Dynamic path planning method and system for self-service beverage robot

Abstract

The application discloses a dynamic path planning method and system of a self-service beverage robot, and particularly relates to the field of service robot path planning, and is used for solving the problem that path jitter frequently occurs and liquid surface stability is difficult to be considered in a high-density micro-motion environment of a self-service beverage robot, and is achieved by the following steps: an event vision unit is used to capture a hand micro-motion trigger sequence in real time and map spatial density, an arc segment decomposition unit is used to extract an instantaneous tangent vector and a micro-motion energy spectrum, a risk assessment unit is used to generate a trajectory tension threshold domain through multi-scale window cross analysis, a liquid surface turbulence potential and a spline tortuosity consumption double-dimensional index are introduced, a multi-layer integrated learning is used to predict a steady-state optimization coefficient, and an optimal coefficient is output to a continuous pose in an execution stage, frequent micro-disturbance is responded to with ultra-high sensitivity, liquid surface stability and path geometric cost are considered at the same time, and the smooth and coherent, efficient switching of a cup delivery path is realized, and fine oscillation and response delay are inhibited.

Description

Technical Field

[0001] This invention relates to the field of service robot path planning, and more specifically, to a dynamic path planning method and system for self-service beverage robots. Background Technology

[0002] The self-service beverage bar is designed with extremely narrow channels for retrieving cups. When customers retrieve cups, their arms often make rapid and uneven small swings near the robot's grasping path. The cups are already filled with beverages, and the liquid surface is particularly sensitive to instantaneous acceleration. Therefore, the robot must maintain a safe distance from customers in a crowded environment while maintaining a stable tray posture to accurately serve cups. The entire scenario is dynamic and constrained.

[0003] The paper "A Method and System for Dynamic Path Planning of Robots Based on an Improved Artificial Potential Field Algorithm" (Publication No. CN119311010A) proposes to first estimate the obstacle velocity vector using Kalman filtering, then use a velocity obstacle model to determine collisions, and finally superimpose multiple layers of exponential repulsive fields within the obstacle avoidance area to guide the path. This scheme assumes that the obstacle velocity is continuous and predictable. However, the high-frequency micro-movements of the customer's hand instantly change the velocity estimate, causing the robot's velocity vector to frequently fall into or leave the velocity obstacle domain, and the repulsive layer trigger threshold is repeatedly crossed. The multi-layer repulsive field amplifies this jump due to exponential gain, resulting in alternating activation and failure of the planning, and the path exhibits fine jitter. The accumulated jitter causes tray swaying and cup feeding delay, and conventional filtering cannot eliminate the oscillations because it needs to maintain a rapid obstacle avoidance response. The problem stems from the mismatch between the micro-movement time scale and the algorithm update cycle scale, forming a stability gap that is difficult to suppress directly within the existing potential field framework.

[0004] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a dynamic path planning method and system for a self-service beverage robot. This method uses an event vision unit to capture hand micro-motion trigger sequences in real time and map spatial density. An arc segment decomposition unit extracts instantaneous tangent vectors and micro-motion energy spectra. A risk assessment unit generates a trajectory tension threshold through multi-scale window cross-analysis. A dual-dimensional index of liquid surface surging potential and spline tortuosity loss is introduced. Multi-layer ensemble learning predicts steady-state optimization coefficients, and continuous poses are output with optimal coefficients during the execution phase. This achieves ultra-high sensitivity to frequent micro-amplitude disturbances while simultaneously considering liquid surface stability and path geometric costs. The method achieves smooth and efficient switching of the cup delivery path and suppresses minor oscillations and response delays, thus solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The dynamic path planning system for self-service beverage robots includes:

[0008] Event vision unit: captures the hand-triggered event stream and generates a spatial density map as a dynamic constraint benchmark;

[0009] Arc segment decomposition unit: Cuts the event flow along the density gradient, extracts the instantaneous tangent vector set and calculates the micro-motion energy spectrum to provide quantitative input for subsequent evaluation;

[0010] Risk assessment unit: Analyzes the micro-motion energy spectrum in a multi-scale window, determines the trajectory tension threshold, and marks the high-frequency disturbance region to limit the path generation boundary;

[0011] Trajectory generation unit: Construct a family of acceleration smooth splines within the tension threshold, calculate the liquid surface surging potential and spline tortuosity loss respectively, derive steady-state optimization coefficients through machine learning, and select candidate trajectories with the lowest amplitude risk accordingly;

[0012] The execution scheduling unit interpolates the path with the highest steady-state optimization coefficient into a continuous pose in real time, and the driver uses this to complete the smooth cup delivery.

[0013] In a preferred embodiment, the event vision unit includes the following processing:

[0014] The system uses an event camera to capture real-time changes in light intensity caused by customer hand movements, generating an event stream that includes timestamps, spatial locations, and polarities. The image plane is divided into a grid, and the number of events in each grid cell within the most recent unit time is counted. The event density per unit time and per unit area is calculated to generate an initial spatial density map. The initial spatial density map is smoothed using a Gaussian filter to generate a smoothed spatial density map, which serves as a dynamic constraint benchmark. The active areas of hand movements are quantified using the smoothed event density values ​​of the grid cells.

[0015] In a preferred embodiment, the arc segment decomposition unit includes the following processing:

[0016] Analyze the smoothed spatial density map, calculate the density gradient vector of each grid cell, and identify the active boundaries of hand movements. Use grid cells whose density gradient vector amplitude exceeds a preset threshold as cutting points, and cut the event flow along the normal direction of the density gradient vector to extract local trajectory arcs. Perform linear regression on the spatial coordinates of events within the local trajectory arcs to generate an instantaneous tangent vector set, reflecting the instantaneous direction of hand movements. Calculate the micro-motion frequency and micro-motion amplitude of events within the local trajectory arcs to generate a micro-motion energy spectrum and quantify the dynamic intensity of hand movements.

[0017] In a preferred embodiment, the risk assessment unit includes the following processing steps:

[0018] A set of time windows is defined, and multi-scale analysis of the micro-motion energy spectrum is performed. For each time window in the set, the set of arc segment energy values ​​generated within the corresponding time window length before the current moment is extracted, and the average energy intensity and fluctuation amplitude of the arc segment energy value set are calculated. Based on the average energy intensity and fluctuation amplitude, the trajectory tension threshold is calculated, and the maximum value among the trajectory tension thresholds of all time windows in the set is selected as the trajectory tension threshold range. Arc segments in the micro-motion energy spectrum whose energy values ​​exceed the trajectory tension threshold range are screened, and the spatial locations of these arc segments are recorded as high-frequency disturbance regions.

[0019] In a preferred embodiment, the trajectory generation unit includes the following processing:

[0020] A set of candidate B-spline paths with continuous acceleration is generated within the trajectory tension threshold. For each candidate path, discrete points are uniformly selected along the path. The triaxial acceleration pulse mode at the discrete point is calculated and multiplied by the width of the liquid surface in the cup to obtain the instantaneous impact momentum. The instantaneous impact momentum is integrated along the path and normalized with the capacity of the cup to obtain the surging potential of the liquid surface. The cumulative curvature of the candidate path is calculated and multiplied by the average tangential angular velocity to obtain the spline tortuosity loss.

[0021] In a preferred embodiment, the trajectory generation unit further includes the following processing:

[0022] The liquid surface surging potential and spline tortuosity loss are input into the multilayer ensemble learner, which outputs steady-state optimization coefficients. The path with the highest steady-state optimization coefficient is selected as the final path.

[0023] In a preferred embodiment, the execution scheduling unit includes the following processing:

[0024] The discrete optimal path output by the trajectory generation unit is transformed into a continuous path curve through five Hermite interpolations. The continuous path curve is discretized into a pose command stream by sampling with a fixed control period. The pose command stream is converted into joint angle or wheel speed commands through inverse kinematics. A closed-loop control method is used to generate motor control signals. Combined with a proportional-derivative controller, the robot's motion is adjusted to achieve precise execution of commands.

[0025] In a preferred embodiment, the execution scheduling unit further includes the following processing:

[0026] The inverse kinematics solution process includes: calculating the mapping relationship from the end-effector pose to the joint parameters based on the geometric parameters and motion constraints of the robot's mechanical structure, ensuring that the calculation results satisfy kinematic consistency; then, measuring the robot's actual position and posture at each time step using sensors to form actual pose data; and then comparing the actual pose with the commanded pose at the corresponding time step, calculating the difference between the two, and obtaining the tracking error.

[0027] In a preferred embodiment, the execution scheduling unit further includes the following processing:

[0028] The calculation method of the proportional-derivative controller is as follows: First, the tracking error is multiplied by a preset proportional gain to obtain the proportional term, which is a dimensionless parameter; then, the rate of change of the tracking error with time is calculated, i.e., the derivative of the error, and the rate of change is multiplied by a preset differential gain to obtain the differential term; finally, the proportional term and the differential term are added together to obtain the final control signal.

[0029] Dynamic path planning methods for self-service beverage robots include:

[0030] S1: Capture the hand-triggered event stream and generate a spatial density map as a dynamic constraint baseline;

[0031] S2: Cut the event flow along the density gradient, extract the instantaneous tangent vector set and calculate the micro-motion energy spectrum to provide quantitative input for subsequent evaluation;

[0032] S3: Analyze the micro-motion energy spectrum in a multi-scale window, determine the trajectory tension threshold, and mark the high-frequency disturbance region to limit the path generation boundary;

[0033] S4: Construct a family of acceleration smooth splines within the tension threshold, calculate the liquid surface surging potential and spline tortuosity loss respectively, derive the steady-state optimization coefficient through machine learning, and select the candidate trajectory with the lowest amplitude risk accordingly.

[0034] S5: The path with the highest steady-state optimization coefficient is interpolated in real time to the continuous pose, and the driving end uses this to complete the smooth cup delivery.

[0035] The technical effects and advantages of the dynamic path planning method and system for the self-service beverage robot of this invention are as follows:

[0036] This invention captures hand micro-motion trigger sequences in real time and maps spatial density using an event vision unit. It then extracts instantaneous tangent vectors and micro-motion energy spectra using an arc segment decomposition unit. A risk assessment unit generates a trajectory tension threshold through multi-scale window cross-analysis. By introducing two-dimensional indices—liquid surface surging potential and spline tortuosity loss—and predicting steady-state optimization coefficients through multi-layer ensemble learning, it outputs continuous poses with optimal coefficients during execution. This achieves ultra-high sensitivity to frequent micro-disturbances while balancing liquid surface stability and path geometric costs, resulting in a smooth, coherent, and efficient cup delivery path that suppresses minor oscillations and response delays. The entire chain is tightly integrated from perception to execution, requiring no subsequent compensation or filtering iterations, significantly reducing algorithm load. This provides precise obstacle avoidance and low-sway risk assurance for self-service beverage robots in narrow passages, effectively improving service throughput and user experience. Attached Figure Description

[0037] Figure 1This is a schematic diagram of the dynamic path planning system for the self-service beverage robot of the present invention.

[0038] Figure 2 This is a flowchart illustrating the dynamic path planning method for the self-service beverage robot of the present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Example 1: Figure 1 The present invention provides a dynamic path planning system for a self-service beverage robot, comprising:

[0041] Event vision unit: captures the hand-triggered event flow and generates a spatial density map as a dynamic constraint benchmark.

[0042] Arc segment decomposition unit: Cuts the event flow along the density gradient, extracts the instantaneous tangent vector set and calculates the micro-motion energy spectrum, providing quantitative input for subsequent evaluation.

[0043] Risk assessment unit: Analyze the micro-motion energy spectrum in a multi-scale window, determine the trajectory tension threshold, and mark the high-frequency disturbance region to limit the path generation boundary.

[0044] Trajectory generation unit: Constructs a family of acceleration smooth splines within the tension threshold, calculates the liquid surface surging potential and spline tortuosity loss respectively, derives steady-state optimization coefficients through machine learning, and selects the candidate trajectory with the lowest amplitude risk accordingly.

[0045] The execution scheduling unit interpolates the path with the highest steady-state optimization coefficient into a continuous pose in real time, and the driver uses this to complete the smooth cup delivery.

[0046] In the dynamic path planning scenario of self-service beverage robots, when customers retrieve cups in narrow aisles, their arms often exhibit rapid and varying amplitude micro-movements near the robot's grasping path. These high-frequency micro-movements pose significant challenges to robot path planning and liquid surface stability: on the one hand, the robot must avoid the customer's hand in real time to ensure a safe distance; on the other hand, the cup is already filled with beverage, and the liquid surface is extremely sensitive to instantaneous acceleration, so path vibrations may cause the liquid to surge or even overflow. Traditional methods, such as potential field-based dynamic planning, often fail to simultaneously ensure obstacle avoidance speed and tray stability when faced with such discontinuous and unpredictable high-frequency disturbances due to response lag or oscillation amplification. The event vision unit of this invention aims to solve the accuracy and real-time problem of micro-movement perception in this dynamic environment through high temporal resolution event stream capture and spatial density quantization, providing an accurate dynamic constraint benchmark for subsequent path planning.

[0047] The event vision unit transforms changes in light intensity caused by customer hand movements into dynamic information that can be used for path planning. The following details the specific computational process and technical logic of each step, ensuring the completeness and accuracy from event flow capture to dynamic constraint benchmark generation.

[0048] 1-1, Event stream capture;

[0049] First, an event camera is used to capture changes in light intensity caused by the customer's hand movements. Unlike traditional cameras, event cameras do not record complete images at fixed time intervals. Instead, they continuously monitor the brightness changes of each pixel and generate event data when a significant change in brightness occurs. Each event data point contains four pieces of information: the specific time of the brightness change, accurate to milliseconds; the position of the brightness change on the image plane, represented by the horizontal and vertical axes; the polarity of the brightness change, indicating whether the brightness is increasing or decreasing; and the corresponding pixel identifier. The event camera uses internal sensors to detect the brightness values ​​of pixels in real time. When the brightness value of a pixel exceeds a preset threshold, it immediately records the time, location, and polarity of the change, forming an independent event data unit.

[0050] In self-service beverage robot scenarios, rapid hand movements by customers can occur within a very short timeframe. Traditional cameras, with their fixed frame rates, struggle to capture these instantaneous changes, while event cameras can record each brightness change event with high temporal resolution. These event data are then arranged chronologically to form a continuous event stream. Each data point in the event stream, centered on time, location, and polarity, directly reflects the light intensity changes caused by hand movements. This event stream provides the raw input for subsequent spatial density map generation, ensuring the complete preservation of dynamic information about hand movements.

[0051] 1-2, Spatial density map generation;

[0052] Based on event streams, spatial density maps are generated through spatial aggregation and data smoothing to quantify active areas of customer hand movements. The generation of the spatial density map consists of two sub-steps: calculation of the initial spatial density map and smoothing processing.

[0053] 1-2-1, Calculation of the initial spatial density map;

[0054] First, the image plane is divided into multiple grid cells, each with a fixed area. For example, the entire image plane is divided into several equal parts horizontally and vertically, forming a regular grid matrix. Next, a time window is selected, for example, all event data within the last 50 milliseconds, and the number of events occurring within each grid cell is counted. Specifically, for each data item in the event stream, its grid cell is determined based on its horizontal and vertical coordinates, and the event count for that grid cell is incremented by one. After the statistics are completed, the event density of each grid cell is calculated by dividing the number of events in the grid cell by the length of the time window, and then dividing by the area of ​​the grid cell, to obtain the event frequency per unit time and unit area. This frequency is the event density of that grid cell, and the event densities of all grid cells together constitute the initial spatial density map.

[0055] The initial spatial density map visually reflects the distribution of hand movements on the image plane, with grid cells of higher event density corresponding to areas of frequent hand movements. However, due to potential noise interference or isolated events recorded by the event camera, the initial spatial density map may contain local fluctuations, which could affect the accuracy of subsequent processing if used directly. Therefore, further smoothing is required to optimize the data.

[0056] 1-2-2, Smoothing of spatial density maps;

[0057] To improve the stability and reliability of the data, Gaussian filtering is applied to smooth the initial spatial density map. Gaussian filtering generates a smoothed spatial density map by weighted averaging of the event density values ​​of each grid cell and its surrounding grid cells. The specific calculation process is as follows: For each grid cell on the image plane, neighboring grid cells within a certain range around that grid cell are selected, for example, all grid cells within a square area with a perimeter twice the grid's side length and the central grid cell as the origin. Next, the weight of each neighboring grid cell is calculated using a Gaussian function. The weight is inversely proportional to the distance from the neighboring grid cell to the central grid cell; the closer the distance, the greater the weight. The calculation logic is as follows: for each neighboring grid cell, the weight is equal to the value of the Gaussian function, which uses distance as the independent variable and the standard deviation as a preset parameter, such as one-third of the grid's side length. Then, the event density value of each neighboring grid cell is multiplied by its corresponding weight, summed, and then divided by the sum of all weights to obtain the smoothed event density value of the central grid cell. This process is repeated for each grid cell on the image plane to finally generate the smoothed spatial density map.

[0058] The smoothed spatial density map retains the main distribution characteristics of hand movements while reducing density fluctuations caused by noise and isolated events, making the data more reflective of the overall trend of hand movements. This result provides reliable input data for subsequent path planning.

[0059] 1-3, Dynamic Constraint Benchmark;

[0060] The smoothed spatial density map is defined as the dynamic constraint benchmark, used to guide the extraction of hand movement trends in subsequent path planning. The core of the dynamic constraint benchmark lies in the density value distribution of its grid cells. These density values ​​not only indicate the active areas of hand movement, but also quantify the degree of activity through the magnitude of the density values.

[0061] Specifically, each grid cell in the smoothed spatial density map corresponds to a smoothed event density value, calculated using the aforementioned Gaussian filtering. The entire smoothed spatial density map is treated as a two-dimensional distribution map, where the position of each grid cell is determined by the x-coordinate and y-coordinate, and its density value reflects the hand movement activity at that location. The subsequent arc segment decomposer utilizes this distribution map to identify regions of significant density change by analyzing the density value differences between adjacent grid cells. The identification process is as follows: for each grid cell, the difference between its density value and that of its adjacent grid cells is calculated. If the difference exceeds a preset threshold, a density gradient is considered to exist, indicating a significant change in hand movement in that region. Along these density gradient directions, features representing hand movement trends, such as tangent vectors and energy distribution, are extracted.

[0062] By transforming the event stream into a smoothed spatial density map and defining it as a dynamic constraint baseline, the event vision unit achieves the conversion from raw data to structured information. This process ensures that the robot can perceive the subtle movements of the customer's hand in real time and transmit the perception results to the path planning module in a quantified form, avoiding planning errors caused by data instability or perception delays.

[0063] The complete processing of the event vision unit includes three steps: event stream capture, spatial density map generation, and dynamic constraint baseline definition. The event camera first captures event data triggered by the customer's hand movements, forming an event stream containing information on time, location, and polarity. Then, an initial spatial density map is generated through mesh generation and event density calculation, and smoothed using Gaussian filtering to produce a smoothed spatial density map. Finally, the smoothed spatial density map is defined as the dynamic constraint baseline, providing dynamic distribution information of hand movements for path planning. This process fully utilizes the high temporal resolution of the event camera and the quantization capability of the spatial density map to ensure that the robot can accurately adapt to the dynamic changes in the customer's hand.

[0064] The event vision unit captures the event stream triggered by the customer's hand micro-movements with millisecond-level precision using an event camera, and generates a smoothed spatial density map as a dynamic constraint benchmark for hand movements. This benchmark provides the basis for the robot to perceive the rapid and varying amplitude micro-swings of the customer's arm within the narrow cup-collecting channel. However, relying solely on the spatial density map cannot fully characterize the instantaneous directionality and energy distribution of hand movements, especially when dealing with high-frequency micro-movements, lacking precise quantification of hand movement trends and energy intensity. Therefore, an arc decomposition unit is introduced to extract key information reflecting the dynamic characteristics of hand movements through in-depth processing of the event stream, providing precise quantitative input for subsequent risk assessment and path planning.

[0065] The processing of the arc segment decomposition unit aims to transform the smoothed spatial density map generated by the event vision unit into an instantaneous tangent vector set and micro-motion energy spectrum to support risk assessment and robot path planning.

[0066] 2-1, Density gradient calculation;

[0067] In narrow cup-collecting channels, the active boundaries of customers' micro-hand movements need to be quickly identified to provide information on spatial variation trends and support subsequent event flow analysis. Density gradient calculation determines the direction and intensity of hand movements by analyzing the rate of change of the smoothed spatial density map.

[0068] The calculation process begins with the smoothed spatial density map generated from the event visual units. The density map is divided into multiple grid cells, each with a defined smoothed event density value. The horizontal gradient component is calculated based on the density difference between adjacent grid cells: the smoothed event density value of the right-hand adjacent grid cell is subtracted from the smoothed event density value of the left-hand adjacent grid cell, and the difference is divided by twice the horizontal spacing between grid cells to obtain the horizontal density change rate. The vertical gradient component is calculated similarly: the smoothed event density value of the upper adjacent grid cell is subtracted from the smoothed event density value of the lower adjacent grid cell, and the difference is divided by twice the vertical spacing between grid cells to obtain the vertical density change rate. The density gradient vector of each grid cell consists of both horizontal and vertical gradient components, representing the direction and magnitude of the event density change within that grid cell.

[0069] Through the above calculations, the density gradient vector can reflect the spatial variation trend of hand movements, pointing to the region where the event density increases the fastest, thereby quantifying the active boundary of hand movements and providing accurate spatial basis for event flow segmentation.

[0070] 2-2, Event Flow Segmentation and Arc Extraction;

[0071] Segmenting the event flow based on density gradient direction allows hand movements to be decomposed into locally continuous trajectory segments, facilitating the analysis of their dynamic characteristics. Arc segment extraction, by identifying the continuity of motion trends, provides data support for the calculation of instantaneous tangent vectors and micro-motion energy spectra.

[0072] In practice, the magnitude of the density gradient vector is first analyzed. For the density gradient vector of each grid cell, the sum of the squares of its horizontal and vertical gradient components is calculated, and the square root is taken to obtain the magnitude of the density gradient vector. This magnitude is compared with a preset threshold. If the magnitude exceeds the preset threshold, the corresponding grid cell is marked as a cutting point.

[0073] At each cut point, the normal direction of the density gradient vector is determined. The normal direction is defined by the perpendicular direction of the density gradient vector. If the density gradient vector consists of horizontal and vertical components, the normal direction is obtained by swapping the components and adjusting their signs. The event stream is cut along the normal direction, and continuous event sequences parallel to the normal direction are extracted to form arcs. Each arc contains a series of event data, recording the trigger time, spatial coordinates, and polarity information.

[0074] After the arc segment extraction is completed, the event stream is divided into multiple segments reflecting the local trajectory of hand movement, ensuring that subsequent analysis can accurately focus on continuous dynamic features.

[0075] 2-3, Extraction of instantaneous tangent vector set;

[0076] Instantaneous directional information of hand movements is crucial in robot path planning and can be used to avoid high-frequency micro-motion regions. Instantaneous tangent vector sets generate a dataset reflecting the real-time direction of hand movements by analyzing the spatiotemporal distribution of events within an arc segment.

[0077] The processing unfolds for the event position sequence of each arc segment. First, the spatial coordinates of all events within the arc segment are collected and arranged in chronological order of trigger time. Linear regression is then performed on these spatial coordinates to determine an optimally fitted straight line. Specifically, the average position of all event spatial coordinates is calculated as a reference point; then, the direction angle of the line is determined by minimizing the sum of the squared distances from each event position to the line. The direction angle represents the angle between the fitted line and the horizontal axis. Based on this direction angle, a unit tangent vector is generated, representing the direction of motion of the arc segment. The horizontal component of the unit tangent vector is the cosine of the direction angle, and the vertical component is the sine of the direction angle. The unit tangent vectors of all arc segments are aggregated to form an instantaneous tangent vector set, reflecting the instantaneous direction of hand movement in different local regions.

[0078] The generation of instantaneous tangent vector sets provides a real-time quantitative description of the hand movement direction, enabling the robot to adjust its path based on dynamic trends.

[0079] 2-4, Calculation of micro-motion energy spectrum;

[0080] The frequency and amplitude of hand micro-movements directly affect liquid surface stability and need to be quantified using energy indicators. Micro-motion energy spectrum analysis calculates the dynamic intensity of hand movements by analyzing the spatiotemporal characteristics of the arc segment, providing crucial data for risk assessment.

[0081] The calculation process begins with the event trigger time of each arc segment. The trigger times of all events within the arc segment are extracted, arranged chronologically, and the time intervals between adjacent events are calculated to generate a time interval sequence. The arithmetic mean of this sequence is calculated to obtain the average time interval. The micro-motion frequency is defined as the reciprocal of the average time interval, representing the number of events occurring per unit time. Next, the micro-motion amplitude is calculated as follows: the average position of the spatial coordinates of all events within the arc segment is determined; for each event, the Euclidean distance from its spatial coordinates to the average position is calculated, which is the square root of the sum of the squares of the horizontal and vertical distances; the root mean square of the Euclidean distances of all events is taken as the micro-motion amplitude. The micro-motion energy is defined as the product of the micro-motion frequency and the micro-motion amplitude, representing the dynamic intensity of the arc segment. The micro-motion energy values ​​of all arc segments are summarized to form a micro-motion energy spectrum, reflecting the energy distribution of hand movements in different local areas.

[0082] Micro-motion energy spectrum, by integrating frequency and amplitude information, accurately quantifies the dynamic characteristics of hand micro-movements, providing a reliable basis for assessing movement risks.

[0083] The complete processing of the arc segment decomposition unit includes four steps: density gradient calculation, event flow cutting and arc segment extraction, instantaneous tangent vector set extraction, and micro-motion energy spectrum calculation. First, the smoothed spatial density map is analyzed to calculate the density gradient vector and identify the active boundaries of hand movements. Then, the event flow is cut along the density gradient direction to extract local trajectory arc segments. Next, linear regression is performed on the arc segments to generate instantaneous tangent vector sets. Finally, the micro-motion energy spectrum is calculated by analyzing the spatiotemporal characteristics of the arc segments. This process transforms the high-frequency micro-movements of the customer's hand within the narrow cup-collecting channel into quantifiable direction and energy data, ensuring that the robot can accurately perceive movement trends and intensity, supporting smooth cup delivery in crowded environments.

[0084] The event vision unit captures the event stream triggered by the customer's hand micro-movements with millisecond-level precision and generates a spatial density map. The arc decomposition unit then cuts the event stream along the density gradient, extracts the instantaneous tangent vector set, and calculates the micro-movement energy spectrum, providing high-precision dynamic information input for subsequent processing. However, relying solely on this perception and decomposition data is insufficient to directly identify high-risk areas in hand micro-movements and set safety constraints for robot path planning. Therefore, the risk assessment unit becomes crucial. By analyzing the multi-scale characteristics of the micro-movement energy spectrum, it identifies potential high-frequency disturbance areas and defines path generation boundaries to ensure that the robot can both avoid risks and maintain liquid surface stability when responding to rapid hand micro-movements within a narrow cup-collecting channel.

[0085] The risk assessment unit's processing aims to identify high-risk areas in hand micro-movements based on the micro-motion energy spectrum generated by the arc segment decomposition unit, through multi-scale analysis and threshold setting, and to provide dynamic constraints for robot path planning.

[0086] 3-1, Multi-scale window settings;

[0087] In the narrow cup-collecting channel, the frequency and amplitude of customers' hand micro-movements change rapidly, making it difficult for analysis on a single time scale to simultaneously capture instantaneous disturbances and overall trends. Therefore, it is necessary to define a series of time windows of different lengths to perform hierarchical analysis of the micro-motion energy spectrum, in order to take into account both the sensitivity to local high-frequency changes and the stability of long-term motion trends.

[0088] In practice, a set of time windows is first defined, containing multiple preset time window lengths. These time window lengths increase from short to long; for example, the shortest time window length is set to milliseconds to capture rapidly changing instantaneous disturbances, while the longer time window length is set to hundreds of milliseconds to observe the overall trend of motion. Each time window corresponds to a specific time scale. The system extracts energy data within the corresponding time range from the micro-motion energy spectrum based on these time windows for subsequent feature analysis. The extraction process is as follows: determine the current moment, then trace back the time range corresponding to the time window length, and collect the micro-motion energy spectrum data generated within that range.

[0089] By setting up this multi-scale window, the dynamic characteristics of hand movements can be analyzed at different time granularities, ensuring that subsequent energy feature extraction can fully reflect the diversity of micro-movements and improve the accuracy of risk assessment.

[0090] 3-2, Analysis of micro-motion energy spectrum;

[0091] The micro-motion energy spectrum contains energy information of each arc segment generated by the arc segment decomposition unit, but without processing, it cannot directly reflect the dynamic intensity and fluctuation characteristics of hand movements. In order to quantify the intensity distribution of hand micro-movements and provide numerical basis for risk assessment, it is necessary to analyze the micro-motion energy spectrum by calculating the energy characteristics within each time window.

[0092] In practice, for each time window in the time window set, the set of arc segment energy values ​​generated within the time range of that time window before the current moment is extracted. The extraction process is as follows: first, the current moment is determined, then the time range of that time window is traced back, and the energy values ​​of all arc segments generated within this time period are collected. Next, the average energy intensity within that time window is calculated. Specifically, the energy values ​​of all collected arc segments are added together to obtain the total energy, and then the total energy is divided by the total number of arc segments to obtain the average value. In addition, the energy fluctuation amplitude is also calculated. Specifically, for the energy value of each arc segment, the absolute difference between its value and the average energy intensity is calculated, i.e., the absolute value is obtained by subtracting the average energy intensity from the arc segment energy value; then the absolute differences of all arc segments are added together to obtain the total difference, and then the total difference is divided by the total number of arc segments to obtain the fluctuation amplitude.

[0093] By calculating the average energy intensity and fluctuation amplitude, the dynamic characteristics of the micro-motion energy spectrum at different time scales can be quantified, providing multi-dimensional input for subsequent threshold calculations. This ensures that the evaluation process can simultaneously consider the intensity and irregularity of motion, thereby improving the adaptability of path planning.

[0094] 3-3, Determination of trajectory tension threshold;

[0095] Robot paths need to remain stable under the dynamic influence of hand micro-movements, but simple energy features are insufficient to directly guide path constraints. To set the dynamic boundaries of the path, limit the robot's range of motion, and avoid jitter caused by micro-movements, it is necessary to calculate the trajectory tension threshold by integrating energy features from multiple scale windows.

[0096] In practice, for each time window in the time window set, a trajectory tension threshold is calculated based on the average energy intensity and fluctuation amplitude within that time window. The calculation method involves a weighted sum of the average energy intensity and fluctuation amplitude, where the fluctuation amplitude is multiplied by a preset adjustment coefficient to control the sensitivity of the trajectory tension threshold to the fluctuation amplitude. The adjustment coefficient is preset by the system and is used to adjust the response of the trajectory tension threshold to motion irregularities. After calculating the trajectory tension threshold for each time window in the time window set, the maximum value is selected from all the trajectory tension thresholds and determined as the final trajectory tension threshold range.

[0097] By setting a trajectory tension threshold, a unified dynamic constraint is provided for path planning, ensuring that the robot can maintain stability and safety under the most drastic motion changes, thereby improving the reliability of the cup delivery process.

[0098] 3-4, High-frequency disturbance zone marking;

[0099] High-frequency disturbances in hand micro-movements pose a significant threat to liquid surface stability due to their high frequency and large amplitude, and need to be explicitly avoided in path planning. To identify and locate these risk areas and provide precise avoidance targets for path generation, it is necessary to mark the high-frequency disturbance areas by comparing the arc energy in the micro-movement energy spectrum with the trajectory tension threshold.

[0100] In practice, the energy value of each arc segment is checked one by one in the micro-motion energy spectrum, and arc segments with energy values ​​exceeding the trajectory tension threshold are selected. The selection process is as follows: compare the energy value of each arc segment with the trajectory tension threshold. If the energy value of an arc segment is greater than the trajectory tension threshold, the spatial position of that arc segment is recorded as a high-frequency disturbance region. The spatial position of an arc segment is represented by its center coordinates. The center coordinates are calculated by collecting the spatial coordinates of all events within the arc segment, adding the values ​​of these spatial coordinates in each dimension, dividing by the total number of events, and obtaining the average value, which is the center coordinate.

[0101] By marking high-frequency disturbance areas, clear avoidance targets are provided for the path generation unit, ensuring that the robot path can bypass high-risk areas of hand movement, thereby improving the smoothness of the cup delivery process and the stability of the liquid surface.

[0102] The complete processing of the risk assessment unit includes four steps: multi-scale window setting, micro-motion energy spectrum analysis, trajectory tension threshold determination, and high-frequency disturbance zone marking. First, a set of time windows is defined, and the micro-motion energy spectrum is analyzed hierarchically. Then, the average energy intensity and fluctuation amplitude are calculated within each time window. Next, the trajectory tension threshold is calculated based on these characteristics, and the trajectory tension threshold range is determined. Finally, by comparing the arc segment energy with the trajectory tension threshold range, high-frequency disturbance zones are marked. This process provides the robot with precise risk boundaries and avoidance areas within narrow cup-collecting channels, ensuring that path planning can effectively address the impact of high-frequency micro-movements of customers' hands, while maintaining liquid surface stability and tray posture stability.

[0103] In the dynamic path planning scenario of a self-service beverage robot, the robot needs to cope with the dynamic disturbances caused by the high-frequency micro-movements of customers' hands within a narrow cup-collecting channel, while ensuring the stability of the liquid surface in the cup to prevent liquid spillage or delivery delays due to path vibration. The initial event vision unit captures the hand-triggered event flow and generates a spatial density map. The arc decomposition unit extracts the instantaneous tangent vector set and micro-motion energy spectrum along the density gradient. The risk assessment unit analyzes the micro-motion energy spectrum based on a multi-scale window to determine the trajectory tension threshold and mark high-frequency disturbance areas, providing precise boundary constraints for path generation. However, relying solely on this perception and assessment data is insufficient to directly generate a path that avoids high-frequency disturbance areas while maintaining liquid surface stability. Therefore, the trajectory generation unit must construct a smooth path within the given tension threshold and integrate dynamic and geometric indicators to ultimately select the optimal trajectory to meet the requirement of smooth cup delivery within the narrow channel.

[0104] The trajectory generation unit's processing aims to generate smooth paths, ensuring that the paths are within the trajectory tension threshold. At the same time, candidate paths are evaluated through dynamic and geometric indices, and the optimal path is finally selected.

[0105] 4-1, Construction of a family of smooth acceleration splines;

[0106] In narrow cup-retrieving channels, the robot's path needs to avoid liquid surface sloshing caused by sudden acceleration changes, while ensuring it does not enter high-frequency disturbance regions. To this end, a spline path with continuous acceleration is constructed using mathematical methods to generate a set of smooth candidate paths that satisfy the trajectory tension threshold constraint, providing diverse options for subsequent evaluation.

[0107] First, the trajectory tension threshold output by the risk assessment unit is used as the boundary to define the path generation region, ensuring that candidate paths do not cross high-frequency disturbance areas. Next, a set of candidate paths is generated using B-spline curves, with each path parameterized by a set of control points. To ensure path smoothness, a fourth-order B-spline is chosen, characterized by its continuous second derivative across the entire domain, thus resulting in gradual acceleration changes along the path. The calculation process is as follows: based on the boundary conditions of the trajectory tension threshold, the initial distribution of control points is determined; then, by adjusting the spatial positions of the control points, multiple paths are generated, each defined by a fourth-order B-spline curve; finally, a set of candidate paths is formed, where all paths satisfy the constraints of the trajectory tension threshold and exhibit continuous acceleration changes.

[0108] The candidate path set generated in this way not only mathematically guarantees the smooth change of acceleration, but also, combined with the boundary conditions of the risk assessment unit, ensures the feasibility and diversity of the paths, providing reliable input for subsequent dynamic and geometric evaluations.

[0109] 4-2, Calculation of liquid surface surging potential;

[0110] Acceleration pulses along the path directly affect liquid surface stability; therefore, it is necessary to quantify their impact effect on the liquid surface to provide a dynamic basis for path selection. The surging potential of the liquid surface is calculated by combining the acceleration distribution of candidate paths with the cup parameters to determine the overall momentum of the liquid under impact.

[0111] For each candidate path, a series of discrete points are uniformly selected along the path. At each discrete point, the triaxial acceleration pulse of the path is calculated, i.e., the components of acceleration in the horizontal forward / backward, horizontal left / right, and vertical directions. These components are determined by the second derivative of the path, in meters per second squared. The calculation process is as follows: First, the acceleration components at each discrete point are calculated; then, the magnitude of the acceleration pulse is calculated by adding the squares of the acceleration components in the three directions and taking the square root of the result to obtain the scalar acceleration value, in meters per second squared. Next, the width of the liquid surface is defined as the width of the cup's cross-section, in millimeters; the scalar acceleration value is multiplied by the width of the liquid surface to obtain the instantaneous impact momentum, in meters per second squared.

[0112] To assess the impact of the entire path on the liquid surface, the instantaneous impact momentum is integrated along the path. The calculation process is as follows: the path is divided into a series of small segments; for each segment, the product of the instantaneous impact momentum on that segment and the length of that segment is calculated, in meters per second squared (m·mm² / s²); then, the products of all segments are summed to obtain the total impact momentum, in m·mm² / s². Finally, the total impact momentum is normalized to the volume of the cup by dividing the total impact momentum by the volume of the cup (in cubic millimeters) to obtain the surging potential of the liquid surface, in meters per second squared (m² / s²).

[0113] The potential for liquid surface surging is calculated by coupling path acceleration with the physical parameters of the cup, accurately reflecting the potential impact of each path on the stability of the liquid surface, and providing a dynamic quantitative standard for subsequent path optimization.

[0114] 4-3, Spline tortuosity loss calculation;

[0115] The geometric curvature and angular velocity variations of a path affect drive power consumption and smoothness; therefore, these characteristics need to be quantified to provide a geometric basis for path selection. Spline tortuosity loss, combined with curvature accumulation and angular velocity, is used to assess path complexity.

[0116] For each candidate path, calculate the curvature at the path parameters. The curvature is determined by the ratio of the second derivative to the first derivative of the B-spline curve, reflecting the degree of path curvature at that point. The calculation process is as follows: Select a series of discrete points along the path; at each discrete point, calculate the ratio of the second derivative to the first derivative to obtain the curvature at that point; then integrate the curvature along the entire path. Specifically, multiply the curvature of each discrete point by the path length between adjacent discrete points, and then sum all the products to obtain the cumulative curvature, expressed in radians, representing the total curvature of the path.

[0117] Simultaneously, the average tangential angular velocity of the path is calculated. Tangential angular velocity represents the rate of change of the tangential direction of the path with time, measured in radians per second. The calculation process is as follows: First, determine the total execution time of the path, measured in seconds; then calculate the tangential angular velocity at each discrete point by dividing the rate of change of the first derivative of the path over time by the rate of change of the path length; next, integrate the tangential angular velocity over the execution time by multiplying the tangential angular velocity at each discrete point by the time interval between adjacent discrete points, and then summing all the products; finally, divide the integral result by the total time to obtain the average tangential angular velocity, measured in radians per second.

[0118] Finally, the cumulative curvature is multiplied by the average tangential angular velocity to obtain the spline tortuosity cost, measured in radians squared per second. The spline tortuosity cost, through the product of curvature and angular velocity, quantifies the geometric cost and execution smoothness of the path, providing an independent geometric standard for subsequent comprehensive evaluation.

[0119] 4-4, Derivation of steady-state optimality coefficients;

[0120] To select the optimal path from the candidate path set, it is necessary to comprehensively consider the surface surging potential and spline tortuosity loss, balancing dynamic stability and geometric smoothness. Steady-state optimization coefficients are adaptively derived using a machine learning model to achieve path selection.

[0121] A multi-layer ensemble learner is constructed and trained offline using historical path data and liquid surface response data to learn the mapping relationship between liquid surface surging potential, spline tortuosity loss, and path stability.

[0122] The construction process of the multi-layer ensemble learner is as follows: First, historical path data and liquid surface response data are collected as a training set. Each record includes the liquid surface surging potential, spline tortuosity loss, and corresponding liquid surface stability label for the path. Then, the training set is divided into multiple subsets, and multiple base learners are trained using different algorithms (such as support vector machines, neural networks, and decision trees). Next, a meta-learner is trained by stacking the outputs of each base learner as input, using a logistic regression model. Finally, the parameters of the base learners and the meta-learner are adjusted to ensure that the multi-layer ensemble learner can accurately predict path stability and output steady-state optimization coefficients based on the liquid surface surging potential and spline tortuosity loss of the candidate paths during runtime.

[0123] The computation process is as follows: First, collect data on the surface surging potential, spline tortuosity loss, and corresponding surface stability of historical paths. Then, input this data into a multi-layer ensemble learner, and adjust the model parameters through multiple iterations to accurately predict path stability. During runtime, the surface surging potential and spline tortuosity loss of each candidate path are input into the trained multi-layer ensemble learner. Based on the input data, the learner outputs a steady-state optimization coefficient, which is a dimensionless value; a higher value indicates better overall path stability. The path with the highest steady-state optimization coefficient is selected as the final path and output to the execution scheduling unit.

[0124] The steady-state optimization coefficient integrates dynamic and geometric indicators in a data-driven manner to ensure that the selected path achieves the optimal balance between liquid surface stability and execution smoothness.

[0125] The complete process of the trajectory generation unit includes four steps: constructing a family of smoothed acceleration splines, calculating the surface surging potential, calculating spline tortuosity loss, and deriving steady-state optimization coefficients. The system first generates a set of smooth candidate paths within the trajectory tension threshold; then, it quantifies the impact of the path on surface stability using the surface surging potential and evaluates the path's geometric complexity using spline tortuosity loss; finally, it uses a multi-layer ensemble learner to derive steady-state optimization coefficients and select the optimal path. This process ensures that the robot generates a path within the narrow cup-retrieving channel that avoids high-frequency disturbances while maintaining surface stability, providing reliable input to the execution scheduling unit and achieving smooth cup delivery.

[0126] The trajectory generation unit, through the collaborative processing of the event vision unit, arc decomposition unit, and risk assessment unit, derives the path with the highest steady-state optimization coefficient based on the dual-dimensional indices of liquid surface surging potential and spline tortuosity loss using multi-layer ensemble learning. This path is output as discrete points, representing the candidate trajectory with the lowest amplitude risk. However, the robot drive requires continuous pose commands to achieve smooth cup delivery, avoiding liquid surface sloshing and tray instability, while also handling disturbances from high-frequency micro-movements of the customer's hands within the narrow cup-retrieving channel. Therefore, the execution scheduling unit must transform the discrete optimal path into continuous poses and generate an executable command stream to ensure the smoothness and real-time performance of the cup delivery process.

[0127] The complete processing of the execution scheduling unit aims to transform the optimal path output by the trajectory generation unit into an executable instruction stream, ensuring that the robot smoothly delivers cups in narrow cup-collecting channels while responding to disturbances from the high-frequency micro-movements of customers' hands.

[0128] 5-1, Path interpolation;

[0129] The optimal path output by the trajectory generation unit consists of a series of discrete spatial location points arranged in the order of the cup delivery task. To ensure that the robot can achieve a smooth motion trajectory during movement, these discrete spatial location points need to be connected into a continuous path curve to ensure the continuity of the robot's position, velocity, and acceleration, thereby reducing the impact of instantaneous acceleration on the liquid surface.

[0130] A continuous path curve is generated using the quintic Hermite interpolation method. Quintic Hermite interpolation is a mathematical technique that constructs a smooth quintic polynomial curve based on given discrete spatial locations and their corresponding velocities and accelerations. In practice, the time node corresponding to each discrete spatial location is first determined, i.e., the time when the robot arrives at that discrete spatial location during the cup delivery task. Next, using information provided by the trajectory generation unit, the velocity and acceleration values ​​for each discrete spatial location are set. The velocity value is preset based on the micro-motion energy spectrum to accommodate the disturbance region of the customer's high-frequency micro-movements; the acceleration value is determined based on the constraint condition of the liquid surface surging potential to limit the impact of instantaneous acceleration on the liquid surface. A quintic polynomial curve is constructed between each pair of adjacent discrete spatial locations, ensuring that the curve's position at each discrete spatial location matches the input discrete spatial location coordinates, the velocity matches the preset velocity value, the acceleration matches the preset acceleration value, and a smooth transition is achieved at the connection points between adjacent discrete spatial locations. The process of constructing a fifth-order polynomial curve includes: first, calculating the positional difference between two adjacent discrete spatial points, in millimeters; then, calculating the time difference between time points, in seconds; next, determining the coefficients of the polynomial curve based on the velocity and acceleration values, ensuring that the position, velocity, and acceleration of the curve at the start and end points all meet preset conditions, and guaranteeing the smoothness of the curve through the properties of a fifth-order polynomial. This method generates a continuously differentiable path curve throughout the entire cup-delivery task time, with each positional component of the path curve in millimeters and the time component in seconds.

[0131] The continuous path curve generated by five Hermite interpolations can achieve a smooth transition in the time domain, avoiding abrupt changes between discrete spatial locations, providing stable input data for subsequent pose command generation, and ensuring the smoothness of robot motion.

[0132] 5-2, Pose command generation;

[0133] The continuous path curve needs to be transformed into a discrete pose command stream to meet the real-time execution requirements of the robot's drive. The pose command includes not only the robot's position information but also its orientation information to ensure the stability of the tray in narrow cup-collecting channels and to accommodate interference from the customer's micro-movements.

[0134] First, a fixed control cycle is set, for example, each cycle lasting 10 milliseconds. The control cycle defines the time interval for the robot to execute commands. Based on the total duration of the cup-delivery task and the control cycle, a time step sequence is determined, i.e., a series of discrete time points, each corresponding to a pose command. At each time step, the position component at the corresponding time point is obtained by sampling from the continuous path curve. The specific sampling process is as follows: based on the time value of the time step, the corresponding position coordinates are calculated in the fifth-order polynomial expression of the continuous path curve. Next, the attitude component at that time point is calculated, representing the spatial orientation of the tray. The calculation of the attitude component is based on the geometric characteristics of the path curve: first, the tangential vector of the path curve at that time point is calculated, i.e., the velocity direction obtained by differentiating the fifth-order polynomial curve; then, the length of the tangential vector is normalized to obtain the forward direction; next, the normal vector of the path curve is calculated, and the normal direction is determined by differentiating the tangential vector again and combining it with the curvature characteristics, reflecting the curvature of the path curve; finally, combined with the preset tray stabilization normal, a three-dimensional rotation matrix or quaternion form of attitude component is constructed. The specific construction process is as follows: the tangential vector is used as the forward direction axis, the normal vector as the bending direction axis, and the pallet stabilization normal as the vertical direction axis. These three axes are orthogonal to each other, forming a complete spatial orientation description. The system combines the position and attitude components of each time step into a pose command, and arranges them in chronological order to form a pose command stream.

[0135] The pose command flow discretizes the continuous path curve into an executable sequence of instructions, ensuring that the robot can move in real time according to the instructions, while maintaining the stability of the tray through posture adjustment and reducing the possibility of liquid sloshing.

[0136] 5-3, Drive-end control;

[0137] The pose command stream needs to be converted into control signals for the robot's drive end to drive the robotic arm or mobile platform to complete the cup delivery task. This process must ensure the precise execution of the commands and use a feedback mechanism to handle real-time disturbances caused by the customer's micro-movements.

[0138] The pose command stream is sent to the driver one by one according to the control cycle, and the driver receives one pose command at each time step. Through inverse kinematics (IK), the pose commands are converted into joint angles or wheel speeds, depending on the robot's mechanical structure. For a robotic arm, IK converts the position and attitude components into angle values ​​for each joint, in radians; for a mobile platform, IK converts the position and attitude components into wheel speeds, in millimeters per second. The IK process includes: calculating the mapping relationship from the end-effector pose to joint parameters based on the robot's geometric parameters and motion constraints, ensuring kinematic consistency. Next, a closed-loop control method is used to generate motor control signals. The core of closed-loop control is to monitor the robot's actual pose in real time and compare it with the pose in the command, calculating the path tracking error. Specifically, sensors measure the robot's actual position and attitude at each time step, forming actual pose data; then, the actual pose is compared with the commanded pose at the corresponding time step, and the difference is calculated to obtain the tracking error. The tracking error is input into a proportional-derivative (PDC) controller to generate control signals.

[0139] The calculation method for the proportional-derivative (PD) controller is as follows: First, the tracking error is multiplied by a preset proportional gain to obtain the proportional term, which is a dimensionless parameter. Then, the rate of change of the tracking error over time, i.e., the derivative of the error, is calculated, and this rate of change is multiplied by a preset differential gain to obtain the differential term. Finally, the proportional term and the differential term are added together to obtain the final control signal. The drive motor adjusts its motion according to the control signal, correcting the robot's pose in real time to ensure that the tracking error is controlled within the range required for liquid surface stability.

[0140] The closed-loop control mechanism enables the robot to accurately follow the pose command flow through error feedback and dynamic adjustment, maintaining the stability and reliability of the cup delivery process even under the disturbance of high-frequency micro-movements of the customer's hands.

[0141] The complete processing of the scheduling unit includes three steps: path interpolation, pose command generation, and drive-end control. First, the system transforms the discrete optimal path into a continuous path curve through five Hermite interpolations. Then, it discretizes the continuous path curve into a pose command stream. Finally, through inverse kinematics and closed-loop control, it converts the pose command stream into control signals for the robot's drive end, achieving smooth cup delivery. This process ensures that the robot, within the narrow cup-collecting channel, follows the optimal path with the highest steady-state selection coefficient, responding to disturbances from the customer's high-frequency micro-movements while maintaining liquid surface stability and tray posture stability, providing technical support for efficient service.

[0142] Example 2: Figure 2 The present invention provides a dynamic path planning method for a self-service beverage robot, comprising:

[0143] S1: Capture the hand-triggered event stream and generate a spatial density map as a dynamic constraint baseline.

[0144] S2: Cut the event flow along the density gradient, extract the instantaneous tangent vector set and calculate the micro-motion energy spectrum to provide quantitative input for subsequent evaluation.

[0145] S3: Analyze the micro-motion energy spectrum in a multi-scale window to determine the trajectory tension threshold and mark the high-frequency disturbance region to limit the path generation boundary.

[0146] S4: Construct a family of acceleration smooth splines within the tension threshold, calculate the liquid surface surging potential and spline tortuosity loss respectively, derive the steady-state optimization coefficient through machine learning, and select the candidate trajectory with the lowest amplitude risk accordingly.

[0147] S5: The path with the highest steady-state optimization coefficient is interpolated in real time to the continuous pose, and the driving end uses this to complete the smooth cup delivery.

[0148] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0149] It should be noted that the system of the present invention can be deployed on the device itself to realize embedded applications, or it can run on a PC or other terminal with a user interface, thereby meeting a variety of hardware environments and usage requirements.

[0150] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0151] It should be noted that, in this document, the use of relational terms such as "first" and "second" is merely to distinguish one entity or operation from another, and does not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0152] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A dynamic path planning system for a self-service beverage robot, characterized in that, Comprise: Event vision unit: capture hand trigger event stream, and generate spatial density map as dynamic constraint reference; Arc segment decomposition unit: cut event stream along density gradient, extract instantaneous tangent vector set and calculate micro-motion energy spectrum, provide quantitative input for subsequent evaluation; Risk assessment unit: analyze micro-motion energy spectrum in multi-scale window, determine trajectory tension threshold domain and mark high-frequency disturbance area to limit path generation boundary; Trajectory generation unit: construct acceleration smooth spline family within tension threshold domain, respectively calculate liquid surface overturning potential and spline tortuosity consumption, derive stable state optimization coefficient through machine learning, and select candidate trajectory with lowest amplitude risk accordingly; Execution scheduling unit: real-time interpolate stable state optimization coefficient highest path into continuous pose, drive end to complete smooth cup delivery accordingly; Event vision unit includes the following processing content: Real-time capture light intensity change caused by customer hand movement by using event camera, generate event stream containing timestamp, spatial position and polarity; divide image plane into grid, count the number of events in each grid unit in the nearest unit time, calculate event density in unit time and unit area, generate initial spatial density map; Apply Gaussian filter smoothing processing to initial spatial density map, generate smoothed spatial density map as dynamic constraint reference; quantify hand movement active area through smoothed event density value of grid unit; Arc segment decomposition unit includes the following processing content: Analyze smoothed spatial density map, calculate density gradient vector of each grid unit, identify active boundary of hand movement; take grid unit with density gradient vector amplitude exceeding preset threshold as cutting point, cut event stream along normal direction of density gradient vector, extract local trajectory arc segment; Perform linear regression on spatial coordinates of events in local trajectory arc segment, generate instantaneous tangent vector set, reflect instantaneous direction of hand movement; calculate micro-motion frequency and micro-motion amplitude of events in local trajectory arc segment, generate micro-motion energy spectrum, quantify dynamic intensity of hand movement.

2. The dynamic path planning system of a self-service beverage robot according to claim 1, characterized in that, Risk assessment unit includes the following processing content: Define time window set, perform multi-scale analysis on micro-motion energy spectrum; for each time window in time window set, extract arc segment energy value set generated in corresponding time window length time range before current time, calculate energy average intensity and fluctuation amplitude of arc segment energy value set; Calculate trajectory tension threshold based on energy average intensity and fluctuation amplitude, select maximum value of trajectory tension thresholds of all time windows in time window set as trajectory tension threshold domain; screen arc segments with energy value exceeding trajectory tension threshold domain in micro-motion energy spectrum, record spatial positions of these arc segments as high-frequency disturbance area.

3. The dynamic path planning system of a self-service beverage robot according to claim 2, characterized in that, Trajectory generation unit includes the following processing content: Generate B-spline candidate path set with continuous acceleration within trajectory tension threshold domain; for each candidate path, uniformly select discrete points along the path, calculate three-axis acceleration impulse modulus at discrete points and multiply by cup liquid surface width to obtain instantaneous impact momentum, integrate instantaneous impact momentum along the path and normalize with cup capacity to obtain liquid surface overturning potential; calculate curvature cumulative quantity of candidate path and multiply by average tangential angular velocity to obtain spline tortuosity consumption.

4. The dynamic path planning system of a self-service beverage robot according to claim 3, characterized in that, The trajectory generation unit further includes the following processing content: The liquid surface overturning potential and the spline bending loss are input into a multi-layer integrated learner to output a steady-state optimal coefficient, and a path with the highest steady-state optimal coefficient is selected as the final path.

5. The dynamic path planning system of the self-service beverage robot according to claim 4, characterized in that, The execution scheduling unit includes the following processing content: The discrete optimal path output by the trajectory generation unit is converted into a continuous path curve through quintic Hermite interpolation, the continuous path curve is discretized into a pose instruction stream through fixed control period sampling, the pose instruction stream is converted into joint angle or wheel speed instructions through inverse kinematics, motor control signals are generated through a closed-loop control method, and the robot motion is adjusted using a proportional-differential controller to achieve accurate execution of instructions.

6. The dynamic path planning system of a self-service beverage robot according to claim 5, characterized in that, The execution scheduling unit further includes the following processing content: The process of inverse kinematics solving includes: calculating the mapping relationship from the end pose to the joint parameters according to the geometric parameters and motion constraints of the robot mechanical structure, to ensure that the calculation result meets the kinematic consistency; then, the actual position and attitude of the robot at each time step are measured through the sensor to form the actual pose data; then the actual pose and the instruction pose at the corresponding time step are compared to calculate the difference between the two, and the tracking error is obtained.

7. The dynamic path planning system of a self-service beverage robot according to claim 6, characterized in that, The execution scheduling unit further includes the following processing content: The calculation method of the proportional-differential controller is: first, multiply the tracking error by a preset proportional gain to get the proportional term, and the proportional gain is a dimensionless parameter; then calculate the rate of change of the tracking error with time, that is, the derivative of the error, and then multiply the rate of change by a preset differential gain to get the differential term; finally, add the proportional term and the differential term to get the final control signal.

8. A dynamic path planning method for a self-service beverage robot, for implementing the dynamic path planning system for a self-service beverage robot according to any one of claims 1 to 7, characterized in that, Comprise: S1: Capture the hand trigger event stream and generate a spatial density map as a dynamic constraint reference; S2: Cut the event stream along the density gradient, extract the instantaneous tangent vector set and calculate the micro-motion energy spectrum to provide quantitative input for subsequent evaluation; S3: Analyze the micro-motion energy spectrum in a multi-scale window to determine the trajectory tension threshold domain and mark the high-frequency disturbance area to limit the path generation boundary; S4: Construct an acceleration smoothing spline family within the tension threshold domain, calculate the liquid surface overturning potential and the spline bending loss respectively, derive the steady-state optimal coefficient through machine learning, and select the candidate trajectory with the lowest amplitude risk accordingly; S5: Real-time interpolation of the path with the highest steady-state optimal coefficient into a continuous pose to drive the end to complete the smooth cup delivery.

Images