A robot complex task autonomous planning method based on adaptive learning

Abstract

The application relates to the technical field of robot task planning, and provides a robot complex task autonomous planning method based on adaptive learning, which comprises the following steps: acquiring hospital ward scene state data and writing the scene state according to a unified timestamp; analyzing a task request to generate a task chain object, establishing regional compliance, passage and elevator constraints, and determining a risk threshold value by historical sliding window crowd density statistics; splitting the task chain into skill nodes, writing in entry conditions, completion criteria and failure compensation actions, and constructing a skill graph; recording results to update confidence during execution, and only selecting a node subset with compensation and meeting constraints when confidence is conservative; distributing road sections according to scene states to allow risks to form risk barriers, and jointly searching on the skill graph to generate a plan; when a disturbance occurs, firstly locally compensate, and only when the completion is not met or the barrier will be broken, globally re-plan; recording decisions and states to an audit log and backfilling updates.

Description

Technical Field

[0001] This invention relates to the field of robot task planning technology, specifically a method for autonomous planning of complex robot tasks based on adaptive learning. Background Technology

[0002] In the application of nursing service robots in hospital wards, the robots need to autonomously plan multi-stage task chains under operating conditions such as continuous addition of tasks, congested passageways, and uncertain elevator waiting times. At the same time, constraints such as controllable safety risks and non-boundary of compliant areas are imposed, so that the robots can simultaneously withstand multiple constraint couplings and strong dynamic changes in the environment within the same task chain.

[0003] Current technologies typically employ a combination of rule orchestration and task schedulers to generate a task execution order from the current task list according to preset rules, and then overlay global path planning onto the task order to obtain cross-regional and cross-floor travel routes. During the execution phase, the local obstacle avoidance module outputs local detour trajectories when the corridor is congested or people block the view. When a task is interrupted or a temporary task is interrupted, the interrupted task is written into the task queue and local path replanning is triggered, or the order and route of the affected task segments are recalculated.

[0004] However, the replanning and re-execution mechanism in the above-mentioned technical solutions requires rewriting the existing plan every time a new order is inserted, congestion occurs, or elevator waiting conditions are triggered. This results in the planning results being rewritten multiple times within a short time window, causing planning jitter. Planning jitter causes task chains that are already halfway executed to roll back and be rearranged, resulting in high task rollback costs. At the same time, the online learning phase relies on only a small amount of interactive data for incremental updates. The model parameter updates are driven by small sample statistics and form a bias, which makes the output trigger replanning conditions more frequently, further amplifying planning jitter and making task chains more prone to timeouts and on-site intervention. Summary of the Invention

[0005] In view of the existing problems mentioned above, a method for autonomous planning of complex robot tasks based on adaptive learning is proposed.

[0006] The technical solution adopted by this invention to solve the above-mentioned technical problems is: a robot complex task autonomous planning method based on adaptive learning, comprising:

[0007] Step S1: Obtain hospital ward scene status data and write the data into the scene status based on a unified timestamp; parse the task request into a task chain object, generate regional compliance constraints, access constraints, and elevator constraints corresponding to the task chain object, and determine the risk threshold from the population density statistics of the historical sliding window;

[0008] Step S2: Decompose the task chain object into skill nodes, write the entry condition, completion criterion and failure compensation action for each skill node, and connect the skill nodes into a skill graph according to the reachability relationship; during the execution process, record the execution result of each skill node and update the confidence. When the confidence information meets the conservative condition, restrict the selectable skills to a subset of nodes with compensation ability and satisfy the constraints.

[0009] Step S3: Based on the scenario state, allocate permissible risks to each segment of the task chain and form risk barriers. Perform joint search on the skill graph to generate an initial execution plan. When a disturbance event is detected, adjust adjacent skill nodes through a local compensation mechanism. When local adjustment alone cannot meet the task completion requirements or will break through the risk barriers, trigger global replanning. Write the planning decision, execution status, and triggering conditions into the audit log to form a closed-loop feedback and update the scenario state model and skill confidence.

[0010] As a preferred implementation method, the specific steps of scene perception and task modeling are as follows:

[0011] In each control cycle, the robot obtains spatial topology and operational status information from the hospital ward scene status data source. Using the local controller clock as a unified time reference, it aligns the acquisition timestamps carried by each data source, backfills observations earlier than the current cycle with the most recent valid value, and writes invalid status for observations later than the current cycle or missing observations and records the failure reason and source identifier to form a scene status record.

[0012] The received task request is parsed into a task chain object according to preset fields, and each stage is bound to the corresponding spatial node and the allowed passage area. Constraint sets are generated item by item based on the task chain object. Area compliance constraints are generated according to the restricted area and disinfection rules, and the reachable area boundary is restricted. Passage constraints are generated according to the corridor passage width and crowd density, and the passage segment status is marked. Elevator constraints are generated according to the elevator availability status, and the set of optional elevators and the reasons for unavailability are given.

[0013] The robot performs field-by-field parsing of the message, generates a stage sequence in a fixed order and writes it into the task chain object. When any field is missing, point mapping fails, cross-layer requirements are inconsistent with the point floor, or timeliness requirements are invalid, the robot writes the task chain object into an unexecutable state and records the failure reason code and the name of the missing field. At the same time, it returns a resend prompt to the scheduler, and obtains a task chain object that can be directly used for subsequent constraint generation and planning execution.

[0014] Extract sample sequences from the population density samples within the historical sliding window according to a fixed window length, remove invalid states, sort them from smallest to largest, take the density value corresponding to the preset quantile as the risk threshold, write the risk threshold into the scene status and associate it with the road segment status.

[0015] As a preferred implementation, the specific steps for risk upper bound modeling and constraint generation are as follows:

[0016] After acquiring and aligning the hospital ward scene status data in each control cycle, the robot extracts safety event records related to road passage from the audit log. It then categorizes these records into buckets based on passage width range, crowd density range, and passage mode. For each bucket, it counts the total number of passages and the number of collision events. Under a given confidence level, it calculates a statistical upper bound for the collision incidence rate of each bucket and uses this upper bound as the collision risk upper bound field for that bucket. This upper bound, along with the confidence flag and sample size, is written into the risk upper bound mapping table. When the sample size of a bucket is less than the minimum limit, the bucket is marked as low confidence, and the maximum collision risk upper bound of all density buckets within the same passage width range is used as a conservative estimate. If the sample size of all buckets within that width range is insufficient, the global maximum collision risk upper bound is used and marked as extremely low confidence.

[0017] As a preferred implementation, the specific steps of skill graph construction and confidence management are as follows:

[0018] After obtaining the task chain object, the stage sequence is discretized and mapped according to the preset skill library; the robot sequentially matches the corresponding skill for each stage in the task chain object and generates a skill node, and writes three types of reproducible fields for each skill node: the entry condition field is used to limit the pre-state of triggering the node, the completion criterion field is used to determine the closed loop completion of the node, and the failure compensation action field is used to prioritize recovery under disturbance without rewriting the overall plan; and compensation branches are preset for elevator-related nodes and corridor-related nodes;

[0019] Skill nodes are connected into a skill graph based on the reachability relationships in the scene state. The reachability relationship is based on the condition that the endpoint of a node can reach the starting point of the next node without violating regional compliance constraints, access constraints, and elevator constraints. An execution record entry is established for each skill node during the execution process.

[0020] After completing the skill graph construction, a connectivity check and anomaly handling loop are executed. First, using the area compliance constraints, access constraints, elevator constraints, and risk barriers in the current scenario as hard constraints, edge and node pruning is performed on the skill graph. Nodes that violate restricted areas and disinfection rules are written to the unavailable state, edges with the corresponding road segment status as prohibited are deleted, elevator edges with malfunctions or emergency occupancy are deleted, and nodes with a confidence state of disabled are removed from the candidate set. The starting node set and the ending node set are determined, and a reachability traversal is performed on the pruned directed skill graph with the starting node set as the source. For each visited node... Write the reachable field to 1 and record its predecessor node identifier; if the reachable field of any terminating node is 1 after the traversal, the skill graph connectivity check is considered passed, and the predecessor chain is retained for subsequent plan backtracking; if the reachable field of all terminating nodes is 0, it is considered disconnected and enters the exception handling process, and the graph repair is performed according to the preset repair order. After each repair is completed, the connectivity traversal is re-executed and the reachable field is updated; if the connection is still not connected after the repair, the task chain object is written to the non-executable state and a safety degradation action is output.

[0021] The reliability of nodes is updated based on execution records to form confidence information. The confidence information is written into three-state identifiers: trusted, conservative, and disabled. When a node fails continuously for the same reason, repeatedly times out in high-risk sections, or has insufficient coverage of key constraint labels, it is placed in a conservative state. When a node is in a conservative state, the set of optional skills is restricted to a subset of nodes that simultaneously meet the constraints and have failure compensation actions during the planning and local adjustment phases, so that recoverable nodes are used first to complete the closed loop when disturbances occur.

[0022] As a preferred implementation, the specific steps for introducing risk upper bound consistency gating are as follows:

[0023] When mapping the phase sequence of a task chain object to skill nodes, the robot adds a risk upper bound constraint field and a strategy mode field to each skill node. The risk upper bound constraint field is jointly determined by the road segment collision risk upper bound field and the allowable risk upper bound field of the task chain. For skill nodes that are related to close interaction with people, the robot reads the corresponding collision risk upper bound field from the scene state according to the road segment or area covered by the node, and compares it with the allowable risk upper bound field of the task chain to calculate the risk margin. If the risk margin is negative, the skill node is directly marked as unselectable and only alternative nodes with failure compensation actions are retained.

[0024] The strategy mode field includes two categories: normal mode and efficiency mode. An executable set of control constraints is written for each mode. The robot only allows a skill node to switch from normal mode to efficiency mode when the risk margin is greater than a preset threshold. The current active mode identifier and switching trigger condition are written into the strategy mode field. When the risk margin is in a preset low-margin range, the node is forced to maintain normal mode operation to leave sufficient safety margin. When establishing directed edges in the skill graph based on reachability, the robot first checks whether the risk upper bound constraint fields of the start and end nodes meet the connection condition, i.e., the collision risk upper bound of both nodes is not higher than the task's allowed risk upper bound. If the condition is met, a directed edge is established, and the risk upper bound constraint field, strategy mode field, and the current active mode of the start and end nodes are written into the edge's constraint label. If the condition is not met, the edge is not established. In the alternative edge set, alternative paths are recorded in the order of higher risk margin, lower proportion of restricted passage sections, and higher skill node confidence state, and a trigger switching condition is written for each alternative edge.

[0025] During execution, a near miss event count and a collision event identifier are added to the execution record entry of each skill node. When updating the node confidence information based on the execution record, the appearance of the collision event identifier or the near miss event count reaching a preset number within a fixed window is used as a strong trigger condition for the safety side. Once triggered, the confidence status of the node in efficiency mode is disabled and forced to fall back to normal mode. Only a subset of nodes with positive risk margin and failure compensation actions are allowed to be selected.

[0026] As a preferred implementation, the specific steps of the anti-jitter planning under risk constraints are as follows:

[0027] After obtaining the updated scene status and skill map, the passable area is divided into several road segments and risk barriers are generated for the road network path of the task chain. The robot first writes the allowed risk level and allowed passage mode for each road segment according to the road segment status identifier, risk threshold and task chain timeliness requirements in the scene status. The allowed risk level is written as the planning gating condition into the constraint label of the edge of the skill map.

[0028] A joint search is performed on the skill graph to simultaneously determine the task order and the path. The joint search is constrained by the task chain stage sequence as the main order and by the reachability relationship of skill nodes as the edge constraint. Nodes with failure compensation actions and a trustworthy state are given priority. When there are multiple feasible edges, the path with a lower proportion of restricted passage segments and that does not depend on the unavailable elevator state is given priority. An initial execution plan is generated and issued for execution.

[0029] As a preferred implementation, the specific steps of the anti-jitter planning under risk constraints are as follows:

[0030] During execution, the robot continuously monitors for disturbance events. When a disturbance event is detected, the robot does not immediately rewrite the global plan, but instead enters a local compensation mechanism. The adjustment range is the adjacent subgraph formed by the current skill node and its successor nodes. First, the robot executes the preset failure compensation action of the current node and attempts to restore to the next node of the original plan. If the compensation action can be completed without violating the regional compliance constraints and risk barriers, the original plan structure remains unchanged. Only the entry conditions and edge constraint labels of adjacent nodes are updated to reflect the new road segment status or elevator status. When a specified situation occurs, global replanning is triggered.

[0031] When global replanning is triggered, the joint search is re-executed on the skill graph while maintaining the completed node sequence without reversal. The robot writes each planning and adjustment process into the audit log and backfills the execution results to update the road segment state model and skill node confidence state in the scene state, forming a closed-loop feedback.

[0032] As a preferred embodiment, the specific steps for performing local adjustments in a fixed sequence when a disturbance event is detected are as follows:

[0033] Prioritize downgrading the traffic mode of affected road segments from efficiency mode to normal or restricted traffic mode, and update the entry conditions and edge constraint labels of adjacent skill nodes; if there are still necessary road segments with collision risk upper bound field higher than allowable risk upper bound field after downgrading, switch the alternative edge set or call the preset failure compensation action in the adjacent subgraph formed by the current skill node and its successor nodes to bypass the high-risk road segments; global replanning is only triggered when the next handover point cannot be reached after local adjustment, the handover completion criterion cannot be met, or the remaining reachable paths will all exceed the allowable risk upper bound field.

[0034] Beneficial effects

[0035] Compared with the prior art, the present invention has the following advantages:

[0036] 1. By unifying the time alignment of scene states, object-oriented modeling of task chains, and the coordinated generation of regional compliance constraints, access constraints, and elevator constraints, the multi-stage task chains of pickup, delivery, cross-floor, handover, and return are made consistent in closed loop under the same scene state, reducing the mismatch between plans and on-site intervention caused by data delays, missing data, and inconsistent standards.

[0037] 2. Based on the node reliability update and confidence state gating of the execution record, unreliable skill nodes are restricted or removed from the selectable set, so that the planning will prioritize nodes with compensation capabilities and satisfy the constraints, reducing the planning jitter and task rollback costs caused by repeatedly using unreliable strategies. Attached Figure Description

[0038] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0039] Figure 1 This is a flowchart illustrating the present invention;

[0040] Figure 2 This is a comparison diagram of the effects of the present invention and the prior art, where black bars represent the prior art and gray bars represent the present invention. Detailed Implementation

[0041] To make the technical means, creative features, objectives, and effects of this invention easier to understand, the invention is further described below with reference to specific embodiments. However, the following embodiments are merely preferred embodiments of this invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments described herein without creative effort are all within the protection scope of this invention.

[0042] Example 1:

[0043] To achieve the above objectives, please refer to Figures 1 to 2 This invention provides a method for autonomous planning of complex robot tasks based on adaptive learning, the method comprising the following steps:

[0044] Step S1: Obtain hospital ward scene status data and write the data into the scene status based on a unified timestamp; parse the task request into a task chain object, generate regional compliance constraints, access constraints, and elevator constraints corresponding to the task chain object, and determine the risk threshold from the population density statistics of the historical sliding window;

[0045] Step S2: Decompose the task chain object into skill nodes, write the entry condition, completion criterion and failure compensation action for each skill node, and connect the skill nodes into a skill graph according to the reachability relationship; during the execution process, record the execution result of each skill node and update the confidence. When the confidence information meets the conservative condition, restrict the selectable skills to a subset of nodes with compensation ability and satisfy the constraints.

[0046] Step S3: Based on the scenario state, allocate permissible risks to each segment of the task chain and form risk barriers. Perform joint search on the skill graph to generate an initial execution plan. When a disturbance event is detected, adjust adjacent skill nodes through a local compensation mechanism. When local adjustment alone cannot meet the task completion requirements or will break through the risk barriers, trigger global replanning. Write the planning decision, execution status, and triggering conditions into the audit log to form a closed-loop feedback and update the scenario state model and skill confidence.

[0047] In this embodiment, the specific content of step S1, scene perception and task modeling, includes:

[0048] Within each control cycle, the robot acquires spatial topology and operational status information from the hospital ward scene status data source. The control cycle is set to 1 second. The scene status data includes the ward floor plan and connectivity, restricted areas and disinfection rules, coordinates of the junction between nursing stations and wards, elevator availability, corridor width, and observed crowd density. The robot uses the local controller clock as a unified time reference, aligns the collection timestamps carried by each data source, backfills observations earlier than the current cycle with the most recent valid value, and writes invalid status for observations later than the current cycle or missing observations, recording the reason for failure and source identifier, thus forming a traceable scene status record.

[0049] The robot parses the received task requests into task chain objects according to preset fields. The task chain object includes task identifier, priority, type of item to be picked up or delivered, start and end point handover points, cross-level requirements, time requirements, and stage sequence. Each stage is bound to a corresponding spatial node and a permitted passage area. Based on the task chain object, a constraint set is generated item by item. According to the restricted area and disinfection rules, area compliance constraints are generated and the reachable area boundaries are restricted. The restricted area and disinfection rules include area identifier, boundary polygon, permitted passage time period and direction. According to the corridor passage width and crowd density, passage constraints are generated and the passage segment status is marked. According to the elevator availability status, elevator constraints are generated and a set of available elevators and reasons for unavailability are given. The elevator availability status is divided into available status, occupied waiting status, full load restriction status, fault outage status, etc.

[0050] Specifically, taking traffic constraints as an example, the robot first converts the passable areas of the ward's floor plan into a road network topology. The corridor is then divided into several segments based on key nodes such as doorways, corners, intersections, and elevator entrances. A geometric width and crowd density field are written for each segment. Subsequently, based on the robot's dimensions and safety clearance requirements, traffic determination rules are generated for each segment. Specifically, when the geometric width is insufficient for the robot to pass with a safety clearance, the segment is marked as prohibited and removed from the globally searched reachable set. When the geometric width only allows for single-machine passage but not for passing oncoming vehicles or lateral detours, the segment is marked as restricted passage and a constraint action field is written, including one-way entry, prohibition of passing oncoming vehicles, waiting at the segment entrance if necessary, and only allowing passage along the centerline. When the crowd density exceeds the risk threshold, even if the geometric width allows passage, the road segment will be changed from passable to restricted passage, and the constraint action field will be written, including speed limit entry, maintaining a larger avoidance distance, setting a yield point at the road segment entrance and implementing a wait-before-entry policy; when the geometric width allows oncoming traffic and the crowd density does not exceed the risk threshold, the road segment will be marked as passable and allowed to be used as a regular path edge; for road segments with missing geometric width or invalid crowd density, they will be uniformly written as unknown status and treated as restricted passage, while recording the reason for the missing data and the data source, so that the planning stage will use prohibited passage as a hard constraint to remove edges, restricted passage as an entry gating and waiting strategy constraint, and passable passage as a regular usable edge, thereby realizing executable constraints on the road segment status based on corridor passage width and crowd density;

[0051] Specifically, the process involves parsing task requests into task chain objects, receiving task request messages from the nurse station system, mobile ordering terminal, or scheduling server. These messages are in a structured format and contain preset fields. The robot performs field-by-field parsing of the messages, first reading the task identifier, task type, priority, start handover point identifier, end handover point identifier, item type, cross-floor requirements, and timeliness requirements. The handover point identifier is mapped to floor number and planar coordinates in the location database, and the stage branches requiring or not requiring elevator access are determined based on the cross-floor requirements. A stage sequence is generated in a fixed order and written into the task chain object. The aforementioned phase sequence includes heading to the starting point, pickup confirmation, heading to the destination, handover confirmation, and return or standby. Each phase is bound to the start and end positions, the allowed access area identifier, whether it is necessary to pass through the elevator, and the completion criteria, such as reaching within the radius threshold, successful handover scan, or successful manual confirmation. When any field is missing, the point mapping fails, the cross-floor requirement is inconsistent with the point floor, or the time requirement is invalid, the robot writes the task chain object to an unexecutable state and records the failure reason code and the missing field name. At the same time, it returns a resend prompt to the scheduler, thus obtaining a task chain object that can be directly used for subsequent constraint generation and planning execution.

[0052] To achieve executable gating of highly dynamic pedestrian flow, sample sequences are extracted from the crowd density samples within the historical sliding window with a fixed window length, invalid states are removed, and the samples are sorted from smallest to largest. The density value corresponding to a preset quantile is taken as the risk threshold, such as the 90th percentile. This risk threshold is written into the scene state and associated with the road segment state for subsequent disturbance detection and risk barrier generation. This ensures that task chain modeling, constraint generation, and risk threshold are consistent in a closed loop within the same time base and the same scene state.

[0053] In this embodiment, the specific content of step S2, skill graph construction and confidence management, includes:

[0054] After obtaining the task chain object, the stage sequence is discretized and mapped according to a preset skill library, which includes executable skills such as reaching a designated point, entering the elevator waiting area, calling and waiting for the elevator, and taking and exiting the elevator. The robot sequentially matches the corresponding skills for each stage in the task chain object and generates skill nodes. Three types of reproducible fields are written for each skill node: the entry condition field is used to limit the preconditions for triggering the node, including that the current position has reached the end point of the previous node, the area meets compliance constraints, the target area is not restricted and the passable road is not prohibited, and the elevator node is in an available or waiting state; complete. The success criterion field is used to determine the completion of the closed loop at this node, including arrival criteria, action criteria, and time criteria; the failure compensation action field is used to prioritize recovery without rewriting the overall plan under disturbances, including compensation trigger conditions, compensation execution sequence, and compensation termination conditions. The compensation execution sequence is explicitly set to retreat to the nearest safe point, switch to restricted access mode, retry the current node, and preset compensation branches for elevator-related nodes, such as changing to an alternative elevator, changing to waiting on the same floor, or changing to an alternative passage. Preset compensation branches are also set for corridor-related nodes, such as waiting at the entrance, passing along the centerline at a limited speed, and retreating to the yield point when necessary.

[0055] Skill nodes are connected into a skill graph based on reachability relationships in the scenario state. The reachability relationship is based on the condition that the endpoint of a node can reach the starting point of the next node without violating regional compliance constraints, access constraints, and elevator constraints. Specifically, connecting skill nodes into a skill graph involves creating directed edges for nodes that satisfy the reachability relationship and writing constraint labels and alternative edge sets for the edges. To suppress planning jitter, an execution record entry is created for each skill node during execution. The entry includes node identifier, start time, end time, execution result, failure reason code, and constraint label at that time. The execution result is divided into success, failure, and abort.

[0056] After the skill graph is constructed, a connectivity check and anomaly handling loop is executed. First, using the area compliance constraints, access constraints, elevator constraints, and risk barriers in the current scenario as hard constraints, edge and node pruning is performed on the skill graph. Nodes that violate restricted areas and disinfection rules are written to the unavailable state, edges with the corresponding road segment status as prohibited are deleted, elevator edges that are faulty or in emergency use are deleted, and nodes with the confidence state of disabled are removed from the candidate set. The starting node set and the ending node set are determined, and a reachability traversal is performed on the pruned directed skill graph with the starting node set as the source. For example, a breadth-first traversal is performed by expanding the queue edge by edge. For each visited node, the reachable field is written as 1 and its predecessor node identifier is recorded. If the reachable field of any ending node is 1 after the traversal, the skill graph connectivity check is considered to have passed. Meanwhile, the predecessor chain is retained for subsequent plan backtracking; if the reachable field of all terminating nodes is 0, it is determined to be disconnected and enters the exception handling process. Specifically, a disconnection diagnosis record is generated, which includes the disconnection type, the earliest fault stage index, the list of nodes and edges pruned at the fault, and the corresponding reason code set; graph repair is performed according to the preset repair order, which includes: putting conservative nodes with failure compensation actions into the candidate set to replace disabled nodes; changing the restricted access segment from unadjustable to optional but requiring entry waiting mode and rebuilding the corresponding edges; rearranging the candidate elevator set according to the available state and reconnecting the corresponding elevator edges for the elevator stage; after each repair is completed, the connectivity traversal is re-executed and the reachable field is updated; if it is still not connected after repair, the task chain object is written to the non-executable state and a safety degradation action is output;

[0057] The reliability of nodes is updated based on the execution records to form confidence information. The confidence information is written into three-state identifiers: trustworthy, conservative, and disabled. When a node fails repeatedly for the same reason, times out repeatedly in high-risk sections, or has insufficient coverage of key constraint labels, it is placed in a conservative state. When a node is in a conservative state, the set of optional skills is restricted to a subset of nodes that simultaneously meet the constraints and have failure compensation actions during the planning and local adjustment phases, so that recoverable nodes are used first to complete the closed loop when disturbances occur.

[0058] Specifically, the process of updating node reliability based on execution records and forming confidence information involves extracting the N most recent valid execution records for any skill node under the same constraint label, and recording the execution result field of each record as... Success is denoted as 1, failure as 0, and i represents the index of the execution record; the number of successes is denoted as 1. Let F be the length of the longest consecutive segment of failed records among the N records. Scan the records in chronological order, incrementing the count by 1 for each failed record and resetting it to zero for each successful record. Take the maximum value over the entire process. The node reliability score is then calculated using the following formula:

[0059] ,

[0060] Where R represents the node reliability score, and N is set to 20 in this embodiment;

[0061] The system checks whether safety-related failure reason codes appear in the valid execution records. These codes include boundary crossing risk, collision risk, intrusion into restricted areas, and elevator malfunction. Confidence information is updated according to a judgment rule. When a safety-related failure reason code appears in the window, or the same failure reason code appears consecutively a preset number of times (e.g., 3 times), or the number of failures in the window reaches a preset proportion (e.g., 30%), the confidence status of the node is set to conservative, and the triggering reason code and corresponding constraint label are written into the confidence information. When a node is already in a conservative state and triggers any of the aforementioned conditions again, the node is set to disabled, and the reason for disabling and the time of disabling are written. When a node is in a conservative or disabled state, its status is restored step-by-step in the order of disabled, conservative, and confident only if subsequent execution records show consecutive successes without safety-related failure reason codes and covering the same constraint label, and the restoration basis is recorded.

[0062] In this embodiment, the specific content of the anti-jitter planning under the risk constraint in step S3 includes:

[0063] After obtaining the updated scene state and skill map, the robot divides the passable area into several road segments and generates risk barriers for the road network path of the task chain. Specifically, the robot first writes the allowed risk level and allowed passage mode for each road segment according to the road segment status identifier, risk threshold and task chain time requirement in the scene state. Among them, the prohibited passage road segments are directly eliminated as hard constraints, and the restricted passage road segments are written with passage modes such as speed limit, entrance waiting, no passing and center line passage, etc. Passable road segments allow regular passage. On this basis, the robot writes the allowed risk level as the planning gating condition into the constraint label of the edge of the skill map, so that subsequent searches can only select the edge that does not break the risk barrier.

[0064] The robot performs a joint search on the skill graph to simultaneously determine the task order and the path. The joint search takes the task chain stage sequence as the main order constraint and the reachability relationship of skill nodes as the edge constraint. It prioritizes nodes with failure compensation actions and a trustworthy state. When there are multiple feasible edges, it prioritizes the path with a lower proportion of restricted passage segments and does not depend on the unavailable elevator state. It generates an initial execution plan and issues it for execution.

[0065] Specifically, the joint search is performed on the skill graph using a constrained A* graph search algorithm. The process is as follows: the search state is defined as the current stage index, current skill node, and current location resource state. The task chain stage sequence serves as the primary order constraint, allowing only progressive expansion by stage. An open list is initialized and written with the initial state. States with the lowest cost in the open list are dequeued sequentially to expand their reachable successor skill nodes. For each candidate successor edge, a hard constraint check is performed first. Successor states that pass the check are then written with their cumulative cost and predecessor pointer and enqueued. Cost sorting prioritizes nodes that meet risk barriers and compliance constraints, then nodes with reliable states and failure compensation actions, and finally nodes with shorter expected walking and waiting times, using a lexicographical order rule. Expansion stops when the state reaches the termination stage and meets the termination node completion criterion. The skill node sequence and corresponding road segment sequence are backtracked along the predecessor pointer and used as the initial execution plan output.

[0066] During execution, the robot continuously monitors disturbance events, including order insertion, crowd density exceeding risk thresholds, road segment status changing from passable to restricted, elevator status changing from available to unavailable, and continuous triggering of backoff in the same road segment for local obstacle avoidance. When a disturbance event is detected, the robot does not immediately rewrite the global plan but enters a local compensation mechanism. The local compensation mechanism adjusts within the adjacent subgraph formed by the current skill node and its successor nodes. It first executes the preset failure compensation action of the current node and attempts to restore to the next node of the original plan. If the compensation action can be completed without violating regional compliance constraints and risk barriers, the original plan structure remains unchanged, and only the entry conditions and edge constraint labels of adjacent nodes are updated to reflect the new road segment status or elevator status. Global replanning is triggered only when any of the following situations occur: after local compensation, the next handover point of the task chain cannot be reached or the handover completion criterion cannot be met, resulting in the task completion not being established; or local compensation will cause the subsequent necessary road segment to enter a prohibited or restricted passage with a risk level exceeding the risk barrier, resulting in the breach of safety and compliance constraints.

[0067] Specifically, the preset failure compensation actions refer to a set of recovery actions predefined and written with compensation trigger conditions, compensation execution sequences, and compensation termination conditions for each type of skill node. These actions are used to prioritize the recovery of the task chain without rewriting the overall plan structure when a disturbance occurs. They include the following directly executable compensation actions: For corridor passage nodes, execute the action of returning to the nearest yielding point, waiting at the entrance until the road segment status is restored to passable or the crowd density is below the risk threshold, and then passing along the centerline at a limited speed in a restricted passage mode; For elevator-related nodes, execute the action of switching to a backup elevator and re-entering the waiting area. If the backup elevator is unavailable, switch to waiting on the same floor and periodically check the elevator status. After recovery, continue compensation for subsequent nodes; For doorway handover nodes, execute the action of waiting at a safe point outside the handover area, retrying the handover confirmation, and if consecutive failures occur, switch to a backup handover point or report for manual assistance and return to a safe point; For positioning and perception anomaly nodes, execute the action of entering the nearest safe stopping area, repositioning and refreshing the status, and then re-entering the current node after recovery.

[0068] When global replanning is triggered, the joint search is re-executed on the skill graph, and the completed node sequence is kept from being rolled back, reducing rollback costs and suppressing planning jitter caused by frequent rewriting. The robot writes each planning and adjustment process into the audit log, which includes the plan version number, the selected skill node sequence, the selected road segment status, the disturbance event type, the triggering condition, the local compensation action used, whether global replanning was triggered and the corresponding reason code, and backfills the execution results to update the road segment status model and skill node confidence status in the scene status, forming a closed-loop feedback to stably complete complex task chains and reduce planning jitter in the highly dynamic scene of the ward.

[0069] like Figure 2 The figure shows a comparison of the performance of an adaptive learning-based autonomous planning method for complex robot tasks. Black bars represent existing technologies, and gray bars represent the present invention. Gray bars below 1 indicate that the present invention achieves a lower performance on the corresponding metrics compared to existing technologies. As can be seen from the figure, the present invention reduces frequent plan rewriting and thus lowers the number of global replanning triggers through joint search under risk barrier constraints and a local compensation priority mechanism under perturbation. Simultaneously, through confidence gating and risk upper bound consistency constraints, high-risk road segments and unreliable skill nodes are restricted or removed from the selectable set, reducing the number of near misses and collisions and suppressing risk fluctuations in high-traffic scenarios. Furthermore, the present invention reduces time losses caused by waiting and backtracking, lowers the task timeout rate, and shortens the average completion time, thus achieving a stable balance between safety and efficiency.

[0070] Example 2:

[0071] This embodiment 2 provides an improved solution based on embodiment 1. Embodiment 1 constructs a skill graph and implements confidence gating on skill nodes. Combined with the local compensation priority and global replanning restriction mechanism under disturbance events, it reduces the frequent rewriting of plans caused by order insertion, congestion, and elevator uncertainty. However, in situations such as high traffic flow in ward corridors and crowding in elevator halls, if the above threshold is relaxed to a soft constraint to improve efficiency, there is a lack of a stable risk upper bound constraint, which leads to an increase in the risk of collisions in crowded corridors and near intersections and makes it difficult to maintain a stable risk level. Thus, it is difficult to achieve both safety and efficiency stably at the same time.

[0072] In this embodiment, based on step S1 of embodiment 1 (scene perception and task modeling), a stable and executable risk upper bound modeling and constraint generation are further added. The specific content includes:

[0073] After acquiring and aligning the hospital ward scene status data in each control cycle, the robot extracts safety event records related to road segment passage from the audit log. These safety event records include road segment identification, passage mode, current crowd density, corridor width, and the number of times the minimum safe distance is triggered. The event records are then binned and statistically analyzed according to passage width range, crowd density range, and passage mode. For each bin, the total number of passages and the number of collision events are counted. The collision rate upper bound at a 95% confidence level is calculated using the Wilson scoring method and used as the upper bound field for the collision risk of that bin. This upper bound field for collision risk, along with the confidence marker and sample size, is then used to calculate the collision rate upper bound. Write the risk upper bound mapping table, which is defined as follows: priority 1 and time limit less than 5 minutes for emergency tasks correspond to an allowable risk upper bound of 5%; priority 2 to 3 and time limit less than 30 minutes for routine tasks correspond to an allowable risk upper bound of 2%; priority 4 to 5 for daily tasks correspond to an allowable risk upper bound of 1%. When the sample size of a certain bucket is less than 100 times, the bucket is marked as a low confidence state, and the maximum upper bound of the collision risk of all density buckets under the same passage width range is used as a conservative estimate. If the sample size of all buckets under the same width range is insufficient, the global maximum upper bound of the collision risk is used and marked as extremely low confidence.

[0074] Specifically, the rules for dividing the bucket intervals are as follows: the robot pre-sets a passage width threshold sequence and a crowd density threshold sequence, and constructs closed and open intervals accordingly to form a bucket index. The passage mode adopts discrete enumeration values ​​and includes normal mode, restricted passage mode and efficiency mode. For each safety event record, the robot reads the corridor passage width field, crowd density field and passage mode field respectively, determines its width interval index and density interval index according to the threshold sequence, and writes it into the bucket statistics table using the width interval index, density interval index and passage mode as bucket keys. When the width field or density field is missing, invalid or exceeds the range, the record is written into the preset missing bucket and marked as a low confidence state. At the same time, a conservative backoff rule is adopted for the missing bucket to select the maximum collision risk upper limit value in the same width interval or the global range as the statistical upper limit.

[0075] When generating road segment status, the robot determines a bucket index for each road segment based on its geometric width field and real-time crowd density field, and looks up the collision risk upper bound field of that road segment in a table. This collision risk upper bound field is then written into the scene status along with the road segment status. Next, based on the priority and timeliness requirements of the task chain objects, a task-level allowable risk upper bound field is determined. Specifically, a mapping table between priority and timeliness levels and allowable risk upper bounds is pre-stored. This mapping table divides priorities into three levels and timeliness requirements into three levels, and assigns a unique allowable risk upper bound value to each combination. The priority field and the timeliness requirement... After mapping the fields to their corresponding levels, a binary index is used to look up the allowed risk upper bound field in the mapping table and write it into the task chain object, along with a value source identifier and a default fallback identifier. When a priority field or timeliness requirement field is missing, level mapping fails, or the index is not found, the robot uses a preset conservative default allowed risk upper bound value and writes a failure reason code. The rule that the collision risk upper bound field should not exceed the allowed risk upper bound field is solidified as a hard constraint. When the collision risk upper bound field of a road segment exceeds the allowed risk upper bound field, the road segment is marked as prohibited and removed from the reachable set. When the upper bound of the collision risk is not higher than the upper bound of the allowable risk, the road segment status determination result generated based on geometric width and crowd density in Example 1 is retained. If it is restricted passage, restricted passage is maintained and the constraint action fields such as speed limit, entrance waiting, center line passage, and no passing defined in Example 1 are used. If it is passable, passable passage is maintained, thereby realizing a two-level gating mechanism: the first level is the rapid screening of density thresholds in Example 1, and the second level is the fine verification of the risk upper bound in this example. The more conservative result is taken as the final road segment status. On this basis, the robot backfills the actual passage record to the review after each execution. By logging data and modeling the risk upper bound based on the statistical analysis of historical safety events, we can perform refined gating based on the actual collision risk distribution under different combinations of passage width and density. This avoids the overall path extension and task timeout caused by the overly conservative fixed threshold in wide corridor low-density scenarios where passable sections are misjudged as restricted. At the same time, in narrow corridor high-density scenarios, hard constraints on the risk upper bound ensure that the collision risk does not exceed the task's allowed upper bound. Furthermore, confidence flags and sample size flags support conservative rollback for low-sample bins, ensuring that newly deployed robots or newly opened wards can still operate safely when historical data is insufficient.

[0076] In this embodiment, based on step S2 of embodiment 1 (skill graph construction and confidence management), a risk upper bound consistency gating is further introduced to simultaneously and stably achieve both security and efficiency. The specific content includes:

[0077] When mapping the phase sequence of the task chain object to skill nodes, in addition to writing an entry condition field, a completion criterion field, and a failure compensation action field for each skill node, the robot also adds a risk upper bound constraint field and a strategy mode field for each skill node. The risk upper bound constraint field is jointly determined by the road segment collision risk upper bound field obtained in step S1 and the allowable risk upper bound field of the task chain. For skill nodes related to close interaction with people, such as corridor passage, entering the elevator waiting area, waiting for the elevator, taking the elevator and exiting the elevator, and doorway handover, the robot reads the corresponding collision risk upper bound field from the scene state according to the road segment or area covered by the node, and compares it with the allowable risk upper bound field of the task chain to calculate the risk margin. If the risk margin is negative, the skill node is directly marked as unselectable and only the alternative node with failure compensation action is retained.

[0078] The strategy mode field is used to adjust efficiency without exceeding the risk upper bound. This field includes two categories: normal mode and efficiency mode. Each mode has an executable set of control constraints, including a speed limit, minimum yield distance, whether entry into restricted access sections is allowed, whether crossing elevator lobby gathering areas is allowed, and entrance waiting rules. For example, the normal mode's control constraints are: speed limit 0.8 m / s, minimum yield distance 1.0 m, prohibition of entry into restricted access sections, prohibition of crossing elevator lobby gathering areas, and entry obstruction requiring waiting until the road is cleared and prioritizing yielding to all personnel. The efficiency mode's control constraints are: speed limit 1.2 m / s, minimum yield distance 0.5 m, allow entry into restricted access sections but with speed limits, allow crossing elevator lobby gathering areas but with a greater yield distance, and entry obstruction allowing speed-limited passage and coordinated passage with personnel traveling in the same direction. The robot only allows the skill node to switch from normal mode to efficiency mode when the risk margin exceeds a preset threshold, such as 2%. The robot establishes a risk margin mode and writes the current activation mode identifier and switching trigger conditions into the strategy mode field. When the risk margin is between 0 and 2%, the node is forced to maintain normal mode operation to leave sufficient safety margin. When the robot then establishes directed edges of the skill graph according to reachability, it first checks whether the risk upper bound constraint fields of the start and end nodes meet the connection conditions, that is, the collision risk upper bound of the two nodes is not higher than the task's allowed risk upper bound. If the conditions are met, the directed edge is established and the risk upper bound constraint field, strategy mode field, and the current activation mode of the start and end nodes are written into the constraint label of the edge. If the conditions are not met, the edge is not established. In the alternative edge set, alternative paths are recorded in the order of higher risk margin, lower proportion of restricted passage sections, and higher confidence status of skill nodes. Trigger switching conditions are written for each alternative edge, such as the main path node confidence level dropping to conservative or disabled, the main path collision risk upper bound exceeding the allowed upper bound, and the main path failing continuously for a preset number of times. This allows for a quick switch to a risk-controllable alternative path when the main path is unavailable.

[0079] During execution, a near miss event count and a collision event identifier are added to the execution record entry for each skill node. The near miss event count is accumulated by the number of times the minimum distance of the local obstacle avoidance output is lower than the minimum yield distance of the mode. The collision event identifier is written by the collision sensor or emergency stop. When updating the node confidence information based on the above execution records, the appearance of the collision event identifier or the near miss event count reaching a preset number within a fixed window is used as a strong trigger condition for safety. Once triggered, the confidence state of the node in efficiency mode is disabled and it is forcibly reverted to normal mode. Only nodes with a positive risk margin and failure compensation action are allowed to be selected. Subset; when a node still experiences a collision event or a consecutive near miss reaches a preset number in the normal mode, the robot disables the entire node and automatically switches to the alternative edge set in the skill graph; by simultaneously writing the risk upper bound constraint field and the strategy mode field to the skill node and the skill graph edge, and enforcing the gating that does not break the risk upper bound in the confidence update and optional set restriction, efficiency can be improved within the range of positive risk margin in high-traffic corridors and elevator hall gathering scenarios, while providing an executable, auditable, and stable upper bound control for risk increase, thereby overcoming the defect of Example 1 where the collision risk increases and is difficult to maintain stability when the soft constraint is relaxed.

[0080] In this embodiment, based on the anti-jitter planning under the risk constraint in step S3 of embodiment 1, the upper bound field of road segment collision risk obtained in step S1 and the upper bound field of allowable risk of the task chain are further used as hard constraints for stable execution, and an adaptive traffic mode based on risk margin is introduced, the specific contents of which include:

[0081] After obtaining the updated scene state and skill map, when the robot generates risk barriers for the road network paths involved in the task chain, it not only marks the road segments as passable, restricted, or prohibited based on the road segment status identifiers and risk thresholds, but also looks up the collision risk upper bound field for each road segment and checks its consistency with the allowable risk upper bound field. If the collision risk upper bound field is higher than the allowable risk upper bound field, the road segment is directly marked as prohibited and removed from the edge group of the skill map. If the collision risk upper bound field is not higher than the allowable risk upper bound field, the road segment is divided into three risk margin levels: high margin level corresponds to risk margin >3%, medium margin level corresponds to 1% < risk margin ≤3%, and low margin level corresponds to 0 ≤ risk margin ≤1%. Based on the risk margin level, allowable barriers are selected for the road segment. According to historical operational data, the "permissive passage mode" is defined as follows: when the risk margin is greater than 3%, a collision rate of less than 0.1% can be considered low risk; when the risk margin is between 1% and 3%, a collision rate of 0.1% to 0.5% can be considered medium risk; and when the risk margin is less than 1%, a collision rate close to the upper limit of the allowable risk can be considered high risk. Therefore, 3% and 1% are used as thresholds for classifying risk margin levels. Specifically, road sections with high risk margins are allowed to use efficiency mode without violating regional compliance constraints and elevator constraints. For example, increasing the speed limit, shortening the entrance waiting limit, and allowing passage through elevator hall corridors while maintaining centerline passage are allowed. Road sections with low risk margins are forced to use regular mode or restricted passage mode, ensuring that the collision risk does not exceed the upper limit of the allowable risk while improving efficiency.

[0082] When generating the initial execution plan through joint search on the skill graph, the following hard gating conditions are written into the constraint A* extension rule: the upper bound of collision risk is no higher than the upper bound of allowable risk, the traffic mode matches the risk margin level, and the node confidence state allows the mode. Among the candidate edges that meet the hard gating conditions, paths with higher risk margins and shorter expected waiting times are prioritized to achieve stable risk control during high-traffic periods and efficiency compensation for low-risk road segments. During execution, when a disturbance event is detected, the robot performs local adjustments in a fixed order without immediately replanning globally. Specifically, for affected road segments, the traffic mode is downgraded from the efficiency mode to the regular, restricted traffic mode, and the entry conditions and edge constraint labels of adjacent skill nodes are updated. Specifically, for affected road segments, if their road segment status... If the road segment is in a passable state and currently in an efficient mode, the passability mode will be downgraded to a normal mode and control parameters such as the speed limit and minimum yield distance will be updated. If the road segment is in a restricted passability state and currently in an efficient mode, the passability mode will be downgraded to a restricted passability mode and constraints such as waiting at the entrance and prohibiting oncoming traffic will be enforced. If, after downgrading, there are still necessary road segments where the upper limit of the collision risk field is higher than the upper limit of the allowable risk field, the alternative edge set will be switched or the preset failure compensation action will be called to bypass the high-risk road segment in the adjacent subgraph formed by the current skill node and its successor nodes. Global replanning will only be triggered when the next handover point cannot be reached after local adjustment, the handover completion criterion cannot be met, or all remaining reachable paths will exceed the upper limit of the allowable risk field, and the completed node sequence will not be rolled back.

[0083] The decision-making process for each joint search and local adjustment is written into the audit log. In addition to the plan version number, node sequence and disturbance type, the audit log also includes the road segment collision risk upper bound field, the allowed risk upper bound field, the selected traffic mode, the mode switching reason code and the near loss event count. This is used to backfill and update the scenario state model and skill confidence. This ensures that even in scenarios with high traffic flow in the ward corridor and crowded elevator hall, even if soft constraints such as speed and waiting are relaxed to improve efficiency, the collision risk upper bound can still be stably constrained through hard gating of the risk upper bound, adaptive mode degradation and local substitution priority process, thereby simultaneously achieving both safety and efficiency.

[0084] The embodiments of the present invention described above are subject to modification and change of method by those skilled in the art without departing from the embodiments and broader aspects of the present invention. The appended claims are intended to include all such modifications and changes of method that do not depart from the present invention.

Claims

1. A method for autonomous planning of complex robot tasks based on adaptive learning, characterized in that, include: Step S1: Obtain hospital ward scene status data and write the data into the scene status based on a unified timestamp; The task request is parsed into a task chain object, and regional compliance constraints, access constraints and elevator constraints corresponding to the task chain object are generated. The risk threshold is determined from the population density statistics of the historical sliding window. Step S2: Decompose the task chain object into skill nodes, write the entry condition, completion criterion and failure compensation action for each skill node, and connect the skill nodes into a skill graph according to the reachability relationship; during the execution process, record the execution result of each skill node and update the confidence. When the confidence information meets the conservative condition, restrict the selectable skills to a subset of nodes with compensation ability and satisfy the constraints. Step S3: Based on the scenario state, allocate permissible risks to each segment of the task chain and form risk barriers. Perform joint search on the skill graph to generate an initial execution plan. When a disturbance event is detected, adjust adjacent skill nodes through a local compensation mechanism. When local adjustment alone cannot meet the task completion requirements or will break through the risk barriers, trigger global replanning. Write the planning decision, execution status, and triggering conditions into the audit log to form a closed-loop feedback and update the scenario state model and skill confidence.

2. The method for autonomous planning of complex robot tasks based on adaptive learning according to claim 1, characterized in that: The specific steps of scene perception and task modeling are as follows: In each control cycle, the robot obtains spatial topology and operational status information from the hospital ward scene status data source. Using the local controller clock as a unified time reference, it aligns the acquisition timestamps carried by each data source, backfills observations earlier than the current cycle with the most recent valid value, and writes invalid status for observations later than the current cycle or missing observations and records the failure reason and source identifier to form a scene status record. The received task request is parsed into a task chain object according to preset fields, and each stage is bound to the corresponding spatial node and the allowed passage area. Constraint sets are generated item by item based on the task chain object. Area compliance constraints are generated according to the restricted area and disinfection rules, and the reachable area boundary is restricted. Passage constraints are generated according to the corridor passage width and crowd density, and the passage segment status is marked. Elevator constraints are generated according to the elevator availability status, and the set of available elevators and the reasons for unavailability are given.

3. The method for autonomous planning of complex robot tasks based on adaptive learning according to claim 2, characterized in that: The specific steps of scene perception and task modeling also include: The robot performs field-by-field parsing of the message, generates a stage sequence in a fixed order and writes it into the task chain object. When any field is missing, point mapping fails, cross-layer requirements are inconsistent with the point floor, or timeliness requirements are invalid, the robot writes the task chain object into an unexecutable state and records the failure reason code and the name of the missing field. At the same time, it returns a resend prompt to the scheduler, and obtains a task chain object that can be directly used for subsequent constraint generation and planning execution. Extract sample sequences from the population density samples within the historical sliding window according to a fixed window length, remove invalid states, sort them from smallest to largest, take the density value corresponding to the preset quantile as the risk threshold, write the risk threshold into the scene status and associate it with the road segment status.

4. The method for autonomous planning of complex robot tasks based on adaptive learning according to claim 1, characterized in that: The specific steps for risk upper bound modeling and constraint generation are as follows: After acquiring and aligning the hospital ward scene status data in each control cycle, the robot extracts safety event records related to road passage from the audit log. It then categorizes these records into buckets based on passage width range, crowd density range, and passage mode. For each bucket, it counts the total number of passages and the number of collision events. Under a given confidence level, it calculates a statistical upper bound for the collision incidence rate of each bucket and uses this upper bound as the collision risk upper bound field for that bucket. This upper bound, along with the confidence flag and sample size, is written into the risk upper bound mapping table. When the sample size of a bucket is less than the minimum limit, the bucket is marked as low confidence, and the maximum collision risk upper bound of all density buckets within the same passage width range is used as a conservative estimate. If the sample size of all buckets within that width range is insufficient, the global maximum collision risk upper bound is used and marked as extremely low confidence.

5. The method for autonomous planning of complex robot tasks based on adaptive learning according to claim 1, characterized in that: The specific steps for skill graph construction and confidence management are as follows: After obtaining the task chain object, the stage sequence is discretized and mapped according to the preset skill library; the robot sequentially matches the corresponding skill for each stage in the task chain object and generates a skill node, and writes three types of reproducible fields for each skill node: the entry condition field is used to limit the pre-state of triggering the node, the completion criterion field is used to determine the closed loop completion of the node, and the failure compensation action field is used to prioritize recovery under disturbance without rewriting the overall plan. Pre-set compensation branches for elevator-related nodes and corridor-related nodes; Skill nodes are connected into a skill graph based on the reachability relationships in the scene state. The reachability relationship is based on the condition that the endpoint of a node can reach the starting point of the next node without violating regional compliance constraints, access constraints, and elevator constraints. An execution record entry is created for each skill node during the execution process.

6. The method for autonomous planning of complex robot tasks based on adaptive learning according to claim 5, characterized in that: The specific steps for skill graph construction and confidence management also include: After completing the skill graph construction, a connectivity check and anomaly handling loop are executed. First, using the area compliance constraints, access constraints, elevator constraints, and risk barriers in the current scenario as hard constraints, edge and node pruning is performed on the skill graph. Nodes that violate restricted areas and disinfection rules are written to the unavailable state, edges with the corresponding road segment status as prohibited are deleted, elevator edges with malfunctions or emergency occupancy are deleted, and nodes with a confidence state of disabled are removed from the candidate set. The starting node set and the ending node set are determined, and a reachability traversal is performed on the pruned directed skill graph with the starting node set as the source. For each visited node... Write the reachable field to 1 and record its predecessor node identifier; if the reachable field of any terminating node is 1 after the traversal, the skill graph connectivity check is considered passed, and the predecessor chain is retained for subsequent plan backtracking; if the reachable field of all terminating nodes is 0, it is considered disconnected and enters the exception handling process, and the graph repair is performed according to the preset repair order. After each repair is completed, the connectivity traversal is re-executed and the reachable field is updated; if the connection is still not connected after the repair, the task chain object is written to the non-executable state and a safety degradation action is output. The reliability of nodes is updated based on execution records to form confidence information. The confidence information is written into three-state identifiers: trusted, conservative, and disabled. When a node fails continuously for the same reason, repeatedly times out in high-risk sections, or has insufficient coverage of key constraint labels, it is placed in a conservative state. When a node is in a conservative state, the set of optional skills is restricted to a subset of nodes that simultaneously meet the constraints and have failure compensation actions during the planning and local adjustment phases, so that recoverable nodes are used first to complete the closed loop when disturbances occur.

7. A method for autonomous planning of complex robot tasks based on adaptive learning according to any one of claims 4, 5, and 6, characterized in that: The specific steps for introducing risk upper bound consistency gating are as follows: When mapping the phase sequence of a task chain object to skill nodes, the robot adds a risk upper bound constraint field and a strategy mode field to each skill node. The risk upper bound constraint field is jointly determined by the road segment collision risk upper bound field and the allowable risk upper bound field of the task chain. For skill nodes that are related to close interaction with people, the robot reads the corresponding collision risk upper bound field from the scene state according to the road segment or area covered by the node, and compares it with the allowable risk upper bound field of the task chain to calculate the risk margin. If the risk margin is negative, the skill node is directly marked as unselectable and only alternative nodes with failure compensation actions are retained. The strategy mode field includes two categories: normal mode and efficiency mode. An executable set of control constraints is written for each mode. The robot only allows the skill node to switch from normal mode to efficiency mode when the risk margin is greater than the preset threshold. The current active mode identifier and switching trigger condition are written in the strategy mode field. When the risk margin is in the preset low margin range, the node is forced to keep running in normal mode to leave sufficient safety margin. When establishing directed edges in a skill graph based on reachability, first check whether the risk upper bound constraint fields of the start and end nodes meet the connection conditions, i.e., the collision risk upper bounds of the two nodes are not higher than the task's allowed risk upper bound. If the conditions are met, establish the directed edge and write the risk upper bound constraint field, strategy mode field, and the current activation mode of the start and end nodes into the edge's constraint label. If the conditions are not met, do not establish the edge. In the alternative edge set, alternative paths are recorded in the order of greater risk margin, lower proportion of restricted passage sections, and higher confidence status of skill nodes, and trigger switching conditions are written for each alternative edge. During execution, a near miss event count and a collision event identifier are added to the execution record entry of each skill node. When updating the node confidence information based on the execution record, the appearance of the collision event identifier or the near miss event count reaching a preset number within a fixed window is used as a strong trigger condition for the safety side. Once triggered, the confidence status of the node in efficiency mode is disabled and forced to fall back to normal mode. Only a subset of nodes with positive risk margin and failure compensation actions are allowed to be selected.

8. The method for autonomous planning of complex robot tasks based on adaptive learning according to claim 1, characterized in that: The specific steps of the anti-jitter planning under the aforementioned risk constraints are as follows: After obtaining the updated scene status and skill map, the passable area is divided into several road segments and risk barriers are generated for the road network path of the task chain. The robot first writes the allowed risk level and allowed passage mode for each road segment according to the road segment status identifier, risk threshold and task chain timeliness requirements in the scene status. The allowed risk level is written as the planning gating condition into the constraint label of the edge of the skill map. A joint search is performed on the skill graph to simultaneously determine the task order and the path. The joint search is constrained by the task chain stage sequence as the main order and by the reachability relationship of skill nodes as the edge constraint. Nodes with failure compensation actions and a trustworthy state are given priority. When there are multiple feasible edges, the path with a lower proportion of restricted passage segments and that does not depend on the unavailable elevator state is given priority. An initial execution plan is generated and issued for execution.

9. The method for autonomous planning of complex robot tasks based on adaptive learning according to claim 8, characterized in that: The specific steps of the anti-jitter planning under the aforementioned risk constraints are as follows: During execution, the robot continuously monitors for disturbance events. When a disturbance event is detected, the robot does not immediately rewrite the global plan, but instead enters a local compensation mechanism. The adjustment range is the adjacent subgraph formed by the current skill node and its successor nodes. The robot first executes the preset failure compensation action of the current node and attempts to restore to the next node of the original plan. If the compensation action can be completed without violating regional compliance constraints and risk barriers, the original planned structure remains unchanged. Only the entry conditions and edge constraint labels of adjacent nodes are updated to reflect the new road segment status or elevator status. When global replanning is triggered, the joint search is re-executed on the skill graph, and the completed node sequence is kept from being rolled back. The robot writes each planning and adjustment process into the audit log and backfills the execution results to update the road segment status model and skill node confidence status in the scene status, forming a closed-loop feedback.

10. The method for autonomous planning of complex robot tasks based on adaptive learning according to claim 9, characterized in that: The specific steps for performing local adjustments in a fixed order when a disturbance event is detected are as follows: Prioritize downgrading the traffic mode of affected road sections from efficiency mode to normal or restricted traffic mode, and update the entry conditions and edge constraint labels of adjacent skill nodes; If, after downgrading, there are still necessary road segments where the upper limit of the collision risk field is higher than the upper limit of the allowable risk field, then the set of alternative edges will be switched or the preset failure compensation action will be invoked in the adjacent subgraph formed by the current skill node and its successor nodes to bypass the high-risk road segments; global replanning will only be triggered when the next handover point cannot be reached after local adjustment, the handover completion criterion cannot be met, or all remaining reachable paths will exceed the upper limit of the allowable risk field.

Images