Task planning and dynamic obstacle avoidance method and device, equipment and medium

Abstract

The application relates to the technical field of intelligent decision-making, can be applied to business scenarios such as smart home, financial technology and medical health, and discloses a task planning and dynamic obstacle avoidance method, device, equipment and medium, which comprises the following steps: collecting environment data and generating an environment state set, analyzing user task instructions to obtain a task priority, identifying dynamic obstacles and establishing a motion trajectory prediction model, predicting future motion trajectories and judging collision risks, constructing a multi-constraint task planning model to generate an optimal execution path, controlling a robot to move along the optimal execution path and execute a task, adjusting an obstacle avoidance strategy according to risk changes in the movement, collecting execution process data and user feedback to update model parameters. Through real-time environment perception, dynamic risk judgment and execution feedback driven model updating, the application improves the obstacle avoidance accuracy and task execution efficiency of the robot in a complex environment, and realizes adaptive optimization capability.

Description

Technical Field

[0001] This invention relates to the field of intelligent decision-making technology, and in particular to a task planning and dynamic obstacle avoidance method, apparatus, device and medium. Background Technology

[0002] With the popularization of artificial intelligence and autonomous mobility technology, intelligent devices with path planning and autonomous obstacle avoidance capabilities are gradually entering home life, medical care scenarios, and fintech operations. Service robots are no longer limited to a single environment but are widely deployed to meet the needs of tasks such as cleaning, delivery, companionship, and business assistance. However, existing autonomous mobile devices operating in complex real-world scenarios still exhibit many bottlenecks, limiting their large-scale deployment and user experience.

[0003] In the smart home field, existing autonomous mobile robots typically rely on a single type of environmental perception sensor for obstacle detection, such as visual sensors or ultrasonic radar alone. This single-channel perception method is unstable in situations with unstable lighting, diverse obstacle types, or cluttered scenes, making it difficult to accurately identify low, transparent, or small objects. This can lead to the robot misjudging its path or colliding with furniture and clutter. Furthermore, existing obstacle avoidance technologies are mostly based on static environmental cognition, relying solely on current perception data for immediate responses. They cannot effectively identify dynamic obstacles (such as fast-moving children or pets), and are prone to delayed avoidance or path jitter when faced with suddenly appearing moving objects, thus reducing the continuity and success rate of task execution.

[0004] In the healthcare sector, mobile service devices such as ward delivery robots and bedside nursing robots need to operate continuously in environments with frequent movement of wheelchairs, IV stands, and healthcare workers. Existing autonomous walking devices lack the ability to accurately predict dynamic crowd behavior, often resulting in them stopping or detouring in the opposite direction, occupying space resources and even interfering with the normal work of healthcare workers. Furthermore, these systems often employ fixed task planning logic, failing to incorporate resource constraints in healthcare scenarios into their decision-making, such as insufficient battery power, rapid consumption of transport consumables, or changes in temporarily closed areas, leading to frequent task interruptions or even failure to complete tasks on time. The lack of learning and iterative capabilities during task execution makes it difficult for these devices to adapt to changes in different departments, patient density, or time periods.

[0005] In the fintech sector, smart mobile devices are increasingly being used for self-service guidance, document distribution, and transaction assistance at financial branches. However, due to the intermittent flow of people and the easily changing spatial layout of financial branches, devices relying on single-sensor operation often fail to recognize moving customers or temporary display shelves, lacking dynamic predictive capabilities and resulting in low task execution efficiency. Furthermore, existing systems are rigid in handling the order and priority of task instructions, failing to dynamically adjust execution plans based on factors such as work area congestion, remaining battery power, and path obstruction, exacerbating congestion issues for robots during peak business hours. In addition, current financial devices generally lack user feedback-driven adaptive capabilities, unable to adjust their operating behavior or execution paths according to the preferences of branch staff, leading to a fragmented user experience. Summary of the Invention

[0006] The main objective of this invention is to provide a task planning and dynamic obstacle avoidance method, apparatus, device, and storage medium, which aims to solve the technical problems of existing autonomous mobile devices being unable to simultaneously achieve accurate environmental understanding, dynamic obstacle prediction and avoidance, and real-time task planning that combines resource and task priority constraints in complex dynamic scenarios, resulting in low execution efficiency, untimely obstacle avoidance, and a lack of adaptive optimization capabilities.

[0007] To achieve the above objectives, the present invention provides a task planning and dynamic obstacle avoidance method, comprising: Environmental data is collected by multimodal sensors mounted on the robot. The environmental data is preprocessed and multi-source information is fused to generate an environmental state set that includes obstacle information, robot's own state information and scene environment feature information. Obtain user task instructions and parse the user task instructions to determine task priority; Based on the set of environmental states, dynamic obstacles are identified, and corresponding motion trajectory prediction models are established for the dynamic obstacles. The motion data of the dynamic obstacle is input into the motion trajectory prediction model to generate a predicted motion trajectory for a future time period. The predicted motion trajectory is then spatially compared with the planned path determined based on the user task instruction to determine the collision risk level. A multi-constraint task planning model is constructed based on the task priority, the set of environmental states, and the collision risk level. The optimal execution path is generated by solving the multi-constraint task planning model. The robot is controlled to move along the optimal execution path and perform task operations. During the movement, the obstacle avoidance strategy is dynamically adjusted according to the real-time updated collision risk level to perform obstacle avoidance operations. Collect operational status data and user interaction feedback data during task execution, and update the parameters of the multi-constraint task planning model and the motion trajectory prediction model based on the operational status data and the user interaction feedback data.

[0008] Furthermore, to achieve the above objectives, the present invention provides a task planning and dynamic obstacle avoidance device, comprising: The environmental perception module is used to collect environmental data through multimodal sensors mounted on the robot, preprocess the environmental data and fuse multi-source information to generate an environmental state set that includes obstacle information, robot's own state information and scene environment feature information. The task parsing module is used to obtain user task instructions, parse the user task instructions, and determine the task priority; An obstacle recognition module is used to identify dynamic obstacles based on the set of environmental states, and to establish a corresponding motion trajectory prediction model for the dynamic obstacles. The collision prediction module is used to input the motion data of the dynamic obstacle into the motion trajectory prediction model to generate a predicted motion trajectory in the future time period, and to spatially compare the predicted motion trajectory with the planned path determined based on the user task instruction to determine the collision risk level. The path planning module is used to construct a multi-constraint task planning model based on the task priority, the set of environmental states, and the collision risk level, and generate the optimal execution path by solving the multi-constraint task planning model. The motion execution module is used to control the robot to move along the optimal execution path and perform task operations. During the movement, the obstacle avoidance strategy is dynamically adjusted according to the real-time updated collision risk level to perform obstacle avoidance operations. The parameter update module is used to collect running status data and user interaction feedback data during task execution, and update the parameters of the multi-constraint task planning model and the motion trajectory prediction model based on the running status data and the user interaction feedback data.

[0009] Furthermore, to achieve the above objectives, the present invention also provides a computer device, the computer device including a memory, a processor, and a task planning and dynamic obstacle avoidance program stored in the memory and executable on the processor, wherein when the task planning and dynamic obstacle avoidance program is executed by the processor, it implements the steps of the task planning and dynamic obstacle avoidance method as described above.

[0010] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a task planning and dynamic obstacle avoidance program, wherein the task planning and dynamic obstacle avoidance program, when executed by a processor, implements the steps of the task planning and dynamic obstacle avoidance method described above.

[0011] Beneficial Effects: This invention relates to the field of intelligent decision-making technology and can be applied to business scenarios such as smart homes, fintech, and healthcare. It discloses a task planning and dynamic obstacle avoidance method, apparatus, device, and medium, including: collecting environmental data and generating an environmental state set; parsing user task instructions to obtain task priorities; identifying dynamic obstacles and establishing a motion trajectory prediction model; predicting future motion trajectories and assessing collision risks; constructing a multi-constraint task planning model to generate the optimal execution path; controlling the robot to move along the optimal execution path and execute the task; adjusting the obstacle avoidance strategy according to risk changes during movement; and collecting execution process data and user feedback to update model parameters. This invention, by integrating multi-source environmental information, dynamically predicting obstacle movement trends, constructing multi-constraint planning objectives, and continuously utilizing execution data feedback to optimize the system model, enables the robot to operate autonomously in dynamic environments, improving obstacle avoidance accuracy and task planning efficiency, and achieving continuous adaptation and performance enhancement. Attached Figure Description

[0012] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a schematic diagram of an application environment for the task planning and dynamic obstacle avoidance method in one embodiment of the present invention; Figure 2 This is a flowchart illustrating an embodiment of the task planning and dynamic obstacle avoidance method of the present invention; Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the task planning and dynamic obstacle avoidance device of the present invention; Figure 4 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 5 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0013] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0014] The task planning and dynamic obstacle avoidance method provided in this invention can be applied to, for example... Figure 1In this application environment, the client communicates with the server via a network. The server can collect environmental data from the client and generate an environmental state set, parse user task instructions to obtain task priorities, identify dynamic obstacles and establish a motion trajectory prediction model, predict future motion trajectories and assess collision risks, construct a multi-constraint task planning model to generate the optimal execution path, control the robot to move along the optimal execution path and execute tasks, adjust obstacle avoidance strategies according to risk changes during movement, and collect execution process data and user feedback to update model parameters. This invention enables the robot to operate autonomously in dynamic environments by integrating multi-source environmental information, dynamically predicting obstacle movement trends, constructing multi-constraint planning objectives, and continuously using execution data feedback to optimize the system model, thereby improving obstacle avoidance accuracy and task planning efficiency, and achieving continuous adaptation and performance enhancement. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The following detailed description of specific embodiments further illustrates this invention.

[0015] Please see Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the task planning and dynamic obstacle avoidance method provided by the present invention. It should be noted that although a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0016] like Figure 2 As shown, the task planning and dynamic obstacle avoidance method proposed in this invention includes the following steps: S10: Collect environmental data through multimodal sensors mounted on the robot, preprocess the environmental data and fuse multi-source information to generate an environmental state set containing obstacle information, robot's own state information and scene environment feature information; In this embodiment, multimodal sensors are deployed near the robot's shell, chassis, or robotic arm to simultaneously collect environmental data from different physical information sources, including visual imaging, obstacle distance feedback, changes in motion posture, and auxiliary environmental conditions. These sources can be cameras, ranging devices, inertial units, and simple environmental detection elements. Visual imaging acquires the edges, textures, and contours of objects within a region through optical imaging structures. The ranging device calculates the relative position of objects using reflected waves or time differences. The inertial unit records motion speed and tilt angle changes to infer the current driving state, and auxiliary sensing elements perceive light intensity or ground friction.

[0017] Data from different sources may contain noise, errors, or temporal inconsistencies at the moment of acquisition, thus requiring sequential denoising, correction, and baseline alignment. Image data undergoes convolutional filtering to remove high-frequency noise, ranging data is filtered to remove anomalous amplitude jumps through statistical methods, and motion attitude data is buffered against instantaneous errors using dynamic prediction update formulas. The processed data is mapped to a unified coordinate system and timeline to eliminate calibration differences between different sensors, enabling various data types to describe the real-world environment within the same spatial framework.

[0018] After completing spatiotemporal alignment, a weighted fusion strategy is used to accumulate and integrate various types of information in a confidence-based manner. When visual and distance data jointly identify an object region, the occupancy probability is increased; when pose changes are significant, the location confidence is reduced, gradually forming a stable description of the environment from multiple dimensions. The fused environmental state set contains structural information, including obstacle boundary attributes and object dimensions, the robot's current position and speed, battery and energy consumption status, and regional characteristic parameters formed by ground and lighting conditions, supporting the robot's understanding of its external operating space.

[0019] This embodiment combines multi-source sensor inputs and performs noise reduction, calibration, and weighted fusion processing to form an environmental state set, enabling the robot to obtain a realistic, continuous, and spatially consistent description of the operating environment, reducing misjudgments and perception blind spots, and laying a stable and reliable data foundation for its subsequent decision-making behavior.

[0020] S20, Obtain user task instructions and parse the user task instructions to determine task priority; In this embodiment, user task commands are input into the robot command receiving unit via voice pickup hardware, mobile application interface, or touch screen. The same input channel can accept natural language expressions, menu clicks, or button selections, all of which are essentially converted into character sequences or numerical codes. After the input information enters the parsing and processing stage, semantic fields are extracted through natural language processing operations such as word segmentation, intent extraction, and parameter recognition. These fields include task type, target location, and execution time fields. The target location field can be converted into spatial coordinates or room labels using a mapping function, the execution time field can be converted into a judgment condition for immediate or delayed execution, and the task type field corresponds to category tags such as cleaning, inspection, and delivery.

[0021] The parsed attribute elements are input into the priority calculation chain. Data is quantified across three dimensions: urgency, user preference, and task complexity. Urgency stems from the user's required time constraints or delay tolerance; user preference is derived from the weights of historical user actions or interaction parameters; and task complexity arises from the length of the action chain, the size of the target area, or the need for cross-area movement. These three dimensions are arranged and combined to form a judgment matrix, filled with the ratio of the relative importance of each dimension. After numerical processing to form a consistent matrix, a weight sequence is generated. The priority calculation module multiplies the weight sequence by the three-dimensional parameter vector to obtain a single priority value, which serves as the basis for task allocation.

[0022] The command receiving module can be activated through different input methods. For example, a microphone array can be used to process far-field speech, ensuring effective reach of call commands in the home space. Alternatively, a mobile application interface can be prioritized to reduce noise interference. The parsing process can employ a deep semantic recognition model to extract intent or use rule-based short text matching logic to adapt fixed command templates. Priority value generation can adjust conflicting requirements by modifying the judgment matrix; for example, increasing the weight of time requirements in elderly care environments or increasing the weight of user preferences in student dormitory environments. Task complexity can be extracted by reading the indoor map area and estimating based on the number of rooms, or by calculating the task list length.

[0023] This embodiment combines task instruction reception and parsing with priority value generation, enabling the robot to accurately understand the user's intent and digitally express the task execution order and importance, providing a clear decision-making basis for subsequent scheduling behavior and reducing the probability of erroneous execution and mis-scheduling of tasks.

[0024] S30, Identify dynamic obstacles based on the set of environmental states, and establish a corresponding motion trajectory prediction model for the dynamic obstacles; In this embodiment, after the environmental state set enters the obstacle recognition processing link, it contains spatial point information, depth data, and target contour data from visual, ranging, and inertial sensing imaging outputs. The recognition process first extracts the contours of entities marked as obstacles in the set, extracting entity boundaries through edge detection or region segmentation, and converting them into a three-dimensional coordinate sequence based on depth information. Then, it obtains velocity vectors based on the coordinate differences across consecutive time points. The magnitude of the velocity vector is used to determine whether the entity is stationary or moving; entities with velocities approaching zero are marked as static obstacles, while entities with velocities exceeding a threshold are marked as dynamic obstacles.

[0025] Dynamic obstacles are further categorized into attributes, and their motion patterns are inferred through features such as the rate of change of acceleration, the frequency of changes in direction, and the shape of the obstacle. Obstacles with stable motion direction and speed changes limited by the path are classified as predictable, such as an elderly person or a pet moving along a path; obstacles with significant directional jitter or prominent acceleration fluctuations are classified as unpredictable, such as a child suddenly running or an object that has rolled away.

[0026] Different trajectory prediction structures are assigned to different categories. For objects with strong regularity, a time-series model is used, predicting the expected position in one or more future time slices by inputting several consecutive sets of position and velocity data. For unpredictable objects, a probabilistic model is used, generating a probability range through velocity direction distribution and a noise function, covering the future region with probability density rather than limiting a unique trajectory. Each dynamic obstacle is bound to an instance model, stored in memory or a cache area, and its internal state parameters are replaced with each update to adapt to the latest motion trends.

[0027] Image segmentation networks can be used to extract boundaries from obstacle pixel regions, or frame-by-frame depth thresholding can be used to form candidate regions. Velocity calculation can employ simple difference or multi-frame fusion to reduce jitter errors. Motion category differentiation can be achieved using support vector machines based on feature vectors composed of velocity, acceleration, and contour shape, or clustering can be used to automatically form category divisions without annotation. The trajectory prediction model can choose a long short-term memory network to process smooth motion trajectories, or a Gaussian distribution-based motion dispersion to infer future location space. If computational resources are limited, a probabilistic model covering all dynamic obstacles can be used to reduce storage pressure; for finer control, different structural models can be configured for each category.

[0028] This embodiment identifies dynamic obstacles and assigns matching trajectory prediction models to different motion patterns, enabling the robot to perceive the potential location area of ​​a moving target in advance. This allows for a time buffer for subsequent spatial planning, helps reduce collision risks, and improves route continuity.

[0029] S40, input the motion data of the dynamic obstacle into the motion trajectory prediction model to generate a predicted motion trajectory for a future time period, and compare the predicted motion trajectory with the planned path determined based on the user task instruction to determine the collision risk level; In this embodiment, after a trajectory prediction structure has been established for each dynamic obstacle, the position coordinates, motion direction, and velocity amplitude obtained through continuous sampling are organized into a time series input. The input action is executed independently for each obstacle and written into the internal cache of its bound prediction model. The prediction model absorbs several historical frame data in a sliding window manner, uses the feature change trend to infer the motion trend over a future period, and generates the possible location regions in several future time slices as predicted motion trajectories. The predicted trajectory can be a single path segment structure, or it can contain multiple branch trajectories or probability coverage areas.

[0030] The predicted trajectory is simultaneously compared spatially with the planned path. This comparison is based on 3D coordinates and distance measurement. By sampling discrete points along the planned path segment by segment or parametrically scanning according to the trajectory curve, the shortest possible spatial distance between the potential future location of a dynamic obstacle and the robot's potential path is calculated. This distance serves as a spatial proximity indicator. The spatial distance is further combined with additional factors such as the robot's future arrival time estimate, the consistency of the obstacle's movement direction, and safe reaction time to convert it into a collision risk value. This risk value is then processed through threshold segmentation, grading, or continuous mapping to convert it into a collision risk level, representing the severity of potential interference.

[0031] Motion data input can be pushed frame-by-frame or batch-updated to accelerate inference efficiency. Predicted trajectory generation can utilize point prediction based on time-series networks or generate multiple candidate segments spatially using probabilistic models. In the spatial comparison phase, Euclidean distance can be used to directly assess the nearest point, or Manhattan distance or Chebyshev distance can be introduced to adapt to different environmental layouts. Risk level determination can employ fixed thresholds or dynamically adjust the distinction range based on environmental noise and scene congestion to adapt to different home layouts and dynamic load conditions.

[0032] This embodiment provides real-time motion data to the trajectory prediction model, generates future motion trends, and performs spatial distance quantification with the planned route. This enables the understanding of potential conflict trends before contact, transforming sudden obstacles into controllable factors in advance, reducing the probability of collisions and improving path continuity.

[0033] S50, construct a multi-constraint task planning model based on the task priority, the set of environmental states and the collision risk level, and generate the optimal execution path by solving the multi-constraint task planning model; In this embodiment, task priority, environmental state set, and collision risk level serve as inputs to the planning decision and are transformed into mathematical constraints and objective function elements. Task priority is used to distinguish between multiple tasks or different execution urgency levels of the same task, influencing path selection bias through weighting. The environmental state set, carrying obstacle locations, robot resource status, and scene category, is broken down into planning factors such as static obstacle restrictions, dynamic obstacle avoidance range, ground passage cost, and resource capacity boundaries. Collision risk level is incorporated into the path feasible region in the form of a dynamic risk boundary, marking high-risk areas as prohibited areas, setting medium-risk areas as high-cost areas, and treating low-risk areas according to the normal path cost.

[0034] The multi-constraint task planning model integrates the above factors to form a constraint system, including path traversability set constraints, obstacle safety distance constraints, energy and material capacity limits, and execution time limits. The model also constructs a cost function, using path length, number of turns, execution time consumption, and task bias resulting from priority as optimization indicators, and assigning weights to each indicator. The solution process searches for candidate trajectories that satisfy all constraints from the starting position to the target position, defining the trajectory with the minimum cost as the optimal execution path. The entire process establishes a deterministic mapping relationship between input factors and output paths, enabling scenario adaptability in path planning.

[0035] Constraint construction can employ hard constraints to directly eliminate impassable areas, or soft constraints to add cost value to high-risk areas. The weights of the cost function indicators can be fixed or adjusted in real-time according to the task level. The solution process can use a discrete search method based on heuristic grid search, or a continuous spatial planning method based on graph search, such as introducing smooth interpolation to generate curved trajectories. If the environment updates frequently, the planning process can use a rolling window to iterate and solve repeatedly to maintain path effectiveness.

[0036] This embodiment maps task priority, scenario information, and dynamic risks into an executable path constraint system, and obtains the feasible trajectory with the minimum cost by solving the problem. This makes the driving route more in line with the task requirements and real-time environmental conditions, reduces invalid movement and path conflicts, and improves the continuity and efficiency of task completion.

[0037] S60, control the robot to move along the optimal execution path and perform task operations, and dynamically adjust the obstacle avoidance strategy according to the real-time updated collision risk level to perform obstacle avoidance operations during the movement; In this embodiment, controlling the robot to move along the optimal execution path and perform task actions requires converting discrete position points or continuous trajectories along the path into speed and direction commands acceptable to the chassis. The desired displacement is obtained from the path points using a differential method, and combined with the steering angle and travel speed to form a control quantity, which is then distributed by the motion controller to the drive wheels, motors, or tracks, thereby generating externally observable displacement. Simultaneously, the task toolchain, such as the vacuuming assembly, mop assembly, delivery and transport mechanism, or gripping mechanism, is triggered according to the task commands, enabling movement and task processing to proceed synchronously.

[0038] During movement, the system continuously receives updated collision risk level results. The risk level is treated as a dynamic input over time, sharing the same control cycle as the trajectory tracking control. Risk levels trigger adjustments to the action strategy, including reducing speed, increasing the turning radius, increasing the spatial distance from obstacles, or temporarily halting propulsion. The controller maps the risk level to parameter adjustment signals, appropriately altering the controller's internal speed limit, acceleration limit, and feasible path range, thereby instantly changing the motion performance. Obstacle avoidance actions are executed locally near the current path, maintaining macroscopic route adherence while avoiding immediate collisions, achieving a balance between tracking control and dynamic obstacle avoidance.

[0039] Path tracking can be based on closed-loop control, using differential feedback to constrain the robot's deviation from the trajectory, or it can use model predictive control to anticipate multiple future action stages and select the control sequence with the smallest deviation. Risk level interpretation can employ threshold switching or continuous mapping to convert risk values ​​into linear or nonlinear adjustment factors for scaling speed or adjusting the allowable contact radius. Obstacle avoidance adjustments can use speed suppression, local path perturbation to generate a new steering vector near the original trajectory, or emergency pause to maintain stillness until the risk is eliminated.

[0040] This embodiment integrates path tracking instructions and risk guidance adjustments into a unified motion control process, enabling the robot to maintain task execution continuity while responding with low latency and avoiding environmental interference, reducing the possibility of collisions, and improving movement stability and task completion success rate.

[0041] S70, collect running status data and user interaction feedback data during task execution, and update the parameters of the multi-constraint task planning model and the motion trajectory prediction model based on the running status data and the user interaction feedback data.

[0042] In this embodiment, the operational status data collected during task execution refers to a set of data that reflects the execution performance during actual action, including metrics such as task time, path tracking offset, obstacle avoidance trigger frequency, and power or consumable consumption ratio. This type of data originates from the robot's onboard sensing unit and control module, which continuously records movement command execution errors, energy changes, and speed changes.

[0043] User interaction feedback data originates from users' proactively expressed opinions or experience evaluations, triggered by voice input, mobile interface clicks, or panel buttons, used to express satisfaction, perceived performance deviations, or personal preferences. Operational status data and user interaction feedback data are then linked through timestamps, spatial markers, task numbers, or path segment identifiers, enabling subjective feelings and objective measurements to be mapped to the same task segment for coordinated judgment.

[0044] The above data together form the basis for measuring the quality of the execution results. By using a preset evaluation method, the direction of updating the model parameters is derived, which in turn affects the variables that influence the planning output in the multi-constraint task planning model, such as resource constraint weights, obstacle avoidance distance parameters, or path selection bias values. It also affects the internal parameters in the motion trajectory prediction model that are related to historical trajectory fitting or speed change trends, so that the model has new adaptability before the next task run.

[0045] Operational status data can be periodically output by the chassis controller or indirectly calculated by sensor nodes, such as determining deviation by accumulating yaw angle and mileage difference. User interaction feedback data can be collected using a fixed evaluation scale or obtained numerically through sentiment analysis of free text. Parameter updates can directly replace old values, use incremental correction to approximate the new parameter range at a fixed ratio, or use decay updates to assign higher weights to recent tasks. Adjustable weights in multi-constraint task planning models can be path length preference, execution time preference, or resource conservation preference. Updated objects in motion trajectory prediction models can be the length of the time window, sensitivity to speed changes, or the range of historical sample selection.

[0046] This embodiment introduces two-dimensional feedback—objective performance and user experience—and maps it to changes in model parameters. The robot gradually improves its ability to adapt to changes in the home environment and individual preferences, and in subsequent executions, it can generate paths and obstacle avoidance behaviors that are more in line with user expectations and more resource-efficient.

[0047] In one embodiment, step S10 above includes: S101 collects environmental image data, obstacle distance data, robot motion posture data, and environmental auxiliary parameter data through vision sensors, ranging sensors, inertial sensors, and environmental auxiliary sensors to obtain environmental data; S102, Gaussian filtering is applied to the environmental image data in the environmental data, median filtering is applied to the obstacle distance data in the environmental data, and Kalman filtering is applied to the robot motion posture data in the environmental data to obtain denoised environmental data. S103, based on the preset environmental map model, the denoised environmental data is mapped to a unified spatial coordinate system and time reference for spatiotemporal calibration to obtain preprocessed environmental data; S104, Based on evidence theory fusion processing, the preprocessed environmental data is fused from multiple sources to generate an environmental state set containing obstacle information, robot's own state information and scene environment feature information.

[0048] In this embodiment, multimodal sensors are mounted on a fixed position on the robot's body or on a gimbal structure to continuously collect environmental data during movement. Multimodal refers to at least one visual sensor, at least one ranging sensor, at least one inertial sensor, and at least one environmental auxiliary sensor. The visual sensor can be a 2D camera, a depth camera, or a structured light device, used to generate environmental image data. The image data includes a capture timestamp, lens intrinsic parameters, and basic exposure parameters during acquisition. The ranging sensor can be a laser ranging unit, an ultrasonic ranging unit, or a millimeter-wave radar, acquiring obstacle distance data through scanning or periodic emission. The distance data is indexed by angle, direction, or emission channel. The inertial sensor includes a combination of an accelerometer and a gyroscope, outputting the robot's linear acceleration and angular velocity in real time. Time integration allows for the derivation of the robot's motion posture data. The environmental auxiliary sensor can include a light intensity sensor, a temperature and humidity sensor, or a ground material recognition unit, used to provide environmental auxiliary parameter data to assist in determining lighting conditions, humidity changes, and ground friction characteristics.

[0049] After environmental data acquisition, the computing unit performs channel-specific denoising on data from different sources according to sensor type. Environmental image data is processed using Gaussian filtering, which smooths local pixel intensity through convolutional kernels to suppress random noise during imaging, preserving obstacle contours while controlling edge blur. Obstacle distance data is processed using median filtering; multiple distance values ​​within the same ranging direction or time window are sorted by magnitude, and the median value is selected as the distance output to suppress abnormal spikes introduced by single mismeasurements. Robot motion posture data is processed using Kalman filtering, combining inertial measurement and wheel speed feedback to represent the system state as attitude and velocity vectors. Prediction and correction steps reduce integral errors and sensor random drift. The resulting filtered environmental data is aligned to a unified timeline and meets the accuracy requirements for subsequent fusion in terms of noise levels.

[0050] In the spatiotemporal calibration phase, a unified spatial coordinate system and time reference are first established based on a pre-defined environmental map model. The environmental map model can originate from offline-constructed floor plans, 3D grid maps, or topological maps, internally defining room boundaries, main passageways, and reference positions for fixed furniture. Image coordinates, ranging directions, and inertial attitudes in the denoised environmental data are transformed within this coordinate system. Image pixels are converted into 3D ray or point cloud coordinates using camera intrinsic and extrinsic parameters. Ranging data is combined with corresponding direction vectors to obtain the spatial location of obstacles, and inertial attitude is combined with chassis pose to obtain the robot's position and orientation on the map. The time reference is managed by a unified clock. All data frames are mapped to a unified time scale through timestamp interpolation or extrapolation, ensuring that multi-source information at the same moment corresponds to the same environmental state, thus forming preprocessed environmental data.

[0051] In the multi-source information fusion stage, evidence theory is used to map preprocessed environmental data from the visual, ranging, inertial, and environmental auxiliary channels into elements within the same environmental state set. Obstacle information is aggregated into individual obstacle entities within the spatial neighborhood by matching visual detection results with ranging spatial points. The obstacle type is determined jointly by shape features, size estimation, and historical identification tags. The position coordinates are obtained through weighted multi-source position estimation, and the size is derived from multi-view contour envelopes or ranging distribution ranges. The robot's own state information is obtained by fusing inertial attitude, chassis encoders, and map reference points to obtain its position coordinates and motion speed in the map coordinate system. Simultaneously, the remaining battery power and cleaning fluid balance are read from the power management module and consumable monitoring module and marked with these values ​​to the current state. Scene environment feature information is obtained by calculating ground type and light intensity based on the brightness distribution of the environmental image and environmental auxiliary parameters. Combined with the map structure and obstacle distribution, the spatial topology is reconstructed, such as the connectivity of corridors, narrow passages, and open areas. Evidence theory assigns confidence levels to different data sources during the fusion process. When a sensor observation conflicts with other sensors, the impact of unreliable data is reduced through conflict allocation rules, making the environmental state set more stable and robust.

[0052] This embodiment integrates multimodal acquisition, filtering and denoising for different data types, spatiotemporal calibration based on a unified map model, and multi-source fusion based on evidence theory into a unified processing flow. This allows the robot to obtain a structured set of environmental states in complex scenarios. The set of environmental states includes not only geometric information such as position and distance, but also contextual information such as power, consumables, ground type, and spatial topology. Compared with direct reading data from a single sensor, this is closer to the real scene, thus providing more reliable and complete input conditions for subsequent dynamic obstacle recognition and task planning.

[0053] In one embodiment, step S20 above includes: S201, activate the robot's voice interaction module, mobile terminal application interface or touch interaction interface, and receive user task instructions through the voice interaction module, the mobile terminal application interface or the touch interaction interface. S202, use natural language processing or instruction protocol parsing to parse the user task instruction, and extract the task type, task target location and task time requirement from the user task instruction; S203, based on the task type, the target location of the task, and the time requirement of the task, determine the urgency, user preference, and task complexity; S204, construct a judgment matrix for the urgency, user preference and task complexity using the analytic hierarchy process, perform weighted processing on the urgency, user preference and task complexity based on the judgment matrix, and determine the quantified value as the task priority based on the weighted processing result.

[0054] In this embodiment, user task instructions are input into the robot control system through a human-machine interaction channel, which includes a voice interaction module, a mobile terminal application interface, and a touch interface. The voice interaction module, which may consist of a microphone array, a voice acquisition front-end, and a voice recognition engine, is used to acquire user voice signals, perform noise suppression and acoustic feature extraction, and transcribe the voice content into text-based task instructions. The mobile terminal application interface is deployed in a smartphone or tablet application and communicates with the robot control system via a network connection. When the user selects a task type, marks a target location, or sets an execution time in the application, the interface packages and sends structured instruction data. The touch interface is typically integrated into the robot's display screen, receiving user operations through components such as buttons, lists, and map point selection, and internally encapsulating them into a unified task instruction structure. All these various interaction methods are unified into a single task instruction data structure at the input layer for subsequent parsing and processing.

[0055] The task instruction parsing stage converts raw instructions into structured parameters through natural language processing or instruction protocol parsing mechanisms. For voice or text-based task instructions, processing techniques such as word segmentation, part-of-speech tagging, named entity recognition, and intent recognition are employed to identify task type words such as cleaning, delivery, and inspection, as well as target location information such as room names, geographic location markers, and preset point identifiers. Time requirements such as immediate execution, scheduled execution, or completion within a specific time window are also parsed. For structured instructions generated by mobile terminals or touch interfaces, the parsing process reads predefined fields and directly maps the task type field, target location field, and time field to internal parameters, eliminating the need for complex language analysis. After parsing, the task type describes the business category the robot needs to perform, the target location describes the spatial location the robot needs to reach or cover, and the time requirement describes the start time, end time, or duration constraints.

[0056] Urgency, user preference, and task complexity are derived based on task type, target location, and time requirements. Urgency is calculated based on whether the task has a hard deadline, whether it is related to safety or health, and the urgency of the time requirement relative to the current time. For example, patrol tasks close to the current time or tasks involving the delivery of medical supplies are assigned a higher urgency level. User preference is determined by combining historical task records and explicit preferences in the current instructions. For example, users who frequently set living room cleaning to be performed early or frequently choose low-noise mode for bedroom cleaning can be assigned a higher preference weight for the corresponding room or time period. Task complexity is evaluated based on target location distribution, path length, expected obstacle density, and the number of operational steps inherent in the task type. For example, whole-house cleaning tasks covering multiple rooms are more complex than single-point delivery tasks, and tasks traversing narrow corridors or densely furnished areas are more complex than tasks performed in open spaces.

[0057] To integrate urgency, user preference, and task complexity into a single priority value, an analytic hierarchy process (AHP) is introduced to construct a judgment matrix. First, the relative importance of the three factors is defined at the weight level. Pairwise comparisons of urgency relative to user preference and task complexity are established using expert rules or strategy configurations. For example, a set of pairwise comparison values ​​can characterize relationships such as urgency being more important than user preference, or urgency and task complexity being similar. After establishing the judgment matrix, weight coefficients for the three factors are generated using eigenvector generation or normalization. A stable weight distribution is obtained after consistency checks. Subsequently, the urgency score, user preference score, and task complexity score of a specific task instance are normalized to a unified scale. These scores are then weighted and summed with the weight coefficients, or subjected to other linear combinations, to calculate a continuous value as the task priority. This value can be mapped to a discrete level interval for scheduling and ranking, or it can be directly used as the priority weight for the corresponding task in the optimization model for subsequent path planning and resource allocation.

[0058] This embodiment receives task instructions uniformly across multiple interaction channels and structures the task type, target location, and time requirements through natural language processing or field parsing mechanisms. It then constructs a judgment matrix by combining urgency, user preferences, and task complexity, and calculates quantified priorities. The user's intent is transformed into a numerical expression usable for computational scheduling. This hierarchical parsing and weighted processing, from semantic tasks to priority values, transforms task importance assessment from a single rule to a multi-factor comprehensive approach. In complex home scenarios, this allows for a more rational arrangement of robot execution order, prioritizing time-sensitive or user-focused tasks.

[0059] In one embodiment, step S30 above includes: S301, parse the obstacle information in the environmental state set, and extract the outline feature data and real-time motion speed data of the obstacles; S302, the contour feature data and the real-time motion speed data are processed using support vector machine classification logic, and obstacles with zero speed are marked as static obstacles, and obstacles with speed greater than zero are identified as dynamic obstacles. S303, Analyze the interaction attributes of the dynamic obstacles, mark obstacles with interactive avoidance characteristics as controllable dynamic obstacles, and mark obstacles with random motion characteristics as uncontrollable dynamic obstacles; S304, For the controllable dynamic obstacle, a long short-term memory network model is loaded to establish a corresponding motion trajectory prediction model; for the uncontrollable dynamic obstacle, a probability distribution model based on velocity vector is loaded to establish a corresponding motion trajectory prediction model.

[0060] In this embodiment, the obstacle information contained in the environmental state set has undergone coordinate alignment and time synchronization before entering the recognition stage. Based on this, the data record corresponding to each obstacle is first parsed, converting image segmentation results, depth information, or point cloud contours into contour feature data, and converting continuous positional changes into real-time motion velocity data. Contour feature data may include bounding box size, aspect ratio, contour area, edge direction distribution, and the number of cells occupied in the grid, used to characterize the obstacle's geometry and spatial occupancy. Real-time motion velocity data is obtained by subtracting the position coordinates of adjacent frames under a unified time reference, yielding the velocity magnitude and direction of motion in planar or three-dimensional space. A velocity change rate can also be added to characterize acceleration or deceleration trends. In this way, each obstacle is mapped to a feature vector composed of contour feature data and real-time motion velocity data, facilitating subsequent classification and modeling processing.

[0061] The Support Vector Machine (SVM) classification logic uses the aforementioned feature vectors to distinguish between static and dynamic obstacles. During training, this logic learns the decision boundary using labeled samples. During runtime, it receives contour feature data and real-time motion velocity data. Objects with near-zero velocity and whose displacement fluctuations fall within the noise range across multiple frames are classified as stationary, while objects with greater than zero velocity and a stable motion trend over time are classified as moving. To improve robustness, velocity confidence and observation frame thresholds can be introduced during classification to avoid misclassifying stationary furniture as moving objects due to instantaneous detection errors. The classification output is written back to the environmental state set, explicitly distinguishing static and dynamic obstacles in the data structure and providing filtering conditions for subsequent trajectory prediction targeting only dynamic obstacles.

[0062] After obtaining the set of dynamic obstacles, it is necessary to further analyze the interaction attributes of each object, subdividing the dynamic obstacles into controllable and uncontrollable dynamic obstacles. The determination of interaction attributes can be inferred by combining target recognition results, tag information, and historical avoidance records. For example, the recognition model can determine whether the target belongs to an adult, child, pet, mobile appliance, or indoor mobile platform, and statistically analyze whether the object tends to actively avoid the robot or moves along a fixed path during past operations. Objects judged to be able to form cooperative avoidance behavior with the robot or follow predetermined movement patterns, such as medical carts working in cooperation with robots, medicine delivery carts running on fixed tracks, and intelligent logistics vehicles moving along preset paths in home scenarios, can be marked as controllable dynamic obstacles. Objects with highly random movement behavior and significant influence from external stimuli, such as pets, toy balls in a child's hand, or objects that are accidentally pushed over and roll, are marked as uncontrollable dynamic obstacles. Interaction attributes are recorded in the dynamic obstacle entries through fields, enabling the system to use different prediction tools based on different movement patterns.

[0063] For controllable dynamic obstacles, a Long Short-Term Memory (LSTM) network model is loaded as a motion trajectory prediction model for each object or similar objects. The LTM network receives historical position coordinate sequences and velocity sequences arranged chronologically. Through a gating structure, it preserves long-term trends and suppresses instantaneous noise in the time dimension, thus outputting a series of possible future positions or paths within a given prediction time window. The input sequence can be updated by sliding within a fixed-length time window, ensuring the model maintains stable predictive ability while continuously receiving new observations. To match different scenarios, different network sizes and parameters can be configured according to categories. For example, a longer time window is used for infusion carts in medical environments to reflect slow and stable motion, while a shorter time window is used for family members or caregivers in home environments to improve sensitivity to short-term motion changes. This time-series-based modeling approach is suitable for handling objects with relatively predictable motion trajectories.

[0064] For uncontrollable dynamic obstacles, a velocity vector-based probability distribution model is loaded for each object as a trajectory prediction model. This model takes the current position, velocity magnitude, velocity direction, and environmental constraints as input, assuming that the movement trend largely follows the current velocity vector within a short timeframe, while also incorporating disturbances caused by environmental interference and the object's own randomness. By applying variance control to the velocity vector, a probability distribution for the future position is constructed. A unimodal or multimodal Gaussian distribution can be used to describe the possible areas within a future period, or a sampling method can be used to generate multiple candidate trajectory point clouds. Scene environmental feature information from the environmental state set is used to constrain the range of the probability distribution. For example, the probability density is compressed to zero within obstacle areas formed by walls, furniture, or medical equipment, while the probability along the passageway is appropriately increased in doorways and passageways. In this way, even if the target's trajectory is difficult to predict precisely, a spatial probability characterization of high-risk and low-risk areas can be provided, offering a reference for subsequent avoidance strategies.

[0065] This embodiment extracts contour feature data and real-time motion speed data from the environmental state set, and uses support vector machine classification logic to distinguish between static and dynamic obstacles. Furthermore, it further subdivides dynamic obstacles into controllable and uncontrollable types based on interaction attributes. Dynamic objects are no longer treated as a single category with consistent behavior, but are represented hierarchically based on motion predictability and collaborative capabilities. By binding a long short-term memory network model to controllable dynamic obstacles and a velocity vector-based probability distribution model to uncontrollable dynamic obstacles, the system can select matching trajectory prediction tools for different motion patterns. This results in more refined path predictions in predictable time-series scenarios and probability distributions covering high-risk areas in highly random scenarios, thereby improving the accuracy and robustness of dynamic obstacle trajectory prediction. This provides more stable and discriminative input for subsequent decision-making, achieving a better balance between safety and efficiency in obstacle avoidance behavior.

[0066] In one embodiment, step S40 above includes: S401, for the controllable dynamic obstacles in the dynamic obstacles, the historical position coordinate sequence of the controllable dynamic obstacles is collected as controllable motion data, and the controllable motion data is input into the long short-term memory network model in the motion trajectory prediction model to obtain the controllable predicted motion trajectory. S402, for the uncontrollable dynamic obstacles in the dynamic obstacles, the current movement speed and direction of the uncontrollable dynamic obstacles are collected as uncontrollable motion data, and the uncontrollable motion data is input into the velocity vector-based probability distribution model in the motion trajectory prediction model to obtain the uncontrollable predicted motion trajectory. S403, Detect the Euclidean distance in three-dimensional space between the planned path and the controllable predicted motion trajectory, and between the planned path and the uncontrollable predicted motion trajectory; S404, determine the minimum Euclidean distance from all detected Euclidean distances; S405, based on the preset trajectory prediction time, the robot's safe response time and the minimum Euclidean distance, the collision risk value is determined using an exponential decay function, and the collision risk value decreases as the trajectory prediction time increases; S406, Determine the collision risk level based on the collision risk value.

[0067] In this embodiment, using the already distinguished controllable and uncontrollable dynamic obstacles as input, continuous calculations are performed around four levels: motion data acquisition, trajectory prediction, path distance calculation, and risk quantification, to transform the spatial motion behavior at continuous moments into a collision risk level that can be used for decision-making.

[0068] For controllable dynamic obstacles, the first step is to organize the historical position coordinate sequence under a unified time reference. Position coordinates can be represented using three-dimensional coordinates, arranging the spatial positions perceived in each frame into a time series. The length of the historical position sequence can be set based on typical motion cycles in the scene, for example, covering motion processes of several seconds to tens of seconds. This position sequence is then input as controllable motion data into a Long Short-Term Memory (LSTM) network model. The LTM network controls the flow of information in the time dimension through input gates, forget gates, and output gates, modeling both long-term trends and short-term fluctuations in the position sequence, and extracting temporal features suitable for describing behavioral habits and local action changes. During the calculation process, the network updates its internal state hourly and outputs a series of spatial coordinates corresponding to future time points within a preset prediction duration, forming a controllable predicted motion trajectory. This trajectory can be represented as an ordered three-dimensional point sequence, or a continuous curve can be generated by interpolation based on the point sequence for subsequent distance calculation.

[0069] For uncontrollable dynamic obstacles, the system focuses more on short-term motion trends and uncertainty boundaries than on long-term behavior patterns. To this end, the current velocity magnitude and direction are extracted from the latest perceived frame, and velocity stability can be estimated by combining this with positional changes from the most recent few frames. This information constitutes uncontrollable motion data, which is input into a velocity vector-based probability distribution model. This model uses the current position as the starting point and the current velocity vector as the desired motion direction, constructing a probability field for the future position by applying variance and offset distributions around the velocity vector. Multiple time slices can be discretized in space, and within each time slice, several candidate position points are sampled in different directions, and probability weights are assigned to each candidate point, thus forming a probabilistic representation of the uncontrollable predicted motion trajectory. This prediction can be represented as a point cloud or set of regions with probability density, reflecting the likelihood of different positions appearing within the predicted time period.

[0070] After obtaining the controllable and uncontrollable predicted motion trajectories, a spatial comparison with the robot's planned path is required. The planned path can be viewed as a polyline or smooth curve parameterized by time or path length in three-dimensional space. During distance detection, the planned path is discretized into several sampling points, and the three-dimensional Euclidean distance is calculated by pairing these points with each predicted point in the controllable predicted motion trajectory. Similarly, the Euclidean distance is calculated between the sampling points of the planned path and representative candidate positions in the uncontrollable predicted motion trajectory. The Euclidean distance is calculated based on the square root of the sum of the squares of the three-dimensional coordinate differences, reflecting the linear interval between any two spatial positions. To avoid missing local near-field situations, the sampling interval can be set according to the robot's size and safety distance, ensuring that potential collision scenarios near any point on the path fall within the calculation range.

[0071] The minimum Euclidean distance is selected from the set of all calculated Euclidean distances. This minimum value represents the closest approach between the robot's planned path and the possible location of any dynamic obstacle within the predicted timeframe. Then, combining the trajectory prediction time and the robot's safe response time, the minimum Euclidean distance is mapped to a collision risk value. The trajectory prediction time represents the length of the future time window covered by the prediction, while the safe response time represents the time required for the robot to execute an effective deceleration or stop action after receiving risk information. An exponential decay function is used here as a time-weighted control tool. When the prediction time is much longer than the safe response time, even if the minimum distance is small, the risk value is reduced through a time factor. When the prediction time is close to or slightly longer than the safe response time, the time factor has less impact on the risk value, emphasizing the distance itself. When the prediction time is insufficient to cover the safe response time, even if the distance is large, the risk value can be increased through the function shape to highlight the potential for insufficient reaction time. The parameters of the exponential decay function can be configured according to the braking performance and control cycle of different robot platforms, controlling the sensitivity of the risk value to time by adjusting the decay rate.

[0072] After calculating the collision risk value, it is necessary to classify the risk level. Multiple risk levels can be pre-set in intervals, such as dividing the entire risk range into high-risk, medium-risk, and low-risk intervals. More granular classifications can also be added based on different application environments. The boundary values ​​can be calibrated by combining long-term operational statistics, empirical thresholds, and simulated collision experiments in a test environment. During operation, each calculated collision risk value is compared with these interval boundaries to determine the corresponding collision risk level. This risk level is then written back to the decision module for subsequent constraint setting and obstacle avoidance strategy selection.

[0073] This embodiment distinguishes between controllable and uncontrollable dynamic obstacles. It uses historical position coordinate sequences to drive a long short-term memory network model to generate controllable predicted motion trajectories, and the current velocity vector to drive a probability distribution model to generate uncontrollable predicted motion trajectories. Then, it performs Euclidean distance detection between the two types of predicted trajectories and the planned path in three-dimensional space and extracts the minimum Euclidean distance. It introduces trajectory prediction time and safety response time to construct a time-weighted exponential decay function, jointly mapping spatial proximity and time margin to a collision risk value, which is further divided into graded collision risk levels. This continuous processing transforms the potential conflict between the dynamic obstacle's motion behavior and the robot's future motion into quantifiable level information that can be directly used for decision control. In complex environments, it can identify high-risk proximity situations earlier and more accurately distinguish different risk levels, thus providing a more stable and refined risk input for subsequent path replanning and obstacle avoidance behavior selection.

[0074] In one embodiment, step S50 above includes: S501, based on the obstacle information in the set of environmental states, the predicted motion trajectory and the collision risk level, define obstacle avoidance constraints, which require the robot's path points to maintain a safe distance from obstacles. S502, Based on the robot's own state information in the set of environmental states, define resource constraints. The resource constraints require that the estimated energy consumption of the task is less than the robot's remaining power, and the estimated consumable consumption of the task is less than the robot's remaining consumables. S503, Define an efficiency constraint condition, which requires that the expected execution time of the task is less than a time threshold; S504, construct a cost function with the optimization objectives of minimizing path length and task execution time, and assign corresponding weight coefficients to the path length term and task execution time term in the cost function according to the task priority; S505, a rasterized map is established based on the scene environment feature information in the set of environment states. An improved heuristic path search is used to search the rasterized map for a sequence of nodes that satisfies the obstacle avoidance constraint, the resource constraint, and the efficiency constraint and minimizes the value of the cost function. The node sequence is then connected to generate the optimal execution path.

[0075] In this embodiment, the system needs to unify task priority, environmental state set, and collision risk level into a planning model. By constructing a constraint set and objective function, the optimal execution path that can be executed in a real environment is obtained. The environmental state set includes obstacle information, robot's own state information, and scene environment feature information. The task priority comes from user requirement analysis, and the collision risk level comes from trajectory prediction and spatial comparison in the previous stage. These three types of inputs jointly determine the feasibility and quality of the path.

[0076] A multi-constraint task planning model can be understood as a discrete path decision-making model. A path in discrete space is represented by a series of nodes and the connections between them. Each node corresponds to a grid cell or feature point in a rasterized map, and each connection corresponds to a single movement of the robot between adjacent positions. The model introduces a sequence of path nodes as decision variables, a set of constraints to limit the range of values ​​for the node sequences, and a cost function to evaluate the merits of different node sequences. During the solution process, only node sequences that satisfy all constraints are considered candidate paths. The optimal execution path is then selected from the candidate paths through comparison using the cost function.

[0077] Obstacle avoidance constraints are based on obstacle information, predicted motion trajectories, and collision risk level definitions within the environmental state set. Obstacle information includes the obstacle's type, location coordinates, and size. The predicted motion trajectory indicates the possible locations of dynamic obstacles within the predicted timeframe. The collision risk level reflects the risk quantification obtained during the prediction process. When constructing obstacle avoidance constraints, each path point on the planned path needs to be distance-detected relative to both the static obstacle geometry and the predicted dynamic obstacle location, ensuring the distance is not less than the set obstacle avoidance safety distance. The obstacle avoidance safety distance can be dynamically adjusted based on obstacle type and collision risk level. For example, a larger safety distance can be set for fragile furniture and easily damaged items, and the lower limit of the distance can be increased for high-risk situations, keeping the path as far away from danger zones as possible. For dynamic obstacles, the predicted trajectory can be discretized on the time axis and then matched with nodes in different time slices on the path, incorporating time information into the obstacle avoidance constraints to prevent the robot from getting too close to obstacles in both time and space.

[0078] Resource constraints are constructed based on the robot's own state information within the environmental state set. This state information includes the robot's current position, speed, remaining battery power, and remaining consumables. During path planning, the energy and consumable consumption required for each segment of movement along the path are estimated using kinematic or empirical models. The consumption of each segment is then summed to obtain the estimated energy and consumable consumption for the task. Resource constraints require that the estimated energy consumption be less than the remaining battery power, and the estimated consumable consumption be less than the remaining consumables. Energy consumption can be calculated based on metrics such as energy consumption per unit distance and additional energy consumption during acceleration and deceleration. Consumable consumption can be estimated based on the amount of consumables used per unit time or per unit work area. When resource constraints cannot be met, the corresponding path is considered infeasible in the planning model and is no longer considered in subsequent searches, thus preventing the robot from interrupting its task due to insufficient battery power or consumables midway through the path.

[0079] Efficiency constraints limit task execution time. The estimated execution time can be calculated based on path length, robot maximum speed, acceleration / deceleration capabilities, and turning characteristics, and can also be adjusted based on current environmental congestion levels. Time thresholds can be derived from user task time requirements, system scheduling strategies, or safety rules, such as limiting dwell time in narrow passages or the total execution time for high-priority tasks. Efficiency constraints require the estimated execution time to be less than the time threshold. If a candidate path is too long or has too many turns leading to exceeding the time limit, that path is considered infeasible, thus preventing the robot from choosing excessively circuitous or inefficient paths.

[0080] After the constraint set is determined, a cost function needs to be constructed to evaluate the merits of different feasible paths. The cost function uses path length and task execution time as optimization objectives, typically employing a weighted sum to linearly combine the path length and time terms. Path length reflects the distance traveled, while execution time reflects task timeliness; their weights can be adaptively adjusted based on task priority. When a task has a high priority, the weight of the time term can be increased, making the planning process more inclined to shorten execution time; conversely, when a task has a low priority or focuses more on energy utilization, the weight of the path length term can be increased, favoring shorter, less energy-intensive paths. A mapping table or continuous mapping function can be established between task priority and weight coefficients, using interpolation to obtain weight configurations for different priorities, avoiding insufficient adaptability caused by fixed weights.

[0081] Scene environmental features from the environmental state set are used to build a rasterized map. These features include ground type, light intensity, and spatial topology. Ground type can be used to distinguish different materials such as carpet, tile, and wood flooring. By setting grid cost values, the impact of different materials on movement energy consumption and slip risk can be reflected. For example, increasing the cost value in slippery areas can encourage paths to avoid these areas. Light intensity can be used to mark the reliability of visual perception. In areas with insufficient light, the cost value can be increased or the sensor weight can be decreased to reduce the frequency of these areas as path sampling points. Spatial topology represents room layout, passageway connections, and traversable areas. An initial grid layout is generated from the topology, and then obstacle and ground information are overlaid to form a rasterized map that reflects the actual environment. Map resolution can be adjusted in different implementations. Higher resolution results in more detailed paths, but also increases computational cost.

[0082] Based on a gridded map, an improved heuristic path search is introduced as a solution tool. Heuristic path search estimates the potential cost of unexplored nodes by incorporating a heuristic function during node expansion, thereby reducing the search range and accelerating convergence. In implementation, the heuristic function can simultaneously consider factors such as target point distance, current grid cost, nearby obstacle density, and local collision risk indicators, making the search more consistent with the actual environmental risk distribution. During path search, each candidate node is evaluated to determine if it meets obstacle avoidance constraints, resource constraints, and efficiency constraints. Nodes that do not meet the constraints are discarded and do not participate in subsequent searches. At the end of the search, the node sequence with the smallest cost function value is selected, and the node sequences are connected in spatial order to form the optimal execution path. Interpolation or smoothing can be performed between nodes if necessary to make the path more suitable for robot trajectory tracking.

[0083] This embodiment incorporates task priority, environmental state set, and collision risk level into a unified planning model. It filters path feasibility through three types of constraints: obstacle avoidance, resource, and efficiency. Then, it uses a cost function associated with task priority to run an improved heuristic path search on a gridded map constructed based on scene environmental features. This can ensure that the path maintains a safe distance from obstacles, while controlling power and consumable consumption to not exceed the robot's resource limit and limiting execution time to an acceptable range. It selects the optimal execution path that is more suitable for the current task requirements and environmental conditions from multiple dimensions. Compared with the processing method that selects paths based solely on geometric distance or a single time indicator, it is more conducive to reducing collision risk and task interruption probability, improving resource utilization efficiency, and enhancing the overall stability of task completion.

[0084] In one embodiment, step S60 above includes: S601, drive the motion control module of the robot to track the optimal execution path, identify the task type of the user task instruction, and start the corresponding functional module to perform the task operation according to the task type; S602, monitor the collision risk level in real time during movement; S603, when the real-time monitored collision risk level is higher than the first preset threshold, execute an obstacle avoidance strategy that immediately stops movement and issues a warning; S604, when the real-time monitored collision risk level is between the first preset threshold and the second preset threshold, an obstacle avoidance strategy of reducing driving speed and dynamically fine-tuning driving trajectory is executed. S605, when the real-time monitored collision risk level is lower than the second preset threshold, execute an obstacle avoidance strategy that maintains the current speed and continues to drive along the optimal execution path while continuously monitoring the obstacle status.

[0085] In this embodiment, the robot needs to switch between different obstacle avoidance strategies based on the real-time updated collision risk level while maintaining movement along the optimal execution path and completing the task operation, thereby achieving a trade-off between motion safety and task execution in a dynamic environment.

[0086] When the motion control module tracks the optimal execution path, it first obtains the optimal execution path, composed of discrete path points or trajectory segments, from the planning stage and buffers these path points into the motion control queue in chronological order. Within a control pulse cycle, the motion control module retrieves the target position, target orientation, and target velocity parameters from the queue. Combining these parameters with the current actual position and attitude, it generates linear velocity and angular velocity commands through velocity planning and trajectory interpolation, and sends them to the chassis drive unit or joint drive unit to drive the motor to perform rotational or linear motion. During path tracking, feedback of position and attitude deviations can be introduced to correct the velocity commands for the next cycle, gradually converging the actual motion trajectory to the optimal execution path.

[0087] While performing path tracking, the system needs to identify the task type parsed from the user's task instructions, such as floor cleaning, room inspection, and item delivery. The system reads the task type field from the task parsing cache and maps the task types to specific functional modules before execution. For example, the cleaning task type corresponds to the sweeping and water spraying components, the delivery task type corresponds to the robotic arm and door control components, and the inspection task type corresponds to the sensor acquisition and data upload components. At the start of operation, the system activates the corresponding functional modules based on the task type, including issuing start commands to the execution units, loading necessary parameter configurations, performing self-check processes, and conducting safety interlock checks, ensuring that the functional modules can synchronously execute task operations when the robot enters the corresponding work area.

[0088] As the robot moves along the optimal execution path, it needs to periodically acquire the collision risk level. The collision risk level is calculated and output by the preceding trajectory prediction and spatial comparison process. It can be expressed as a continuous value mapped to a level range, or directly represented by discrete level labels indicating high risk, medium risk, and low risk. During operation, the control system reads the latest collision risk level from the risk assessment module at fixed time intervals, compares the current value with a first preset threshold and a second preset threshold, and provides a basis for subsequent strategy selection.

[0089] When the real-time monitored collision risk level exceeds a first preset threshold, the system determines that there is a high probability of collision and needs to immediately enter emergency protection mode. In this mode, the motion control module terminates path tracking logic, issues zero-speed commands to the chassis drive or joint drive, clears the unexecuted motion queue, and triggers mechanical braking or braking torque limiting when conditions permit, shortening the stopping distance. Simultaneously, to issue risk warnings to surrounding personnel and users, the system controls the voice broadcast unit, buzzer, or indicator lights to execute predefined warning modes, such as continuous voice reminders, high-frequency sounds, or bright flashing lights, to reduce the probability of personnel continuing to approach the robot or obstructing the movement space. Emergency protection mode remains operational until the collision risk level drops to a safe range or user intervention occurs.

[0090] When the real-time monitored collision risk level falls between a first preset threshold and a second preset threshold, the system determines that it is currently in an intermediate range where potential interference exists but a collision is not yet certain. This range employs a strategy combining speed adjustment and local trajectory fine-tuning. For speed adjustment, the upper limits of linear velocity and angular velocity can be scaled proportionally based on the specific values ​​of the collision risk level within this range, allowing the robot to move at a lower speed when approaching potential obstacle areas, thus providing more reaction time for perception and decision-making. For trajectory fine-tuning, a small local grid or set of sampling points is constructed within the local neighborhood of the optimal execution path. Based on the current obstacle distribution and passable area, a local driving trajectory is recalculated, and the local trajectory is smoothly spliced ​​with the globally optimal execution path. This increases the lateral safety distance or detour radius of the robot without deviating from the task objective direction, reducing the speed and degree of relative approach to dynamic obstacles.

[0091] When the real-time monitored collision risk level is below the second preset threshold, the system determines that it is currently in a safe zone and can continue to execute the movement according to the original plan. At this time, the current speed configuration remains unchanged, the motion control module continues to perform position tracking and attitude control along the optimal execution path, and the functional modules continue to execute operations according to the established task logic. Within this safe zone, in order to prevent a sudden change in environmental conditions from causing a rapid increase in the risk level, the system still maintains continuous monitoring of the obstacle status, periodically reads the obstacle position, speed, and category information, and refreshes the collision risk level in real time. Once a significant change in the obstacle status is detected, such as a sharp decrease in distance or a sharp increase in speed, the risk level will be re-compared with the threshold in the next cycle, and switching between different obstacle avoidance strategies will be performed.

[0092] Obstacle avoidance strategies across different risk ranges require coordination through state management logic. During robot operation, a finite set of states can be maintained for the current strategy, such as emergency stop, medium-risk deceleration, and normal driving. Each time a risk assessment result is generated, a decision is made based on the relationship between the new risk level and the threshold to determine whether to transition between states. During state transitions, the motion control module executes corresponding control output switching. For example, when transitioning from normal driving to emergency stop, speed planning data is immediately discarded; when returning from medium-risk deceleration to normal driving, the speed upper limit is gradually restored to avoid large speed fluctuations that could cause attitude instability. Through the constraints of state management logic, the obstacle avoidance strategy maintains controllability and predictability when switching between different risk ranges.

[0093] Example Description: In a smart home environment, the robot initiates its operation and first establishes real-time awareness of its surroundings. Multimodal sensors continuously collect image frames, distance and depth values, inertial attitude parameters, and environmental parameters. The raw observation data is then fed into specific filtering modules: image data undergoes Gaussian filtering to reduce brightness abrupt changes and texture noise; distance sequences are filtered by median filtering to remove abnormal reflections; attitude sequences are input into a Kalman filter to smooth acceleration and angular velocity errors; and environmental auxiliary parameters are directly used for subsequent mapping. The processed data is projected onto a unified coordinate system and a unified time reference, ensuring that observations from different sources are synchronized and consistent according to spatial and temporal relationships. Then, through evidence fusion logic, a set of environmental states is formed, including obstacle locations, the robot's own state, and the environmental topology.

[0094] The robot receives task input submitted by the user, and the interaction module can come from any channel, such as a voice interface, mobile terminal, or touch panel. The information parsing module performs semantic analysis or protocol parsing on the command text, extracting the task type, target area range, and completion deadline, and then structures the extracted results for subsequent judgment. The system derives the task urgency and execution complexity based on the task objective, user preference tags, and time constraints, and obtains a quantified priority value through hierarchical weight calculation. This priority is directly associated with the subsequent path planning module.

[0095] The environmental state set is immediately used for obstacle recognition and classification after generation. The algorithm analyzes the pixel contours of obstacles and the corresponding velocity vector changes. The support vector machine model labels environmental objects with zero velocity as static and objects with velocity greater than zero as dynamic. Dynamic objects are further distinguished by their interaction attributes: those with predictable motion characteristics or yielding behavior are marked as controllable, while those with irregular motion patterns or large trajectory changes are marked as uncontrollable. Controllable objects are loaded with a time series learning model, using a long short-term recursive structure to generate future movement predictions; uncontrollable objects are loaded with a probabilistic model composed of velocity vectors and variance estimates to infer future distribution areas from current velocity and direction.

[0096] The trajectory prediction phase iteratively processes the aforementioned identification results. Dynamic obstacles are input into the trained temporal network or velocity probability model according to their categories, and the system outputs future position predictions or probability coverage areas for different time periods. The spatial distance between the planned path and the predicted trajectory is calculated using three-dimensional Euclidean distance, and the minimum distance is used in subsequent evaluations. The system combines a preset prediction time window, a safe response delay, and the minimum distance value to derive collision risk values ​​through an exponential decay function, and generates corresponding risk levels according to numerical ranges. Risk results are continuously updated, providing real-time data for planning, solving, and execution control.

[0097] After acquiring task priorities, environmental state sets, and collision risk levels, the robot constructs multi-dimensional planning constraints using these inputs. Obstacle avoidance conditions require all locations on the path to maintain at least a safe distance from any static or dynamic object; resource conditions require sufficient remaining power and cleaning supplies to cover the planned area and support return; efficiency conditions limit the total execution time to no more than the calculated window. The path objective function is weighted between length and execution time, with the weight adjusted by task priority. Higher priorities favor time efficiency, while lower priorities favor path simplification. The system performs heuristic path search on a grid map generated from environmental information, selecting the node sequence that satisfies all constraints and has the minimum cost function value, and connecting this sequence into a continuous execution path.

[0098] The robot control module begins movement according to the planned path and invokes task execution functions, such as activating the cleaning component. During execution, the system continuously monitors the risk level. When the risk level exceeds a safety threshold, the controller pauses motor output and provides feedback via a prompt. When the risk level is within a certain range, the robot reduces its speed and avoids moving objects by deviating from the path. When the risk level is low, it maintains its original speed and continues to track obstacle states. The path tracking module and obstacle avoidance strategy selection module work in a closed loop, ensuring coordination between the robot's travel path, obstacle avoidance adjustments, and target task execution as the environment dynamically changes.

[0099] During task execution, the robot simultaneously collects operational status metrics, including path deviation error, energy consumption change rate, obstacle avoidance trigger count, and execution time. After the execution process is complete, the user provides a satisfaction rating via an interactive interface. The system integrates the collected data with user feedback; the operational status is used to calibrate planning model parameters, such as optimizing safety distances or reducing unnecessary detours; dynamic object trajectory samples collected during the task can be used to update the trajectory prediction model, improving its prediction accuracy in future environments. Parameter update logic is automatically triggered based on data performance, without affecting the real-time performance of the task, while simultaneously reducing planning errors in subsequent task execution.

[0100] In operational scenarios targeting healthcare service environments such as home care and elderly care centers, the robot first performs environmental understanding upon startup. The multimodal sensing components activate image acquisition, distance measurement, inertial sensing, and environmental state detection devices to collect data on the floor plan, assistive devices, patient positions, lighting conditions, and movable objects within the service area. The obtained raw data undergoes characteristic correction through independent filter groups. Image signals are filtered using Gaussian filtering to remove uneven lighting and texture interference, distance readings are filtered using median filtering to eliminate abnormal jump values, and the inertial attitude sequence is input into a Kalman frame to reduce jitter and inertial drift. The processed parameters are calibrated according to the same coordinate and time reference, and a fusion processing method is used to generate an environmental state set containing moving obstacles, static facilities, robot operating status, and indoor environmental conditions, providing foundational data for subsequent task scheduling.

[0101] The robot then receives task input from patients or caregivers, including voice requests, mobile terminal triggers, or clicks on the bedside interactive screen. The instruction parsing module extracts the semantic structure from the input, identifying the mobile service category (e.g., medication delivery, meal delivery, IV fluid replenishment), target area (e.g., specific room, bedside location), and execution time limit. It also calculates task urgency (e.g., medication reminder), service preference (e.g., whether the patient is currently resting), and execution complexity (path traversing dense medical aids or assistive devices). The system constructs a weight matrix and calculates quantified priority values ​​to determine whether a task should be executed immediately or prioritized to cover other tasks.

[0102] Environmental status data enters the obstacle identification process, analyzing assistive devices, furniture, patient activity trajectories, and walking medical staff or residents in the area. A support vector classification model distinguishes between bedside tables, wheelchair parking spots, and rehabilitation training equipment with zero speed, and patients or caregivers with speeds greater than zero. For people continuously moving in a defined direction or slowly pushed walkers, the system marks them as predictable moving objects; for those exhibiting sudden turns, random stops, or movements associated with rehabilitation training, they are marked as uncontrollable moving objects. Predictable objects are input into a time-series model to form a future movement trend line, while uncontrollable objects are input into a speed probability model to determine the coverage area of ​​their future locations.

[0103] Dynamic obstacle information is transmitted to the trajectory prediction module. For predictable objects, continuous historical position sequences are collected and input into a recursive model to generate a sequence of movement points for future time periods; for uncontrollable objects, a probability cloud region is generated based on the current velocity direction and state change trend. The prediction results are then compared with the robot's planned path for the task, and the minimum possible approach distance between the path and each predicted trajectory is determined through 3D distance calculation. Combining the planning time window and the robot's maneuver response delay, a collision risk value is calculated using an exponentially decreasing function, and the risk is classified and quantified.

[0104] The task planning module constructs constraints based on priority, environmental conditions, and risk levels. Obstacle avoidance requires the path to maintain a safe distance from equipment, auxiliary facilities, and moving personnel; resource constraints require the robot to have sufficient remaining battery power and carried resources to complete the task and possess return capability; efficiency constraints require the total execution time to not exceed the task's set time window. The system constructs a cost function with execution time and path length as the common minimization objectives and dynamically assigns weights. Through gridded processing of the indoor activity area, the path search algorithm explores node combinations that meet all the above constraints and outputs a sequence of path nodes, which, when continuously connected, constitutes an executable path.

[0105] The robot's execution module activates the motion control system to advance along the planned path and invokes task-triggered components, such as opening storage compartments, releasing items, or interacting with cabinets. The obstacle avoidance control module continuously receives and assesses the risk level. If the risk value exceeds the safe range, it immediately stops advancing and issues a warning to avoid contact with the elderly, moving people, or equipment. If the risk is in the intermediate range, it reduces its speed and performs a minimal repositioning. If the risk is low, it maintains the route and continues real-time monitoring. The combination of path tracking and obstacle avoidance control ensures accurate task completion.

[0106] During execution, the robot continuously records operational deviations, detours, execution time, changes in power consumption and consumables, and the number of safe shutdowns, and combines this data with user or caregiver feedback to form backhaul data. Based on task performance, the system makes appropriate adjustments to the planning model parameters or calibrates the trajectory prediction model to improve accuracy in predicting crowd movement patterns and the impact of equipment placement in healthcare scenarios.

[0107] In fintech service scenarios combining distributed offices and unmanned business halls, customer service assistance robots perform an environmental data collection process upon activation. Multimodal sensing components acquire images, distance, pose, and environmental state parameters from the branch's self-service area, waiting area, information area, and manual service counter area, forming a raw data set including personnel distribution, furniture layout, accessible area outlines, battery status, and changes in ambient noise and lighting. Gaussian filtering, median filtering, and Kalman filtering are used to process visual signals, distance readings, and pose sequences, respectively, eliminating data disturbances caused by large-area glass reflections, personnel movement, or ground textures. After unified spatiotemporal mapping and fusion processing, an environmental state set is formed, including dynamic customer locations, queue density, the robot's current remaining capacity, and scene structural features.

[0108] The robot receives business task instructions from customers or the system dispatcher, including assisting customers in printing business vouchers, guiding customers to designated windows, delivering contract materials, delivering financial information to VIP customers, or delivering documents to the back-end review point. The task acquisition module collects instructions through a speech recognition engine, mobile app interface, or touchscreen terminal, and performs semantic structure analysis on the task text to extract task category, target area (e.g., wealth management area, contract printing area, counter X), time requirements, and service priority restrictions; for example, urgent document delivery may require high priority processing. The module calculates the task's urgency, customer preferences (e.g., whether non-interactive service is needed), and path complexity. A quantified priority value is obtained through a hierarchical weight matrix to determine whether the task should be executed immediately or inserted at the top of the task queue.

[0109] An environmental state set is used to identify objects within the area that may affect movement. The system analyzes data from the movement trajectories of people, wheelchairs, temporary queuing partitions, promotional displays, and customer service personnel. A support vector classification model distinguishes between stationary equipment (such as promotional stands, ATMs, and booth surfaces) and customers or staff with significant movement speed. People exhibiting a stable walking trend or moving around a fixed point are labeled as predictable dynamic objects; customers exhibiting sudden acceleration, aimless wandering, or irregular movement due to conducting business are labeled as uncontrollable dynamic objects. These two types of objects are associated with time-series models and velocity probability models, respectively, laying the foundation for subsequent prediction processes.

[0110] The robot collects historical or real-time movement sequence data. For predictable dynamic objects, the continuous position sequence is input into a time recursive model, which outputs the activity range of the corresponding personnel over several future time periods. For uncontrollable dynamic objects, a probability diffusion range is derived based on the current speed, direction, and movement pattern. The predicted trajectory and the robot's planned execution path are compared using 3D spatial distance detection, and the minimum distance value for all objects is calculated. Combining the task execution time window and equipment response latency, a risk score is generated using an exponentially decreasing function to classify the risk level. This is used to determine whether the path may be blocked by customer flow or whether it is necessary to avoid highly congested areas in advance.

[0111] The planning module constructs a constraint system based on task priority, resource status, and risk information. Obstacle avoidance conditions require the path to avoid high-density crowds or moving objects; resource conditions require the equipment to have sufficient power to support round-trip scheduling without having to return to a refueling station midway; efficiency conditions require the service completion time to be within the business window requirements. A cost function is constructed with the goal of minimizing path length and processing time, and task priority weights are assigned. The hall layout is gridded, dynamically blocking areas are marked as constraint units, and an improved heuristic search is used to find node combinations that satisfy all constraints, forming an optimal executable path.

[0112] The robot initiates its journey along a designated path and performs specified business assistance actions. These include opening the top storage area to deliver contracts, activating a mechanical retrieval device to retrieve vouchers from a printer tray, or delivering packaged documents to the customer service area. During its journey, the robot continuously receives risk level updates. If the risk exceeds a threshold, it immediately stops moving and issues a warning message to avoid contact with customers; if the risk is in the intermediate range, it reduces its speed and detours around congested areas; if the risk is in the safe range, it maintains its original path and continues monitoring.

[0113] During task execution, the system records path deviations, delays, obstacles encountered, power consumption, and customer feedback, such as satisfaction ratings or confirmations of business assistance effectiveness. By comprehensively analyzing the operational status and feedback data, the system adjusts constraint parameters or corrects deviations in the trajectory model, gradually improving path prediction and dynamic detour capabilities. This enables the robot to maintain stable operation under varying daily passenger volumes, personnel flow patterns, and service peaks.

[0114] This embodiment integrates path tracking, task function module execution, and a graded obstacle avoidance strategy based on collision risk level into a single operation process. During robot movement, it continuously monitors the collision risk level and switches between three behaviors—emergency stop, deceleration fine-tuning, and speed maintenance—based on the relationship with a first and a second preset threshold. This allows for timely control outputs when environmental conditions change rapidly. In high-risk scenarios, it quickly stops and issues warning messages; in medium-risk scenarios, it increases safety buffering by reducing speed and adjusting the trajectory; and in low-risk scenarios, it maintains task continuity and execution efficiency. Simultaneously, continuous monitoring of obstacle states ensures that a more stringent protection strategy can be immediately implemented when the risk increases. Therefore, without relying on a single static rule, it significantly reduces the probability of collisions, decreases the number of task interruptions, and improves the stability and adjustability of the entire robot's operation.

[0115] In one embodiment, a task planning and dynamic obstacle avoidance device is provided, which corresponds one-to-one with the task planning and dynamic obstacle avoidance methods described in the above embodiments. (Refer to...) Figure 3 , Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the task planning and dynamic obstacle avoidance device of the present invention. The modules include: environmental perception module 10, task parsing module 20, obstacle recognition module 30, collision prediction module 40, path planning module 50, motion execution module 60, and parameter update module 70. Detailed descriptions of each functional module are as follows: The environment perception module 10 is used to collect environmental data through multimodal sensors mounted on the robot, preprocess the environmental data and fuse multi-source information to generate an environmental state set containing obstacle information, robot's own state information and scene environment feature information. Task parsing module 20 is used to obtain user task instructions and parse the user task instructions to determine task priority; The obstacle recognition module 30 is used to identify dynamic obstacles based on the set of environmental states, and to establish a corresponding motion trajectory prediction model for the dynamic obstacles. The collision prediction module 40 is used to input the motion data of the dynamic obstacle into the motion trajectory prediction model to generate a predicted motion trajectory in the future time period, and to spatially compare the predicted motion trajectory with the planned path determined based on the user task instruction to determine the collision risk level. The path planning module 50 is used to construct a multi-constraint task planning model based on the task priority, the set of environmental states and the collision risk level, and generate the optimal execution path by solving the multi-constraint task planning model. The motion execution module 60 is used to control the robot to move along the optimal execution path and perform task operations. During the movement, the obstacle avoidance strategy is dynamically adjusted according to the real-time updated collision risk level to perform obstacle avoidance operations. The parameter update module 70 is used to collect running status data and user interaction feedback data during task execution, and update the parameters of the multi-constraint task planning model and the motion trajectory prediction model based on the running status data and the user interaction feedback data.

[0116] For specific limitations regarding task planning and dynamic obstacle avoidance devices, please refer to the aforementioned limitations on task planning and dynamic obstacle avoidance methods, which will not be repeated here. Each module in the aforementioned task planning and dynamic obstacle avoidance device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0117] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When executed by the processor, the computer program implements the functions or steps of a task planning and dynamic obstacle avoidance method on the server side.

[0118] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When executed by the processor, the computer program implements client-side functions or steps of a task planning and dynamic obstacle avoidance method.

[0119] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Environmental data is collected by multimodal sensors mounted on the robot. The environmental data is preprocessed and multi-source information is fused to generate an environmental state set that includes obstacle information, robot's own state information and scene environment feature information. Obtain user task instructions and parse the user task instructions to determine task priority; Based on the set of environmental states, dynamic obstacles are identified, and corresponding motion trajectory prediction models are established for the dynamic obstacles. The motion data of the dynamic obstacle is input into the motion trajectory prediction model to generate a predicted motion trajectory for a future time period. The predicted motion trajectory is then spatially compared with the planned path determined based on the user task instruction to determine the collision risk level. A multi-constraint task planning model is constructed based on the task priority, the set of environmental states, and the collision risk level. The optimal execution path is generated by solving the multi-constraint task planning model. The robot is controlled to move along the optimal execution path and perform task operations. During the movement, the obstacle avoidance strategy is dynamically adjusted according to the real-time updated collision risk level to perform obstacle avoidance operations. Collect operational status data and user interaction feedback data during task execution, and update the parameters of the multi-constraint task planning model and the motion trajectory prediction model based on the operational status data and the user interaction feedback data.

[0120] In one embodiment, a computer-readable storage medium is provided, which may be non-volatile or volatile, and a computer program is stored thereon, which, when executed by a processor, performs the following steps: Environmental data is collected by multimodal sensors mounted on the robot. The environmental data is preprocessed and multi-source information is fused to generate an environmental state set that includes obstacle information, robot's own state information and scene environment feature information. Obtain user task instructions and parse the user task instructions to determine task priority; Based on the set of environmental states, dynamic obstacles are identified, and corresponding motion trajectory prediction models are established for the dynamic obstacles. The motion data of the dynamic obstacle is input into the motion trajectory prediction model to generate a predicted motion trajectory for a future time period. The predicted motion trajectory is then spatially compared with the planned path determined based on the user task instruction to determine the collision risk level. A multi-constraint task planning model is constructed based on the task priority, the set of environmental states, and the collision risk level. The optimal execution path is generated by solving the multi-constraint task planning model. The robot is controlled to move along the optimal execution path and perform task operations. During the movement, the obstacle avoidance strategy is dynamically adjusted according to the real-time updated collision risk level to perform obstacle avoidance operations. Collect operational status data and user interaction feedback data during task execution, and update the parameters of the multi-constraint task planning model and the motion trajectory prediction model based on the operational status data and the user interaction feedback data.

[0121] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0122] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0123] It should be noted that if any software tools or components not belonging to this company appear in the embodiments of this application, they are merely illustrative examples and do not represent actual use. The embodiments described above are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

[0124] The user personal information involved in this application embodiment is all authorized (knowing and consenting) by the relevant parties or fully authorized by all parties, and the executing entity can obtain it through various open, legal and compliant means. The collection, storage, use, processing, transmission, provision and disclosure of the information, data and signals involved all comply with the relevant laws and regulations of the relevant countries and regions, and do not violate public order and good morals.

Claims

1. A task planning and dynamic obstacle avoidance method, characterized in that, Includes the following steps: Environmental data is collected by multimodal sensors mounted on the robot. The environmental data is preprocessed and multi-source information is fused to generate an environmental state set that includes obstacle information, robot's own state information and scene environment feature information. Obtain user task instructions and parse the user task instructions to determine task priority; Based on the set of environmental states, dynamic obstacles are identified, and corresponding motion trajectory prediction models are established for the dynamic obstacles. The motion data of the dynamic obstacle is input into the motion trajectory prediction model to generate a predicted motion trajectory for a future time period. The predicted motion trajectory is then spatially compared with the planned path determined based on the user task instruction to determine the collision risk level. A multi-constraint task planning model is constructed based on the task priority, the set of environmental states, and the collision risk level. The optimal execution path is generated by solving the multi-constraint task planning model. The robot is controlled to move along the optimal execution path and perform task operations. During the movement, the obstacle avoidance strategy is dynamically adjusted according to the real-time updated collision risk level to perform obstacle avoidance operations. Collect operational status data and user interaction feedback data during task execution, and update the parameters of the multi-constraint task planning model and the motion trajectory prediction model based on the operational status data and the user interaction feedback data.

2. The task planning and dynamic obstacle avoidance method as described in claim 1, characterized in that, Environmental data is collected by multimodal sensors mounted on the robot. This environmental data is preprocessed and fused with multi-source information to generate an environmental state set containing obstacle information, the robot's own state information, and scene environmental feature information, including: Environmental data is obtained by collecting environmental image data, obstacle distance data, robot motion posture data, and environmental auxiliary parameter data through visual sensors, ranging sensors, inertial sensors, and environmental auxiliary sensors. Gaussian filtering is applied to the environmental image data in the environmental data, median filtering is applied to the obstacle distance data in the environmental data, and Kalman filtering is applied to the robot motion posture data in the environmental data to obtain denoised environmental data. Based on a preset environmental map model, the denoised environmental data is mapped to a unified spatial coordinate system and time reference for spatiotemporal calibration to obtain preprocessed environmental data. Based on evidence theory fusion processing, the preprocessed environmental data is fused from multiple sources to generate an environmental state set that includes obstacle information, robot state information, and scene environment feature information.

3. The task planning and dynamic obstacle avoidance method as described in claim 1, characterized in that, Obtaining and parsing user task instructions to determine task priority includes: Activate the robot's voice interaction module, mobile terminal application interface, or touch interaction interface to receive user task instructions through the voice interaction module, the mobile terminal application interface, or the touch interaction interface; The user task instructions are parsed using natural language processing or instruction protocol parsing, and the task type, task target location, and task time requirements are extracted from the user task instructions. Based on the task type, the target location of the task, and the time requirement of the task, determine the urgency, user preferences, and task complexity; An analytic hierarchy process (AHP) is used to construct a judgment matrix for the urgency, user preference, and task complexity. Based on the judgment matrix, the urgency, user preference, and task complexity are weighted, and the quantified value is determined as the task priority according to the weighting result.

4. The task planning and dynamic obstacle avoidance method as described in claim 1, characterized in that, Identify dynamic obstacles based on the set of environmental states, and establish corresponding motion trajectory prediction models for the dynamic obstacles, including: Analyze the obstacle information in the environmental state set, and extract the outline feature data and real-time motion speed data of the obstacles; The contour feature data and the real-time motion speed data are processed using support vector machine classification logic. Obstacles with zero speed are marked as static obstacles, and obstacles with a speed greater than zero are identified as dynamic obstacles. Analyze the interactive attributes of the dynamic obstacles, and mark obstacles with interactive avoidance characteristics as controllable dynamic obstacles, and obstacles with random motion characteristics as uncontrollable dynamic obstacles. For the controllable dynamic obstacle, a long short-term memory network model is loaded to establish a corresponding motion trajectory prediction model; for the uncontrollable dynamic obstacle, a probability distribution model based on velocity vector is loaded to establish a corresponding motion trajectory prediction model.

5. The task planning and dynamic obstacle avoidance method as described in claim 1, characterized in that, The motion data of the dynamic obstacle is input into the motion trajectory prediction model to generate a predicted motion trajectory for a future time period. The predicted motion trajectory is then spatially compared with the planned path determined based on the user task instructions to determine the collision risk level, including: For controllable dynamic obstacles among the dynamic obstacles, the historical position coordinate sequence of the controllable dynamic obstacles is collected as controllable motion data. The controllable motion data is input into the long short-term memory network model in the motion trajectory prediction model to obtain the controllable predicted motion trajectory. For uncontrollable dynamic obstacles among the dynamic obstacles, the current speed and direction of the uncontrollable dynamic obstacles are collected as uncontrollable motion data. The uncontrollable motion data is input into the velocity vector-based probability distribution model in the motion trajectory prediction model to obtain the uncontrollable predicted motion trajectory. Detect the Euclidean distance in three-dimensional space between the planned path and the controllable predicted trajectory, as well as between the planned path and the uncontrollable predicted trajectory; Determine the minimum Euclidean distance from all detected Euclidean distances; Based on the preset trajectory prediction time, the robot's safe response time, and the minimum Euclidean distance, the collision risk value is determined using an exponential decay function, and the collision risk value decreases as the trajectory prediction time increases. The collision risk level is determined based on the collision risk value.

6. The task planning and dynamic obstacle avoidance method as described in claim 1, characterized in that, A multi-constraint task planning model is constructed based on the task priority, the set of environmental states, and the collision risk level. The optimal execution path is generated by solving the multi-constraint task planning model, including: Based on the obstacle information in the set of environmental states, the predicted motion trajectory, and the collision risk level, obstacle avoidance constraints are defined, which require the robot's path points to maintain a safe distance from obstacles. Based on the robot's own state information in the set of environmental states, resource constraints are defined, which require that the estimated energy consumption of the task is less than the robot's remaining power, and the estimated material consumption of the task is less than the robot's remaining materials. Define an efficiency constraint that requires the expected execution time of the task to be less than a time threshold. Construct a cost function with the optimization objectives of minimizing path length and task execution time, and assign corresponding weight coefficients to the path length and task execution time terms in the cost function according to the task priority. A rasterized map is built based on the scene environment feature information in the set of environment states. An improved heuristic path search is used to search the rasterized map for a sequence of nodes that satisfies the obstacle avoidance constraints, the resource constraints, and the efficiency constraints and minimizes the value of the cost function. The sequence of nodes is then connected to generate the optimal execution path.

7. The task planning and dynamic obstacle avoidance method as described in claim 1, characterized in that, Controlling the robot to move along the optimal execution path and perform task operations, dynamically adjusting the obstacle avoidance strategy according to the real-time updated collision risk level during the movement to perform obstacle avoidance operations, including: The motion control module that drives the robot tracks the optimal execution path, identifies the task type of the user task instruction, and activates the corresponding functional module to perform the task operation according to the task type. The collision risk level is monitored in real time during movement; When the real-time monitored collision risk level is higher than the first preset threshold, an obstacle avoidance strategy is implemented to immediately stop movement and issue a warning. When the real-time monitored collision risk level is between the first preset threshold and the second preset threshold, an obstacle avoidance strategy is implemented that reduces the driving speed and dynamically fine-tunes the driving trajectory. When the real-time monitored collision risk level is lower than the second preset threshold, an obstacle avoidance strategy is executed to maintain the current speed and continue driving along the optimal execution path while continuously monitoring the obstacle status.

8. A task planning and dynamic obstacle avoidance device, characterized in that, The task planning and dynamic obstacle avoidance device includes: The environmental perception module is used to collect environmental data through multimodal sensors mounted on the robot, preprocess the environmental data and fuse multi-source information to generate an environmental state set that includes obstacle information, robot's own state information and scene environment feature information. The task parsing module is used to obtain user task instructions, parse the user task instructions, and determine the task priority; An obstacle recognition module is used to identify dynamic obstacles based on the set of environmental states, and to establish a corresponding motion trajectory prediction model for the dynamic obstacles. The collision prediction module is used to input the motion data of the dynamic obstacle into the motion trajectory prediction model to generate a predicted motion trajectory in the future time period, and to spatially compare the predicted motion trajectory with the planned path determined based on the user task instruction to determine the collision risk level. The path planning module is used to construct a multi-constraint task planning model based on the task priority, the set of environmental states, and the collision risk level, and generate the optimal execution path by solving the multi-constraint task planning model. The motion execution module is used to control the robot to move along the optimal execution path and perform task operations. During the movement, the obstacle avoidance strategy is dynamically adjusted according to the real-time updated collision risk level to perform obstacle avoidance operations. The parameter update module is used to collect running status data and user interaction feedback data during task execution, and update the parameters of the multi-constraint task planning model and the motion trajectory prediction model based on the running status data and the user interaction feedback data.

9. A computer device, characterized in that, The computer device includes a memory, a processor, and a task planning and dynamic obstacle avoidance program stored in the memory and executable on the processor. When the task planning and dynamic obstacle avoidance program is executed by the processor, it implements the steps of the task planning and dynamic obstacle avoidance method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a task planning and dynamic obstacle avoidance program, which, when executed by a processor, implements the steps of the task planning and dynamic obstacle avoidance method as described in any one of claims 1-7.

Images