategy, or the state of uncertain control strategy.

[0036] Specifically, the control strategy selection module includes a hard condition filtering unit, a semantic matching unit, and a decision output unit. The hard condition filtering unit excludes control strategies that do not meet the boundary hard conditions based on the applicable conditions and safety constraints in each dynamic capability card. The semantic matching unit determines the semantic applicability of each control strategy based on the degree of matching between the task instruction and the capability descriptions and strategy selection reasons in the remaining dynamic capability cards. The decision output unit forms a candidate set of control strategies that simultaneously meet the boundary hard conditions and a preset applicability threshold; when the candidate set is empty, it outputs a state of no applicable strategy; when there are two or more strategies in the candidate set and the difference between the highest and second-highest applicability is less than or equal to a preset uncertainty threshold, it outputs a strategy uncertainty state; otherwise, it outputs the strategy with the highest applicability as the target control strategy. When there is no applicable strategy or multiple applicable strategies cannot be uniquely determined, the control strategy selection module triggers a safety processing procedure, which includes at least one of supplementary sensing, replanning, deceleration, or safe stopping.

[0037] The control output and execution module is used to invoke the target control strategy and complete processing according to the output level. When the target control strategy outputs a task execution plan, the trajectory generator or lower-level control strategy converts the task execution plan into a reference trajectory or control target. The robot controller calculates the underlying control instructions based on the reference trajectory or control target and the robot's real-time state. When the target control strategy directly outputs the underlying control instructions, the robot controller no longer performs control quantity conversion. Regardless of whether the underlying control instructions are calculated by the robot controller or directly generated by the target control strategy, the control output and execution module performs safety constraint checks on the underlying control instructions, including joint limits, speed limits, torque limits, contact constraints, or collision constraints. After the checks pass, the control output and execution module sends the underlying control instructions to the robot driver or actuator control interface.

[0038] The execution feedback module is used to acquire the robot's state, contact state, and task progress after execution, and provides feedback to the next control cycle, safety check, or task completion determination. The execution feedback is also used to determine whether the current control strategy is still applicable to the changed robot state and environmental conditions; if it is no longer applicable, it triggers a re-selection of the control strategy.

[0039] In this embodiment, each robot control strategy corresponds to a dynamic capability card. The dynamic capability card records at least the applicable conditions and safety constraints of the corresponding control strategy.

[0040] Furthermore, each robot control strategy corresponds to a dynamic capability card. The dynamic capability card may contain natural language descriptions and structured fields to explain what inputs the control strategy requires, what operating conditions it is applicable to, what control outputs it produces, what safety constraints it must satisfy, and the robot state required to activate or switch to that strategy. The dynamic capability card records at least one of the following fields: control strategy identifier, control task type, capability description, input state variables, boundary values, control output type, non-boundary factors, strategy initialization or switching conditions, and the reason for strategy selection. The recorded content of each field and its control function are shown in Table 1: Table 1 The following is combined Figure 2 The process of registering the new control strategy and maintaining the dynamic capability card in this embodiment is described in detail: Step S301: Receive the new control policy, policy description, representative task, and verification task.

[0041] Step S302: Analyze the tasks, robot state inputs, and applicable working conditions for the newly added control strategy.

[0042] Step S303: Generate the initial dynamic capability card for the new control strategy.

[0043] Step S304: Compare the newly added dynamic capability card with the existing dynamic capability cards of the control strategy.

[0044] Step S305: Detect whether there is overlap in task type, status input, environmental conditions, and applicable boundaries. The control strategy applicable condition overlap detection module determines that overlap exists when the same task simultaneously satisfies the boundary conditions of both the newly added dynamic capability card and the existing dynamic capability card, or when there is an intersection in the numerical range, category set, inclusion relationship, or logical condition of semantically equivalent boundary factors that can be satisfied simultaneously. Furthermore, if the boundary factors and boundary values ​​within the same dynamic capability card do not correspond, or if multiple boundary conditions contradict each other and cannot be satisfied simultaneously, this is treated as a boundary consistency error.

[0045] Step S306: When there is no overlap, perform field integrity checks, boundary factor and boundary value consistency checks, strategy selection rationale consistency checks, and coverage checks on the representative tasks and verification tasks of the newly added control strategy on the initial dynamic capability card.

[0046] Step S307: If the aforementioned check passes, add the new control policy and dynamic capability card to the control policy library and dynamic capability card registry; if the check fails, terminate the registration of the new control policy, do not add it to the control policy library, and continue to use the dynamic capability card registry before the update.

[0047] Step S308: When overlap exists, only the newly added control strategy and the related overlapping control strategies are identified as objects to be updated.

[0048] Step S309: Partially modify the dynamic capability card of the object to be updated, modifying only the fields required to eliminate applicable condition conflicts or maintain internal consistency of the capability card; fields not involved in conflicts remain unchanged. The dynamic capability card partial update module only modifies the dynamic capability card record, without modifying the existing control policy body and its parameters, rules, and calling interfaces; existing control policies continue to maintain their original callable state. Simultaneously, control policy capability cards unrelated to the newly added control policy remain unchanged.

[0049] Step S310: Perform field integrity checks, boundary factor and boundary value consistency checks, strategy selection rationale consistency checks, coverage of representative and verification tasks for newly added control strategies, coverage of associated verification tasks for affected existing control strategies, and residual boundary overlap checks on the modified dynamic capability card.

[0050] Step S311: If the check passes, submit the newly added or modified dynamic ability card to the dynamic ability card registry; if the check fails, do not submit the newly added or modified card, and continue to use the dynamic ability card registry before the update.

[0051] The following is combined Figure 3 This embodiment provides a detailed explanation of the control strategy selection and robot control execution process. Control strategy selection can be triggered at the start of a task, when a task phase changes, when environmental conditions change significantly, when the current control strategy fails, or when execution feedback exceeds the applicable boundaries of the current dynamic capability card. Once the target control strategy is determined, it generates control outputs based on the robot's real-time state according to a preset real-time control cycle. Unless a re-selection trigger condition occurs, it is unnecessary to repeatedly perform large language model semantic parsing or dynamic capability card matching in each underlying control cycle.

[0052] After the control strategy is selected, the system calls the determined target control strategy to generate a task plan, reference trajectory, control target, or low-level control instructions. The task plan is converted into a reference trajectory or control target by the trajectory generator or low-level control strategy, and the robot controller then calculates the low-level control instructions based on the real-time status. Regardless of whether the low-level control instructions are calculated by the robot controller or directly generated by the target control strategy, they are sent to the robot driver or actuator control interface after safety constraint checks. The driver then drives the robot actuator and continues control or reselects the control strategy based on the execution feedback.

[0053] Specifically, the following steps are included: Step S401: Receive robot task instructions.

[0054] Step S402: Obtain the current state of the robot.

[0055] Step S403: Obtain environmental perception information.

[0056] Step S404: Determine whether control strategy selection trigger conditions have occurred, such as task start, task phase change, significant change in environmental conditions, failure of the current control strategy, or execution feedback exceeding the applicable boundary of the current dynamic capability card.

[0057] Step S405: If no trigger condition for reselecting the control strategy occurs, maintain the current control strategy and enter the next real-time control cycle of the control strategy based on the latest robot state; if a trigger condition occurs, convert the task, robot state, and environmental information into control strategy query conditions.

[0058] Step S406: Read the input requirements, applicable conditions, control output form, security constraints, and policy initialization or switching conditions of each control policy from the dynamic capability card registry.

[0059] Step S407: The hard condition screening unit in the control strategy selection module performs boundary hard condition matching on each dynamic capability card. Control strategies that do not meet the task type, state input, environmental conditions, boundary values, or security constraints are directly excluded. The semantic matching unit in the control strategy selection module determines the semantic applicability of the remaining control strategies based on the boundary matching results, the degree of matching between the task and capability description, and the consistency between the task and the reason for the strategy selection. Allowable changes in non-boundary factors are not used as exclusion conditions.

[0060] Step S408: The decision output unit in the control strategy selection module will form a candidate set of control strategies that simultaneously meet the boundary hard conditions and the preset applicability threshold.

[0061] Step S409: When the candidate set is empty, the decision output unit outputs a state of no applicable control strategy (uncovered state); when there are two or more candidates and the difference between the highest and second highest applicability is less than or equal to a preset uncertainty threshold, it outputs a state of uncertain control strategy (ambiguous state); otherwise, it determines the target control strategy with the highest applicability. When the state of no applicable control strategy or uncertain control strategy is output, no candidate strategy is invoked, and the process of supplementary sensing, replanning, deceleration, or safe stop is entered; when the target control strategy is determined, steps S410 to S418 are executed.

[0062] Step S410: When the current control strategy differs from the target control strategy, check whether the robot's current state meets the initial state requirements of the target control strategy.

[0063] Step S411: If the initial state requirements are met, initialize the target control strategy with the robot's current state, and complete the strategy switching through transition trajectory, smooth control quantity processing, or intermediate stable state; if not met, first execute deceleration, attitude adjustment, motion stop, or return to a safe state.

[0064] Step S412: Input the robot's real-time status, task objectives, and environmental parameters into the target control strategy.

[0065] Step S413: Generate one or more of the following from the target control strategy: task execution plan, reference trajectory, target pose, foot trajectory, end effector trajectory, joint target, or low-level control command.

[0066] Step S414: When the target control strategy outputs a task execution plan, the trajectory generator or lower-level control strategy converts the task execution plan into a reference trajectory or control target; the robot controller calculates joint position, joint velocity, or joint torque commands based on the reference trajectory or control target and the robot's real-time state. When the target control strategy directly outputs lower-level control commands, control quantity conversion is no longer performed, but subsequent safety constraint checks are still performed.

[0067] Step S415: Perform joint limit, speed limit, torque limit, contact constraint or collision constraint checks on the underlying control commands.

[0068] Step S416: Send the low-level control command that has passed the check to the robot driver or actuator control interface, so that the driver can drive the motor, hydraulic actuator, gripper or other robot actuator.

[0069] Step S417: Obtain the robot status, contact status, and task progress feedback after execution.

[0070] Step S418: Based on the execution feedback, continue the next control cycle of the current strategy, trigger a new control strategy selection, or determine whether the task is completed.

[0071] In one implementation, a large language model is used for task semantic parsing, dynamic capability card candidate modification, and control strategy semantic applicability judgment. The input of the large language model includes natural language task, robot state summary, environmental information, current dynamic capability card, and description of the newly added control strategy; the output includes structured task conditions, candidate applicability boundaries, candidate capability card modification schemes, and control strategy applicability judgment.

[0072] It should be noted that the large language model does not directly output joint position, joint velocity, or joint torque commands to the robot actuator, nor does it need to be executed in every low-level real-time control cycle. The final task plan, reference trajectory, control target, or low-level control command is generated by the selected robot control strategy; the task plan is converted into a reference trajectory or control target by the trajectory generator or lower-level control strategy, and the robot controller then calculates the low-level control command based on the real-time state. All low-level control commands are sent to the robot driver or actuator control interface after safety constraint checks.

[0073] Compared to existing technologies, this invention configures dynamic capability cards for each robot control strategy, explicitly recording its applicable conditions, input states, control output forms, and safety constraints using structured fields. This dynamic capability card registry serves as a unified benchmark for strategy selection and maintenance. Based on this, the control strategy selection module performs boundary hard condition matching and semantic applicability judgment with each dynamic capability card according to task instructions, robot state information, and environmental perception information. This accurately identifies the applicable strategy for the current working condition, effectively avoiding problems such as foot placement deviation, robot instability, or joint command over-limits caused by incorrect strategy selection. Simultaneously, by recording safety constraints in the capability cards and performing joint limit, speed limit, torque limit, contact constraint, or collision constraint checks on the underlying control commands in the control output and execution module, it ensures that commands are only sent to the robot driver when safety conditions are met, forming a full-link safety guarantee between strategy selection and underlying execution. When no applicable strategy exists or multiple strategies are difficult to distinguish, the control strategy selection module outputs an "applicable" or "uncertain" state and suppresses strategy invocation, triggering supplementary sensing, replanning, deceleration, or safe stop processes. This avoids the safety hazards caused by the system forcibly outputting control quantities when information is insufficient or the operating conditions exceed the strategy boundaries. These measures together form a safety closed loop from task input, strategy selection, control output to execution feedback, ensuring that the strategy selection result directly serves the robot's real-time control.

[0074] Based on this, the dynamic capability card local update module of this invention enables only relevant capability cards with overlapping applicable conditions to be partially modified when adding a new control strategy, while capability cards with unrelated strategies remain unchanged. Furthermore, the maintenance process only modifies capability card records without rewriting the existing control strategy body, its parameters, rules, and calling interfaces. This avoids the problem of existing strategy selection boundaries drifting due to retraining or updating selectors in traditional solutions, ensuring the stability of the original strategy calling relationships during strategy library expansion. In the strategy switching stage, by checking the initial state requirements of the target control strategy, the target strategy is initialized with the current state only when the conditions are met, and the switching is completed using transition trajectories, smooth control input processing, or intermediate stable states, effectively reducing the risk of system instability caused by sudden changes in control commands. In specific application scenarios of legged robots, the above mechanism ensures the consistency and stability of control strategy selection under different working conditions such as flat ground walking, low obstacle crossing, and stair climbing, and can adjust the strategy in a timely manner based on environmental perception results, ensuring the safe operation of the robot in complex open environments. Therefore, this invention forms a complete control strategy selection and maintenance mechanism from strategy matching, selection decision, safe output, local maintenance to strategy switching, which significantly reduces the risk of boundary drift and erroneous control caused by the expansion of the strategy library, and provides a safe, stable and maintainable strategy management solution for robot systems in the continuously evolving control strategy library environment.

[0075] Example 2 The following uses the selection and execution of control strategies for a legged robot as an example to illustrate the specific application of this invention.

[0076] Taking the legged robot control strategy library as an example, the system already has a flat ground walking control strategy and a low obstacle crossing control strategy, and a new stair climbing control strategy has been added. The flat ground walking strategy is suitable for flat ground; the low obstacle crossing strategy is used for single or discontinuous low obstacles; and the stair climbing strategy is used for multiple consecutive steps.

[0077] When registering a new stair climbing control strategy, the system generates an initial dynamic ability card based on the strategy description, representative tasks, and verification tasks. Overlap detection of applicable conditions revealed that both low obstacle crossing and stair climbing involve changes in vertical height, leg lifts, and adjustments to foot placement. The system only partially updates these two dynamic ability cards: limiting low obstacle crossing to single or discontinuous low obstacles, and limiting stair climbing to multiple consecutive steps; the dynamic ability card for the flat ground walking control strategy remains unchanged and is unrelated to the stair task. After modification, task coverage and boundary conflicts are checked using the original low obstacle task and the newly added stair task.

[0078] When the task is "climb three consecutive steps," the system receives the instruction to climb. The environmental perception module acquires the height, width, number, and continuity of each step, as well as the robot's relative position to the first step; the robot state acquisition module acquires the robot's position and posture, joint angles, joint velocities, foot contact state, and current motion state. The task and state parsing module converts the above information into query conditions.

[0079] After reading the dynamic capability card, the hard condition filtering unit in the control strategy selection module finds that the flat ground walking control strategy does not meet the continuous height change condition, and the low obstacle crossing control strategy does not meet the multiple consecutive steps condition, thus excluding these two strategies. The semantic matching unit matches the remaining strategies; the decision output unit determines the stair climbing control strategy as the target strategy. The system loads the stair climbing control strategy. If the robot is currently executing the flat ground walking control strategy, it first determines whether the robot is in a stable support state, whether the body posture is within the allowable range, and whether the foot contact meets the initialization requirements of the stair climbing strategy.

[0080] When the initialization requirements are met, the system initializes the stair climbing strategy based on the current body posture and foot position, and generates a transition trajectory from a flat-ground gait to the first step climbing action. When the initialization requirements are not met, the robot first decelerates and enters a stable standing state before activating the stair climbing control strategy. During strategy switching, the reference trajectories or joint targets of the old and new strategies are smoothly connected to avoid abrupt changes in joint position, velocity, or torque commands.

[0081] The stair climbing control strategy generates one or more of the following: body reference trajectory, foot swing trajectory, foot landing position, contact sequence, and joint targets. The robot controller further calculates joint position, joint velocity, or joint torque commands based on the robot's real-time state, the aforementioned trajectory, and the joint targets. After the joint limit, velocity, torque, contact, and collision checks are passed, the commands are sent to the leg actuators, which then drive the joint motors and other actuators.

[0082] During execution, the system continuously collects data on body posture, foot contact, and joint status as feedback to update the status input for the next control cycle and determine whether the target platform has been reached. This feedback also determines whether the stair climbing strategy still meets the current operating conditions; when the environment changes significantly or the current state exceeds the applicable boundaries of its dynamic capability card, a new control strategy selection is triggered. When the system detects that the body posture deviation exceeds the allowable range, the expected foot contact fails, joint commands exceed limits, or a collision may occur, the system pauses the current action, reduces the movement speed, or performs a safety stop.

[0083] When the environmental perception result is a single low obstacle rather than a series of steps, the dynamic ability card matching result changes to a low obstacle traversal control strategy. This strategy generates the foot lifting trajectory and landing position required for one traversal, without invoking the stair climbing control strategy. Therefore, the local maintenance results of the dynamic ability card can directly affect the selection of specific control strategies, control trajectory generation, and execution of safety procedures, while maintaining the stability of the applicable range of irrelevant control strategies such as flat-ground walking.

[0084] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A robot control strategy selection and maintenance system based on dynamic capability cards, characterized in that, include: The dynamic capability card registry is used to store the dynamic capability cards corresponding to each control strategy. The dynamic capability cards at least record the applicable conditions and security constraints of the corresponding control strategy. A control strategy library is used to store various robot control strategies; The control strategy application condition overlap detection module is used to compare the application conditions of the new control strategy with those of the existing control strategies when adding a new control strategy, and to identify whether there is any overlap. The Dynamic Capability Card Partial Update Module is used to update only the dynamic capability cards corresponding to newly added control strategies and control strategies whose applicable conditions overlap with them, while other dynamic capability cards remain unchanged. The control strategy selection module is used to determine the target control strategy from the control strategy library based on the matching results of the robot's state information and environmental perception information with each dynamic capability card. When there is no applicable strategy or multiple applicable strategies cannot be uniquely determined, the corresponding inapplicable state or uncertain state is output, and the safety processing procedure is triggered. The control output and execution module is used to call the target control strategy to generate control output, and after performing a safety constraint check on the control output, send it to the robot driver or actuator control interface.

2. The system according to claim 1, characterized in that, The dynamic capability card also records at least one of the following fields: control strategy identifier, control task type, capability description, input state variable, boundary value, control output type, non-boundary factor, strategy initialization or switching conditions, and reason for strategy selection.

3. The system according to claim 1, characterized in that, The overlapping conditions for the control strategy application are as follows: The overlapping conditions that the detection module determines exist are: When the same task simultaneously satisfies the boundary conditions of the newly added dynamic capability card and the existing dynamic capability card, or when there is an intersection in the numerical range, category set, inclusion relationship or logical condition of the semantically equivalent boundary factors that can be satisfied simultaneously, it is determined that there is an overlap; the control strategy application condition overlap detection module is also used to treat the situation where the boundary factors and boundary values ​​within the same dynamic capability card do not correspond, or multiple boundary conditions are contradictory and cannot be satisfied simultaneously as a boundary consistency error.

4. The system according to claim 1, characterized in that, When the dynamic capability card partial update module makes partial modifications, it only modifies the fields required to eliminate applicable condition conflicts or maintain the internal consistency of the capability card, while fields not involved in conflicts remain unchanged. The dynamic capability card partial update module only modifies the dynamic capability card record and does not modify the existing control policy body and its parameters, rules, and calling interfaces. When the modified dynamic capability card fails the integrity or consistency check, the dynamic capability card partial update module does not submit the modification and continues to use the dynamic capability card registry before the update.

5. The system according to claim 1, characterized in that, The control strategy selection module includes: The hard condition screening unit is used to exclude control strategies that do not meet the boundary hard conditions based on the applicable conditions and safety constraints in each dynamic capability card. The semantic matching unit is used to determine the semantic applicability of each control strategy based on the degree of matching between the task instructions and the capability descriptions and strategy selection reasons in the remaining dynamic capability cards. The decision output unit is used to form a candidate set of control strategies that simultaneously meet the boundary hard conditions and the preset applicability threshold. When the candidate set is empty, the output state is "no applicable strategy". When there are two or more strategies in the candidate set and the difference between the highest applicability and the second highest applicability is less than or equal to the preset uncertainty threshold, the output state is "strategy uncertainty". Otherwise, the strategy with the highest applicability is output as the target control strategy.

6. The system according to claim 1, characterized in that, When the current control strategy differs from the target control strategy, the control output and execution module is also used to check whether the robot's current state meets the initial state requirements of the target control strategy. If it does, the target control strategy is initialized with the robot's current state, and the strategy switching is completed through transition trajectory, smoothing of control quantity, or intermediate stable state. If it does not meet the requirements, deceleration, attitude adjustment, motion stop, or return to a safe state is executed first, and then the strategy switching is executed. In the control output and execution module, when the target control strategy outputs a task execution plan, the trajectory generator or lower-level control strategy converts the task execution plan into a reference trajectory or control target. The robot controller calculates the underlying control instructions based on the reference trajectory or control target and the robot's real-time state. When the target control strategy directly outputs the underlying control instructions, the robot controller no longer performs control quantity conversion. The underlying control instructions are sent to the robot driver or actuator control interface after safety constraint checks.

7. The system according to claim 1, characterized in that, It also includes a task input module, a robot status acquisition module, and an environment perception module; the task input module is used to receive robot task instructions, the robot status acquisition module is used to acquire the robot's current status information, and the environment perception module is used to acquire environmental perception information.

8. The system according to claim 1, characterized in that, It also includes a task and state parsing module, which is used to convert task instructions, robot state information and environmental perception information into structured query conditions required for control strategy selection, and to generate the structured information required for initial dynamic capability cards based on the description of the new control strategy, representative tasks and verification tasks when a new control strategy is registered.

9. The system according to claim 1, characterized in that, It also includes an execution feedback module, which is used to obtain the robot's status and task progress feedback after execution and provide the feedback to the next control cycle; when the feedback information exceeds the applicable boundary of the current dynamic capability card, the execution feedback module triggers the control strategy selection module to reselect the control strategy.

10. The system according to claim 1, characterized in that, The safety processing procedure includes at least one of supplemental sensing, replanning, deceleration, or safe stopping.

Images