Coffee making skill learning method and apparatus, electronic device, and readable medium

Abstract

Embodiments of the present application provide a coffee making skill learning method and device, electronic equipment and readable medium. The method is executed by a first robot running an artificial intelligence (AI) agent and includes: obtaining demonstration data of a human making a coffee drink; obtaining a human coffee drink making action time sequence and action parameters of each operation action in the action time sequence according to the demonstration data; the action parameters include the motion trajectory and mechanical parameters of each action; determining a first skill model according to the action time sequence and the action parameters; the first skill model includes at least one of the following: an operation time sequence of the robot making coffee, operation parameters of each operation in the operation time sequence, and perception data and / or trigger conditions for executing each operation, wherein the first skill model is used for the robot to imitate the human to make the coffee drink; and applying the first skill model online and distributing model parameters of the first skill model to at least one second robot.

Description

Technical Field

[0001] This invention relates to the field of beverage processing technology, and in particular to a coffee-making skill learning method, a coffee-making skill learning device, an electronic device, and a computer-readable medium. Background Technology

[0002] When using robots to make coffee (beverages), engineers typically spend a significant amount of time on trajectory planning, parameter debugging, and code writing, resulting in extremely low efficiency. Furthermore, the quality of the programming is highly dependent on the engineer's expertise. At the same time, the wealth of valuable operational experience and skills accumulated by human baristas is difficult to effectively transfer to robots through traditional programming methods. Summary of the Invention

[0003] This invention provides a method, apparatus, electronic device, and computer-readable storage medium for learning coffee making skills.

[0004] The first aspect of this invention provides a coffee-making skill learning method, characterized in that it is executed by a first robot equipped with an artificial intelligence (AI) agent, comprising: acquiring demonstration data of humans making coffee beverages; acquiring, based on the demonstration data, a sequence of actions of the human making coffee beverages and action parameters of each action in the sequence of actions; the action parameters including the motion trajectory and mechanical parameters of each action; determining a first skill model based on the sequence of actions and the action parameters; the first skill model including at least one of the following: a sequence of actions of the robot making coffee, operation parameters of each operation in the sequence of actions, and perception data and / or triggering conditions for performing each operation, wherein the first skill model is used by the robot to imitate humans making coffee beverages; applying the first skill model online and distributing the model parameters of the first skill model to at least one second robot.

[0005] A second aspect of this invention provides a coffee-making skill learning device, comprising: a first acquisition module for acquiring demonstration data of humans making coffee beverages; a second acquisition module for acquiring, based on the demonstration data, a sequence of actions of the human making coffee beverages and action parameters of each action in the sequence of actions; the action parameters including the motion trajectory and mechanical parameters of each action; a determination module for determining a first skill model based on the sequence of actions and the action parameters; the first skill model including at least one of the following: a sequence of actions of a robot making coffee, operation parameters of each operation in the sequence of actions, and perception data and / or triggering conditions for performing each operation, wherein the first skill model is used for a robot to imitate a human making coffee beverages; and an execution module for applying the first skill model online and distributing the model parameters of the first skill model to at least one second robot.

[0006] A third aspect of the present invention discloses an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; the memory is used to store computer programs; and the processor is used to execute the program stored in the memory to implement the method described in the embodiments of the present invention.

[0007] A fourth aspect of the present invention also discloses one or more computer-readable media having instructions stored thereon that, when executed by one or more processors, cause the processors to perform the method as described in the embodiments of the present invention.

[0008] The embodiments of the present invention have the following advantages: In this embodiment, a first robot equipped with an AI agent directly acquires demonstration data of human coffee-making techniques and analyzes this data to obtain a first skill model for imitation. Compared to the method of relying on programmers to repeatedly communicate with human baristas and repeatedly adjust operational experience to convert it into programming code before writing it into the robot, this eliminates the need for manual translation and code writing, significantly improving the efficiency of the first robot learning coffee-making skills. It also reduces the cost of skill transfer and model updates, enabling faster and more accurate replication of the human coffee-making process. After the first robot learns the first skill model, in addition to its own application, it will share it with one or more second robots. This sharing among robots further enhances the efficiency of the robots learning coffee-making skills. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating the steps of a coffee-making skill learning method provided in an embodiment of the present invention; Figure 2AThis is a schematic diagram of a fully automated robot making coffee in an embodiment of the present invention; Figure 2B This is a schematic diagram of a fully automated robot making coffee in an embodiment of the present invention; Figure 3 This is a flowchart illustrating a coffee-making skill learning method provided in an embodiment of the present invention; Figure 4 This is a flowchart illustrating a coffee-making skill learning method provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a coffee-making skills learning device provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of a first robot provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of a computer-readable medium provided in an embodiment of the present invention. Detailed Implementation

[0010] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0011] Reference Figure 1 This diagram illustrates a flowchart of a coffee-making skill learning method provided in an embodiment of the present invention. The method can be executed by a first robot running an artificial intelligence (AI) agent. Exemplarily, the method may specifically include: S1110: Obtain demonstration data on human coffee beverage preparation; S1120: Based on the demonstration data, obtain the action sequence of the human making coffee beverage and the action parameters of each operation in the action sequence; the action parameters include the motion trajectory and mechanical parameters of each action. S1130: Determine a first skill model based on the action sequence and the action parameters; the first skill model includes at least one of the following: an operation sequence for the robot to make coffee, operation parameters for each operation in the operation sequence, and perception data and / or triggering conditions for performing each operation, wherein the first skill model is used for the robot to imitate humans in making coffee beverages; S1140: Apply the first skill model online and distribute the model parameters of the first skill model to at least one second robot.

[0012] In some embodiments, the first robot and the second robot are different coffee-making robots. These robots can automatically make coffee based on instructions, without human assistance. Typically, coffee-making robots operate with AI systems capable of independently completing the entire process from order placement to product output, distinguishing them from ordinary coffee-making equipment such as coffee machines that assist humans in making coffee. Figure 2A and Figure 2B The diagram shown is a schematic of a fully automated coffee-making robot, but the actual implementation is not limited to this diagram.

[0013] For example, a coffee-making robot can perform closed-loop control based on environmental perception, including a multi-degree-of-freedom collaborative robotic arm as an end effector. The robotic arm's motion trajectory is generated in real time by an upstream motion planning algorithm. This motion planning algorithm receives input from a machine vision subsystem, which uses depth cameras and image recognition technology to locate, identify, and monitor the status of objects (such as coffee cups and ingredient containers) within the workspace. Simultaneously, the coffee-making robot's operating system integrates various industrial sensors, including pressure sensors, flow meters, weighing modules, and temperature probes, for real-time acquisition of process parameters during grinding, extraction, and steaming. The central control unit of the coffee-making robot (typically an industrial PC or a high-performance embedded system) runs a software model of the coffee-making process (for example, including but not limited to the aforementioned AI agent). Thus, the coffee-making robot can parse customer order instructions into a series of executable action sequences and, based on sensor feedback data, dynamically adjust the actuators (such as pumps, valves, and motors) using proportional-integral-derivative (PID) control or other control algorithms to achieve precise control of flavor parameters. The entire workflow, from cup collection and cleaning to product output, is autonomously scheduled by the system, forming a complete closed loop of perception, decision-making, and execution.

[0014] In some embodiments, an AI agent may refer to an intelligent agent that is based on a large language model (LLM), possesses autonomous perception, planning, memory, tool invocation, and decision-making execution capabilities, and can proactively complete complex tasks in complex environments.

[0015] In some embodiments, the AI ​​agent may run on a server, cloud computing platform, local computer, or terminal device with strong computing power. In this disclosure embodiment, the AI ​​analysis engine may run on a coffee-making robot, or on a control device for multiple coffee-making robots within a target area, or on a server that distributes coffee recipes and / or collects coffee making or sales data to coffee-making robots across the network.

[0016] In some embodiments, the demonstration data for human coffee beverage preparation can be data from manual coffee preparation. This demonstration data can be used by the first robot to extract the actions, steps, and parameters of each action required for coffee preparation.

[0017] In some embodiments, acquiring demonstration data of humans making coffee beverages may include receiving demonstration data of humans making coffee beverages from other devices, such as downloading instructional videos of baristas making coffee beverages from a web server. These instructional videos are a type of demonstration data.

[0018] In some embodiments, acquiring exemplary data on human preparation of coffee beverages includes: acquiring first sensor data sensed by a sensor during the human preparation of coffee beverages.

[0019] In some embodiments, one or more sensors collect and obtain corresponding first sensor data during the process of a barista making coffee beverages. These sensors can be wearable sensors, such as inertial measurement units, accelerometers, angle sensors, pressure sensors, etc., worn on the barista's wrist, arm, or fingers, to directly collect data such as the barista's hand gestures, force applied, and movement trajectory; they can also be non-contact, airborne sensors, such as vision sensors (cameras), infrared sensors, laser rangefinders, etc., placed near the pour-over area, to collect pour-over related data such as water pouring position, water flow height, kettle tilt angle, and pouring time without direct contact with the barista or equipment through shooting or sensing methods; or they can be contact sensors installed on coffee equipment, such as weight sensors, flow sensors, temperature sensors, etc., installed on pour-over kettles, filter cups, and electronic scales, to collect data such as water flow rate, water temperature, and changes in portafilter weight. The raw data collected by the above-mentioned various sensors that are related to the human actions and process of making coffee beverages constitutes the first sensor data.

[0020] Thus, acquiring the first sensor data of humans making coffee beverages using the above methods enables accurate and objective collection of key parameters throughout the entire process, reducing subjective errors and information omissions caused by manual observation and recording. Through the collaborative collection of multiple types of sensors, including wearable, non-contact, and / or those mounted on equipment used during coffee making, comprehensive data can be captured on multi-dimensional features such as barista hand movements, water flow status, temperature, weight, and flow rate. This allows for the complete reconstruction of detailed information about standard demonstration movements, providing reliable data support for subsequent development of standard coffee-making movement models, intelligent teaching guidance, or automated replication of high-quality hand-drip processes. Simultaneously, it improves the efficiency and consistency of demonstration data collection, reducing reliance on experienced baristas.

[0021] In some embodiments, the acquisition of first sensor data sensed by the sensor during the human preparation of the coffee beverage includes at least one of the following: Acquire mechanical data of the human making coffee beverage recorded by a mechanical sensor; the mechanical data is used to indicate at least one of the operating force, the direction of the force, and the duration of the force. Acquire motion trajectory data of the human making coffee beverage, collected by joint sensors; Acquire visual data of the human making coffee beverages collected by a visual sensor; Acquire flow data of the human-made coffee beverage collected by a flow sensor.

[0022] In some embodiments, force-related data related to the human process of making coffee beverages are collected using mechanical sensors. This mechanical data includes the force applied by the barista when holding the pour-over kettle and pouring water, as well as the direction of the force.

[0023] In some embodiments, joint sensors are used to collect motion trajectory data of the joints of the arm, wrist, and fingers when a barista is making coffee drinks, in order to reflect the movement path and posture changes of the hand and limbs.

[0024] In some embodiments, visual sensors arranged in the pour-over operation area are used to acquire visual data of the entire process of human coffee making in a non-contact, remote manner, such as water pouring position, water flow pattern, pot tilt angle, coffee grounds status, and other visual information.

[0025] In some embodiments, flow sensors are used to collect flow data during the process of making coffee beverages, mainly flow rate-related parameters such as the water flow velocity of hot water from a pour-over kettle and the amount of water output per unit time.

[0026] In some embodiments, the first sensor is not limited to the examples described above. For example, the first sensor may also include a temperature sensor for detecting water temperature, etc.

[0027] Thus, by employing multi-dimensional data acquisition using mechanical sensors, joint sensors, vision sensors, and / or flow sensors, key information such as force, direction, movement trajectory, visual characteristics, and water flow rate during human coffee-making operations can be accurately and quantitatively obtained. This objectively reproduces the detailed characteristics of standard demonstration operations, reducing subjective bias and data gaps caused by manual recording. The collaborative acquisition of data from at least two types of sensors can cover multiple levels, including action, posture, vision, and fluid parameters, making the demonstration data more complete and comprehensive. This provides a reliable data foundation for subsequently building standard operating models for human coffee-making, achieving intelligent teaching and action error correction, and improving the accuracy of hand-drip replication, while also enhancing the automation and accuracy of data acquisition.

[0028] In some embodiments, an AI agent analyzes and processes previously collected demonstration data of human coffee making, extracts and learns the corresponding operational rules and features from the demonstration data, thereby generating a first skill model. This first skill model is used to control the first robot to execute the coffee making process, enabling the robot to imitate the actions and methods of a professional barista in making coffee drinks, and ultimately presenting a coffee making effect similar to that of a barista's hand-drip coffee.

[0029] In some embodiments, the primary skill model can be an AI model, specifically designed to be integrated with human coffee-making scenarios. This primarily includes the following common and suitable AI models, all used to accurately translate barista skills: The first skill model can be a deep learning model, such as a convolutional neural network (CNN), a recurrent neural network (RNN), or a long short-term memory network (LSTM). CNN is mainly used to analyze the image data collected by the visual sensor and extract visual features such as the water pouring position, the tilt angle of the kettle, and the state of the coffee grounds. RNN or LSTM is mainly used to process time-series data, such as the duration of force in mechanical data, the temporal changes in motion trajectory, and the real-time fluctuations in flow rate data, to learn the temporal patterns of the barista's operation (such as the speed of water pouring and the gradual change in force).

[0030] The first skill model can be a reinforcement learning model. By learning from demonstration data and combining the taste feedback of humans making coffee drinks (which can be used as auxiliary data), the model parameters are continuously optimized, so that the robot can fine-tune its actions according to real-time conditions (such as the dryness of the coffee grounds and the deviation of the water flow) when performing operations, so as to approach the optimal operation effect of the barista.

[0031] The first skill model can be a regression or classification model, such as logistic regression or support vector machine (SVM). It is mainly used to quantify and analyze mechanical data and flow data, establish a correlation between the barista's operating force, direction of force, water flow rate and other parameters with the hand-drip effect, output standardized control parameters, and guide the robot to accurately execute the corresponding actions.

[0032] It should be noted that the first skill model is not limited to AI models; it can also be a non-AI model composed of specific control instructions and descriptive information. Its core purpose is to provide the robot with executable operational guidelines, and specific forms include: The first skill model may include action sequence instructions: breaking down the entire hand-drip barista process into a series of continuous, executable action steps, arranged chronologically to form an action sequence. For example, the steps involved in this action sequence instruction may include: Step 1: Hold the pour-over kettle at a position X cm directly above the filter cup (corresponding to the position parameters extracted from the motion trajectory data); Step 2: Tilt the kettle body with force YN (corresponding to the operating force in the mechanical data) for Z seconds (corresponding to the duration of the force) and control the water flow rate to n ml / s (corresponding to the flow rate data). Step 3: Pour water in a uniform circular motion along the edge of the filter cup at a speed of m cm / s (corresponding to motion trajectory data). The robot only needs to perform the actions in sequence to simulate the barista's operation.

[0033] The first skill model may include a fixed set of motion parameters. Key operational parameters extracted from demonstration data are organized into a standardized parameter set, eliminating the need for complex model calculations. The robot directly calls these parameters to execute operations. For example, parameters such as "total water filling time 20 seconds, maximum force 8N, minimum force 3N, average flow rate 12ml / s, maximum tilt angle of the kettle 30°, and water filling trajectory radius 4cm" constrain the robot's movements, ensuring operational consistency.

[0034] The first skill model can be action description and constraint information: in the form of a combination of text and numerical values, the action requirements and constraints of robot operation are clearly defined, such as "when pouring water, the direction of force should always be towards the center of the filter cup, the force fluctuation should not exceed ±1N, the water flow should be continuous and uninterrupted, and the tilt angle of the pot should be gradually adjusted according to the dryness and wetness of the powder layer (15° when the powder is dry, and adjusted to 30° after the powder layer is wet)". This kind of descriptive information combined with basic parameters constitutes a simple, easy-to-understand, and directly executable non-AI skill model.

[0035] Whether it's an AI model or a non-AI model, the first skill model analyzes and transforms data from barista's hand-drip demonstrations, providing the first robot with executable operational guidelines to simulate the effect of a human barista making coffee drinks.

[0036] In this way, by using AI agents to analyze demonstration data and generate first skill models (including both AI and non-AI models), it is possible to flexibly adapt to different robot control scenarios, effectively transforming baristas' hand-drip experience and operational skills into standardized content that robots can recognize and execute. This enables the digital replication of human coffee-making skills, breaks the "experience dependence," and allows robots to complete hand-drip operations without real-time human intervention.

[0037] For the first skill model in the form of AI model, it has strong learning and generalization capabilities. It can deeply explore the hidden patterns in the demonstration data (such as the relationship between the barista's operating force and taste, and subtle changes in the timing of actions). It can not only accurately replicate the barista's standard operation, but also make adaptive adjustments based on real-time operation deviations (such as water flow fluctuations and abnormal coffee layer), thereby improving the stability and consistency of the taste of the robot making coffee drinks. At the same time, it can continuously learn new demonstration data to continuously optimize the operation effect and adapt to hand-drip scenarios with different taste requirements.

[0038] For first-skill models that are not in the form of AI models, such as action sequences and action parameters, the structure is simple and the execution efficiency is high. They do not require complex model calculations and hardware support and can be quickly deployed to various simple robotic devices, reducing the implementation cost and technical threshold of intelligent humans making coffee drinks. At the same time, these models are intuitive and easy to understand, making it easy for humans to adjust parameters and optimize actions. They can also be quickly adapted to different sizes of hand-drip equipment (such as filter cups and hand-drip kettles of different capacities), improving the flexibility of use.

[0039] The two forms of first skill models complement each other and can be flexibly selected according to actual application scenarios (such as large-scale mass production, small store use, and personalized customization). This ensures the stability and standardization of the effects of human-made coffee drinks while reducing the difficulty and cost of technical implementation. It provides reliable technical support for the popularization of intelligent human-made coffee drinks, while reducing the reliance on professional baristas and enabling the rapid replication and large-scale application of high-quality hand-drip techniques.

[0040] In some embodiments, parsing the demonstration data to obtain a first skill model includes: Based on the demonstration data, obtain the action sequence of the human making coffee beverage and the action parameters of each operation in the action sequence; the action parameters include the motion trajectory and / or mechanical parameters of each action. A first skill model is determined based on the action timing sequence and the action parameters; the first skill model includes at least one of the following: the operation timing sequence of the robot making coffee, the operation parameters of each operation in the operation timing sequence, and the perception data and / or triggering conditions for performing each operation.

[0041] In some embodiments, the action sequence can be understood as: a sequence of actions for making coffee arranged in chronological order.

[0042] For example, on the one hand, we can trace all the actions a barista performs from the start of a pour-over coffee to the end, in chronological order, forming a temporal sequence of actions. A common pour-over sequence is as follows: reaching for the kettle, moving it directly above the filter, tilting the kettle to pre-pour water, moving the spout at a steady pace to complete the full pour, stopping the tilt, turning off the water flow, and returning the kettle to its storage location. This series of actions arranged in chronological order constitutes the temporal sequence of actions involved in making coffee.

[0043] On the other hand, the motion parameters corresponding to each operation in the action time sequence are extracted from the demonstration data. These motion parameters are mainly the motion trajectory and / or mechanical parameters of each action.

[0044] In some embodiments, the action parameter includes, but is not limited to, at least one of the following: Action parameters for gripping a pour-over kettle: For example, the hand moves upward from the initial position on the table, then moves horizontally to the right, and finally reaches the position directly above the handle of the pour-over kettle; Example of mechanical parameters: When gripping the handle of the pour-over kettle, the applied gripping force is XN, the direction of the gripping force is perpendicular to the surface of the handle, and the duration is Y seconds, ensuring a firm grip on the kettle body without damaging the handle.

[0045] The motion parameters for moving the pour-over kettle are as follows: For example, from the kettle's storage location, move horizontally at a speed of Z cm per second to directly above the filter cup, ensuring the spout always faces the filter cup without deviation. During the movement, maintain a lifting force of m N, with the force direction vertically upwards to balance the kettle's weight and prevent hot water from spilling out, for a duration of n seconds.

[0046] The parameters for the pouring action of a tilting kettle are as follows: For example, the kettle body is slowly tilted from a vertical position at a speed of p° per second to an angle Q° with the tabletop. During the tilting process, the spout remains aligned with the center of the filter cup, and this angle is maintained after tilting. For instance, when tilting the kettle, a pushing force of wN is applied, directed diagonally downwards towards the filter cup, lasting for E seconds. After tilting, a holding force of RN is applied to maintain the tilt angle until the pouring is complete.

[0047] Action parameters for the spout pouring action, example of the motion trajectory: The spout moves clockwise in a circular motion along the edge of the filter cup from the center position at a speed of 2cm per second, with a radius of 3cm. After one circular motion, it slowly moves from the edge of the filter cup back to the center, with a distance of 3cm. Example of mechanical parameters: When moving the spout, the applied control force is 1.5N, in the same direction as the movement direction, with the force fluctuation not exceeding ±0.2N to ensure smooth movement and even water flow covering the coffee grounds, lasting for 8 seconds.

[0048] The parameters for stopping the pouring and returning the kettle to its original position. For example, the kettle is slowly returned to a vertical position from a tilted state at a speed of I per second, and then moved horizontally back to its storage location at a speed of U cm per second, placing it vertically downwards. For example, when returning the kettle to its original position, a pulling force of 0 N is applied, directed diagonally upwards, for a duration of T seconds; when returning it to its original position, a downward force of J N is applied to ensure the kettle is placed stably for a duration of K seconds.

[0049] In some embodiments, a first skill model is determined based on the action timing sequence and the various action parameters obtained in the first step.

[0050] After acquiring the action sequence and action parameters of a human demonstrating coffee making, the actions in the aforementioned action sequence are mapped to operations that the first robot can perform, based on the first robot's capability parameters. This creates an operation sequence, which is then mapped to the robot's coffee-making operation parameters based on the first robot's capability parameters and the action parameters. Thus, the action sequence corresponds to the operation sequence; the action parameters correspond to the operation parameters. It is worth noting that the actions in the action sequence and the operations in the operation sequence can correspond one-to-one, or multiple actions in the action sequence can correspond to one operation in the operation sequence, or one action in the action sequence can correspond to multiple operations in the operation sequence. Overall, the mapping relationship between the action sequence and the operation sequence is based on the breakdown of the functions and effects achieved by the human making coffee, which are then executed by the first robot to achieve the corresponding functions and effects. For example, when a human makes coffee, the action of pouring water from a kettle might correspond to the first robot opening a hot water tap. When the first robot opens the hot water tap, it needs to control the opening degree, opening angle, and other operation parameters of the hot water tap based on the action parameters of the kettle's water-pouring action.

[0051] In summary, firstly, the operational sequence of the robot's coffee-making process involves transforming the barista's sequential actions into steps that the robot can recognize and execute, maintaining the correct order of movements. Secondly, the operational parameters for each operation within the sequence are the extracted motion trajectories and mechanical parameters, which are then translated into specific standards for the robot's execution of each operation. These include the robot's grip strength and speed when grasping the pour-over kettle, and the angle and force used to tilt the kettle, ensuring precise robot movements. Thirdly, the strategy for executing the operational sequence comprises the overall rules governing the robot's execution of the entire sequence. This includes rules for adjusting movements when encountering abnormal water flow, correcting pouring time deviations, or fine-tuning motion parameters based on the amount of coffee grounds, ensuring the robot can stably complete the pour-over operation and simulate the barista's pour-over effect.

[0052] In this way, by obtaining the first skill model through the above method, the hand-drip skill can be digitally replicated. By extracting the action sequence and various action parameters, the details of the barista's hand-drip process can be objectively quantified, reducing subjective errors in manual recording and easily converting it into a first skill model that the robot can imitate and replicate. Furthermore, relying on the operation sequence and parameters in the first skill model ensures standardized robot actions, improving the consistency and stability of the taste of coffee beverages made by humans. This reduces the difficulty and learning cost of implementing intelligent hand-drip robotics, eliminates the need for manual programming, reduces reliance on professional baristas, and facilitates the large-scale application of high-quality hand-drip techniques.

[0053] In some embodiments, S1120 may specifically include: The demonstration data preprocessing may include: first, cleaning the collected demonstration data of human coffee making (including mechanical data, motion trajectory data, visual data, etc.) to remove abnormal data (such as invalid values ​​mistakenly collected by sensors and abnormal motion data caused by barista operation errors); then, standardizing the data to ensure the accuracy and consistency of the data, providing a reliable foundation for subsequent analysis.

[0054] Action sequence extraction can include: based on preprocessed demonstration data, using imitation learning algorithms and / or action primitive algorithms to extract the complete action sequence of a barista making coffee beverages, clarifying the order of each operation, and labeling the start and end times of each action to form a complete sequence of actions and timing. Action primitive algorithms can break down continuous pour-over actions into independent basic action units (such as grasping, moving, tilting, etc.), while imitation learning algorithms learn the barista's action logic, ensuring that the extracted sequence closely matches actual operating habits.

[0055] Action parameter extraction can include: for each operation in the action time sequence (i.e., the basic action unit after action primitive decomposition), continuing to use imitation learning algorithms and / or action primitive algorithms to extract the corresponding motion trajectory parameters and mechanical parameters (such as grip strength and movement trajectory for grasping actions, tilt angle and water flow force for water injection actions, etc.), and quantifying and organizing the parameters to form a parameter set for each action, ensuring that the parameters can be directly used for robot control. Imitation learning algorithms can learn the parameter patterns of the same action in different demonstrations, while action primitive algorithms clarify the parameter range of each basic action, improving the accuracy of parameter extraction. Furthermore, for each action in the action time sequence, extracting the corresponding motion trajectory parameters and mechanical parameters (such as grip strength and movement trajectory for grasping actions, tilt angle and water flow force for water injection actions, etc.), and quantifying and organizing the parameters to form a parameter set for each action, ensuring that the parameters can be directly used for robot control.

[0056] The generation of the first skill model in S1130 may include: based on the action timing sequence and action parameters extracted by the imitation learning algorithm and / or action primitive algorithm, further transforming them into content that the robot can recognize and execute according to the robot's capability set, determining the first skill model (including robot operation sequence, operation parameters and execution strategy), and completing the generation and debugging of the model; among which, the imitation learning algorithm can optimize the coherence of the action timing, and the action primitive algorithm can solidify the parameter standards of the basic actions, ensuring that the model can guide the robot to accurately replicate the hand-push action.

[0057] For example, imitation learning algorithms can be used to enable AI agents to learn from demonstration data (a barista's hand-drip technique), autonomously imitating action logic, operational habits, and temporal relationships without requiring manual pre-setting of action rules. They can automatically capture commonalities and subtle patterns in operations. Action primitive algorithms can be used to break down continuous, complex actions (a complete hand-drip process) into independent, reusable basic action units (action primitives), clearly defining the boundaries and temporal nodes of each unit for accurate extraction of action parameters. Imitation learning algorithms focus on learning and imitation, capturing action logic and temporal relationships, and adapting to the continuity of operations; action primitive algorithms focus on decomposition and standardization, breaking down actions, clarifying parameter boundaries, and adapting to accurate parameter extraction. The two algorithms work synergistically and complement each other, ensuring both the continuity of action temporality and the accuracy of parameter extraction.

[0058] In some cases, imitation learning algorithms and action primitive algorithms are used simultaneously to process demonstration data of human coffee making. The action primitive algorithm breaks down the barista's continuous hand-drip action into actions and clarifies the boundaries of each action. The imitation learning algorithm learns the temporal logic and parameter patterns of each unit, and together they complete the extraction of action temporal sequences and action parameters, providing accurate support for the generation of the first skill model.

[0059] In some embodiments, such as Figure 3 As shown, S1140 may include: S1141: Conduct a skill quality assessment of the first skill model; S1142: Based on the results of the skill quality assessment, apply the qualified first skill model online and distribute the model parameters of the qualified first skill model to at least one second robot.

[0060] In some embodiments, S1141 may include: comparing the output effect of the first skill model with the demonstration data of human coffee making, qualitatively verifying the effectiveness, stability and adaptability of the model, eliminating problems such as overfitting, operation process deviation, and unreasonable parameter settings that may occur during the training process of the first skill model, and providing a reliable basis for subsequent online deployment and distribution.

[0061] For example, S1141 may include: first defining the evaluation criteria, focusing on the needs of human beings in making coffee beverages, and determining evaluation dimensions including process integrity, operational accuracy and / or effect stability for evaluation.

[0062] The process integrity assessment may include: determining whether the entire process of human coffee beverage making in the demonstration data is completely replicated, without missing core operational steps such as grinding, distributing, brewing, and collecting liquid.

[0063] The accuracy of operation needs to be verified, including whether the key parameters such as the powder-to-water ratio control, brewing speed adjustment, and water temperature control are consistent with the demonstration data and have no obvious deviation when the first robot is driven to perform the hand-brew operation according to the first skill model.

[0064] The stability evaluation of the effect may include: after multiple sets of repeated tests, verifying that when the model drives the robot to make coffee, there is no operational lag or process disorder, and the taste, concentration and other effects of the finished coffee do not fluctuate significantly.

[0065] In some embodiments, a simulated scenario of human coffee beverage making with the same source as the demonstration data is constructed, providing coffee beans and brewing equipment of the same specifications. The first robot, driven by the first skill model, continuously executes multiple sets of human coffee beverage making operations, recording the operation process, parameter control, and finished product effect of each set. The model performance is observed in marginal scenarios (such as slight deviations in coffee bean grind size or slight fluctuations in water temperature), while reserving space for expansion of evaluation dimensions. Evaluation items such as coffee bean category adaptation and brewing time control can be added according to actual application needs.

[0066] When conducting a qualitative assessment of feasibility, if the first skill model meets all the above evaluation criteria, it is deemed qualified; if there are problems such as missing core operation steps, significant deviations in key parameters, or excessive fluctuations in the finished product effect, it is deemed unqualified, and the model is returned to the S1120 stage for optimization and adjustment to ensure that the first skill model entering the subsequent steps can accurately replicate the human demonstration process of making coffee beverages.

[0067] In S1140, the first skill model is implemented through two paths. First, the first skill model is put into application. The qualified first skill model is directly deployed to the core first robot running with AI agent. The model and robot hardware (brewing equipment, control module, sensors, etc.) are adapted and debugged to ensure that the model can drive the robot to perform human coffee beverage making operations normally. After going online, routine trial operation is carried out to monitor the robot's operation process, parameter control and finished product effect in real time to ensure that it is consistent with the human coffee beverage making standards in the demonstration data.

[0068] Secondly, the first skill model parameter distribution involves extracting specified parameters from a qualified first skill model, including key elements such as the coffee-to-water ratio control logic, brewing speed adjustment parameters, water temperature control thresholds, and operational sequence. These parameters are then synchronously distributed to at least one second robot via a robot cluster management system. During distribution, the integrity and accuracy of parameter transmission are ensured to prevent parameter loss or deviation. After parameter distribution, each second robot undergoes adaptation and debugging to verify its consistency with the first robot in executing the human coffee-making process, operational precision, and final product effect under the first skill model parameter guidance. This ensures that the human coffee-making skills of multiple robots are unified, eliminating the need for each robot to learn a specific human coffee-making technique individually, further improving learning efficiency.

[0069] In some embodiments, the method further includes: The first skill model and the second robot after the parameters are distributed are continuously monitored. If problems such as operational deviation, decline in finished product effect, or abnormal parameters occur, the first skill model is optimized and adjusted in a timely manner based on relevant feedback information and demonstration data to achieve dynamic adjustment of the first skill model, taking into account both deployment efficiency and application standardization.

[0070] Thus, through the multi-dimensional and detailed evaluation of S1141, qualified first-skill models capable of accurately replicating the human coffee-making demonstration process are effectively selected, reducing the risk of robot malfunctions and substandard products, and ensuring that skill replication meets demonstration standards. During model deployment, multiple robots do not need to undergo repeated model training; parameter distribution enables multiple second robots to quickly master human coffee-making skills, significantly reducing model training and deployment costs and improving the skill synchronization efficiency of the robot cluster. Both detailed implementation steps clearly define standardized operating procedures, providing a solid foundation for subsequent iterations and upgrades of the first-skill model and multi-skill collaboration (such as adding fancy human coffee-making techniques). Simultaneously, it ensures that the model deployed by the first robot is accurate and reliable, and that the skills of multiple second robots remain consistent with the first robot, effectively adapting to the application needs of large-scale coffee making and improving the overall practicality and stability of the application.

[0071] In some embodiments, S1141 may include at least one of the following: Obtain the coffee-making effect of the first robot based on the first skill model; Acquire second sensor data from the first robot during the coffee-making process based on the first skill model; Based on the coffee-making results and / or the data from the second sensor, a skill quality assessment of the first skill model is performed.

[0072] In some embodiments, the coffee-making effect of the first robot based on a first skill model is obtained: After the robot completes the human-made coffee beverage according to the first skill model, the coffee-making effect is obtained through professional testing equipment and human observation and tasting. This includes not only the extraction degree, mouthfeel intensity, and flavor layers of the coffee, but also the visual presentation effect of the coffee, such as the color of the coffee, which can be divided into light brown, dark brown, etc., to determine whether the color is uniform and whether it conforms to the standard color demonstrated by the barista. If latte art is involved, the presentation effect of the latte art also needs to be tested, including the completeness of the latte art pattern, the smoothness of the lines, and the symmetry of the pattern, to determine whether the latte art is clear and consistent with the demonstration effect. At the same time, the color of the coffee soup is also tested to observe whether the soup is clear and free of impurities, so as to comprehensively grasp the final presentation effect of the coffee beverage made by the robot and humans.

[0073] In some embodiments, second sensor data is acquired during the coffee-making process of the first robot based on the first skill model: Throughout the entire process of the robot performing the human's coffee-making operation, its various onboard sensors synchronously collect relevant data in real time. This data constitutes the second sensor data, which corresponds to the first sensor parameters during the human's coffee-making process. Specifically, the second sensor data includes whether the robot's robotic arm vibrates during operation, the magnitude of the vibration, whether the force applied by the robotic arm exceeds a preset reasonable range, and also includes deviations in the robotic arm's movement trajectory, fluctuations in applied force, stability of water flow, and errors in the tilt angle of the kettle. For example, it collects data on whether the robot vibrates when moving the spout, whether the vibration amplitude is within the allowable range, and whether the force applied when grasping the pour-over kettle and tilting the kettle is too large or too small, or exceeds the previously determined force range, ensuring comprehensive capture of detailed data during the robot's operation.

[0074] In some embodiments, the skill quality of the first skill model is evaluated based on the coffee making effect and / or second sensor data: the obtained coffee making effect (including extraction, taste, color, latte art, etc.) and / or second sensor data (including operational jitter, force range, trajectory deviation, etc.) are used as evaluation criteria to establish standardized evaluation standards to judge the skill quality of the first skill model. For example, if the coffee color is uniform, the latte art pattern is complete and smooth, the extraction degree is within a specified range (e.g., 18%-30%), and the second sensor data shows no obvious jitter in the robotic arm and the force does not exceed the preset range (e.g., trajectory deviation ≤0.5cm, force fluctuation ≤±0.3N), then the first skill model is evaluated as qualified; if the coffee color is uneven, the latte art is blurry, or the extraction is over-extracted or under-extracted, or the second sensor data shows obvious robotic arm jitter or force exceeding a certain range, then the first skill model is determined to have defects and requires subsequent optimization and adjustment.

[0075] In this way, the practicality and accuracy of the first skill model can be accurately determined, and deviations in the model can be detected in a timely manner (such as unreasonable motion parameters or inconsistent timing logic). This provides a clear direction for model optimization. By combining the dual evaluation of production effect and process sensing parameters, the limitations of a single evaluation can be avoided, ensuring that the evaluation results are objective and reliable, guaranteeing the stability of the quality of robotic hand-pouring, and reducing the cost of model debugging. No manual monitoring is required throughout the process. Through automated data collection and standardized evaluation, the model quality verification can be completed quickly, improving the efficiency of the deployment of intelligent hand-pouring systems.

[0076] In some embodiments, obtaining the coffee-making effect of the first robot based on the first skill model includes at least one of the following: Acquire video data of the first robot making coffee based on the first skill model; the video data is at least used to reflect the appearance data of the coffee made by the first robot based on the first skill model. Obtain the taste data of the coffee made by the first robot based on the first skill model.

[0077] In some embodiments, during the process of the first robot making coffee according to the first skill model, a camera (e.g., a high-definition camera) is deployed. The camera's shooting angle is aimed at the core area of ​​coffee making, including the brewing head, the coffee cup receiving area, and / or the finished coffee placement area, ensuring that the entire coffee making process and the finished product state can be captured completely. When the first skill model drives the first robot to start the human coffee beverage making process, the camera starts recording, recording the processes of grinding, distributing, hot water brewing, coffee dripping, and / or finished product formation, generating continuous video data. This video data focuses on reflecting coffee appearance-related information, including but not limited to the coffee's color (such as the uniformity of dark brown and light brown), the fineness and distribution of surface foam, the uniformity of the coffee dripping rate during brewing, and the smoothness of the finished coffee surface. Subsequently, the above appearance data can be extracted through video playback, frame analysis, etc., and compared with the appearance standards in the human coffee beverage demonstration data obtained in step S1110 to determine whether the coffee made by the first robot meets the preset requirements in appearance, providing an intuitive visual basis for the quality assessment of the first skill model.

[0078] In some embodiments, after the first robot completes the coffee beverage preparation for humans, the finished coffee is transferred to a professional taste testing device. This device can accurately detect core taste indicators of coffee, including multiple tastes such as acidity, bitterness, body, aftertaste, and / or freshness. It collects relevant data through sensors and converts it into quantitative values, generating a taste testing report as core data support for the coffee preparation effect. Simultaneously, it can be combined with human tasting, where professional coffee tasters provide subjective evaluations of the finished coffee's taste, supplementing subtle taste differences that the taste testing device cannot capture, ensuring the comprehensiveness and accuracy of the taste data. This taste data will be compared with the taste standards of human-made coffee beverages in the demonstration data to verify whether the coffee prepared by the first robot under the first skill model is consistent with the demonstration effect and meets the preset taste requirements, providing key performance evidence for the skill quality assessment of the first skill model.

[0079] In practical applications, one of the above-mentioned implementation methods can be selected to obtain the coffee-making effect according to the evaluation requirements of the first skill model. Alternatively, two implementation methods can be used simultaneously, combining video data (appearance data) and taste data to comprehensively and accurately reflect the coffee-making effect of the first robot based on the first skill model. This provides reliable data support for the skill quality evaluation of S1141, ensuring the accuracy and comprehensiveness of the evaluation results. At the same time, it takes into account the adaptability and scalability of the scenario, and parameters such as video recording accuracy and taste detection indicators can be adjusted according to actual evaluation requirements.

[0080] In some embodiments, the skill quality assessment of the first skill model based on the coffee-making effect and / or the second sensor data includes at least one of the following: Based on the difference between the second sensor data and the first sensor data, a skill quality assessment of the first skill model is performed; the first sensor data is the sensor data of the human making coffee beverages. Based on the data from the second sensor, the control effect of the coffee made by the first robot based on the first skill model is determined, and the skill quality of the first skill model is evaluated based on the control effect.

[0081] In some embodiments, the first sensor data is standard sensor data collected during a human demonstration of making coffee beverages, covering real-time changes in key parameters such as water temperature, water flow rate, coffee-to-water ratio, brewing pressure, and / or powder distribution intensity during the demonstration operation, and serves as the benchmark for training the first skill model.

[0082] In some embodiments, the second sensor data is real-time operational data collected by the first robot, equipped with temperature sensors, flow sensors, pressure sensors, and / or force sensors, when the first robot performs the human-made coffee beverage preparation based on the first skill model. This data directly reflects the actual operating parameters of the first robot. The magnitude of the difference between the two data directly reflects the accuracy of the first skill model in replicating the demonstration process. The smaller the difference, the closer the model's control of the robot is to the demonstration standard, and the higher the skill quality. If the difference exceeds a preset threshold, it indicates that the model has skill deviations and needs optimization.

[0083] For example, the data from the first and second sensors are first preprocessed to unify the data format and calibrate the data accuracy, ensuring that the comparison dimensions are consistent (e.g., both collect water temperature data per second and water flow rate data per milliliter per minute). Then, for each specified parameter in human coffee brewing, the deviations between the second and first sensor data are compared. A reasonable deviation threshold is set; for specified parameters (e.g., water temperature, coffee-to-water ratio), the threshold is set within a small range, while for unspecified parameters (e.g., coffee distribution speed), the threshold can be appropriately relaxed. Finally, a qualitative and quantitative evaluation is performed based on the overall deviation. If the differences in all specified parameters are within the preset threshold and the differences in unspecified parameters are controllable, the first skill model is deemed to be of acceptable quality. If the differences in specified parameters exceed the threshold, or multiple unspecified parameters show significant deviations, resulting in substandard coffee brewing, the process returns to step S1120 for optimization and adjustment of the first skill model.

[0084] In addition, deviation trend analysis is set up. If the deviation between the second sensor data and the first sensor data gradually decreases in multiple tests, it indicates that the model has good stability and iterative potential, and can be used as an auxiliary basis for evaluation.

[0085] In some embodiments, the second sensor data is not only a real-time record of the robot's operating parameters, but also reflects the control effectiveness, such as the control capability of the first skill model. For example, the first skill model is used to control the first robot to replicate the human demonstration process of making coffee drinks, thereby producing coffee with a good taste. The stability and accuracy of the second sensor data directly correspond to the control effectiveness of the model. The better the control effectiveness, the more accurate the model's output of operating instructions to the robot, and the higher the skill quality. Conversely, poor control effectiveness indicates that the model has control logic deviations and needs to be optimized and improved.

[0086] In practical implementation, features related to control effectiveness are first extracted from the second sensor data, including parameter stability, response timeliness, and / or commands. Parameter stability is mainly achieved by analyzing the fluctuation range of the second sensor data. For example, observe whether the water temperature sensor data remains stable during brewing, without sudden increases or decreases; whether the water flow speed sensor data is uniform, without obvious pauses or abrupt changes. If the parameter fluctuations are within the preset range, it indicates that the model's control of the robot is stable. Response timeliness mainly evaluates whether the parameter changes collected by the robot's sensors are timely after the model outputs operation commands. For example, after the model commands to adjust the water temperature, can the temperature sensor quickly capture the water temperature change without significant delay? If the delay time exceeds the preset value, it indicates that the model's control responsiveness is insufficient. Command execution accuracy mainly evaluates whether the robot can accurately execute the operation commands output by the model. For example, if the model commands a powder-to-water ratio of 1:15, is the actual powder-to-water ratio recorded in the second sensor data consistent with the command value, and is the deviation within an acceptable range? Subsequently, based on the three main categories of characteristics mentioned above, the control effect level of the first robot was determined, divided into three levels: excellent, qualified, and unqualified. If the parameters are stable, the response is timely, the execution accuracy is high, and the coffee-making effect (video data, taste data) meets the standards, then the control effect is excellent, and the quality of the first skill model is qualified. If the parameter fluctuations are small, the response is basically timely, the execution accuracy meets the basic requirements, and the coffee-making effect meets the standards, then the control effect is qualified, and the model quality is qualified. If the parameter fluctuations are too large, the response delay is obvious, and the execution accuracy is insufficient, resulting in the coffee-making effect not meeting the standards, then the control effect is unqualified, the model quality is unqualified, and it needs to be returned for optimization.

[0087] In some embodiments, second sensor data from different coffee bean types and brewing scenarios are combined to analyze the model's control adaptability in different scenarios. For example, different coffee brewing scenarios may have different threshold values. If the first skill model can maintain good control performance in multiple scenarios, it indicates that the model has strong versatility and stability, further improving the comprehensiveness of the evaluation results.

[0088] In practical applications, one of the above-mentioned implementation methods can be selected for quality assessment based on the evaluation requirements of the first skill model. Alternatively, both methods can be combined: the difference between the second and first sensor data can be used to determine the model's replication accuracy, while the second sensor data can be used to determine the model's control effect. This, combined with the coffee-making effect (appearance and taste), forms a multi-dimensional and comprehensive quality assessment system, ensuring the accuracy and reliability of the assessment results. Furthermore, considering the scalability of the technical solution, new sensor parameter evaluation dimensions can be added, deviation thresholds adjusted, and control effect evaluation standards optimized according to changes in the actual application scenario. This adapts to the needs of robots of different specifications and humans making different types of coffee beverages, providing a flexible and feasible implementation path for the quality assessment of the first skill model.

[0089] In some embodiments, the method further includes: Coffee making skill information is generated based on the first model parameters of the first skill model; the coffee making skill information includes, in addition to the first model parameters, at least one of the following: skill name, skill version information, and skill combination information. The coffee-making skill information is added to a skill library, which supports querying, rolling back, and / or updating different coffee-making skill versions.

[0090] In some embodiments, generating coffee-making skill information based on the first model parameters of the first skill model may include: integrating the first model parameters of the first skill model with manageable and identifiable auxiliary skill information to form standardized skill information. This retains the core control logic (i.e., the first model parameters) that drives the robot to make coffee while supplementing auxiliary information to meet the needs of skill management, identification, and collaborative applications, ensuring the completeness and usability of the skill information and providing a foundation for its subsequent inclusion in a skill database for standardized management.

[0091] The primary model parameters directly determine the precision and quality of robotic coffee making, encompassing key control parameters throughout the entire coffee-making process and serving as the core support for coffee-making skill information. In addition to the primary model parameters, the coffee-making skill information may also include at least one auxiliary piece of information, each serving the standardized management and application of the skill. The skill name is used to uniquely and intuitively identify the core purpose of the skill, avoiding confusion with other skills and facilitating subsequent queries and invocations; the skill version information is used to distinguish different iterative versions of the same skill, recording the model update trajectory and supporting skill tracing, querying, and rollback; the skill combination information is used to support the collaborative application of this skill with other related skills, expanding the skill's application scenarios and improving the overall standardization and efficiency of coffee making.

[0092] Adding coffee-making skill information to the skill library to build a standardized and centralized skill storage and management system, and to achieve unified management of coffee-making skill information, is essentially to provide a standardized management carrier for the first skill model and subsequent iterative versions. The core objective is to reduce skill management costs, improve skill reuse and iteration efficiency, and ensure the standardization of skill application.

[0093] The skill library, a structured database specifically storing coffee-making skill information, has the core function of supporting the query, rollback, and / or update of different coffee-making skill versions to achieve standardized management and stable application. The query function uses a multi-dimensional retrieval mechanism to quickly locate skill information and support efficient skill retrieval. The rollback function retains backups of skill information for each version, allowing restoration to a qualified version when a new version of the skill encounters anomalies, ensuring the stability of coffee making. The update function supports iterative optimization of skills, updating skill information based on model optimization results to synchronize the skill library with model iterations. It also allows users to choose to retain or overwrite old versions, forming a version management mode that adapts to different needs.

[0094] In some embodiments, the first skill model comprises two parts: a skill knowledge graph and an action sequence encoding. These two parts work together to enable the coffee-making robot to replicate human coffee-making techniques. The skill knowledge graph, in natural language, carries core knowledge such as the logic and parameter standards of human coffee-making, serving as the logical basis for skill execution. The action sequence encoding transforms the knowledge graph into a sequence of instructions that the robot can recognize, driving the robot to complete specific operations. For example, the knowledge graph ensures the standardization of skill logic, while the action sequence encoding ensures the executableness of operations. The combination of these two elements provides the model with both knowledge support and the ability to accurately drive the robot to make coffee, adapting to the needs of skill information generation and subsequent management.

[0095] In some embodiments, the method further includes: Receive the second model parameters of the second skill model sent by the third robot; Based on the equipment difference information between the first robot and the third robot, the second model parameters are adjusted to obtain the third model parameters suitable for the first robot; The second skill model is run based on the parameters of the third model.

[0096] In some embodiments, by adapting the equipment differences between the first robot and the third robot, the skill models of different robots can be transferred and reused, reducing repetitive training and improving the group learning efficiency of the robots.

[0097] In some embodiments, the device difference information includes, but is not limited to, at least one of the following: robot hardware specifications and / or performance parameters. In a specific implementation, the first robot first receives the second model parameters sent by the third robot through a communication module, synchronously collects the device difference information between itself and the third robot, then adjusts the parameters based on the difference information, and finally runs the adapted third model parameters to achieve normal application of the second skill model on the first robot.

[0098] In some embodiments, adjusting the second model parameters to obtain third model parameters suitable for the first robot based on the device difference information between the first robot and the third robot includes: Based on the equipment differences between the first robot and the third robot, the parameters in the second model parameters related to the joint limits and / or end effector specifications of the first robot are adjusted to obtain the third model parameters.

[0099] In some embodiments, joint limits are constraints on the range of motion of the first robot, directly determining the range of motion that the robot can perform. The specifications of the end effector affect the operational precision of coffee making (such as coffee-to-water ratio control, brewing intensity control, etc.). Both are hardware differences between different robots and are also key factors affecting the adaptability of skill models.

[0100] By making targeted adjustments to the two types of parameters, it can be ensured that the parameters of the third model are strictly adapted to the hardware constraints of the first robot, reducing problems such as joint movement exceeding limits and actuator operation deviations. At the same time, it ensures that the core functions of the second skill model are not lost, ensuring that it can run stably and accurately on the first robot, and realizing the efficient transfer and reuse of the skill model.

[0101] Thus, by receiving the second skill model sent by the third robot, the skill model can be transferred and reused across robots, eliminating the need for the first robot to repeatedly train the second skill model, significantly reducing model training costs and improving skill implementation efficiency. Secondly, by specifically adjusting the joint limits and end effector parameters, the incompatibility issues between different robot devices are resolved, ensuring the stability and accuracy of the skill model's operation and preventing operational anomalies or skill failures. Thirdly, the skill range of the first robot is broadened, allowing it to acquire the second skill from the third robot without autonomous training, improving the first robot's coffee-making capabilities and scenario adaptability. In addition, it provides support for skill collaboration in multi-robot clusters, promotes the flow of skill resources among different robots, drives the large-scale application and efficient management of coffee-making skills, and further improves the full-process control system for robot coffee-making skills.

[0102] In some embodiments, such as Figure 4 As shown, the method further includes: Detect the trigger events for skill learning and select the learning method.

[0103] In some embodiments, the triggering events for skill learning include, but are not limited to, at least one of the following: Received a skill update command from the user; The preset learning cycle is reached, and iterative optimization is initiated. The operating efficiency of the first robot decreased; The coffee's flavor consistency is not up to standard. Product sales continued to decline; The rate of negative user reviews is increasing; A demand for launching new coffee categories has been detected. The service scenarios for the first robot have changed; Other coffee-making robots have developed better brewing strategies; Specific changes in environmental parameters (such as temperature, humidity, and coffee bean type) will lead to increased coffee production costs or make it impossible to continue making a certain type of coffee. We're entering the seasonal change period.

[0104] In some embodiments, the learning method includes, but is not limited to, at least one of the following: Reinforcement learning, with trial and error as its core, involves an agent continuously interacting with the production environment to receive reward feedback, autonomously optimizing its action strategies and control parameters, and gradually approaching the optimal production solution. For example, reinforcement learning may include: testing in a simulation environment and optimizing skills based on a reward function.

[0105] Demonstration learning, based on demonstrations by human expert (barista) demonstrators, uses sensors to capture and memorize motion data to learn from the demonstration data. It can fit the expert operation model through supervised learning and quickly replicate standard production skills.

[0106] Video learning involves extracting visual features and analyzing motions from video data of coffee making (e.g., instructional videos or practical footage). By learning the operational process through image recognition and temporal modeling, and extracting motion sequences, skills acquisition based on visual data can be achieved.

[0107] Transfer learning adapts pre-trained coffee-making models, feature representations, or control strategies from source scenarios and / or devices to the current target device, enabling rapid skill reuse and fine-tuning with limited data. For example, the first robot transfers existing skills for parameter fine-tuning.

[0108] In some embodiments, the learning method is selected according to the received instructions.

[0109] In other embodiments, the learning method is selected based on the triggering event. For example, the strategy for selecting the learning method corresponding to different triggering events may include, but is not limited to, at least one of the following: When receiving instructions or launching a new coffee category, you can prioritize learning from baristas' demonstrations. You can quickly learn the standard production process by directly relying on the practical data of professional baristas, ensuring that the production techniques and taste meet professional standards. When the learning cycle is reached and continuous optimization of production precision is required, reinforcement learning can be used to continuously try and iterate parameters through interaction with the actual production environment, thereby improving the stability and precision of coffee production. When service scenarios change or when it is necessary to quickly supplement production experience in multiple scenarios, video learning can be used to complete visual learning based on a large number of coffee making tutorial videos and practical cases, and quickly adapt to the production needs of different scenarios. When optimizing strategies for similar equipment and reusing mature skills, transfer learning can be employed to transfer proven coffee-making skills and parameter models to the current equipment, reducing learning costs and enabling rapid skill upgrades. For example, when the first robot receives a second skill model from a third robot, transfer learning can be initiated to update the second model parameters of the second skill model to third model parameters suitable for the first robot.

[0110] The skill learning and / or learning method selection here can be a preceding step of S1110. The demonstration data acquired in S1110 can be the skill learning of the first robot under the condition of demonstration learning.

[0111] Based on the selected learning method, a skill knowledge graph and action codes are generated. In some embodiments, the skill knowledge graph may be a natural language representation, while the action codes may be a machine-generated language representation of the learned skill knowledge graph.

[0112] Skills verification testing, exemplified by, is a simulation environment assessment.

[0113] If the test fails, the parameters are adjusted and the training is repeated to adjust the learned skill model. If the test passes, it will be implemented under safety supervision in a small-sample real-world environment. This skill verification test achieves the skill quality assessment described in S1141 above.

[0114] If successful, the skill is officially deployed (equivalent to an application going live) and added to the skill library; Continuous performance monitoring and quality assessment; If the quality deteriorates, acquire a retraining set of skills to augment the data. If the performance is stable, skills are shared, and multiple robots are updated synchronously, the skill model acquired by one robot can be synchronously distributed to multiple robot applications. If the execution fails, the cause of failure is analyzed, the training data is updated, and the relevant steps for generating the skill knowledge graph and action codes are returned.

[0115] This disclosure provides a method for autonomous skill learning for a coffee robot AI agent, enabling the robot to efficiently learn new skills from demonstrations by human baristas and transfer the learned skills to other individual robots, achieving rapid skill diffusion and continuous optimization. Specifically, it may include the following steps: Step 1: A human barista performs a demonstration, and the teaching data acquisition module records the data throughout the process.

[0116] Step 2: The skill learning engine processes the teaching data and extracts the initial skill model.

[0117] Step 3: The robot performs its skills in actual operation and collects execution data.

[0118] Step 4: The skills quality assessment submodule assesses the execution quality and calculates the quality score.

[0119] Step 5: The reinforcement learning algorithm optimizes the skill model parameters based on the quality score.

[0120] Step 6: Repeat steps 3-5 until the skill quality reaches the preset threshold (e.g., quality score > 90 points).

[0121] Step 7: Qualified skills are stored in the skill library, and the cross-robot skill transfer module distributes the skills to other robots.

[0122] Step 8: The robot automatically adapts the parameters to receive the skills and completes the skill deployment.

[0123] Example 1: Learning skills to make coffee drinks A barista is invited to demonstrate coffee brewing, with motion capture recording the barista's pouring technique (pouring angle, speed, and / or force variations). An AI agent analyzes the motion data, extracting key features of the circular pouring motion (a spiral pouring path, expansion from the center outwards, uniform speed, and / or stable force). A control code generation module automatically generates trajectory control code for the robotic arm, which is then verified in a digital twin environment before being deployed to the physical robot. The robot successfully replicates the champion's pour-over technique, and if user blind testing shows a satisfaction level reaching a first threshold, the evaluation is considered successful. For example, this first threshold could include, but is not limited to, 80%, 85%, or 90%.

[0124] Example 2, Learning Latte Art Skills The process involves recording a master latte art artist's tulip latte art demonstration. An AI agent analyzes the movement trajectory (position, speed, and force) of the latte art needle, extracts key feature points (such as the position and force of each fold in the pattern), and generates control code. After verification in a simulation environment, the code is deployed to the robot. The robot successfully reproduces the tulip latte art, achieving a pattern completeness that meets a second threshold. For example, the second threshold may include, but is not limited to, 85%, 90%, or 95%.

[0125] Example 3, Skill Iteration and Optimization After the robot performed the hand-pour skill, the vision sensor detected insufficient water uniformity (over-poured water in some areas). The AI ​​agent analyzed the cause (insufficient speed control precision at the end of the robotic arm), automatically adjusted the speed parameters in the control code, re-verified in the simulation environment, and updated the skill library in the cloud for synchronized optimization across the entire network.

[0126] It should be noted that, for the sake of simplicity, the method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily essential to the embodiments of the present invention.

[0127] like Figure 5 As shown in the figure, a coffee-making skill learning device provided in this embodiment of the invention may include the following modules: The first acquisition module 5110 is used to acquire demonstration data of humans making coffee beverages; The second acquisition module 5120 is used to acquire, based on the demonstration data, the action sequence of the human making coffee beverage and the action parameters of each operation in the action sequence; the action parameters include the motion trajectory and mechanical parameters of each action. The determining module 5130 is used to determine a first skill model based on the action timing sequence and the action parameters; the first skill model includes at least one of the following: an operation timing sequence of the robot making coffee, operation parameters of each operation in the operation timing sequence, and perception data and / or triggering conditions for performing each operation, wherein the first skill model is used for the robot to imitate humans in making coffee beverages; The execution module 5140 is used to apply the first skill model online and distribute the model parameters of the first skill model to at least one second robot.

[0128] In some embodiments, the acquisition module is specifically used to acquire first sensor data sensed by the sensor when the human is making a coffee beverage.

[0129] In some embodiments, the acquisition module is specifically used for at least one of the following: Acquire mechanical data of the human making coffee beverage recorded by a mechanical sensor; the mechanical data is used to indicate at least one of the operating force, the direction of the force, and the duration of the force. Acquire motion trajectory data of the human making coffee beverage, collected by joint sensors; Acquire visual data of the human making coffee beverages collected by a visual sensor; Acquire flow data of the human-made coffee beverage collected by a flow sensor.

[0130] In some embodiments, the parsing module is specifically used to process the demonstration data based on an imitation learning algorithm and / or an action primitive algorithm, and extract the action sequence of the human making coffee beverage and the action parameters of each operation in the action sequence.

[0131] In some embodiments, the execution module includes: An evaluation submodule is used to evaluate the skill quality of the first skill model; The processing submodule is used to apply the qualified first skill model online and / or distribute the model parameters of the qualified first skill model to at least one second robot based on the results of the skill quality assessment.

[0132] In some embodiments, the evaluation submodule is configured to perform at least one of the following: Obtain the coffee-making effect of the first robot based on the first skill model; Acquire second sensor data from the first robot during the coffee-making process based on the first skill model; Based on the coffee-making results and / or the data from the second sensor, a skill quality assessment of the first skill model is performed.

[0133] In some embodiments, the evaluation submodule is further configured to perform at least one of the following: acquiring video data of the first robot making coffee based on the first skill model; the video data being used to reflect at least the appearance data of the coffee made by the first robot based on the first skill model; and acquiring the taste data of the coffee made by the first robot based on the first skill model.

[0134] In some embodiments, the evaluation submodule is further configured to evaluate the skill quality of the first skill model based on the difference between the second sensor data and the first sensor data; the first sensor data is sensor data of the human making coffee beverages; Based on the data from the second sensor, the control effect of the coffee made by the first robot based on the first skill model is determined, and the skill quality of the first skill model is evaluated based on the control effect.

[0135] In some embodiments, the apparatus further includes: A generation module is used to generate coffee making skill information based on the first model parameters of the first skill model; the coffee making skill information includes, in addition to the first model parameters, at least one of the following: skill name, skill version information, and skill combination information. An add module is used to add the coffee making skill information to the skill library, wherein the skill library supports querying, rolling back and / or updating different coffee making skill versions.

[0136] In some embodiments, the apparatus further includes: The receiving module is used to receive the second model parameters of the second skill model sent by the third robot; An adjustment module is used to adjust the second model parameters based on the equipment difference information between the first robot and the third robot to obtain third model parameters suitable for the first robot; The running module is used to run the second skill model based on the third model parameters.

[0137] In some embodiments, the adjustment module is specifically used to adjust the parameters in the second model parameters that are related to the joint limits and / or end effector specifications of the first robot based on the equipment difference information between the first robot and the third robot, so as to obtain the third model parameters.

[0138] As the apparatus embodiment is basically similar to the method embodiment, it is described in a relatively simple manner. For relevant details, please refer to the description of the method embodiment.

[0139] In some embodiments, this AI agent may comprise five main parts: a teaching data acquisition module, a skill learning engine, a skill quality assessment module, a skill library management system, and a cross-robot skill transfer module. The teaching data acquisition module records the operational data of the human barista, the skill learning engine extracts skill models, the skill quality assessment module verifies the learning effect, the skill library stores and manages all skills, and the cross-robot skill transfer module distributes skills to other individual robots.

[0140] The teaching data acquisition module is used to collect operational data through multiple sensors during human barista demonstrations: force / torque sensor (records the force and direction of operation), joint angle sensor (records the movement trajectory of the robotic arm), vision sensor (records visual information of the operation process), and flow sensor (records changes in liquid flow).

[0141] The skill learning engine is used to extract skill models (including motion trajectories, force control strategies, and perceptual conditions) from teaching data based on imitation learning and movement primitives algorithms, and to continuously optimize skill parameters in actual operation through reinforcement learning.

[0142] The skills quality assessment module is used to evaluate the quality of skills learning through multiple dimensions: visual assessment (pattern similarity, latte art shape scoring), sensor assessment (liquid flow consistency, force control accuracy), and sensory assessment (professional evaluator scoring), and comprehensively calculates the skills quality score.

[0143] The skill library management system is used to store and manage all learned skills (skill name, skill model parameters, quality score, version history), and supports skill version management (retaining historical versions and supporting rollback) and skill combination (combining multiple basic skills into complex skills).

[0144] The cross-robot skill transfer module is used to distribute skill models from the skill library to other individual robots via the cloud, automatically adapting to the differences in kinematic parameters of different robots (such as joint limits and end effector specifications) to achieve rapid skill transfer and deployment.

[0145] In addition, embodiments of the present invention also provide an electronic device, which may be included in or correspond to the first robot. For example... Figure 6 As shown, it includes a processor 601, a communication interface 602, a memory 603, and a communication bus 604, wherein the processor 601, the communication interface 602, and the memory 603 communicate with each other through the communication bus 604. Memory 603 is used to store computer programs; When the processor 601 executes the program stored in the memory 603, it performs the following steps: acquiring demonstration data of a human making coffee; acquiring, based on the demonstration data, a sequence of actions of the human making coffee and action parameters of each action in the sequence of actions; the action parameters include the motion trajectory and mechanical parameters of each action; determining a first skill model based on the sequence of actions and the action parameters; the first skill model includes at least one of the following: a sequence of actions of a robot making coffee, operation parameters of each operation in the sequence of actions, and perception data and / or triggering conditions for performing each operation, wherein the first skill model is used by the robot to imitate a human making coffee; applying the first skill model online and distributing the model parameters of the first skill model to at least one second robot.

[0146] In some embodiments, acquiring exemplary data on human preparation of coffee beverages includes: acquiring first sensor data sensed by a sensor during the human preparation of coffee beverages.

[0147] In some embodiments, the acquisition of first sensor data sensed by the sensor during the human's preparation of the coffee beverage includes at least one of the following: acquiring mechanical data of the human's preparation of the coffee beverage recorded by a mechanical sensor; the mechanical data indicating at least one of the operating force, the direction of the force, and the duration of the force; acquiring motion trajectory data of the human's preparation of the coffee beverage collected by a joint sensor; acquiring visual data of the human's preparation of the coffee beverage collected by a vision sensor; and acquiring flow data of the human's preparation of the coffee beverage collected by a flow sensor.

[0148] In some embodiments, parsing the demonstration data to obtain a first skill model includes: obtaining, based on the demonstration data, a sequence of actions of the human making coffee and action parameters of each action in the sequence of actions; the action parameters include the motion trajectory and / or mechanical parameters of each action; determining a first skill model based on the sequence of actions and the action parameters; the first skill model includes at least one of the following: a sequence of actions of the robot making coffee, operation parameters of each operation in the sequence of actions, and perception data and / or triggering conditions for performing each operation.

[0149] In some embodiments, obtaining the action sequence of the human making coffee beverage and the action parameters of each operation in the action sequence based on the demonstration data includes: processing the demonstration data based on an imitation learning algorithm and / or an action primitive algorithm to extract the action sequence of the human making coffee beverage and the action parameters of each operation in the action sequence.

[0150] In some embodiments, the step of applying the first skill model online and distributing the model parameters of the first skill model to at least one second robot includes: performing a skill quality assessment of the first skill model; and, based on the result of the skill quality assessment, applying a qualified first skill model online and distributing the model parameters of the qualified first skill model to at least one second robot.

[0151] In some embodiments, the skill quality assessment of the first skill model includes: obtaining the coffee-making effect of the first robot based on the first skill model; and obtaining second sensor data of the first robot during the coffee-making process based on the first skill model. Based on the coffee-making results and / or the data from the second sensor, a skill quality assessment of the first skill model is performed.

[0152] In some embodiments, obtaining the coffee-making effect of the first robot based on the first skill model includes at least one of the following: Acquire video data of the first robot making coffee based on the first skill model; the video data is at least used to reflect the appearance data of the coffee made by the first robot based on the first skill model. Obtain the taste data of the coffee made by the first robot based on the first skill model.

[0153] In some embodiments, the skill quality assessment of the first skill model based on the coffee-making effect and / or the second sensor data includes at least one of the following: Based on the difference between the second sensor data and the first sensor data, a skill quality assessment of the first skill model is performed; the first sensor data is the sensor data of the human making coffee beverages. Based on the data from the second sensor, the control effect of the coffee made by the first robot based on the first skill model is determined, and the skill quality of the first skill model is evaluated based on the control effect.

[0154] In some embodiments, the method further includes: generating coffee making skill information based on the first model parameters of the first skill model; the coffee making skill information includes, in addition to the first model parameters, at least one of the following: skill name, skill version information, and combination information of skill combinations; adding the coffee making skill information to a skill library, wherein the skill library supports querying, rolling back, and / or updating different coffee making skill versions.

[0155] In some embodiments, the method further includes: receiving second model parameters of a second skill model sent by a third robot; adjusting the second model parameters to obtain third model parameters suitable for the first robot based on device difference information between the first robot and the third robot; and running the second skill model based on the third model parameters.

[0156] In some embodiments, adjusting the second model parameters to obtain third model parameters suitable for the first robot based on the device difference information between the first robot and the third robot includes: adjusting the parameters in the second model parameters related to the joint limits and / or end effector specifications of the first robot based on the device difference information between the first robot and the third robot to obtain the third model parameters.

[0157] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not indicate that there is only one bus or one type of bus.

[0158] The communication interface is used for communication between the aforementioned terminal and other devices.

[0159] The memory may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0160] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0161] like Figure 7 As shown, in another embodiment of the present invention, a computer-readable storage medium 701 is also provided, which stores instructions that, when executed on a computer, cause the computer to perform the coffee-making skill learning method described in the above embodiment.

[0162] In another embodiment of the present invention, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute the coffee-making skill learning method described in the above embodiments.

[0163] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0164] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do 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 said element.

[0165] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0166] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A method for learning coffee making skills, characterized in that, Executed by a first robot equipped with an artificial intelligence (AI) agent, including: To obtain demonstrative data on human coffee beverage production; Based on the demonstration data, obtain the action sequence of the human making coffee beverage and the action parameters of each operation in the action sequence; the action parameters include the motion trajectory and mechanical parameters of each action. A first skill model is determined based on the action sequence and the action parameters; the first skill model includes at least one of the following: an operation sequence for the robot to make coffee, operation parameters for each operation in the operation sequence, and perception data and / or triggering conditions for performing each operation, wherein the first skill model is used for the robot to imitate humans in making coffee beverages; The first skill model is deployed online and its model parameters are distributed to at least one second robot.

2. The method according to claim 1, characterized in that, The acquisition of demonstration data on human coffee beverage making includes: acquiring first sensor data sensed by the sensor when the human is making coffee beverage.

3. The method according to claim 2, characterized in that, The first sensor data obtained by the acquisition sensor during the human preparation of the coffee beverage includes at least one of the following: Acquire mechanical data of the human making coffee beverage recorded by a mechanical sensor; the mechanical data is used to indicate at least one of the operating force, the direction of the force, and the duration of the force. Acquire motion trajectory data of the human making coffee beverage, collected by joint sensors; Acquire visual data of the human making coffee beverages collected by a visual sensor; Acquire flow data of the human-made coffee beverage collected by a flow sensor.

4. The method according to claim 1, characterized in that, The step of obtaining the action sequence of the human making coffee beverage and the action parameters of each operation in the action sequence based on the demonstration data includes: The demonstration data is processed using an imitation learning algorithm and / or an action primitive algorithm to extract the action sequence of the human making coffee drinks and the action parameters of each operation in the action sequence.

5. The method according to any one of claims 1 to 3, characterized in that, The process of applying the first skill model online and distributing the model parameters of the first skill model to at least one second robot includes: Perform a skill quality assessment of the first skill model; Based on the results of the skill quality assessment, the qualified first skill model is deployed online and the model parameters of the qualified first skill model are distributed to at least one second robot.

6. The method according to claim 5, characterized in that, The skill quality assessment of the first skill model includes: Obtain the coffee-making effect of the first robot based on the first skill model; Acquire second sensor data from the first robot during the coffee-making process based on the first skill model; Based on the coffee-making results and / or the data from the second sensor, a skill quality assessment of the first skill model is performed.

7. The method according to claim 6, characterized in that, The process of obtaining the coffee-making effect of the first robot based on the first skill model includes at least one of the following: Acquire video data of the first robot making coffee based on the first skill model; the video data is at least used to reflect the appearance data of the coffee made by the first robot based on the first skill model. Obtain the taste data of the coffee made by the first robot based on the first skill model.

8. The method according to claim 6, characterized in that, The assessment of the skill quality of the first skill model based on the coffee-making effect and / or the second sensor data includes at least one of the following: Based on the difference between the second sensor data and the first sensor data, a skill quality assessment of the first skill model is performed; the first sensor data is the sensor data of the human making coffee beverages. Based on the data from the second sensor, the control effect of the coffee made by the first robot based on the first skill model is determined, and the skill quality of the first skill model is evaluated based on the control effect.

9. The method according to any one of claims 1 to 3, characterized in that, The method further includes: Coffee making skill information is generated based on the first model parameters of the first skill model; the coffee making skill information includes, in addition to the first model parameters, at least one of the following: skill name, skill version information, and skill combination information. The coffee-making skill information is added to a skill library, which supports querying, rolling back, and / or updating different coffee-making skill versions.

10. The method according to any one of claims 1 to 3, characterized in that, The method further includes: Receive second model parameters sent by the third robot, wherein the second model parameters are model parameters of the second skill model learned by the third robot; Based on the equipment difference information between the first robot and the third robot, the second model parameters are adjusted to obtain the third model parameters, wherein the third model parameters are applicable to the first robot; The second skill model is run based on the parameters of the third model.

11. The method according to claim 10, characterized in that, The step of adjusting the second model parameters to obtain the third model parameters based on the equipment difference information between the first robot and the third robot includes: Based on the equipment differences between the first robot and the third robot, the parameters in the second model parameters related to the joint limits and / or end effector specifications of the first robot are adjusted to obtain the third model parameters.

12. A coffee-making skill learning device, characterized in that, include: The first acquisition module is used to acquire demonstration data on how humans make coffee drinks; The second acquisition module is used to acquire, based on the demonstration data, the action sequence of the human making coffee beverage and the action parameters of each operation in the action sequence; the action parameters include the motion trajectory and mechanical parameters of each action; A determining module is configured to determine a first skill model based on the action timing sequence and the action parameters; the first skill model includes at least one of the following: an operation timing sequence of the robot making coffee, operation parameters of each operation in the operation timing sequence, and perception data and / or triggering conditions for performing each operation, wherein the first skill model is used for the robot to imitate humans in making coffee beverages; The execution module is used to apply the first skill model online and distribute the model parameters of the first skill model to at least one second robot.

13. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; The memory is used to store computer programs; When the processor executes a program stored in the memory, it implements the method as described in any one of claims 1-11.

14. A computer-readable medium having instructions stored thereon that, when executed by one or more processors, cause the processors to perform the method as described in any one of claims 1-11.

Images