An agent-driven self-evolution robot navigation method and device

Abstract

The application discloses an agent-driven self-evolution robot navigation method and device. A deterministic running base delivers a navigation target, task progress, a robot and environment state, capability description, skill retrieval result, capability calling permission, resource budget, safety constraint and feedback to an agent. The agent generates a task semantic capability calling decision round by round; the base makes a fact and compliance determination on the calling and does not replace the selection when the calling is rejected. When the budget, emergency stop or safety event occurs, the base directly terminates or stops and informs the agent. The agent initiates skill creation, modification, merging, degradation or deactivation; and a statistical module delivers a skill effect evaluation. Physical motion requests are reviewed by a local safety execution module.

Description

Technical Field

[0001] This invention belongs to the fields of robot navigation, robot agent control, and robot capability evolution technology, and particularly relates to an agent-driven self-evolving robot navigation method and device. Background Technology

[0002] Existing robot navigation technologies include at least three technical approaches.

[0003] The first type is the workflow-driven navigation system. This type of system uses deterministic code to predefine task stages such as localization, mapping, path planning, interaction, and motion control, as well as the necessary processing branches in each state. The model can generate local content at reserved nodes, but the deterministic workflow still defines the task processing procedure. Without the corresponding branches configured, the system struggles to form new processing sequences.

[0004] The second category is large-model-driven navigation systems. In these systems, agents can progressively select tools and complete the current task based on observations. However, they typically lack agent-initiated capability evolutionary meta-capabilities and do not formally distinguish between the agent's task decision-making power and the deterministic base's safety veto power. Furthermore, for calls rejected by the base, these systems do not specify a runtime relationship where a preset process does not replace the agent's selection after rejection.

[0005] The third category is large-scale model skill library systems. There are schemes that extract operational experience into skills and inject them into subsequent tasks, but their evolution is executed by a fixed pipeline at the end of the round. The triggering and selection of evolution are not decided by the agent, and the separation of task decision-making power and safety veto power is not formalized.

[0006] Therefore, existing technologies cannot simultaneously achieve the following in physical robot navigation: the intelligent agent forming a task handling process round by round; the intelligent agent initiating capability evolution; and the local deterministic layer independently retaining the right to veto physical motion. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of existing technologies by proposing an agent-driven, self-evolving robot navigation method and apparatus. Unlike systems where task stages are controlled by preset processes and the agent generates local content at each stage, this invention generates the navigation process itself through the agent. The deterministic runtime platform does not generate task handling procedures; it is only responsible for capability exposure, invocation constraints, fact recording, compliance judgment, feedback injection, and physical security enforcement.

[0008] This invention employs three types of decentralization: Task semantic decisions, such as what to do, which callable capabilities and their parameters to select, whether to abandon non-safety-critical sub-goals, whether skills are worth merging, whether to initiate capability evolution, and whether goals are achieved, are made by the intelligent agent; Factual and compliance judgments, such as whether calls conflict, whether interfaces are valid, whether capability call permissions are compliant, and whether static checks have been passed, are made by the corresponding modules in the deterministic runtime base and the results are delivered to the intelligent agent; Safety enforcement, such as resource budget exhaustion, ontology health status timeout, manual emergency stop, safety constraint triggering, capability call permission revocation, and dangerous skill isolation, is directly executed by the deterministic runtime base or local safety execution module, which can preempt capability calls and deliver the termination or isolation facts and reasons after execution.

[0009] The objective of this invention is achieved through the following technical solution: a self-evolving robot navigation method driven by an intelligent agent, which is based on an intelligent agent and a deterministic operating platform, and specifically includes the following steps:

[0010] (1) The deterministic runtime base delivers runtime decision context to the agent;

[0011] (2) The agent selects callable capabilities round by round according to the runtime decision context, and generates and initiates capability call decisions;

[0012] (3) The deterministic operating platform makes compliance determination on the capability call permission, resource budget, parameter mode and security constraints of the capability call decision, executes the capability call that passes the compliance determination, records and delivers execution feedback to the agent; if the compliance determination fails, it rejects the execution and delivers the termination facts and reasons to the agent.

[0013] (4) The deterministic operating base records the execution feedback, current robot state changes and environmental state changes associated with the capability call, and delivers the record to the agent; the agent updates the current task problem state according to the record, and redetermines the next round of callable capabilities and their call parameters according to the updated current task problem state, thereby forming a capability call sequence that is not fixed before the task is executed;

[0014] (5) The intelligent agent decides whether to initiate a capability evolution operation based on the skill effect evaluation information, skill content and historical handling trajectory, and makes a compliance judgment based on the deterministic operation base.

[0015] When a forced termination event occurs, the call constraint and execution module in the deterministic operation base terminates subsequent capability calls; when there is ongoing physical motion, the local safety execution module executes an explicit stop instruction; after termination or stop, the deterministic operation base delivers the corresponding facts and reasons to the agent.

[0016] Further, in step (1), the runtime decision context includes the overall navigation target, the current sub-target or task progress, the current robot state, the environment state, the machine-readable description of the callable capabilities, the skill retrieval results, the capability call permissions and resource budget, the security constraints, and the feedback from the previous round of capability call.

[0017] Further, in step (2), the capability invocation decision includes the selected callable capability, invocation parameters, expected state change, whether to adjust the current sub-target, whether to abandon non-safety critical sub-target, whether the navigation target is completed, and the determination of the continuation condition or task semantic termination.

[0018] Furthermore, the capability invocation sequence is formed round by round during task execution, rather than being selected from a preset task workflow or anomaly branch table.

[0019] Furthermore, when compliance is not approved, the deterministic operating platform simply refuses to execute, records the reason, and returns the reason to the agent. When resource budget is exhausted, the ontology's health status times out, a manual emergency stop occurs, security constraints are triggered, or capability call permissions are revoked, the deterministic operating platform directly terminates subsequent capability calls. If physical motion is in progress, the local security execution module directly executes an explicit stop command. When a forced termination event occurs, the call constraint and execution module in the deterministic operating platform terminates subsequent capability calls; if physical motion is in progress, the local security execution module executes an explicit stop command; after termination or stoppage, the deterministic operating platform delivers the corresponding facts and reasons to the agent. The deterministic operating platform does not select another capability for the agent, nor does it generate alternative handling procedures corresponding to the task category.

[0020] Further, in step (4), when the recorded expected robot state change does not occur, the existing skills do not meet the success conditions, or the failure or termination conditions have been met, the deterministic operating platform delivers the corresponding task problem state and execution feedback to the agent; the agent adds the task problem state and execution feedback to the runtime decision context, and uses the same decision loop of the round-by-round generation of capability call decisions to regenerate subsequent capability call decisions that are different from the previous capability call decisions based on the updated runtime decision context.

[0021] Furthermore, in step (5), the deterministic operating platform records the retrieval, adoption, actual execution, problem solving, task contribution, average cost, security rejection, mismatch, continuous failure and recent use information of the skill, and delivers skill effect evaluation information to the agent.

[0022] Furthermore, the local safety execution module in the deterministic operating base reviews all physical motion requests; the local safety execution module only executes physical motion requests that pass the review, or rejects physical motion requests that fail the review and records the reason for rejection; when a safety constraint is triggered, the local safety execution module directly executes an explicit stop instruction, and delivers the termination facts and reasons to the agent after execution.

[0023] Furthermore, the agent determines whether to initiate capability evolution operations to create, modify, merge, downgrade, or deactivate a skill based on the novelty of the problem, historical repetition, handling trajectory, and skill effect evaluation. Problem novelty and historical repetition are primarily used to determine whether capability evolution should be triggered; skill content, skill effect evaluation information, and historical handling trajectory are used to determine what evolution operation to take for a specific skill. When the agent initiates a capability evolution operation to create a skill, it compares the changes in the task problem state before and after each capability call in the handling trajectory with the problem resolution conditions, retains the capability calls that change the task problem state to the problem resolution conditions, converts scenario-specific constants into parameter constraints, extracts applicable conditions, success conditions, and failure or termination conditions, and generates candidate skills containing skill identifiers, problem characteristics, applicable conditions, handling content, success conditions, failure or termination conditions, and usage records.

[0024] The static inspection and version registration module assesses candidate skills based on their interface, capability call permissions, call constraints, and security compliance. Candidate skills that pass the assessment are registered and set to a searchable state, participating in skill retrieval in subsequent decision-making cycles or navigation tasks according to applicable conditions. The retrieval results are added to the runtime decision context. When modifying a skill, the agent decides whether to retain, downgrade, or disable the old version, and the static inspection and version registration module records the corresponding version status. If a assessment fails, the non-compliance facts and reasons are returned to the agent, which then decides whether to propose a revision.

[0025] Further, the deterministic operating platform determines the skill retrieval result according to preset matching conditions, which include problem matching conditions, ontology matching conditions, and environment matching conditions. The problem matching condition is that the problem category is consistent or the problem feature similarity reaches a threshold. The ontology matching condition is that the robot type, capability call permission, and necessary robot state meet the skill applicability conditions. The environment matching condition is that the environment state falls within the skill applicability range and there is no environmental constraint conflict. The deterministic operating platform delivers the skill retrieval result to the agent. When the skill retrieval result is empty, the existing skill does not meet the success condition, or the failure or termination condition has been met, the agent uses the decision loop in step (4) and can change the current sub-goal, adjust the call parameters, select other callable capabilities, combine multiple callable capabilities, or call human-computer interaction capabilities.

[0026] Furthermore, the skills statistics module records information on skill retrieval, adoption, actual execution, problem-solving, task contribution, average cost, security rejection, mismatch, consecutive failures, and recent use, and delivers skill effectiveness evaluation information. Based on this skill effectiveness evaluation information, skill content, and historical processing trajectory, the agent generates capability evolution decisions to modify, merge, downgrade, disable, or retain skills. Low success rates, mismatches, or low adoption rates are only used as evaluation information for the agent's adjustment choices. When the static check and version registration module determines that a skill contains unauthorized calls, instructions that bypass the local security execution module, or dangerous calls, it directly sets the skill to a frozen and unretrieval state; the agent can only submit revision or restoration requests, and the freeze is only lifted after the module's compliance determination is passed. Retrieval Adoption Rate and execution success rate The threshold rules are optional implementation methods and do not constitute a fixed task processing flow.

[0027] Secondly, the present invention also provides an agent-driven self-evolving robot navigation device, including a memory, a processor, and a communication interface. The memory stores navigation tasks, robot states, environmental states, machine-readable descriptions of callable capabilities, a skill library, and usage records. The processor executes programs to form an agent decision-making module and a deterministic operating base.

[0028] The agent decision-making module is used to select callable capabilities based on the runtime decision context, generate capability call decisions, make task semantic termination judgments, and initiate skill capability evolution operations.

[0029] The deterministic operating base includes:

[0030] The call constraint and execution module is used to make compliance judgments on capability call decisions, execute capability calls that pass the compliance judgment, record and deliver execution feedback, and directly terminate subsequent capability calls when a termination event occurs;

[0031] The problem status recording module is used to record execution feedback, robot status changes, and environmental status changes;

[0032] The static inspection and version registration module is used to make compliance judgments on candidate skills, register versions, and set the searchable or frozen status of skills.

[0033] The skills usage statistics module is used to record usage information and deliver skills effectiveness evaluation information;

[0034] The local security execution module is used to review, execute, or reject physical motion requests, and to directly execute stop commands when a termination event occurs;

[0035] The communication interface is used to exchange observation and physical motion requests with the mobile robot body.

[0036] Thirdly, the present invention also provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the above-described agent-driven self-evolving robot navigation method.

[0037] Fourthly, the present invention also provides a computer program product, including a computer program or instructions, which, when executed by a processor, implement the above-mentioned agent-driven self-evolving robot navigation method.

[0038] The beneficial effects of this invention are as follows: without pre-setting corresponding processing branches, the agent can continue to form a capability call sequence for navigation tasks; the same capability set can form a cross-module call sequence that was not previously configured; skill accumulation can reduce repeated processing when similar problems recur; skill effect evaluation information can prevent inefficient or unsafe skills from occupying search results for a long time; all entity movements still satisfy local safety constraints. The increased autonomy does not hand over physical safety control to an uncertain model: the agent has the decision-making power to do what and try what, while the local deterministic layer retains the veto power over whether it can actually be executed. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the overall decentralized architecture of the present invention.

[0040] Figure 2 This is a closed-loop flowchart of disposal, solidification, reuse, and updating under the same decision-making cycle of the present invention;

[0041] Figure 3 This is a flowchart of the skill generation, compliance determination, registration, retrieval, usage evaluation, update, freezing, and recovery process of the present invention.

[0042] Among them, 100. Intelligent agent decision-making module, 200. Deterministic operation base, 210. Call constraint and execution module, 300. Problem status recording module, 400. Static inspection and version registration module, 500. Skill usage statistics module, 600. Local security execution module, and 700. Entity mobile robot body. Detailed Implementation

[0043] The technical details and principles of the present invention will be further described below with reference to the accompanying drawings.

[0044] like Figure 1 As shown, the robot navigation device in this embodiment includes an intelligent agent decision-making module 100, a deterministic operating base 200, and a physical mobile robot body 700.

[0045] The deterministic operating platform 200 includes a call constraint and enforcement module 210, a problem status recording module 300, a static inspection and version registration module 400, a skill usage statistics module 500, and a local security execution module 600. The agent decision-making module 100 generates the navigation process, makes a task semantic termination judgment, and initiates skill capability evolution operations. The call constraint and enforcement module 210 makes factual and compliance judgments on capability calls, executes calls that pass the judgment, and delivers feedback. The problem status recording module 300 records execution feedback and status changes. The static inspection and version registration module 400 makes static compliance judgments and registers versions or sets skill states. The skill usage statistics module 500 records usage information and delivers effect evaluations. The local security execution module 600 reviews, executes, or rejects physical movement requests and executes preemptible security enforcement.

[0046] like Figure 1 and Figure 2 As shown, in a wheel-by-wheel navigation embodiment, the deterministic runtime base 200 delivers to the agent decision module 100 the overall navigation goal, the current wheel sub-goal or task progress, the current wheel robot state, the environmental state, a machine-readable description of the callable capabilities, skill retrieval results, capability call permissions, resource budget, safety constraints, and feedback from the previous round's capability call. The agent decision module 100 selects capabilities and generates capability call records based on the runtime decision context delivered by the deterministic runtime base 200, and then initiates capability call decisions. Each capability call record includes a capability identifier, parameters, expected state changes, whether to adjust the current sub-goal, and a continuation condition or task semantic termination judgment. The call constraint and execution module 210 makes deterministic judgments on whether the capability exists, whether the parameters conform to the parameter pattern, whether the capability call permissions are compliant, and whether the resource budget and safety constraints are met; if the judgment passes, the capability call decision is executed; if the judgment fails, the call is rejected. Execution feedback or rejection reasons are delivered to the agent decision module 100.

[0047] After each capability invocation, the problem status recording module 300 records the execution feedback, current robot state changes, and environmental state changes. The deterministic operation base 200 delivers these records to the agent decision-making module 100. Based on this, the agent decision-making module 100 updates the current task problem status and re-determines the next available capability and its invocation parameters. The capability invocation decision sequence is thus formed round by round during execution. The deterministic operation base 200 does not select another available capability based on the task category, nor does it replace rejected capability invocation decisions with a preset handling procedure.

[0048] like Figure 2As shown, when the problem state recording module 300 records at least one of the following facts: the expected state change has not occurred; the adopted and executed skill meets the failure or termination condition during execution; or the success condition is still not met after the skill execution is completed; the deterministic runtime base 200 delivers the problem state and feedback to the agent decision module 100. The agent decision module 100 reintegrates the problem state and feedback information into the runtime decision context and continues to use the same round-by-round decision loop to regenerate subsequent capability invocation decisions that are different from the previous capability invocation decisions. In an optional embodiment, the subsequent capability invocation decision may change the current sub-goal, adjust the invocation parameters, select other callable capabilities, combine multiple callable capabilities, or invoke human-computer interaction capabilities. This process is not taken over by the exception handling module, nor is the agent's next choice restricted by an exception branch table.

[0049] In one embodiment of positioning drift handling, the deterministic operating base 200 delivers the actual pose, map matching results, and execution feedback, while the problem state recording module 300 records the pose residual. The agent decision module 100 updates the positioning drift task problem state based on the stated facts and navigation target, and sequentially selects re-observation, map matching, and re-positioning capabilities. The agent decision module 100 regenerates the next capability call decision based on the pose residual after each call; after the pose residual drops to a preset range and does not exceed this range again for L consecutive decision cycles, the agent decision module 100 determines whether the positioning drift problem is resolved and decides whether to initiate a capability evolution operation to create a positioning recovery skill, where L is a pre-registered positive integer.

[0050] In one embodiment of a temporary channel closure, the deterministic operation base 200 delivers channel status, alternative route accessibility information, and capability invocation permissions. The problem status recording module 300 records the continuous passable distance of the alternative route and the duration of channel openness. The agent decision module 100 updates the current task problem status based on the information and selects environmental query, wait, detour, and remote interaction capabilities within the same loop. The agent decision module 100 generates the next invocation decision based on subsequent feedback until a passable alternative route or channel openness record is obtained. After the fact that the continuous passable distance of the alternative route reaches a preset length or the channel openness lasts for a preset time is delivered, the agent decision module 100 determines whether the local problem has been resolved and decides whether to initiate a capability evolution operation to create a temporary channel closure handling skill.

[0051] In a cross-floor handling embodiment, the deterministic operating base 200 grants access to the elevator status, target floor, button location, and human-machine interaction capabilities. The intelligent agent decision module 100 updates the button unreachability status based on records indicating the target button is outside the mechanical operating range, and selects button location, reachability judgment, human-machine interaction interface, response detection, and elevator status detection capabilities. After a person or system with access to the capabilities completes the target floor selection, the intelligent agent decision module 100 generates subsequent invocation decisions based on the elevator status; after the target floor selection is completed and the elevator is moving towards the target floor, the intelligent agent decision module 100 determines whether the button unreachability problem is resolved. This embodiment only illustrates that the human-machine interaction interface can be used as an invoked capability and does not limit the applicable scenarios of the invention.

[0052] like Figure 3 As shown, in one skill generation embodiment, the agent decision module 100 initiates a capability evolution operation to create skills, and generates candidate skills according to the following steps:

[0053] 1. Identify the problem to be solved and write down the problem category, associated objects, and status differences;

[0054] 2. Extract the observation state, robot state, and environmental state before and after the problem is resolved;

[0055] 3. Compare the changes in the problem state before and after each capability call with the problem resolution conditions, and retain the calls that contributed to the change in the problem state towards the resolution conditions;

[0056] 4. Delete failed attempts that did not contribute, or add them to the negative notes;

[0057] 5. Rewrite the scene-specific constants as parameter constraints;

[0058] 6. Extract applicable conditions from robot status, environment status, capability access permissions, and capability access prerequisites;

[0059] 7. Extract the conditions for resolving the problem and the conditions for failure or termination;

[0060] 8. Generate marked text format, object representation, or configuration data format skills that include problem characteristics, applicable conditions, handling content, success conditions, failure or termination conditions, and usage records according to a fixed field structure;

[0061] 9. The candidate skills are delivered to the static inspection and version registration module 400; the module 400 makes a judgment on the format, interface reference, capability call permission, call constraint and security compliance, registers the version of the candidate skills that pass the judgment and sets them to a searchable state, and delivers the non-compliance facts and reasons or registration results to the intelligent agent decision module 100.

[0062] An example of a configuration data format for the fixed field structure is as follows:

[0063] skill_id: corridor-closure-v1

[0064] problem_feature: The passage is temporarily closed.

[0065] applicability:

[0066] robot_type: mobile_robot

[0067] capability_call_permissions: [environment_query, detour, remote_interaction]

[0068] environment_constraint: Alternative routes are reachable or remote interaction can be requested.

[0069] procedure:

[0070] - capability: environment_query

[0071] parameter_constraint: {target: blocked_corridor}

[0072] - capability: detour

[0073] parameter_constraint: {minimum_clear_distance: D}

[0074] success_condition: Alternative routes are continuously passable or passages are open.

[0075] termination_condition: The agent determines that the current sub-objective has no continuing value or has been completed.

[0076] usage_record: {version: v1, searchable: true}

[0077] In one embodiment of language model-generated skills, the agent decision module 100 provides the language model with a fixed field structure, observation states before and after problem resolution, robot state, environmental state, contributing calls, negative attention items, problem resolution conditions, callable capability descriptions, parameter patterns, and capability call permissions. The agent decision module 100 initiates candidate skill generation; the language model outputs candidate skills corresponding to the fixed field structure. The static checking and version registration module 400 makes compliance determinations regarding field completeness, field type, interface references, parameter patterns, capability call permission scope, termination conditions, call limits, deadlines, and requirements for not bypassing the local security execution module 600. If the structure is inconsistent, fields are missing, references to non-existent capabilities are made, or physical motion instructions that bypass the local security execution module 600 are included, the static checking and version registration module 400 refuses registration and delivers the non-compliance facts and reasons to the agent decision module 100. Candidate skills that pass the compliance determination are version registered and set to a searchable state. The agent decision module 100 decides whether to propose a revision of the candidate skills.

[0078] like Figure 3 As shown, in one embodiment of skill usage statistics, the skill usage statistics module 500 records the number of times each skill is retrieved within a statistics window consisting of the most recent M skill retrievals. Number of times adopted Actual number of executions Number of times successfully resolved The data includes task contribution, average cost, number of security rejections, number of false matches, number of consecutive failures, and recent usage time, where M is a positive integer. The skill usage statistics module 500 delivers the retrieval adoption rate to the agent decision-making module 100. and execution success rate ,in ( When it is 0, take 0). ( (When the value is 0, take 0), and the skill effect evaluation information corresponding to the indicator.

[0079] when Less than the minimum search sample size or Less than the minimum execution sample size At this time, the skill usage statistics module 500 only records usage information and delivers observation period information; the observation period refers to the stage where the current skill's retrieval sample size or actual execution sample size has not yet reached the minimum sample size required for statistical evaluation, and only continues to accumulate the skill's usage information without triggering capability evolution operations based on the usage information; the agent decision module 100 does not initiate disposal operations based on this; when the actual execution sample size is sufficient and the execution success rate is low, the agent decision module 100 can decide whether to initiate a capability evolution operation to rewrite the skill; when the mismatch rate is high, the agent decision module 100 can decide whether to initiate a capability evolution operation to tighten the applicable conditions or adjust the retrieval index; after the statistical sample size reaches the requirement, based on the evaluation information that the skill is frequently retrieved but rarely used, the agent decision module 100 can decide whether to initiate a capability evolution operation to reduce the retrieval priority. The retrieval priority is a sorting attribute stored in association with the skill and its version, used to determine the return order of candidate skills that are in a searchable state and meet the matching conditions of question, ontology, and environment. In one optional implementation, priority is represented by a non-negative integer; the larger the value, the higher the ranking. When priorities are the same, they are ranked from highest to lowest based on the similarity of problem features. In the above implementation, after the adjustment request is verified by the static check and version registration module, the corresponding priority is reduced by 1, with a minimum of 0, and takes effect in subsequent searches. Each search returns at most the top K skills, where K is a preset positive integer. Lowering the priority may cause the skill to move to a lower ranking or no longer appear in the top K results, but it does not change its searchable status, applicable conditions, or calling permissions, and is not equivalent to disabling or freezing. When the static check and version registration module 400 determines that a skill contains instructions that exceed authority, bypass the local security execution module 600, or are dangerous, it directly sets the skill to a frozen and unsearchable state, records it, and delivers the freezing facts and reasons to the agent decision module 100. The agent decision module 100 can submit a revision or restoration request; only after the static check and version registration module 400 passes the compliance determination will the static check and version registration module 400 unfreeze the skill and record the corresponding version status. Skills whose recent usage time exceeds the preset duration will be evaluated separately based on their recent usage time and will not be treated as execution failures.

[0080] In one implementation of skill merging, the fact that the problem characteristics and applicable conditions of the skills to be merged meet preset similarity conditions is delivered by the deterministic runtime base 200. The agent decision module 100 determines whether the skills to be merged are worth merging and generates merge candidates. The static inspection and version registration module 400 makes compliance determinations on the call conflicts, interface references, capability call permissions, and security constraints of the merge candidates, delivers the determination results to the agent decision module 100, and registers the versions and sets the searchable status of the merge candidates that pass the compliance determination.

[0081] like Figure 1As shown, in one embodiment of tool semantic adaptation, the invocation constraint and execution module 210 in the deterministic operation base 200 records the invocation intent, raw return data, and robot state, and delivers execution feedback. The execution feedback includes at least the execution state, cause category, and actual execution effect; for example, the execution feedback for a forward invocation can record the expected forward distance, actual pose change, execution state, and obstacle cause. This execution feedback is used by the agent decision module 100 to update the current task problem state and generate the next capability invocation decision.

[0082] In one tool adapter embodiment, when the basic tool fails to return the required execution feedback according to the capability invocation decision, the agent decision module 100 initiates the creation of a tool adapter based on the interface description, allowed parameter range, and sample invocation records. The static inspection and version registration module 400 makes a compliance determination on the tool adapter's format, import, interface testing, and compliance with the requirement of not bypassing the local secure execution module 600; during testing, no physical motion requests are made to the physical mobile robot body 700. The tool adapter that passes the compliance determination is registered by module 400 to the callable capability set, and the deterministic runtime base 200 delivers the corresponding new machine-readable capability description to the agent decision module 100.

[0083] like Figure 1 As shown, in a local security execution embodiment, the local security execution module 600 verifies the capability invocation permissions, ontology health status, motion type, speed limit, spatial range, and deadline corresponding to the physical motion request; the specific verification rules are as follows: (1) Capability Invocation Permission Determination: Check whether the capability requested belongs to the capability set granted to the current task, and whether the request exceeds the authorized scope of the capability. This determination passes when the capability requested is within the authorized set, the permission is valid and has not been revoked, and the request does not exceed the authorized scope.

[0084] (2) Body health status determination: Read the health status and update time reported locally by the robot, check whether there is a fault mark prohibiting movement, and whether the difference between the current time and the most recent health status update time exceeds the preset timeout. If there is no fault prohibiting movement and the health status has not exceeded the timeout, this determination is passed.

[0085] (3) Motion type determination: Check whether the motion type in the request belongs to the motion type allowed by the safety constraint configuration, and whether it can be supported by the current robot body and its operating state. If the motion type is allowed and the current robot can execute it, this determination is passed.

[0086] (4) Speed ​​Limit Determination: Read the linear velocity and angular velocity in the request and compare them with the speed limit of the current robot and the corresponding motion type. If the speed parameters in the request are valid and the magnitude of the linear velocity and the absolute value of the angular velocity do not exceed the corresponding limit, the determination passes. Requests that exceed the limit will not be executed, and the agent will decide whether to adjust the parameters and re-initiate the request based on the reason for rejection.

[0087] (5) Spatial Range Determination: Based on the robot's current pose, shape envelope, and motion trajectory corresponding to this request, determine the spatial range occupied by the robot when executing the request. The motion trajectory is provided by the corresponding motion control interface, or determined based on the motion parameters in the request and the pre-registered robot kinematic model. This determination passes when the space occupied by the robot on the trajectory is within the allowed movement area and does not intersect with the prohibited entry area; for requests involving continuous motion processes, not only the target position is checked, but also the space occupied during the motion process is checked.

[0088] (6) Deadline determination: Compare the current time and the expected execution end time of this request with the request deadline. If the request has not expired and the expected execution end time is not later than the deadline, this determination is successful.

[0089] If the review passes, the local security execution module 600 executes the physical motion request; if the review fails, the local security execution module 600 rejects the request, records the reason for rejection, and delivers the reason for rejection. When resource budget is exhausted, the ontology health status times out, manual emergency stop occurs, security constraints are triggered, or capability call permissions are revoked, the local security execution module 600 can preempt the currently executing physical motion and directly execute a clear stop command, delivering the security enforcement facts and reasons after stopping. Thus, the agent decision-making module 100 has the decision-making power over what to do and what to try, while the local security execution module 600 retains the veto power over whether it can actually be executed.

[0090] like Figure 1 As shown, in one termination event embodiment, when the navigation target is completed, the path has no continuing value, or a non-safety-critical sub-target is abandoned, the agent decision module 100 generates a task semantic termination judgment. When resource budget is exhausted, the ontology health status times out, manual emergency stop occurs, safety constraints are triggered, or capability call permissions are revoked, the call constraint and execution module 210 directly terminates subsequent capability calls; if the physical mobile robot body 700 is performing motion, the local safety execution module 600 preempts and executes an explicit stop command. The deterministic running base 200 records and delivers the forced termination fact, reason, and final task status after termination or stop.

[0091] like Figure 1As shown, in one embodiment, the memory stores navigation tasks, observation information, robot state, environmental state, a set of callable capabilities, capability call records, execution feedback, a skill library, usage records, and termination events. A communication interface exchanges observation and physical motion requests with the physical mobile robot body 700. The processor executes programs to form an agent decision-making module 100 and a deterministic operating base 200; the deterministic operating base 200 includes a call constraint and execution module 210, a problem state recording module 300, a static check and version registration module 400, a skill usage statistics module 500, and a local safety execution module 600, to enable the agent to generate the navigation process and initiate skill capability evolution, the deterministic module to make factual and compliance judgments, and the local safety execution module 600 to independently execute safety enforcement.

[0092] The above embodiments are used to explain and illustrate the present invention, but not to limit the present invention. Any modifications and changes made to the present invention within the spirit and scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A self-evolving robot navigation method driven by an intelligent agent, characterized in that, This method is based on intelligent agents and a deterministic operating framework, and specifically includes the following steps: (1) The deterministic runtime base delivers runtime decision context to the agent; (2) The agent selects callable capabilities round by round according to the runtime decision context, and generates and initiates capability call decisions; (3) The deterministic operating platform makes compliance determination on the capability call permission, resource budget, parameter mode and security constraints of the capability call decision, executes the capability call that passes the compliance determination, records and delivers execution feedback to the agent; if the compliance determination fails, it rejects the execution and delivers the termination facts and reasons to the agent. (4) The deterministic operating base records the execution feedback, current robot state changes and environmental state changes associated with the capability call, and delivers the record to the agent; the agent updates the current task problem state according to the record, and redetermines the next round of callable capabilities and their call parameters according to the updated current task problem state, thereby forming a capability call sequence that is not fixed before the task is executed; (5) The intelligent agent decides whether to initiate capability evolution operation based on skill effect evaluation information, skill content and historical handling trajectory, and makes compliance judgment based on deterministic operation base.

2. The agent-driven self-evolving robot navigation method according to claim 1, characterized in that, In step (1), the runtime decision context includes the overall navigation target, the current sub-target or task progress, the current robot state, the environment state, the machine-readable description of the callable capabilities, the skill retrieval results, the capability call permissions and resource budget, the security constraints, and the feedback from the previous round of capability call.

3. The agent-driven self-evolving robot navigation method according to claim 1, characterized in that, In step (2), the capability invocation decision includes the selected callable capability, invocation parameters, expected state changes, whether to adjust the current sub-target, whether to abandon non-safety critical sub-targets, whether the navigation target is completed, and the determination of the continuation condition or task semantic termination.

4. The agent-driven self-evolving robot navigation method according to claim 1, characterized in that, In step (4), when the recorded expected robot state change does not occur, the existing skills do not meet the success conditions, or the failure or termination conditions have been met, the deterministic operating platform delivers the corresponding task problem state and execution feedback to the agent; the agent adds the task problem state and execution feedback to the runtime decision context, and uses the same decision loop for generating capability call decisions round by round to regenerate subsequent capability call decisions that are different from the previous capability call decisions based on the updated runtime decision context.

5. The agent-driven self-evolving robot navigation method according to claim 1, characterized in that, In step (5), the deterministic operating platform records the retrieval, adoption, actual execution, problem solving, task contribution, average cost, security rejection, mismatch, continuous failure and recent use information of skills, and delivers skill effect evaluation information to the agent.

6. The agent-driven self-evolving robot navigation method according to claim 1, characterized in that, The local safety execution module in the deterministic operating base reviews all physical motion requests; the local safety execution module only executes physical motion requests that pass the review, or rejects physical motion requests that fail the review and records the reason for rejection; when a safety constraint is triggered, the local safety execution module directly executes an explicit stop instruction, and delivers the termination facts and reasons to the agent after execution.

7. The agent-driven self-evolving robot navigation method according to claim 1, characterized in that, In step (5), the capability evolution operation includes at least one of creating a skill, modifying a skill, merging a skill, downgrading a skill, or deactivating a skill; When the intelligent agent initiates the capability evolution operation of creating a skill, it compares the changes in the task problem state before and after each capability call in the handling trajectory with the problem resolution conditions, retains the capability calls that change the task problem state to the problem resolution conditions, converts the scene-specific constants into parameter constraints, extracts the applicable conditions, success conditions and failure or termination conditions, and generates candidate skills that include skill identifier, problem characteristics, applicable conditions, handling content, success conditions, failure or termination conditions and usage records. The candidate skills are evaluated for compliance based on format, interface reference, capability call permission scope, call limit, deadline, and whether they bypass the local security execution module. The candidate skills that pass the compliance evaluation are registered and set to searchable status, so that the candidate skills can participate in skill retrieval in subsequent decision cycles or navigation tasks according to applicable conditions. The skill retrieval results are added to the runtime decision context. If a compliance determination fails, the non-compliance facts and reasons are recorded and delivered to the intelligent agent, which then decides whether to propose revisions.

8. The agent-driven self-evolving robot navigation method according to claim 1, characterized in that, The deterministic operating platform determines the skill retrieval results based on preset matching conditions, which include question matching conditions, ontology matching conditions, and environment matching conditions; the question matching conditions are that the question categories are consistent or the question feature similarity reaches a threshold. The on-body matching condition is that the robot type, capability call permission and necessary robot state meet the skill application conditions; the environment matching condition is that the environment state falls within the skill application range and there is no environmental constraint conflict; the deterministic operating base delivers the skill retrieval result to the agent; when the skill retrieval result is empty, the existing skill does not meet the success condition, or the failure or termination condition has been met, the agent uses the decision loop in step (4) and can change the current sub-goal, adjust the call parameters, select other callable capabilities, combine multiple callable capabilities or call human-computer interaction capabilities.

9. The agent-driven self-evolving robot navigation method according to claim 1, characterized in that, The deterministic operating platform includes a skills usage statistics module and a static check and version registration module; When the number of retrievals is less than the minimum retrieval sample size or the number of actual executions is less than the minimum execution sample size, the skill usage statistics module only records the usage information of the corresponding skill and delivers the observation period information to the agent. The observation period refers to the stage in which the current retrieval sample size or actual execution sample size of the skill has not yet reached the minimum sample size required for statistical evaluation, and the skill usage information is only accumulated, without triggering capability evolution operations based on the usage information. When the actual execution sample is sufficient and the execution success rate is lower than the threshold, the agent decides whether to initiate the ability evolution operation of rewriting skills; when the mismatch rate is higher than the threshold, it decides whether to initiate the ability evolution operation of tightening the applicable conditions or adjusting the retrieval index; when it is frequently retrieved but rarely used, it decides whether to initiate the ability evolution operation of reducing the retrieval priority. When the static inspection and version registration module determines that a skill contains instructions that exceed authority, bypass the local security execution module, or are dangerous, it directly sets the skill to a frozen and unsearchable state, records the freezing facts and reasons, and delivers them to the agent. The agent can submit a revision or restoration request, and the static inspection and version registration module will only unfreeze the skill and record the corresponding version status after the compliance determination is passed. Skills whose most recent usage time exceeds a preset time will be evaluated separately according to the most recent usage time and will not be treated as execution failures.

10. A self-evolving robot navigation device driven by an intelligent agent based on the method of any one of claims 1-9, comprising a memory, a processor, and a communication interface, characterized in that, The memory stores navigation tasks, robot status, environmental status, machine-readable descriptions of callable capabilities, a skill library, and usage records; the processor executes programs to form an intelligent agent decision-making module and a deterministic operating base. The agent decision-making module is used to select callable capabilities based on the runtime decision context, generate capability call decisions, make task semantic termination judgments, and initiate skill capability evolution operations. The deterministic operating base includes: The call constraint and execution module is used to make compliance judgments on capability call decisions, execute capability calls that pass the compliance judgment, record and deliver execution feedback, and directly terminate subsequent capability calls when a termination event occurs; The problem status recording module is used to record execution feedback, robot status changes, and environmental status changes; The static inspection and version registration module is used to make compliance judgments on candidate skills, register versions, and set the searchable or frozen status of skills. The skills usage statistics module is used to record usage information and deliver skills effectiveness evaluation information; The local security execution module is used to review, execute, or reject physical motion requests, and to directly execute stop commands when a termination event occurs; The communication interface is used to exchange observation and physical motion requests with the mobile robot body.

Images