A dynamic scene data processing and adaptive decision method and system for embodied intelligence

Abstract

The present application relates to the field of data management and scheduling, in particular to a kind of dynamic scene data processing and adaptive decision method and system for embodied intelligence;Method includes: real-time receiving multimodal perception data, task instruction, weather data and group experience risk atlas;Enhanced environment representation that is fused macroscopic statistical risk and microcosmic geometric information is constructed and updated;Fusion decision analysis is carried out, and a set of coordinated adaptive scheme is generated;The present application generates coordinated adaptive scheme based on enhanced environment representation, real-time task instruction and weather data, so that visual attention focuses, path re-planning and gait parameter adjustment can be executed in a unified decision framework. Therefore, the system can dynamically allocate sensing resources for high-risk areas, adjust the passage path in real time according to the risk cost change, and adjust the motion control parameters according to the estimated road surface properties, to improve the passage stability and task execution continuity of embodied intelligent devices in complex dynamic scenes.

Description

Technical Field

[0001] This invention relates to the field of data management and scheduling, and more specifically to a method and system for dynamic scene data processing and adaptive decision-making for embodied intelligence. Background Technology

[0002] Embodied intelligent systems refer to intelligent systems that can perceive the external environment through sensors, make autonomous analysis and decisions based on task objectives, and interact with the real physical environment through actuators. As embodied intelligent devices such as humanoid robots and mobile robots are gradually applied to open and dynamic scenarios such as commercial districts, logistics warehouses, community roads, and park inspections, the environment they face is no longer a static, regular, and fully predictable laboratory scenario, but a complex scenario that includes multiple dynamic factors such as pedestrians, vehicles, weather changes, road surface changes, temporary obstacles, and changes in group behavior.

[0003] Existing embodied intelligence systems typically acquire information about the surrounding environment using devices such as cameras, lidar, and inertial sensors, and perform static obstacle recognition, basic path planning, and simple obstacle avoidance control. However, in real-world open environments, relying solely on single, real-time perception data for decision-making makes it difficult to accurately reflect the time-varying and hidden nature of environmental risks. For example, the traffic risk at the same geographical location may change significantly under conditions such as rain, snow, low temperatures, nighttime, and strong glare; some road surfaces may only appear as reflective, water-stained, or color-different areas, but may actually correspond to different risk states such as thin ice, standing water, mud, or loose gravel. Existing systems often struggle to correlate and analyze weather data, time data, historical traffic data, and group behavior data with specific geographical locations, leading to discrepancies between the results of dynamic scene data processing and the actual traffic risks. Furthermore, when processing dynamic scenes, most existing embodied intelligence systems still rely primarily on data collected by the device itself, failing to effectively utilize behavioral information from other moving entities in the environment. For example, when pedestrians exhibit behaviors such as detouring, slowing down, walking cautiously, slipping, or falling on a certain road section, it usually implies a potential risk at that location. However, existing systems lack a mechanism to transform this type of behavioral data into reusable risk experience data. Even if a single robot gains risk experience through trial and error in a certain area, this experience is often limited to its current task and cannot be promptly accessed by other robots. This causes multiple embodied intelligent devices to repeatedly undergo the risk exploration process in the same or similar environments, reducing overall operational efficiency and safety.

[0004] Existing data processing and decision-making workflows often suffer from the disconnect between perception, analysis, planning, and control modules. The visual perception module primarily outputs target recognition or obstacle detection results, the path planning module mainly calculates based on geometric distance and static obstacle avoidance rules, and the motion control module performs gait or posture adjustments according to a predetermined path. These modules lack a unified enhanced environmental representation, making it difficult to simultaneously express microscopic geometric information, macroscopic empirical risks, real-time weather impacts, and task priorities. This results in the system's inability to generate consistent visual attention focusing schemes, path replanning schemes, and motion control adjustment schemes in dynamic scenarios. Existing embodied intelligence systems lack a complete data closed-loop mechanism. Feedback data such as successful passage, slippage, instability, sudden stops, and detours during execution are often not standardized, recorded, uploaded, aggregated, and reused for model optimization. Consequently, the system struggles to continuously update the risk probabilities of specific geographical areas under different weather, time, and pedestrian conditions, and it also fails to generate a shared group empirical risk map that can be shared by multiple devices. This affects the adaptive capabilities and scalable deployment capabilities of embodied intelligence systems in long-term operational scenarios. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention proposes a dynamic scene data processing and adaptive decision-making method and system for embodied intelligence. The robot not only perceives the world through its own sensors, but also understands risks by observing human behavior, anticipates risks by sharing group experience, and ultimately avoids risks by combining real-time weather data, thus achieving human-like, socialized safe navigation and interaction.

[0006] The dynamic scene data processing and adaptive decision-making method for embodied intelligence described in this invention is executed by a central control platform and includes the following steps: S1. Receive multimodal sensing data, high-level task instructions, real-time weather data, and population experience risk map data matched with the current environment in real time; S2. Construct and update the local dynamic scene interaction map, and integrate the macro-statistical risk data in the group experience risk map to form an enhanced environmental representation that includes micro-geometric information and macro-experience risk labels; S3. Based on the enhanced environmental characterization, real-time task instructions and weather data, perform fusion decision analysis to generate a set of coordinated adaptive schemes. The schemes include at least: a visual attention focusing scheme based on experience risk guidance, a path replanning scheme to avoid high-risk areas, and a gait parameter adjustment scheme to adapt to the estimated road surface properties. S4. Distribute the adaptive scheme to the corresponding perception, planning and control modules for execution; S5. Monitor the execution process and generate a comprehensive performance evaluation vector; S6. Based on the performance evaluation vector and the newly collected data, generate new experience data fragments and upload them to update the group experience risk map, while triggering local scheme fine-tuning.

[0007] Preferably, the process of constructing the group experience risk map includes: collecting behavioral video data of other moving entities in the environment by the vision acquisition devices of each humanoid robot; processing the video data through a behavior analysis model to extract behavioral pattern labels and their corresponding environmental data labels at specific geographical locations, wherein the behavioral pattern labels include walking steadily, walking cautiously, slipping, falling, and detouring; binding the behavioral pattern labels with environmental data labels and geographical location information to form standardized experience data units; and performing clustering and statistical analysis on the massive standardized experience data units according to geographical location grids and weather data on a cloud terminal processor to calculate the risk probability distribution and mainstream safe path patterns of each grid unit under different weather data, generating and continuously updating the map.

[0008] Preferably, the experience-risk-guided visual attention focusing scheme described in step S3 is as follows: The central control platform analyzes the group experience risk map to identify high-risk points of concern in the vicinity or ahead of the current planned path where the historical risk index exceeds a preset threshold. The central control platform sends instructions to the vision processing module to dynamically adjust the allocation of perception resources and perform continuous, high-resolution scanning and analysis of the high-risk points of concern. The visual processing module feeds back the enhanced analysis results for high-risk points to the central control platform for real-time confirmation and correction of map risk information.

[0009] Preferably, the path replanning scheme for avoiding high-risk areas described in S3 is as follows: The central control platform transforms the gridded risk probability data provided by the group experience risk map into a dynamic cost map in the path planning algorithm. Grids with high historical risk are assigned higher travel costs, while historically safe paths are assigned lower travel costs. Under the premise of satisfying task endpoint constraints and basic geometric obstacle avoidance, the path planning algorithm performs real-time path search with the goal of minimizing the total travel cost. When real-time perception confirms that the actual danger of a high-risk point has been reduced or a new danger not recorded in the map is discovered, the central control platform dynamically adjusts the cost value of that local area and triggers local re-optimization of the path.

[0010] Preferably, when the behavior analysis model identifies slipping or falling behavior, it makes a judgment based on one or more of the following abrupt changes in visual features: the pedestrian's torso tilts relative to the ground at an angular velocity exceeding a threshold; the pedestrian's arms are involuntarily and rapidly waving to maintain balance; the pedestrian's feet make an abnormal, non-forward sliding displacement after contacting the ground; and the pedestrian's center of gravity trajectory shows violent, non-periodic fluctuations.

[0011] Preferably, the generation of new experience data fragments in step S6 includes records of both positive and negative experiences: When the robot smoothly passes through a certain area following the planned safe path, it is recorded as a positive experience verification. When the robot slips, loses its posture, or is forced to stop suddenly at a certain position, it is recorded as a negative experience update; Both positive and negative experiences are uploaded to dynamically update the risk probability value at the corresponding position in the group experience risk map.

[0012] Preferably, the method further includes a cross-robot experience sharing and collaboration mechanism: when multiple humanoid robots are active in the same geographical area within a similar time period, their central control platform exchanges data through a communication network; the robot that first passes through a complex road section can share its real-time constructed enhanced environmental representation and detailed control parameters for successful passage with neighboring companion robots in the form of an experience package; the receiving party's central control platform uses this experience package as high-priority real-time reference information and integrates it into its own decision-making process.

[0013] Preferably, the data from the group experience risk map is further used for business operation analysis and decision support, including: identifying geographical locations where traffic difficulties occur frequently and over a long period of time, and automatically generating safety hazard reports; analyzing the safe traffic flow of each route under different weather conditions, and providing dynamic suggestions for traffic management; and setting differentiated risk tolerance for route planning of different priority tasks under severe weather conditions based on historical risk data.

[0014] An adaptive interaction system for humanoid robots in dynamic scenarios includes robot subsystems located locally on each humanoid robot and a central processing subsystem located in the cloud; the robot subsystems include: The perception module is used to collect multimodal sensor data of the environment and perform behavioral analysis on the video of pedestrian or vehicle movement in the environment to extract behavioral pattern labels. A local data processing module, connected to the perception module, is used to record and temporarily store standardized experience data units bound by behavior pattern labels, current environment data labels, and geographic location information. The central control platform is connected to the perception module, the local data processing module, the robot's motion planning and control module, and the communication module, respectively. The central control platform is used to: receive and integrate real-time perception data, task instructions, weather data, and group experience risk map data from the cloud, generate adaptive solutions, and coordinate the execution of each module. The central processing subsystem includes: The swarm data processor connects to the local data processing modules of each robot subsystem via a communication network. It is used to receive standardized experience data units from all robots, build and continuously update the global swarm experience risk map. Dynamic environmental factor interface, used to access real-time and historical macro weather data; The communication module is used to enable data uploading, map downloading, and direct communication between the robot subsystem and the central processing subsystem.

[0015] Preferably, the perception module includes a visual acquisition unit and a visual processing unit; the visual acquisition unit is a high-definition camera equipped with a controllable gimbal; the visual processing unit has a built-in behavior analysis model for performing the feature judgment as described in claim 5 to identify pedestrian slipping or falling behavior; and, in response to the visual attention focusing scheme instruction based on experience risk guidance issued by the central control platform, dynamically adjusts the gimbal angle and focal length of the visual acquisition unit and the allocation of computing resources of the visual processing unit.

[0016] Preferably, the central control platform integrates a risk perception and path optimization decision-maker; the decision-maker is used to convert the group experience risk map data into a dynamic cost map and run a path planning algorithm to execute a path replanning scheme to avoid high-risk areas.

[0017] Preferably, the communication module supports low-latency local area network communication protocols; when the central control platform of multiple robot subsystems determines that they are in the same adjacent area, a point-to-point or multicast communication link is established through the communication module to realize the cross-robot experience sharing and collaboration mechanism.

[0018] The beneficial effects of this invention are as follows: 1. This invention receives multimodal perception data, high-level task instructions, real-time weather data, and group experience risk map data in real time, and constructs an enhanced environmental representation that includes micro-geometric information and macro-experience risk labels. This enables embodied intelligent systems to no longer rely solely on the current single perception result for judgment, but to combine historical experience, group behavior, and real-time environmental factors to comprehensively analyze dynamic scenarios, thereby improving the completeness and accuracy of environmental representation.

[0019] 2. This invention integrates risk probability data from a group experience risk map with a current local dynamic scene interaction map, enabling the identification of potential risk areas that are difficult to determine visually, such as thin ice, water accumulation, mud, loose road surfaces, or historically high-frequency instability areas. This improves the ability of embodied intelligent devices to predict hidden environmental risks and reduces the risk of slipping, instability, detour failure, or task interruption caused by perceptual misjudgment.

[0020] 3. This invention utilizes an adaptive scheme based on enhanced environmental representation, real-time task instructions, and weather data to coordinate visual attention focusing, path replanning, and gait parameter adjustment within a unified decision-making framework. Consequently, the system can dynamically allocate perception resources for high-risk areas, adjust travel paths in real-time to changes in risk costs, and adjust motion control parameters based on estimated road surface properties, thereby improving the stability and continuity of task execution for embodied intelligent devices in complex dynamic scenarios. Attached Figure Description

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] Figure 1 This is a flowchart of the method in this invention; Figure 2 This is a flowchart of the system in this invention. Detailed Implementation

[0023] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0024] like Figure 1-2 As shown, the dynamic scene data processing and adaptive decision-making method for embodied intelligence described in this invention is executed by a central control platform and includes the following steps: S1. Receive multimodal sensing data, high-level task instructions, real-time weather data, and population experience risk map data matched with the current environment in real time; S2. Construct and update the local dynamic scene interaction map, and integrate the macro-statistical risk data in the group experience risk map to form an enhanced environmental representation that includes micro-geometric information and macro-experience risk labels; S3. Based on the enhanced environmental characterization, real-time task instructions and weather data, perform fusion decision analysis to generate a set of coordinated adaptive schemes. The schemes include at least: a visual attention focusing scheme based on experience risk guidance, a path replanning scheme to avoid high-risk areas, and a gait parameter adjustment scheme to adapt to the estimated road surface properties. S4. Distribute the adaptive scheme to the corresponding perception, planning and control modules for execution; S5. Monitor the stability indicators, path tracking deviations, and compliance with expected risks during the execution process, and generate a comprehensive performance evaluation vector; S6. Based on the performance evaluation vector and the newly collected data, generate new experience data fragments and upload them to update the group experience risk map, while triggering local solution fine-tuning; based on the performance evaluation vector and the newly collected environmental and self-behavioral data during the execution process, generate new experience data fragments and upload them to the cloud to update the group experience risk map, while triggering local solution fine-tuning, forming a two-layer adaptive closed loop from individual to group, and then from group feedback to guide individual; Upgrading robot navigation from relying on instantaneous physical perception to relying on a fusion of historical experience knowledge base and instantaneous perception; the effect is to endow robots with predictive risk avoidance capabilities and collective intelligence evolution capabilities, enabling them to make judgments comparable to experienced humans when facing road conditions with blurred visual features but actual dangers, greatly improving the reliability and safety of operation in real dynamic business environments.

[0025] As one embodiment of the present invention, the construction process of the group experience risk map includes: collecting behavioral video data of other moving entities in the environment by the vision acquisition devices of each humanoid robot, and collecting mobile video data of other pedestrians or vehicles in the environment; processing the video data through a behavior analysis model to extract behavioral pattern labels and their corresponding environmental data labels at specific geographical locations, wherein the behavioral pattern labels include walking steadily, walking cautiously, slipping, falling, and detouring; and the environmental data labels include weather type, temperature, humidity, and time; binding the behavioral pattern labels with the environmental data labels and geographical location information to form standardized experience data units; and performing clustering and statistical analysis on the massive standardized experience data units according to geographical location grids and weather data on a cloud terminal processor to calculate the risk probability distribution and mainstream safe path patterns of each grid unit under different weather data, generating and continuously updating the map; This transforms soft indicators like human behavior into quantifiable and statistically probable risk probabilities. For example, the behavior pattern of multiple people detouring is quantified through statistical analysis as having a safe passage probability of less than 20% at that point. This provides a data foundation for solving the fundamental problem of robots' lack of discernment capabilities and creates a continuously evolving digital map that reflects real-world physical risks and human consensus.

[0026] As one embodiment of the present invention, the visual attention focusing scheme based on experience-risk guidance in step S3 is specifically as follows: The central control platform analyzes the group experience risk map and identifies high-risk points of concern in the vicinity or ahead of the current planned path where the historical risk index exceeds a preset threshold. The central control platform sends instructions to the vision processing module to dynamically adjust the camera pan-tilt angle, focal length, or image processing algorithm's computing resource allocation to continuously and at high resolution scan and analyze the high-risk points of concern. The visual processing module feeds back the enhanced analysis results of high-risk points, such as texture details, reflective properties, and water depth estimation, to the central control platform for real-time confirmation and correction of map risk information. For example, when the robot approaches a section of road prone to icing, even if the overall image appears normal, the central control platform will instruct the camera to carefully examine the microscopic texture of specific areas of the road surface, looking for reflective features of ice crystals. This is equivalent to giving the robot an eye of experience, significantly improving its detection rate of hidden hazards (such as black ice and thin mud).

[0027] As one embodiment of the present invention, the path replanning scheme for avoiding high-risk areas described in S3 is specifically as follows: The central control platform transforms the gridded risk probability data provided by the group experience risk map into a dynamic cost map in the path planning algorithm. Grids with high historical risk are assigned higher travel costs, while historically safe paths are assigned lower travel costs. Under the premise of satisfying task endpoint constraints and basic geometric obstacle avoidance, the path planning algorithm performs real-time path search with the goal of minimizing the total travel cost. When real-time perception confirms that the actual danger of a high-risk point has been reduced or a new danger not recorded in the map is discovered, the central control platform dynamically adjusts the cost value of that local area and triggers local re-optimization of the path.

[0028] As one embodiment of the present invention, the behavior analysis model makes a judgment based on one or more of the following abrupt changes in visual features when identifying slipping or falling behavior: the angular velocity of the pedestrian's torso relative to the ground exceeds a threshold; the pedestrian's arms are involuntarily and rapidly waving to maintain balance; the pedestrian's feet produce abnormal, non-forward sliding displacement after contacting the ground; and the pedestrian's center of gravity trajectory shows violent, non-periodic fluctuations.

[0029] In one embodiment of the present invention, the generation of new experience data fragments in step S6 includes records of positive and negative experiences: When the robot smoothly passes through a certain area following the planned safe path, it is recorded as a positive experience verification. When the robot slips, loses its posture, or is forced to stop suddenly at a certain position, it is recorded as a negative experience update; Both positive and negative experiences are uploaded to dynamically update the risk probability values ​​of corresponding locations in the group experience risk map; it learns not only from humans but also from its own successes and failures; the robot itself becomes a new data collection source, and its walking performance is the best verification and supplement to the historical experience map; this enables the system to continuously improve and adapt to long-term environmental changes; for example, after a road section is rebuilt, the experience data of the first batch of robots successfully and smoothly passing through it will quickly reduce the historical risk value of that road section in the map, guiding subsequent robots to use it with confidence.

[0030] As one embodiment of the present invention, the method further includes a cross-robot experience sharing and collaboration mechanism: when multiple humanoid robots are active in the same geographical area within a similar time period, their central control platform exchanges data through a communication network; the robot that first passes through a complex road section can share its real-time constructed enhanced environmental representation and detailed control parameters for successful passage with neighboring companion robots in the form of an experience package; the receiving central control platform uses this experience package as high-priority real-time reference information and integrates it into its own decision-making process; this solves the problem of real-time collaboration in extreme dynamic scenarios; for example, in a rainstorm, the first robot discovers the water depth and underwater road conditions and immediately shares the information with the robot team behind, which can then directly choose the optimal route without having to re-explore the path; this greatly improves the overall task efficiency and safety of the robot team, and has significant value in commercial scenarios such as logistics warehousing and emergency response.

[0031] As one embodiment of the present invention, the data of the group experience risk map is further used for business operation analysis and decision support, including: identifying geographical locations where traffic difficulties occur frequently and over a long period of time, and automatically generating safety hazard reports; analyzing the safe traffic flow of each route under different weather conditions, and providing dynamic suggestions for traffic management; and setting differentiated risk tolerance for route planning of different priority tasks under severe weather conditions based on historical risk data.

[0032] An adaptive interaction system for humanoid robots in dynamic scenarios includes robot subsystems located locally on each humanoid robot and a central processing subsystem located in the cloud; the robot subsystems include: The perception module is used to collect multimodal sensor data of the environment and perform behavioral analysis on the video of pedestrian or vehicle movement in the environment to extract behavioral pattern labels. A local data processing module, connected to the perception module, is used to record and temporarily store standardized experience data units bound by behavior pattern labels, current environment data labels, and geographic location information. The central control platform is connected to the perception module, the local data processing module, the robot's motion planning and control module, and the communication module, respectively. The central control platform is used to: receive and integrate real-time perception data, task instructions, weather data, and group experience risk map data from the cloud, execute steps S1 to S6 as described in claim 1, generate an adaptive scheme, and coordinate the execution of each module. The central processing subsystem includes: The swarm data processor is connected to the local data processing module of each robot subsystem through a communication network. It is used to receive standardized experience data units from all robots, execute step S24 as described in claim 2, and construct and continuously update a global swarm experience risk map. Dynamic environmental factor interface, used to access real-time and historical macro weather data; The communication module is used to enable data uploading, map downloading, and direct communication between the robot subsystem and the central processing subsystem.

[0033] In one embodiment of the present invention, the perception module includes a visual acquisition unit and a visual processing unit; the visual acquisition unit is a high-definition camera equipped with a controllable pan-tilt unit; the visual processing unit has a built-in behavior analysis model for performing the feature judgment to identify pedestrian slipping or falling behavior; and, in response to the visual attention focusing scheme instruction based on experience risk guidance issued by the central control platform, the pan-tilt angle and focal length of the visual acquisition unit and the allocation of computing resources of the visual processing unit are dynamically adjusted.

[0034] As one embodiment of the present invention, the central control platform integrates a risk perception and path optimization decision-maker; the decision-maker is used to convert the group experience risk map data into a dynamic cost map and run a path planning algorithm to execute the path replanning scheme to avoid high-risk areas.

[0035] As one embodiment of the present invention, the communication module supports low-latency local area network communication protocols; when the central control platform of multiple robot subsystems determines that they are in the same adjacent area, a point-to-point or multicast communication link is established through the communication module to realize the cross-robot experience sharing and collaboration mechanism.

[0036] This invention uses a robot-mounted vision acquisition device to continuously collect and analyze video streams of human walking within its field of vision; it employs computer vision algorithms to extract pedestrian gait stability, trajectory deviation, abnormal movements (such as slipping, stumbling), and the final path chosen; it binds pedestrian behavior to specific location coordinates in a high-precision map, generating empirical data in triplets of location, behavior, and result; for example, if multiple pedestrians are identified circling around a manhole cover or exhibiting a cautious gait, that location is marked as a potential risk point; if a pedestrian slips in a wet, reflective area, it is marked as a high-risk zone; Once a group of pedestrians consistently crosses a certain path, it is marked as a recommended safe path. Each robot uploads the learned experience data, along with the weather environment labels (e.g., after snow, temperature -2℃, cloudy), timestamps, and location information, to the cloud terminal processor. The terminal processor integrates and analyzes the massive amount of data uploaded by robots, removes occasional noise, extracts statistical patterns, and generates a safe passage probability index and recommended path model under different weather conditions for each tiny road segment or point in the city's gridded map, forming a dynamically updated, globally comprehensive group experience risk map. When the robot performs a solo travel task, the central control platform loads the current task, real-time high-definition visual perception information, real-time weather data (from a network API or onboard sensors), and corresponding risk map slices based on group experience. The central control platform performs multi-source information coupling decision-making: First, based on the current weather (e.g., heavy rain), it matches the risk distribution of the road segment in the map under heavy rain history; second, it uses real-time vision to focus on scanning and confirming high-risk areas indicated by the map (e.g., detecting an expansion of puddles); finally, considering the urgency of the task, its own mobility, and detour costs, it plans a route that maximizes the probability of safety and can actively avoid high-risk areas when necessary. This decision-making process is dynamic and continuously updated as the robot moves and the weather changes slightly.

[0037] Let's illustrate this with a real-world example: On a snowy morning, delivery robot No. 1 and inspector No. 2 were active in the community. Their visual acquisition devices (high-definition cameras) continued to operate. Data capture: The camera records a video stream of pedestrians walking in front of it; Analysis: The onboard processor runs a lightweight AI model and identifies the following in the video: a) Pedestrian A suddenly leans back and waves his arms on the marble floor on the east side of the square (signs of slipping); b) Pedestrian B walks cautiously and with small steps in the same area; c) Pedestrian C and most of the subsequent pedestrians choose to detour around the marble area and walk on the asphalt path next to it with a steady gait. Record: The local experience logger generates three records: Record 1: Location P1 (center coordinates of the marble area), Behavior: Slip, Environmental labels: Weather = After snow, Temperature = -1℃, Time = 08:30, Confidence = High; Record 2: Location P1, Behavior: Walking cautiously, Environmental tags: ... (same as before); Record 3: Location P2 (asphalt path), Behavior: Normal walking, Environmental tags: ... Upload and Fusion: These records are uploaded to the cloud processor; the cloud processor performs statistics on dozens of behavioral records associated with location P1 in a weather scenario of post-snowfall and temperature of approximately -1°C: slipping events account for 70%, and cautious walking accounts for 30%; thus, P1 is determined to be a high-risk point under these weather conditions; at the same time, the normal walking rate of location P2 exceeds 95%, and it is determined to be a safe path; this information is updated to the global group experience risk map, which now contains an entry: under weather pattern W1 (post-snowfall, low temperature), within grid G7 (containing P1 and P2), the risk index of P1 is 0.7, and the recommended path is trajectory T1 passing through P2; Real-time adaptive navigation for an individual robot (deliveryman 01 performing a solo mission). At noon, deliveryman No. 01 received a delivery assignment; at this time, the sun was shining, some of the snow had melted, and the temperature had risen to 1°C. Task Initiation and Environment Loading: The central control platform receives the instruction to proceed to the residential building; it immediately invokes: a) real-time weather data (current: sunny, 1°C); b) a slice of population experience risk map corresponding to similar weather patterns after snowfall and low temperatures turning warm; c) real-time visual perception stream; Multi-source information coupled decision making: Map guidance: The central control platform learned from the map that path P1 had a history of high risk under similar weather conditions; even if the current camera view shows that the P1 area is only slightly wet and has no obvious ice, the central control platform will still mark it as a potential risk area that needs to be verified. Visual confirmation and reassessment: The robot vision system performs enhanced scanning of the P1 area according to the instructions of the central control platform; by analyzing texture and reflective properties, and combining temperature information (1℃ may melt the ice but the surface is still slippery), the algorithm determines that the current risk is still higher than that of the normal road surface. Dynamic path generation: The central control platform comprehensively considers: ① mission objectives; ② its own stability (increased probability of falling on slippery surfaces); ③ the additional distance and time cost of detouring through the safe path P2. After calculation by the built-in multi-objective optimization model (such as minimizing the risk value × time cost function), the central control platform decisively decides to abandon the shortest straight path that directly crosses P1, and instead generates a smooth walking trajectory that passes through the safe path P2.

[0038] Execution and Monitoring: The full-body motion control module receives the trajectory command and controls the robot to walk stably on the P2 path. Simultaneously, the vision system continuously monitors the actual condition of the P2 surface to ensure consistency with the prediction.

[0039] Feedback and map update: After successful completion, the robot's stable walking behavior data (location P2, behavior: normal walking, environment label: sunny, 1℃) is uploaded to the cloud as a positive verification, further consolidating the safety label of P2 in this type of weather; if multiple people walk normally in area P1 due to complete dryness in the afternoon, the data uploaded by the robot in the future will gradually reduce its risk index in this weather, realizing dynamic decay and self-correction of the map.

[0040] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for dynamic scene data processing and adaptive decision-making for embodied intelligence, characterized in that, Executed by the central control platform, including the following steps: S1. Receive multimodal sensing data, high-level task instructions, real-time weather data, and population experience risk map data matched with the current environment in real time; S2. Construct and update the local dynamic scene interaction map, and at the same time integrate the macro-statistical risk data in the group experience risk map to form an enhanced environmental representation that includes micro-geometric information and macro-experience risk labels; S3. Based on the enhanced environmental characterization, real-time task instructions and weather data, perform fusion decision analysis to generate a set of coordinated adaptive schemes. The schemes include at least: a visual attention focusing scheme based on experience risk guidance, a path replanning scheme to avoid high-risk areas, and a gait parameter adjustment scheme to adapt to the estimated road surface properties. S4. Distribute the adaptive scheme to the corresponding perception, planning and control modules for execution; S5. Monitor the execution process and generate a comprehensive performance evaluation vector; S6. Based on the performance evaluation vector and the newly collected data, generate new experience data fragments and upload them to update the group experience risk map, while triggering local scheme fine-tuning.

2. The method for dynamic scene data processing and adaptive decision-making for embodied intelligence according to claim 1, characterized in that, The process of constructing the group experience risk map includes: collecting behavioral video data of other moving entities in the environment by the vision acquisition devices of each humanoid robot; processing the video data through a behavior analysis model to extract behavioral pattern labels and their corresponding environmental data labels at specific geographical locations, wherein the behavioral pattern labels include walking steadily, walking cautiously, slipping, falling, and detouring; binding the behavioral pattern labels with environmental data labels and geographical location information to form standardized experience data units; and performing clustering and statistical analysis on the massive standardized experience data units according to geographical location grids and weather data on a cloud terminal processor to calculate the risk probability distribution and mainstream safe path patterns of each grid unit under different weather data, generating and continuously updating the map.

3. The dynamic scene data processing and adaptive decision-making method for embodied intelligence according to claim 1, characterized in that, The experience-risk-guided visual attention focusing scheme described in step S3 is as follows: The central control platform analyzes the group experience risk map and identifies high-risk points of concern in the vicinity or ahead of the current planned path where the historical risk index exceeds a preset threshold. The central control platform sends instructions to the vision processing module to dynamically adjust the allocation of perception resources and perform continuous, high-resolution scanning and analysis of the high-risk points of concern. The visual processing module feeds back the enhanced analysis results for high-risk points to the central control platform for real-time confirmation and correction of map risk information.

4. The dynamic scene data processing and adaptive decision-making method for embodied intelligence according to claim 1, characterized in that, The specific path replanning scheme for avoiding high-risk areas described in S3 is as follows: The central control platform transforms the gridded risk probability data provided by the group experience risk map into a dynamic cost map in the path planning algorithm. Grids with high historical risk are assigned higher travel costs, while historically safe paths are assigned lower travel costs. Under the premise of satisfying task endpoint constraints and basic geometric obstacle avoidance, the path planning algorithm performs real-time path search with the goal of minimizing the total travel cost. When real-time perception confirms that the actual danger of a high-risk point has been reduced or a new danger not recorded in the map is discovered, the central control platform dynamically adjusts the cost value of that local area and triggers local re-optimization of the path.

5. A dynamic scene data processing and adaptive decision-making method for embodied intelligence according to claim 2, characterized in that, When identifying slipping or falling behavior, the behavior analysis model makes a judgment based on one or more of the following abrupt changes in visual features: the pedestrian's torso tilts at an angular velocity relative to the ground exceeding a threshold; the pedestrian's arms are involuntarily and rapidly waving to maintain balance; the pedestrian's feet produce an abnormal, non-forward sliding displacement after contacting the ground; and the pedestrian's center of gravity trajectory shows violent, non-periodic fluctuations.

6. The dynamic scene data processing and adaptive decision-making method for embodied intelligence according to claim 1, characterized in that, Step S6, which involves generating new experiential data fragments, includes records of both positive and negative experiences. When the robot smoothly passes through a certain area following the planned safe path, it is recorded as a positive experience verification. When the robot slips, loses its posture, or is forced to stop suddenly at a certain position, it is recorded as a negative experience update; Both positive and negative experiences are uploaded to dynamically update the risk probability value at the corresponding position in the group experience risk map.

7. The method for dynamic scene data processing and adaptive decision-making for embodied intelligence according to claim 1, characterized in that, The method also includes a cross-robot experience sharing and collaboration mechanism: when multiple humanoid robots are active in the same geographical area within a similar time period, their central control platform exchanges data through a communication network; the robot that first passes through a complex road section can share its real-time constructed enhanced environmental representation and detailed control parameters for successful passage with neighboring companion robots in the form of an experience package; the receiving party's central control platform uses this experience package as high-priority real-time reference information and integrates it into its own decision-making process.

8. The method for dynamic scene data processing and adaptive decision-making for embodied intelligence according to claim 1, characterized in that, The data from the group experience risk map is further used for business operation analysis and decision support, including: identifying geographical locations where traffic difficulties occur frequently and over a long period of time, and automatically generating safety hazard reports; analyzing the safe traffic flow of each route under different weather conditions, and providing dynamic suggestions for traffic management; and setting differentiated risk tolerance for route planning of different priority tasks under severe weather conditions based on historical risk data.

9. A humanoid robot adaptive interaction system for dynamic scenes, applicable to the dynamic scene data processing and adaptive decision-making method for embodied intelligence as described in any one of claims 1-8, characterized in that, This includes robot subsystems located locally on each humanoid robot, and a central processing subsystem located in the cloud; the robot subsystems include: The perception module is used to collect multimodal sensor data of the environment and perform behavioral analysis on the video of pedestrian or vehicle movement in the environment to extract behavioral pattern labels. A local data processing module, connected to the perception module, is used to record and temporarily store standardized experience data units bound by behavior pattern labels, current environment data labels, and geographic location information. The central control platform is connected to the perception module, the local data processing module, the robot's motion planning and control module, and the communication module, respectively. The central control platform is used to: receive and integrate real-time perception data, task instructions, weather data, and group experience risk map data from the cloud, generate adaptive solutions, and coordinate the execution of each module. The central processing subsystem includes: The swarm data processor connects to the local data processing modules of each robot subsystem via a communication network. It is used to receive standardized experience data units from all robots, build and continuously update the global swarm experience risk map. Dynamic environmental factor interface, used to access real-time and historical macro weather data; The communication module is used to enable data uploading, map downloading, and direct communication between the robot subsystem and the central processing subsystem.

10. The humanoid robot adaptive interaction system for dynamic scenes according to claim 9, characterized in that, The perception module includes a visual acquisition unit and a visual processing unit; the visual acquisition unit is a high-definition camera equipped with a controllable gimbal; the visual processing unit has a built-in behavior analysis model for performing feature judgment to identify pedestrian slipping or falling behavior; and, in response to the visual attention focusing scheme instructions issued by the central control platform based on experience risk guidance, it dynamically adjusts the gimbal angle and focal length of the visual acquisition unit and the computing resource allocation of the visual processing unit.

11. The humanoid robot adaptive interaction system for dynamic scenarios according to claim 9, characterized in that, The central control platform integrates a risk perception and path optimization decision-making unit. The decision-maker is used to transform the group experience risk map data into a dynamic cost map and run a path planning algorithm to execute a path replanning scheme to avoid high-risk areas.

12. The humanoid robot adaptive interaction system for dynamic scenarios according to claim 9, characterized in that, The communication module supports low-latency local area network communication protocols. When the central control platform of multiple robot subsystems determines that they are in the same adjacent area, a point-to-point or multicast communication link is established through the communication module to realize the cross-robot experience sharing and collaboration mechanism.

Images