Milk foam injection method and system capable of protecting liquid level and suitable for different cup shapes

Abstract

The invention relates to the technical field of automatic coffee machines, and provides a milk foam injection method capable of protecting liquid level and suitable for different cup types, comprising the following steps: S1, acquiring cup parameters and target injection volume; s2, generating an injection control track according to the obtained parameters; and S3, executing the injection control track and adjusting the injection control track in real time until the target injection volume is reached. The automatic coffee machine can accurately adapt to various cup types, accurate volume injection is achieved by sensing cup type parameters in real time, dynamically adjusting the milk foam injection track and flow speed and combining closed-loop control, a coffee grease layer is effectively protected, and injection consistency, stability and efficiency under different cup types are remarkably improved. The sensing system provided by the invention has high flexibility; in a cost-sensitive scene, only an electronic scale can be used for gravity measurement; a 3D visual system can be used in a scene with a high precision requirement; under the optimal configuration, various sensor data can be fused; the system can automatically select an optimal sensing strategy according to actual hardware configuration.

Description

Technical Field

[0001] The present invention relates to the technical field of automatic coffee machines, and in particular to a method and system for protecting a liquid level and injecting milk foam suitable for different cup shapes. Background Art

[0002] In the application scenarios of automated coffee machines, the diversity of cup shapes poses a severe challenge to the milk froth injection system. Existing automated equipment is mostly designed for standardized cylindrical cups, and its injection trajectory and flow rate parameters are fixed in preset programs, making it difficult to adapt to the increasingly diverse cup shapes on the market - from narrow-mouthed travel cups (diameter 3-5cm) to wide-mouthed lattes (diameter 8-10cm), from straight-walled cups to conical cups (taper 5°-30°). The significant differences in geometric parameters directly lead to unstable injection effects. For example, due to the limited diameter of the narrow-mouthed cup, milk froth is prone to overflow at a fixed flow rate; due to the narrow bottom space of the conical cup, if the injection angle is unchanged, the milk froth will be unevenly distributed in the cup; and due to the large capacity of the wide-mouthed cup, a fixed injection time will result in insufficient liquid level. This reliance on a single cup shape makes automated coffee machines lack flexibility when responding to personalized orders, making it difficult to meet the adaptation needs of multiple types of cups in commercial scenarios.

[0003] Milk foam, the core fluid of automated latte art, exhibits nonlinear characteristics that further exacerbate the difficulty of controlled injection. The viscosity (0.1-0.5 Pa·s) and carbonation content (10%-50%) of milk foam vary dynamically with milk batch, whipping temperature (55-65°C), and storage time, exhibiting typical non-Newtonian fluid characteristics. Existing automated systems often use a fixed "inclination angle-flow rate" relationship, such as a preset "30° inclination angle corresponds to a 2 ml / s flow rate," which fails to account for fluctuations in the actual physical properties of the milk foam. When the carbonation content of the milk foam is too high, the actual flow rate can be 15%-20% lower than the predicted value, resulting in insufficient injection. When the carbonation content is too low, the flow rate can be 10%-15% higher, impacting the liquid surface and causing damage to the crema layer. This passive adaptation to fluid characteristics makes it difficult for automated equipment to ensure consistent injection of milk foam across batches, directly impacting the taste stability of the beverage.

[0004] The lack of real-time perception and closed-loop control capabilities is another shortcoming of existing automated injection systems. In order to achieve precise injection, the system needs to obtain the cup geometry parameters (such as cup diameter, depth, and taper) and volume changes during the injection process in real time, but most automated coffee machines only rely on a preset cup database and lack a dynamic scanning and adjustment mechanism. For example, when there is a ±5mm deviation in the placement of the cup, the fixed trajectory will cause the milk foam injection point to deviate from the center, forming an eccentric liquid surface; when there is a ±3ml error between the actual injection volume and the target value, the system cannot adjust the flow rate in time due to the lack of real-time monitoring, which ultimately leads to the liquid level being too high or too low. This "open-loop" control mode makes the automated equipment less robust when facing actual scenarios such as cup placement errors and fluctuations in milk foam volume, making it difficult to stably reproduce the preset injection effect.

[0005] In the batch production scenario of commercial automated coffee machines, the balance between efficiency and precision is particularly critical. In order to increase the number of orders processed per unit time, the injection process needs to be shortened as much as possible while ensuring accuracy; but in order to protect the coffee oil layer, the initial injection requires a low flow rate (<1ml / s) to "gently" touch the liquid surface, the flow rate needs to be increased in the middle stage (2-3ml / s) to accelerate filling, and the speed needs to be reduced again at the end (<1ml / s) to avoid overflow. Existing automation systems have difficulty in achieving smooth transitions and precise switching of flow rates, and liquid level fluctuations often occur due to sudden changes in flow rate - for example, when switching directly from a high flow rate in the middle stage to a low flow rate at the end, the milk foam flow will impact the liquid surface due to inertia, destroying the integrity of the oil layer. This lack of control over the balance of "efficiency-precision-liquid level protection" limits the application experience of automated coffee machines in scenarios with high order volumes. Summary of the Invention

[0006] In response to the above problems, the purpose of the present invention is to provide a method and system for injecting milk foam that protects the liquid level and is suitable for different cup shapes. It can enable automatic coffee machines to accurately adapt to various cup shapes. By real-time sensing of cup shape parameters, dynamic adjustment of the milk foam injection trajectory and flow rate, and combined with closed-loop control, precise volume injection can be achieved, effectively protecting the coffee oil layer, and significantly improving the consistency, stability and efficiency of injection under different cup shapes.

[0007] The above-mentioned object of the present invention is achieved through the following technical solutions:

[0008] The technical solution of the present invention is not limited to a specific robotic arm configuration and can be applied to:

[0009] 2-5 axis robotic arm: suitable for basic injection tasks;

[0010] 6-axis robotic arm: achieves complete position and posture control and is the preferred embodiment of the present invention;

[0011] 7-axis robotic arm: increases redundant degrees of freedom and improves obstacle avoidance capabilities;

[0012] Robotic arm + auxiliary device: to achieve more complex coordinated movements;

[0013] The perception system of the present invention is highly flexible:

[0014] In cost-sensitive scenarios, only electronic scales can be used for gravity measurement;

[0015] In scenarios where high precision is required, 3D vision systems can be used;

[0016] In the optimal configuration, multiple sensor data can be fused;

[0017] The system can automatically select the optimal perception strategy based on the actual hardware configuration.

[0018] A method for injecting milk foam to different cup shapes while maintaining a liquid surface, comprising the following steps:

[0019] S1: Obtain cup parameters and target injection volume;

[0020] S2: Generate injection control trajectory according to the acquired parameters;

[0021] S3: Execute the injection control trajectory and adjust in real time until the target injection volume is reached.

[0022] Furthermore, in step S2, the generated injection control trajectory is a motion trajectory of a multi-axis robotic arm, and the multi-axis robotic arm is any one of a five-axis to a seven-axis robotic arm.

[0023] Furthermore, the multi-axis robotic arm is a six-axis robotic arm, which realizes precise positioning (X, Y, Z) and posture control (Roll, Pitch, Yaw) of the end effector through the coordinated work of each degree of freedom.

[0024] Furthermore, the multi-axis robotic arm can also cooperate with one or more auxiliary motion devices, each of which has one to three degrees of freedom and is used to carry and move containers.

[0025] Furthermore, in step S1, obtaining cup parameters and target injection volume includes: sensing the milk foam injection environment, scanning the workspace through one or more sensing devices, identifying the cup position, measuring cup parameters including caliber, depth, and shape, and calculating the target injection volume, specifically:

[0026] Scan the cup using a sensing device to extract key geometric information including caliber, depth, bottom diameter, and taper, and create a digital model of the cup based on the key geometric information;

[0027] The user presets the distance between the liquid level and the cup mouth, and calculates the target injection volume based on the key geometric information;

[0028] The cup type is determined based on the key geometric information, and an adaptive control strategy is selected based on the different cup types to generate an adaptive pouring trajectory. For conical cups, the pouring angle is dynamically adjusted to compensate for the volume difference of the cup body as it changes with height. For narrow-mouth cups, a high-precision control strategy is adopted to reduce the milk foam injection flow rate to adapt to the narrow mouth characteristics. For wide-mouth cups, the milk foam injection flow rate is reduced to improve the stability of the injection process.

[0029] Among them, the perception device includes at least one of a 3D vision system, a 2D vision system, and a depth camera; at least one of a gravity sensor, an electronic scale, and a pressure sensor; and at least one of an ultrasonic sensor and a laser ranging sensor.

[0030] Furthermore, in step S2, generating an injection control trajectory based on the acquired parameters includes: planning the milk foam injection trajectory, determining a safe height of the latte art cylinder above the cup mouth at the injection starting position, loading a fluid control model to generate a latte art cylinder tilting angle trajectory sequence, and simultaneously generating a multi-axis robotic arm motion trajectory in Cartesian space, converting each trajectory point in the motion trajectory into an angle value of at least five joints of the robotic arm through inverse kinematics solution, and performing trajectory smoothing processing, specifically as follows:

[0031] Performing algorithm decoupling design on the dumping angle trajectory sequence and the robot arm motion trajectory, using different algorithms to generate the dumping angle trajectory sequence and the robot arm motion trajectory respectively, so that the system can flexibly combine different dumping strategies and motion modes;

[0032] At the same time, when latte art is required at different cup positions or angles, there is no need to regenerate the entire set of trajectories. A rotation matrix can be superimposed on the currently calculated pouring angle trajectory sequence and the robotic arm motion trajectory to achieve posture transformation.

[0033] Furthermore, before step S2, a fluid control model is designed, specifically:

[0034] The fluid control model includes at least one of the following:

[0035] Fluid dynamics models based on physical laws;

[0036] Predictive models based on machine learning;

[0037] Hybrid models of physical and machine learning models;

[0038] Empirical models based on lookup tables;

[0039] When a hybrid model of the physical model and the machine learning model is adopted, the fluid control model is generated by mixing the physical model with the single learning model;

[0040] The physical model, based on the Bernoulli equation and the Coanda effect, achieves low-velocity laminar flow in the initial injection phase by controlling the adherent flow of the liquid on the curved surface of the spout, thus avoiding damage to the coffee crema layer. The flow rate is calculated using the principles of fluid dynamics, taking into account the geometric parameters of the latte art pot, including its diameter, height, spout width, and spout curvature radius, as well as the physical properties of the fluid, including density, viscosity, and surface adhesion. The model also supports parameter configurations for latte art pots of various standard capacities.

[0041] The single-shot learning model uses domain adaptation technology to learn from a single pouring demonstration. By recording the pouring angle and weight change data, the single-shot learning model can quickly capture the behavioral characteristics of a specific fluid and generate a lookup table (LUT) to achieve real-time prediction.

[0042] The adaptive weight fusion ratio of the physical model and the single learning model is customized to perform weighted fusion of the physical model and the single learning model.

[0043] Furthermore, in step S2, the milk foam injection trajectory is planned to determine the safe height of the latte art cylinder above the cup mouth at the injection starting position, and the fluid control model is loaded to generate the latte art cylinder tilting angle trajectory sequence, specifically:

[0044] The target injection volume and expected injection time are used as inputs, and an inverse solver based on the trained hybrid fluid control model is called to reversely calculate an optimal dumping angle trajectory sequence that can achieve the target injection volume through an iterative optimization algorithm. The solution process includes:

[0045] Calculating an average flow rate according to the target injection volume and the expected injection time;

[0046] Design a trapezoidal velocity profile with the peak velocity being a preset multiple of the average flow velocity;

[0047] The flow velocity curve is converted into an angle curve by using the inverse solver of the fluid control model, and the minimum angle and maximum angle of the backflow are set as application constraints.

[0048] Furthermore, step S2 also includes establishing a multi-physics field coupled fluid modeling system to achieve precise control of the multi-physics field coupled fluid control model based on reinforcement learning, specifically:

[0049] Phase 1: Reinforcement Learning in Simulation Environment:

[0050] Build high-fidelity fluid models in the partio+SPlisHSPlasH fluid simulation environment;

[0051] Consider the temperature-viscosity coupling: μ(T,t)=μ0×exp(-k(T-T0))×(1-αt), where both temperature and time affect the fluid viscosity, μ(T,t) is the dynamic viscosity at temperature T and time t, μ0 is the reference viscosity of the fluid at the reference temperature T0 and the initial time t=0, k is the temperature influence coefficient, and α is the time attenuation coefficient;

[0052] Consider the shear thinning effect: the non-Newtonian properties of milk foam at different shear rates;

[0053] Through deep reinforcement learning DRQN / RDPG, the basic policy network is trained in a simulation environment to learn the optimal dumping strategy under different physical parameters;

[0054] Phase 2: Real-world domain adaptation learning:

[0055] Build a real experimental platform: six-axis robotic arm + high-precision electronic scale + 3D vision system;

[0056] Systematically study the behavior characteristics of three fluids, including water, milk, and milk foam, under different conditions;

[0057] Introducing multivariable control: different capacities of latte art pitchers: 350ml, 450ml, and 600ml; different grip positions: at the top of the pitcher at a high center of gravity, in the middle at the balance point, and at the bottom at a low center of gravity; each configuration combination affects the system's dynamic characteristics and fluid behavior;

[0058] Use Graph Networks to capture the complex interactions between fluid particles.

[0059] Furthermore, in step S3, executing the injection control trajectory and adjusting it in real time until the target injection volume is reached includes: performing milk foam injection execution control, moving the robotic arm to a starting position to start executing the pouring trajectory according to the pouring angle sequence, performing real-time monitoring during the pouring process, using Kalman filter volume estimation, stopping the injection when the target volume is reached, and resetting the robotic arm; otherwise, continuing the pouring trajectory after adjusting the pouring angle;

[0060] Among them, real-time monitoring is carried out during the dumping process, and the volume estimation is done using Kalman filtering, specifically:

[0061] defining state variables for estimating the target volume, including the injected volume and the instantaneous flow rate, wherein the injected volume is obtained through external sensor measurements, and the instantaneous flow rate is obtained based on the fluid control model;

[0062] The target volume is estimated using the following formula: V_pred = V_prev + flow_rate × dt, where V_pred is the injected volume predicted at the current moment, V_prev is the injected milk foam volume determined after updating at the previous moment, flow_rate is the instantaneous flow rate, and dt is the time interval;

[0063] When sensor measurements are obtained, the predicted and measured values ​​are fused using Bayesian updating;

[0064] An adaptive noise model is used to dynamically adjust the covariance of process noise and measurement noise according to the flow rate.

[0065] A system for implementing the above-mentioned method for protecting the liquid level to be suitable for milk froth injection into different cup shapes and adapted to the milk froth injection into different cup shapes, comprising:

[0066] At least one multi-axis robotic arm having at least two controllable joints;

[0067] Optional auxiliary motion devices with one or more additional degrees of freedom;

[0068] a combination of one or more sensory devices;

[0069] Control unit, adapted to robot arm configurations with different numbers of axes.

[0070] Furthermore, the liquid level protection is suitable for different cup-shaped milk foam injection systems, and further includes the multi-axis robotic arm executing the following configuration scheme, specifically:

[0071] Automatically detect the type and number of axes of the connected multi-axis robotic arm;

[0072] Select the corresponding kinematic model according to the test results;

[0073] Adaptive adjustment of control algorithm parameters;

[0074] Supports hot-plugging and dynamic reconfiguration.

[0075] Furthermore, the liquid level protection system is adapted to different cup-shaped milk foam injection systems, and further includes:

[0076] Core control module, compatible with different types of actuators;

[0077] Replaceable actuator interface, supporting 2-7 axis multi-axis robotic arms;

[0078] Scalable sensor interface, supporting multiple sensing devices;

[0079] Unified software architecture, adaptive to different hardware configurations.

[0080] Furthermore, the liquid level protection system is adapted to different cup-shaped milk foam injection systems, and further includes:

[0081] The system supports three sensing modes: electronic scale gravity measurement only, 3D visual geometry measurement only, and fusion of the two.

[0082] Use Kalman filter to fuse different sensor data to achieve accurate volume estimation;

[0083] Adaptively adjust the covariance of process noise and measurement noise to optimize the fusion effect.

[0084] A computer-readable storage medium is used to execute the above-mentioned method for maintaining a liquid level suitable for milk foam injection in different cup shapes, and stores an adaptive control program. When executed:

[0085] Identify the type of connected multi-axis robot arm, including the number of axes for 2-7 axes and other type configuration information;

[0086] Load the corresponding kinematic model and control parameters;

[0087] Automatic calibration and optimization of control strategies;

[0088] Implement a unified control interface that is independent of hardware type.

[0089] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0090] (1) The present invention uses a 3D vision system to intelligently adapt to multiple cup types and recognize the cup diameter, depth, taper and other parameters in real time, accurately distinguishing between conical, narrow-mouth, wide-mouth and other cup types, and designs targeted control strategies: conical cups dynamically compensate for the nonlinear change of "height-volume" to avoid bottom accumulation or top overflow; narrow-mouth cups use high-precision flow control (flow rate error ≤ 0.2 ml / s) to adapt to small diameters to prevent blockage and splashing; wide-mouth cups strengthen the Coanda effect to stabilize the flow, and cooperate with the centered spiral trajectory to ensure that the liquid surface flatness is ≤ 0.3 mm, breaking through the limitation of traditional equipment's reliance on a single cup type and covering the mainstream cup scenarios on the market.

[0091] (2) In terms of control accuracy and stability, the hybrid fluid model of the present invention integrates physical laws (Bernoulli equation, Coanda effect) and machine learning capabilities: the physical model ensures the generalization of flow rate calculation, covering characteristics such as the geometric parameters of the latte art cylinder and fluid density; the single learning model captures dynamic characteristics such as milk foam temperature-viscosity coupling and shear thinning from a "single demonstration" and quickly adapts to new batches of milk foam. Superimposed Kalman filter closed-loop control, real-time fusion of model prediction and sensor data, filtering model error (±0.5g / s) and noise (±0.1g), making the volume estimation error <1.5ml, accurately protecting the coffee oil layer, and achieving a significant improvement in injection consistency under different cup shapes.

[0092] (3) The present invention builds strong robustness through collaborative training in simulation and real environments: first, in a high-fidelity fluid simulation environment, the complex physical properties of milk foam (temperature, viscosity, shear, etc.) are simulated, and the "optimal pouring strategy" is pre-trained with the help of deep reinforcement learning; then, in a real experimental platform (six-axis robotic arm + high-precision sensor), the behavioral differences between water, milk, and milk foam, as well as the dynamic influence of different capacities of latte art jars and holding positions are adapted. This allows the system to be retrained when the formula is changed or the equipment parameters are changed, greatly improving its cross-scenario adaptability.

[0093] (4) The present invention adopts trajectory decoupling and flexible posture transformation design, decoupling the pouring angle trajectory (flow rate control) from the robot arm motion trajectory (position control) algorithm, and supports flexible "strategy-motion" combination; when the position or angle of the cup changes, only the rotation matrix needs to be superimposed to quickly adjust the posture, adapting to scenarios such as multiple cups in parallel and tilted tables, and the trajectory update time is <0.1s, which significantly improves the response speed and flexibility of the equipment and meets the efficient and diverse production needs of commercial coffee machines.

[0094] In summary, the present invention systematically breaks through the core pain points of automated coffee latte art from four dimensions: cup shape adaptation, control accuracy, robustness, and scenario flexibility, and provides a full-link solution for the standardized and personalized production of coffee beverages. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] Figure 1 This is an overall flow chart of the method for protecting the liquid level and injecting milk foam into different cup shapes according to the present invention;

[0096] Figure 2 This is a diagram of the architecture of the hybrid physics and machine learning fluid model of the present invention;

[0097] Figure 3 This is a diagram of the multi-physics field coupled fluid modeling and two-stage reinforcement learning architecture of the present invention;

[0098] Figure 4 Schematic diagram of the three-dimensional control principle of the six-axis robotic arm of the present invention;

[0099] Figure 5 This is a diagram of the Kalman filter real-time volume tracking system of the present invention;

[0100] Figure 6 This is the architecture diagram of the hierarchical control system of the present invention. DETAILED DESCRIPTION

[0101] To make the purpose, technical solutions, and advantages of the embodiments of this application more clear, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0102] Those skilled in the art will appreciate that, unless otherwise stated, the singular forms "a," "an," "said," and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of the stated features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0103] First embodiment

[0104] like Figure 1 As shown, this embodiment provides a method for injecting milk foam that protects the liquid level and is suitable for different cup types, starting from cup type recognition, to calculating the required milk foam volume, and then using the core fluid control model to generate the pouring trajectory, and finally performing the injection under real-time monitoring, forming a complete closed-loop (or pseudo-closed-loop) control system to ensure that the injection of different cup types can achieve accurate and consistent effects. The fundamental idea is the combination of prediction and control. The system does not rely on preset fixed programs, but dynamically generates and executes the optimal injection strategy based on the real-time perception of the environment (cup) and a deep understanding of fluid behavior (hybrid model). Specifically, it includes the following steps:

[0105] S1: Obtain cup parameters and target injection volume.

[0106] In step S1, obtaining cup parameters and target injection volume includes: sensing the milk foam injection environment, scanning the workspace through one or more sensing devices, identifying the cup position, measuring cup parameters including caliber, depth, and shape, and calculating the target injection volume, specifically:

[0107] Scan the cup using a sensing device to extract key geometric information including caliber, depth, bottom diameter, and taper, and create a digital model of the cup based on the key geometric information;

[0108] The user presets the distance between the liquid level and the cup opening (for example, 1 cm from the cup opening), and calculates the target injection volume based on the key geometric information;

[0109] The cup type classification of the cup is determined based on the key geometric information, and an adaptive control strategy is selected according to the different cup type classifications to generate an adaptive pouring trajectory. When it is a conical cup, the pouring angle is dynamically adjusted to compensate for the volume difference of the cup body as the height changes. When it is a narrow-mouth cup, a high-precision control strategy is adopted to reduce the milk foam injection flow rate to adapt to the narrow mouth characteristics. When it is a wide-mouth cup, the milk foam injection flow rate is reduced to improve the stability of the injection process.

[0110] Among them, the perception device includes at least one of a 3D vision system, a 2D vision system, and a depth camera; at least one of a gravity sensor, an electronic scale, and a pressure sensor; and at least one of an ultrasonic sensor and a laser ranging sensor.

[0111] S2: Generate an injection control trajectory according to the acquired parameters.

[0112] In step S2, the generated injection control trajectory is the motion trajectory of a multi-axis robotic arm, wherein the multi-axis robotic arm is any one of a five-axis to a seven-axis robotic arm. A preferred solution is to use a six-axis robotic arm, which achieves precise positioning (X, Y, Z) and posture control (Roll, Pitch, Yaw) of the end effector through the coordinated operation of various degrees of freedom. Furthermore, the multi-axis robotic arm can be coupled with one or more auxiliary motion devices, each having one to three degrees of freedom, for carrying and moving the container.

[0113] In step S2, generating an injection control trajectory based on the acquired parameters includes: planning the milk foam injection trajectory, determining a safe height of the latte art cylinder above the cup mouth at the injection starting position, loading a fluid control model to generate a latte art cylinder tilting angle trajectory sequence, and simultaneously generating a multi-axis robotic arm motion trajectory in Cartesian space. Each trajectory point in the motion trajectory is converted into an angle value of at least five joints of the robotic arm through inverse kinematics solution, and trajectory smoothing is performed. Specifically,

[0114] First, the algorithm decouples the dumping angle trajectory sequence and the robotic arm motion trajectory. Different algorithms are used to generate the dumping angle trajectory sequence (such as the pitch angle required for flow rate control) and the robotic arm motion trajectory (such as the elliptical motion trajectory). This decoupling design allows the system to flexibly combine different dumping strategies and motion patterns. For example, while executing an elliptical fusion trajectory, the dumping angle can be independently controlled to maintain a constant flow rate, maximizing the system's flexibility and control accuracy.

[0115] Secondly, when it is necessary to perform latte art at different cup positions or angles, there is no need to regenerate the entire set of trajectories. A rotation matrix can be superimposed on the currently calculated pouring angle trajectory sequence and the robotic arm motion trajectory to achieve the transformation of the posture. For example, for the robotic arm motion trajectory, let the standard latte art posture be T0, and the transformation matrix of the target position be T_offset, then the final execution posture T_final = T_offset × T0. This matrix operation is a standard operation in robotics and has extremely high computational efficiency. This design allows the same set of latte art algorithms and trajectories to adapt to different work scenarios, such as making multiple cups at the same time, tilting the work surface, or special layouts that need to avoid obstacles, greatly improving the practicality and deployment flexibility of the system.

[0116] Furthermore, in order to plan the dumping angle trajectory sequence, the fluid control model is designed before step S2, specifically:

[0117] The fluid control model includes at least one of the following:

[0118] Fluid dynamics models based on physical laws;

[0119] Predictive models based on machine learning;

[0120] Hybrid models of physical and machine learning models;

[0121] Empirical models based on lookup tables;

[0122] When a hybrid model of the physical model and the machine learning model is adopted, the fluid control model is generated by mixing the physical model with the single learning model. We do not use a pure physical model or a pure data-driven model alone, but combine the two.

[0123] The physical model is based on the Bernoulli equation and the Coanda effect. By controlling the adherent flow of the liquid on the curved surface of the spout, low-velocity laminar flow is achieved in the initial stage of injection to avoid damaging the coffee oil layer. The flow rate is calculated based on the principles of fluid dynamics. The calculation takes into account the geometric parameters of the latte art pitcher, including diameter, height, spout width, and spout curvature radius (for example, a 350ml latte art pitcher has a diameter of approximately 75mm, a height of approximately 100mm, a spout width of 30mm, and a spout curvature radius of 20mm), as well as the physical properties of the fluid, including density, viscosity, and surface adhesion. The model also supports parameter configurations for latte art pitchers of various standard capacities.

[0124] The single-shot learning model uses domain adaptation technology to learn from a single pouring demonstration. By recording the pouring angle and weight change data, the single-shot learning model can quickly capture the behavioral characteristics of a specific fluid and generate a lookup table (LUT) to achieve real-time prediction.

[0125] The adaptive weight fusion ratio of the physical model and the single learning model is customized to perform weighted fusion of the physical model and the single learning model.

[0126] like Figure 2 As shown in the architecture diagram of the hybrid physics and machine learning fluid model, the physical model provides strong generalization capabilities and guarantees physical consistency, while the single-shot learning model can quickly learn from a single demonstration and compensate for errors in the physical model, adapting to subtle differences in specific fluids (such as different batches of milk foam). A weighted fusion of the two (e.g., a default of 30% physics + 70% learning) achieves unprecedented prediction accuracy and robustness. The system also supports adaptive weight adjustment, dynamically optimizing the model combination ratio based on actual performance.

[0127] The physical model cleverly exploits the Coanda effect, as the spout of the latte art pitcher has a specific curvature. During pouring, the liquid (milk foam) adheres to the curved surface of the spout due to the Coanda effect, rather than falling directly vertically. The separation point varies with the angle and flow rate. This creates a smooth, controllable laminar flow. Our model explicitly factors in the spout curvature and the Coanda effect, allowing us to maintain a stable flow even at lower flow rates. This allows us to gently "push" the milk foam below the espresso surface during the initial pour, avoiding disrupting the critical crema layer. This is a common technique used for latte art in takeout cups or deep cups.

[0128] One-Shot Learning: This is the core trick to overcoming the generalization challenge of models. The system doesn't need to collect large amounts of data for each new situation. Instead, the operator only needs to perform a standard demonstration pour under new conditions (e.g., a different type of milk), and the system will:

[0129] Automatically detect the backflow starting angle: by analyzing the threshold moment when the flow rate exceeds 0.2g / s, the backflow starting angle of the fluid (usually 60°±5°) is automatically identified; learn fluid characteristics: extract parameters such as the equivalent viscosity and Coanda effect coefficient of the fluid from the pouring angle-flow rate curve; update the prediction model: incorporate the newly learned parameters into the hybrid model, and automatically adjust the weight ratio of the physical model and the learning model; generate a dedicated lookup table: generate an optimized lookup table for the new fluid type to ensure the real-time performance of subsequent operations; this greatly reduces the system's usage threshold and calibration cost, allowing the system to adapt to a new fluid type within 5 minutes.

[0130] Physics-guided machine learning: This is the key to solving the nonlinear challenges of fluids. Rather than letting machine learning models blindly learn from data, we use known physical laws (Bernoulli's equations) as the model's "skeleton." This enables the model to make physically intuitive and reasonable predictions even in areas with sparse training data, avoiding overfitting and absurd outputs.

[0131] Convert volume control into flow rate control: We do not control the final volume directly, but achieve control of the final volume by precisely controlling the flow rate during the process. By integrating the flow rate curve, the target volume can be reached more smoothly and accurately, which is much more advanced than simple "on-off" control. Multi-physics coupled fluid control model based on reinforcement learning: In order to accurately handle the complex behavior of non-Newtonian fluids (such as milk foam), such as Figure 3 As shown in the diagram of the multi-physics field coupled fluid modeling and two-stage reinforcement learning architecture, the present invention establishes a multi-physics field coupled fluid modeling system and achieves precise control through two-stage reinforcement learning:

[0132] Phase 1: Reinforcement Learning in Simulation Environment:

[0133] Build high-fidelity fluid models in the partio+SPlisHSPlasH fluid simulation environment;

[0134] Consider the temperature-viscosity coupling: μ(T,t)=μ0×exp(-k(T-T0))×(1-αt), where both temperature and time affect the fluid viscosity, μ(T,t) is the dynamic viscosity at temperature T and time t, μ0 is the reference viscosity of the fluid at the reference temperature T0 and the initial time t=0, k is the temperature influence coefficient, and α is the time attenuation coefficient;

[0135] Consider the shear thinning effect: the non-Newtonian properties of milk foam at different shear rates;

[0136] Through deep reinforcement learning DRQN / RDPG, the basic policy network is trained in a simulation environment to learn the optimal dumping strategy under different physical parameters;

[0137] Phase 2: Real-world domain adaptation learning:

[0138] Build a real experimental platform: six-axis robotic arm + high-precision electronic scale + 3D vision system;

[0139] Systematically study the behavior characteristics of three fluids, including water, milk, and milk foam, under different conditions;

[0140] Introducing multivariable control: different capacities of latte art pitchers: 350ml, 450ml, and 600ml; different grip positions: at the top of the pitcher at a high center of gravity, in the middle at the balance point, and at the bottom at a low center of gravity; each configuration combination affects the system's dynamic characteristics and fluid behavior;

[0141] Use Graph Networks to capture the complex interactions between fluid particles.

[0142] One-Shot Learning advantages:

[0143] Based on existing physical rules and prior knowledge from simulation training, the system can adapt to new fluid types with just a single demonstration; through domain-adaptive imitation learning, the physical intuition learned in simulation is transferred to the real environment; the policy network is updated in real time: new demonstration data is immediately incorporated into the model through online learning, without the need for retraining; robustness is guaranteed: even when faced with fluids that have never been seen before (such as plant milk), the system can quickly adapt based on physical similarities.

[0144] Furthermore, in step S2, the milk foam injection trajectory is planned to determine the safe height of the latte art cylinder above the cup mouth at the injection starting position, and the fluid control model is loaded to generate the latte art cylinder tilting angle trajectory sequence, specifically:

[0145] The target injection volume and expected injection time are used as inputs, and an inverse solver based on the trained hybrid fluid control model is called to reversely calculate an optimal dumping angle trajectory sequence that can achieve the target injection volume through an iterative optimization algorithm. The solution process includes:

[0146] Calculating an average flow rate according to the target injection volume and the expected injection time;

[0147] Design a trapezoidal velocity profile with the peak velocity being a preset multiple of the average flow velocity (e.g., 2.2 times);

[0148] The flow velocity curve is converted into an angle curve by the inverse solver of the fluid control model, and the minimum angle and maximum angle of backflow are set as application constraints (such as a minimum angle of 60° (backflow starting angle) and a maximum angle of 180°).

[0149] The generated trajectory exhibits a smooth S-shaped curve, with lower flow rates at the beginning and end (to protect the liquid level) and higher flow rates in the middle (to improve efficiency). Subsequent injection execution is strictly closed-loop and calibrated: the robotic arm executes the dumping action strictly according to the generated optimal dumping angle trajectory sequence. During this process, a Kalman filter estimates the injected volume in real time. The system continuously compares the injected volume with the target volume, and once they are equal, the dumping stops immediately.

[0150] Furthermore, for the robot arm motion trajectory, such as Figure 4 The schematic diagram of the three-dimensional control principle of a six-axis robotic arm is shown below. All joints (J1-J6) of the six-axis robotic arm work together to achieve precise positioning (X, Y, Z) and posture control (roll, pitch, yaw) of the end effector (i.e., the latte art bowl containing the milk froth) in Cartesian space. Each trajectory point requires an inverse kinematics solution to convert the Cartesian position into the angle values ​​of each joint. This all-joint control method ensures smooth and precise motion.

[0151] S3: Execute the injection control trajectory and adjust in real time until the target injection volume is reached.

[0152] In step S3, executing the injection control trajectory and adjusting it in real time until the target injection volume is reached includes: performing milk foam injection execution control, moving the robotic arm to the starting position to start executing the pouring trajectory according to the pouring angle sequence, and performing real-time monitoring during the pouring process, using Kalman filter volume estimation, stopping injection when the target volume is reached, and resetting the robotic arm, otherwise continuing to execute the pouring trajectory after adjusting the pouring angle.

[0153] Among them, step S3 performs real-time monitoring during the dumping process and uses Kalman filter volume estimation, such as Figure 5 The Kalman filter real-time volume tracking system is shown in the figure, specifically:

[0154] To ensure the accuracy of the final liquid level, the exact volume of milk foam injected must be known. The system uses a Kalman filter to track the injected volume in real time.

[0155] defining state variables for estimating the target volume, including the injected volume and the instantaneous flow rate, wherein the injected volume is obtained through external sensor measurements, and the instantaneous flow rate is obtained based on the fluid control model;

[0156] The target volume is estimated using the following formula: V_pred = V_prev + flow_rate × dt, where V_pred is the injected volume predicted at the current moment, V_prev is the injected milk foam volume determined after updating at the previous moment, flow_rate is the instantaneous flow rate, and dt is the time interval;

[0157] When sensor measurements are obtained, the predicted and measured values ​​are fused using Bayesian updating;

[0158] An adaptive noise model is used to dynamically adjust the covariance of process noise and measurement noise according to the flow rate.

[0159] The Kalman filter fuses two sources of information: the model's predicted flow rate and an optional external sensor measurement (such as a high-precision electronic scale placed under the cup). This fusion makes volume estimation far more accurate and stable than using a single source of information. It effectively filters out instantaneous errors in model predictions (±0.5g / s) and sensor noise (±0.1g), ensuring precise stopping when the specified volume is reached (error <2ml).

[0160] In addition, if Figure 6 As shown in the hierarchical control system architecture diagram, the control method of this embodiment is based on a hierarchical control system. The upper layer is responsible for task planning, including identifying the cup shape, calculating the target injection volume, generating the motion trajectory of the Cartesian space (X, Y, Z sequence), and generating the optimal pouring posture sequence (Rol l, Pitch, Yaw) through the single-shot learning flow rate predictor (One-Shot Learning Flow Rate Predictor). The predictor can quickly learn the fluid characteristics from a single demonstration and can adapt to different batches of milk foam or new fluid types without a large amount of training data. The lower layer is responsible for motion control, and converts the Cartesian space posture into the angle value of each controllable degree of freedom through the inverse kinematics solver (such as KDL, TRAC-IK, IKFast and other mature algorithms), and cooperates with the trajectory optimizer (such as TrajOpt, CHOMP, B-spline optimization, etc.) to generate a smooth motion trajectory, and finally outputs the motor command to drive the robotic arm for precise execution.

[0161] In addition, the adaptive predictive control algorithm of the present invention, which protects the liquid level and is suitable for the injection of milk foam into different cup shapes, constructs a three-layer architecture of "constraint drive - execution closed loop - model self-optimization":

[0162] The constraint processing layer first defines the core boundaries: physical constraints limit the tilting angle range of the latte art cylinder (such as 0° to 45°), safety constraints lock the milk foam flow rate threshold (such as 0.5 to 3 ml / s), and performance constraints clearly define the accuracy requirement of volume error < 2 ml; the control execution layer plans the Cartesian space trajectory based on the constraints, and converts it into the joint angles of each controllable degree of freedom of the robotic arm through inverse kinematics solution, driving the six axes to collaboratively perform the injection action; the feedback correction layer relies on sensors such as electronic scales and 3D vision to collect data, and through state estimators such as Kalman filtering, it integrates model predictions and measurements in real time, estimates the injected volume and instantaneous flow rate, and dynamically updates the control parameters (such as the tilting angle step and flow rate correction amount) after calculating the error, and corrects the execution deviation in a closed loop; the adaptive learning layer and the control model layer form an iteration: the online learning module relies on simulation reinforcement learning (DRQN / RDPG) and real scene domain adaptation data to continuously optimize the mixed fluid model (integrating the Bernoulli equation, Coanda effect and single learning ability); the model side uses the inverse solver (Angle Finding The optimal dumping angle sequence is reversely deduced by the algorithm, and the trapezoidal velocity curve generator plans the smooth transition of flow rate (such as the trapezoidal profile to avoid liquid surface impact). Finally, the optimal trajectory Angle_t that adapts to the cup shape and fluid characteristics is output, realizing the full closed-loop precise control of "constraint → execution → feedback → learning → model iteration" to ensure the consistency and stability of injection under multiple cup shapes.

[0163] Second embodiment

[0164] This embodiment provides a simplified configuration using a 5-axis robotic arm. In this configuration, the system achieves milk froth injection by constraining certain degrees of freedom. While the posture control capability is reduced, it still meets basic injection requirements and is more cost-effective.

[0165] Third embodiment

[0166] This embodiment provides an advanced configuration using a 7-axis robotic arm. The additional redundant degrees of freedom enable the system to: better avoid obstacles in complex environments, optimize motion trajectories, reduce joint wear, and provide more flexible workspace configurations.

[0167] Fourth embodiment

[0168] This embodiment provides a coordinated configuration of a 6-axis robotic arm and a 3-axis auxiliary platform. The auxiliary platform carries the coffee cup and provides XY plane motion, while the robotic arm focuses on Z-axis motion and pouring control, achieving an optimized distribution of motion.

[0169] Fifth embodiment

[0170] The present invention provides a system for implementing the method for injecting milk froth into different cups to maintain a liquid level suitable for different cup shapes, comprising:

[0171] At least one multi-axis robotic arm having at least two controllable joints;

[0172] Optional auxiliary motion devices with one or more additional degrees of freedom;

[0173] a combination of one or more sensory devices;

[0174] Control unit, adapted to robot arm configurations with different numbers of axes.

[0175] Furthermore, the liquid level protection is suitable for different cup-shaped milk foam injection systems, and further includes the multi-axis robotic arm executing the following configuration scheme, specifically:

[0176] Automatically detect the type and number of axes of the connected multi-axis robotic arm;

[0177] Select the corresponding kinematic model according to the test results;

[0178] Adaptive adjustment of control algorithm parameters;

[0179] Supports hot-plugging and dynamic reconfiguration.

[0180] In addition, the inverse kinematics solution algorithm of the present invention can adapt to different robot arm configurations:

[0181] For a 5-axis robot arm, this is achieved by constraining certain posture degrees of freedom;

[0182] For 6-axis robotic arms, complete 6-degree-of-freedom control is achieved;

[0183] For 7 axes and above, use redundant degrees of freedom to optimize motion trajectory;

[0184] The algorithm automatically detects the type of robot and selects the corresponding solution strategy;

[0185] Furthermore, the liquid level protection system is adapted to different cup-shaped milk foam injection systems, and further includes:

[0186] Core control module, compatible with different types of actuators;

[0187] Replaceable actuator interface, supporting 2-7 axis multi-axis robotic arms;

[0188] Scalable sensor interface, supporting multiple sensing devices;

[0189] Unified software architecture, adaptive to different hardware configurations.

[0190] Furthermore, the liquid level protection system is adapted to different cup-shaped milk foam injection systems, and further includes:

[0191] The system supports three sensing modes: electronic scale gravity measurement only, 3D visual geometry measurement only, and fusion of the two.

[0192] Use Kalman filter to fuse different sensor data to achieve accurate volume estimation;

[0193] Adaptively adjust the covariance of process noise and measurement noise to optimize the fusion effect.

[0194] The present invention further provides a computer-readable storage medium for executing the method for maintaining a liquid level suitable for milk foam injection in different cup shapes as in the first embodiment, the computer-readable storage medium storing an adaptive control program. When executed:

[0195] Identify the type of connected multi-axis robot arm, including the number of axes for 2-7 axes and other type configuration information;

[0196] Load the corresponding kinematic model and control parameters;

[0197] Automatic calibration and optimization of control strategies;

[0198] Implement a unified control interface that is independent of hardware type.

[0199] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiment. All technical solutions based on the concept of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should also be considered within the scope of protection of the present invention.

[0200] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0201] It should be noted that the above embodiments can be freely combined as needed. The above description is only a preferred embodiment of the present invention. It should be pointed out that those skilled in the art can make several improvements and modifications without departing from the principles of the present invention, and such improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for protecting the liquid level and injecting milk foam into different cup shapes, characterized in that: The following steps are involved: S1: Obtain cup parameters and target injection volume; S2: Generate injection control trajectory according to the acquired parameters; S3: Execute the injection control trajectory and adjust in real time until the target injection volume is reached.

2. The method for protecting the liquid level according to claim 1 is suitable for injecting milk foam into different cup shapes, characterized in that: In step S2, the generated injection control trajectory is a motion trajectory of a multi-axis robotic arm, and the multi-axis robotic arm is any one of a five-axis to a seven-axis robotic arm.

3. The method for protecting the liquid level according to claim 2 is suitable for the milk foam injection method of different cup shapes, characterized in that: The multi-axis robotic arm is a six-axis robotic arm that achieves precise positioning (X, Y, Z) and posture control (Roll, Pitch, Yaw) of the end effector through the coordinated work of various degrees of freedom.

4. The method for protecting the liquid level and being suitable for milk foam injection into different cup shapes according to claim 2, characterized in that: The multi-axis robotic arm can also be equipped with one or more auxiliary motion devices, each of which has one to three degrees of freedom for carrying and moving containers.

5. The method for protecting the liquid level according to claim 1 is suitable for injecting milk foam into different cup shapes, characterized in that: In step S1, obtaining cup parameters and target injection volume includes: sensing the milk foam injection environment, scanning the workspace through one or more sensing devices, identifying the cup position, measuring cup parameters including caliber, depth, and shape, and calculating the target injection volume, specifically: Scan the cup using a sensing device to extract key geometric information including caliber, depth, bottom diameter, and taper, and create a digital model of the cup based on the key geometric information; The user presets the distance between the liquid level and the cup mouth, and calculates the target injection volume based on the key geometric information; The cup type is determined based on the key geometric information, and an adaptive control strategy is selected based on the different cup types to generate an adaptive pouring trajectory. For conical cups, the pouring angle is dynamically adjusted to compensate for the volume difference of the cup body as it changes with height. For narrow-mouth cups, a high-precision control strategy is adopted to reduce the milk foam injection flow rate to adapt to the narrow mouth characteristics. For wide-mouth cups, the milk foam injection flow rate is reduced to improve the stability of the injection process. Among them, the perception device includes at least one of a 3D vision system, a 2D vision system, and a depth camera; at least one of a gravity sensor, an electronic scale, and a pressure sensor; and at least one of an ultrasonic sensor and a laser ranging sensor.

6. The method for protecting the liquid level and being suitable for milk foam injection into different cup shapes according to claim 5, characterized in that: In step S2, generating an injection control trajectory based on the acquired parameters includes: planning the milk foam injection trajectory, determining a safe height of the latte art cylinder above the cup mouth at the injection starting position, loading a fluid control model to generate a latte art cylinder tilting angle trajectory sequence, and simultaneously generating a multi-axis robotic arm motion trajectory in Cartesian space. Each trajectory point in the motion trajectory is converted into an angle value of at least five joints of the robotic arm through inverse kinematics solution, and trajectory smoothing is performed. Specifically, Performing algorithm decoupling design on the dumping angle trajectory sequence and the robot arm motion trajectory, using different algorithms to generate the dumping angle trajectory sequence and the robot arm motion trajectory respectively, so that the system can flexibly combine different dumping strategies and motion modes; At the same time, when latte art is required at different cup positions or angles, there is no need to regenerate the entire set of trajectories. A rotation matrix can be superimposed on the currently calculated pouring angle trajectory sequence and the robotic arm motion trajectory to achieve posture transformation.

7. The method for protecting the liquid level and being suitable for milk foam injection into different cup shapes according to claim 6, characterized in that: Before step S2, the fluid control model is designed, specifically: The fluid control model includes at least one of the following: Fluid dynamics models based on physical laws; Predictive models based on machine learning; Hybrid models of physical and machine learning models; Empirical models based on lookup tables; When a hybrid model of the physical model and the machine learning model is adopted, the fluid control model is generated by mixing the physical model with the single learning model; The physical model, based on the Bernoulli equation and the Coanda effect, achieves low-velocity laminar flow in the initial injection phase by controlling the adherent flow of the liquid on the curved surface of the spout, thus avoiding damage to the coffee crema layer. The flow rate is calculated using the principles of fluid dynamics, taking into account the geometric parameters of the latte art pot, including its diameter, height, spout width, and spout curvature radius, as well as the physical properties of the fluid, including density, viscosity, and surface adhesion. The model also supports parameter configurations for latte art pots of various standard capacities. The single-shot learning model uses domain adaptation technology to learn from a single pouring demonstration. By recording the pouring angle and weight change data, the single-shot learning model can quickly capture the behavioral characteristics of a specific fluid and generate a lookup table (LUT) to achieve real-time prediction. The adaptive weight fusion ratio of the physical model and the single learning model is customized to perform weighted fusion of the physical model and the single learning model.

8. The method for injecting milk foam into different cups with a liquid level protection according to claim 7, characterized in that: In step S2, the milk foam injection trajectory is planned to determine the safe height of the latte art cylinder above the cup mouth at the injection starting position. The fluid control model is loaded to generate the latte art cylinder tilting angle trajectory sequence, specifically: The target injection volume and expected injection time are used as inputs, and an inverse solver based on the trained hybrid fluid control model is called to reversely calculate an optimal dumping angle trajectory sequence that can achieve the target injection volume through an iterative optimization algorithm. The solution process includes: Calculating an average flow rate according to the target injection volume and the expected injection time; Design a trapezoidal velocity profile with the peak velocity being a preset multiple of the average flow velocity; The flow velocity curve is converted into an angle curve by using the inverse solver of the fluid control model, and the minimum angle and maximum angle of the backflow are set as application constraints.

9. The method for injecting milk foam into different cups with a liquid level protection function according to claim 7, characterized in that: Step S2 also includes establishing a multi-physics field coupled fluid modeling system to achieve precise control of the multi-physics field coupled fluid control model based on reinforcement learning, specifically: Phase 1: Reinforcement Learning in Simulation Environment: Build high-fidelity fluid models in the partio+SPlisHSPlasH fluid simulation environment; Consider the temperature-viscosity coupling: μ(T,t)=μ0×exp(-k(T-T0))×(1-αt), where both temperature and time affect the fluid viscosity, μ(T,t) is the dynamic viscosity at temperature T and time t, μ0 is the reference viscosity of the fluid at the reference temperature T0 and the initial time t=0, k is the temperature influence coefficient, and α is the time attenuation coefficient; Consider the shear thinning effect: the non-Newtonian properties of milk foam at different shear rates; Through deep reinforcement learning DRQN / RDPG, the basic policy network is trained in a simulation environment to learn the optimal dumping strategy under different physical parameters; Phase 2: Real-world domain adaptation learning: Build a real experimental platform: six-axis robotic arm + high-precision electronic scale + 3D vision system; Systematically study the behavior characteristics of three fluids, including water, milk, and milk foam, under different conditions; Introducing multivariable control: different capacities of latte art pitchers: 350ml, 450ml, and 600ml; different grip positions: at the top of the pitcher at a high center of gravity, in the middle at the balance point, and at the bottom at a low center of gravity; each configuration combination affects the system's dynamic characteristics and fluid behavior; Use Graph Networks to capture the complex interactions between fluid particles.

10. The method for injecting milk foam into different cup shapes while protecting the liquid level according to claim 1, characterized in that: In step S3, executing the injection control trajectory and adjusting it in real time until the target injection volume is reached includes: performing milk foam injection execution control, moving the robotic arm to a starting position to start executing the pouring trajectory according to the pouring angle sequence, performing real-time monitoring during the pouring process, using Kalman filter volume estimation, stopping injection when the target volume is reached, and resetting the robotic arm; otherwise, continuing to execute the pouring trajectory after adjusting the pouring angle; Among them, real-time monitoring is carried out during the dumping process, and the volume estimation is done using Kalman filter, specifically: defining state variables for estimating the target volume, including the injected volume and the instantaneous flow rate, wherein the injected volume is obtained through external sensor measurements, and the instantaneous flow rate is obtained based on the fluid control model; The target volume is estimated using the following formula: V_pred = V_prev + flow_rate × dt, where V_pred is the injected volume predicted at the current moment, V_prev is the injected milk foam volume determined after updating at the previous moment, flow_rate is the instantaneous flow rate, and dt is the time interval; When sensor measurements are obtained, the predicted and measured values ​​are fused using Bayesian updating; An adaptive noise model is used to dynamically adjust the covariance of process noise and measurement noise according to the flow rate.

11. A system for protecting the liquid level from being distorted by milk froth in different cup shapes for executing the method for protecting the liquid level from being distorted by milk froth in different cup shapes according to any one of claims 1 to 10, characterized in that: include: At least one multi-axis robotic arm having at least two controllable joints; Optional auxiliary motion devices with one or more additional degrees of freedom; a combination of one or more sensory devices; Control unit, adapted to robot arm configurations with different numbers of axes.

12. The system for injecting milk foam into different cups with a protective liquid level according to claim 11, characterized in that: It also includes the multi-axis robotic arm executing the following configuration scheme, specifically: Automatically detect the type and number of axes of the connected multi-axis robotic arm; Select the corresponding kinematic model according to the test results; Adaptive adjustment of control algorithm parameters; Supports hot-plugging and dynamic reconfiguration.

13. The system for injecting milk foam into different cups with a protective liquid level according to claim 11, characterized in that: Also includes: Core control module, compatible with different types of actuators; Replaceable actuator interface, supporting 2-7 axis multi-axis robotic arms; Scalable sensor interface, supporting multiple sensing devices; Unified software architecture, adaptive to different hardware configurations.

14. The system for injecting milk foam into different cups with a protective liquid level according to claim 11, characterized in that: Also includes: The system supports three sensing modes: electronic scale gravity measurement only, 3D visual geometry measurement only, and fusion of the two. Use Kalman filter to fuse different sensor data to achieve accurate volume estimation; Adaptively adjust the covariance of process noise and measurement noise to optimize the fusion effect.

15. A computer-readable storage medium, configured to execute the method for maintaining a liquid level suitable for milk foam injection in different cup shapes according to any one of claims 1 to 10, storing an adaptive control program, which, when executed: Identify the type of connected multi-axis robot arm, including the number of axes for 2-7 axes and other type configuration information; Load the corresponding kinematic model and control parameters; Automatic calibration and optimization of control strategies; Implement a unified control interface that is independent of hardware type.

Images