Robot action generation method and device, electronic equipment and medium

Abstract

Embodiments of the present application provide a robot action generation method and device, electronic equipment and medium, belonging to the field of artificial intelligence, and applied to financial and medical scenarios. The method comprises: performing action task analysis on the action control instruction through a robot action generation model to obtain a target action task, and assigning a basic task step length to the target action task; adjusting the basic task step length according to a target image and proprioceptive data to obtain an adjusted task step length; weighting and fusing the basic task step length and the adjusted task step length to obtain a fused task step length; and generating an action based on the fused task step length to obtain a target action. Through action task analysis, basic task step length assignment, task step length adjustment and action generation on the control data of the target robot, the embodiments of the present application can perform differentiated task step length control according to different task requirements, thereby improving the action control effect of the robot.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and is applied to financial scenarios and medical scenarios, and in particular relates to a robot action generation method and device, an electronic device and a medium. BACKGROUND

[0002] A conventional robot action generation method usually generates an action sequence based on a fixed sampling step and an action template through a diffusion policy generated based on a diffusion model. Taking a financial application scenario as an example, a robot in a bank hall receives a control instruction to deliver a file to a customer, and according to the control instruction, the robot captures continuous actions of picking up the file, moving to the customer and delivering the file according to a fixed sampling step (such as 100 steps) through the diffusion policy. The above-mentioned method of generating robot actions by using a fixed sampling step results in poor robot action control effect. Therefore, how to improve the robot action control effect has become a problem to be solved. SUMMARY

[0003] The main purpose of the embodiments of the present application is to provide a robot action generation method and device, an electronic device and a medium, which aims to improve the robot action control effect.

[0004] To achieve the above-mentioned purpose, a first aspect of the embodiments of the present application provides a robot action generation method, which comprises:

[0005] obtaining an action control instruction for controlling a target robot, and obtaining body perception data and a target image of the target robot according to the action control instruction;

[0006] obtaining a pre-trained robot action generation model; wherein the robot action generation model comprises a task analysis network, a step adjustment network and an action generation network; the step adjustment network comprises a task step adjustment sub-network and a state step adjustment sub-network;

[0007] performing action task analysis on the action control instruction by using the task analysis network to obtain a target action task;

[0008] allocating a basic task step for the target action task by using the task step adjustment sub-network;

[0009] adjusting the basic task step by using the state step adjustment sub-network, the target image and the body perception data to obtain an adjusted task step;

[0010] performing weighted fusion on the basic task step and the adjusted task step to obtain a fused task step;

[0011] generating an action based on the action generation network and the fusion task step, to obtain a target action.

[0012] In some embodiments, the adjusting the basic task step based on the state step adjustment sub-network, the target image and the proprioceptive data to obtain an adjusted task step comprises:

[0013] encoding the target image based on the state step adjustment sub-network to obtain visual features;

[0014] extracting features from the proprioceptive data to obtain proprioceptive data features;

[0015] recognizing a motion state of the target robot based on the visual features and the proprioceptive data features, to obtain a target motion state;

[0016] adjusting the basic task step based on the target motion state to obtain the adjusted task step.

[0017] In some embodiments, the target motion state comprises an interactive motion state and a non-interactive motion state.

[0018] The adjusting the basic task step based on the target motion state to obtain the adjusted task step comprises:

[0019] if the target motion state is the interactive motion state, incrementally adjusting the basic task step to obtain the adjusted task step;

[0020] if the target motion state is the non-interactive motion state, decrementally adjusting the basic task step to obtain the adjusted task step.

[0021] In some embodiments, the generating an action based on the action generation network and the fusion task step, to obtain a target action comprises:

[0022] determining a diffusion step number based on the action generation network and the fusion task step;

[0023] obtaining initial noise data;

[0024] gradually diffusing the initial noise data into the action generation network based on the diffusion step number, to obtain a noise action sequence;

[0025] gradually denoising the noise action sequence based on the diffusion step number, to obtain the target action.

[0026] In some embodiments, before the step of determining the diffusion step number based on the action generation network and the fusion task step, the method further comprises:

[0027] obtaining an initial window size and a window weight of a preset smoothing window filter;

[0028] adjusting the initial window size based on the fusion task step to obtain an adjusted window size;

[0029] adjusting the window weight based on the fusion task step to obtain an adjusted window weight;

[0030] obtaining a historical task step, and performing smoothing processing on the fusion task step based on the historical task step, the adjusted window size and the adjusted window weight to obtain a smoothed fusion task step.

[0031] In some embodiments, the step of performing action task analysis on the action control instruction by using the task analysis network to obtain a target action task comprises:

[0032] performing semantic feature extraction on the action control instruction by using the task analysis network to obtain a control instruction feature;

[0033] performing initial action task analysis on the control instruction feature to obtain an initial action task;

[0034] performing task decomposition on the initial action task to obtain an action subtask;

[0035] performing task precision requirement identification on the action subtask to obtain a task precision requirement;

[0036] determining the target action task based on the action subtask and the task precision requirement.

[0037] In some embodiments, the step of performing weighted fusion on the basic task step and the adjusted task step to obtain a fusion task step comprises:

[0038] obtaining a basic step weight of the basic task step, and obtaining an adjusted step weight of the adjusted task step;

[0039] performing step weighted fusion on the basic task step and the adjusted task step based on the basic step weight and the adjusted step weight to obtain the fusion task step.

[0040] To achieve the above object, a second aspect of the embodiment of the present application proposes a robot action generation device, which comprises:

[0041] The action control data acquisition module is configured to acquire an action control instruction for controlling a target robot, and acquire body perception data and a target image of the target robot according to the action control instruction.

[0042] The action generation model acquisition module is configured to acquire a pre-trained robot action generation model, wherein the robot action generation model comprises a task analysis network, a step adjustment network, and an action generation network; the step adjustment network comprises a task step adjustment subnetwork and a state step adjustment subnetwork.

[0043] The action task analysis module is configured to perform action task analysis on the action control instruction by using the task analysis network to obtain a target action task.

[0044] The basic step matching module is configured to assign a basic task step to the target action task by using the task step adjustment subnetwork.

[0045] The step adjustment module is configured to adjust the basic task step by using the state step adjustment subnetwork, the target image, and the body perception data to obtain an adjusted task step.

[0046] The step fusion module is configured to perform weighted fusion on the basic task step and the adjusted task step to obtain a fused task step.

[0047] The action generation module is configured to perform action generation based on the action generation network and the fused task step to obtain a target action.

[0048] To achieve the above object, a third aspect of the embodiments of the present application provides an electronic device, which comprises a memory and a processor, the memory stores a computer program, and the processor implements the method of the first aspect when executing the computer program.

[0049] To achieve the above object, a fourth aspect of the embodiments of the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the method of the first aspect.

[0050] The robot action generation method and device, the electronic equipment and the medium provided in the application obtain the action control instruction for controlling the target robot, the body perception data and the target image, which helps the robot to comprehensively perceive the task demand and the environment state, and provides rich data basis for subsequent action generation. Then, the control data of the target robot is subjected to action task analysis, basic task step distribution, task step adjustment and action generation by the robot action generation model, which can perform differentiated task step control for different task demands, realizes more task steps for high-precision tasks, ensures the accuracy of the generated action, reduces the task steps for low-precision tasks, improves the efficiency of the generated action when the accuracy of the generated action is unchanged, and thus improves the action control effect of the robot. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 is a flowchart of the robot action generation method provided by the embodiment of the application;

[0052] Figure 2 is a flowchart of step S103 in Figure 1

[0053] Figure 3 is a flowchart of step S105 in Figure 1

[0054] Figure 4 is a flowchart of step S304 in Figure 3

[0055] Figure 5 is a flowchart of step S106 in Figure 1

[0056] Figure 6 is another flowchart of the robot action generation method provided by the embodiment of the application;

[0057] Figure 7 is a flowchart of step S107 in Figure 1

[0058] Figure 8 is a structural schematic diagram of the robot action generation device provided by the embodiment of the application;

[0059] Figure 9 is a hardware structural schematic diagram of the electronic equipment provided by the embodiment of the application. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical scheme and advantages of the application clearer, the application is further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the application, and are not used to limit the application.​​​​​

[0061] It should be noted that although the functional modules are divided in the device schematic diagram, and the logical sequence is shown in the flowchart, in some cases, the steps shown or described can be performed in a manner different from the module division in the device or the sequence in the flowchart. The terms "first", "second", and the like in the specification and claims and the above-described drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0062] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification herein is for the purpose of describing the embodiments of the present application only and is not intended to be limiting of the present application.

[0063] First, the meanings of several terms involved in the present application are analyzed:

[0064] Artificial intelligence (AI): is a new technical science that studies, develops theories, methods, technologies and application systems for simulating, extending and expanding human intelligence; artificial intelligence is a branch of computer science, and artificial intelligence attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a similar way to human intelligence. The field of research includes robots, language recognition, image recognition, natural language processing and expert systems. Artificial intelligence can simulate the information process of human consciousness and thinking. Artificial intelligence is also the theory, method, technology and application system of using digital computers or digital computer controlled machines to simulate, extend and expand human intelligence, to perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0065] The embodiments of the present application provide a robot action generation method and device, electronic equipment and medium, aiming to improve the action control effect of the robot.

[0066] The robot action generation method and device, electronic equipment and medium provided by the embodiments of the present application are specifically described through the following embodiments. First, the robot action generation method in the embodiments of the present application is described.

[0067] The embodiments of the present application can acquire and process related data based on artificial intelligence technology. Artificial intelligence (AI) is a theory, method, technology and application system of using digital computers or digital computer controlled machines to simulate, extend and expand human intelligence, to perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0068] The artificial intelligence basic technology generally includes technologies such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction system, mechatronics, etc. The artificial intelligence software technology mainly includes computer vision technology, robot technology, biometric identification technology, speech processing technology, natural language processing technology, and machine learning / deep learning, etc.

[0069] The robot action generation method provided by the embodiments of the present application relates to the technical field of artificial intelligence. The robot action generation method provided by the embodiments of the present application can be applied to a terminal, can also be applied to a server end, and can also be software running in the terminal or the server end. In some embodiments, the terminal can be a smart phone, a tablet computer, a notebook computer, a desktop computer, etc.; the server end can be configured as an independent physical server, can also be configured as a server cluster or a distributed system composed of multiple physical servers, can also be configured as a cloud server providing basic cloud computing services such as cloud service, cloud database, cloud computing, cloud function, cloud storage, network service, cloud communication, middleware service, domain name service, security service, CDN, and big data and artificial intelligence platform; and the software can be an application implementing the robot action generation method, etc., but is not limited to the above forms.

[0070] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld devices or portable devices, tablet devices, multi-processor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present application can also be practiced in a distributed computing environment in which tasks are performed by remote processing devices connected by a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media, including storage devices.

[0071] Figure 1 is an optional flowchart of the robot action generation method provided by the embodiments of the present application, Figure 1 The method in can include but is not limited to steps S101 to S107.

[0072] In step S101, an action control instruction for controlling a target robot is acquired, and ontology perception data and a target image of the target robot are acquired according to the action control instruction.

[0073] In step S102, a pre-trained robot action generation model is obtained; the robot action generation model comprises a task analysis network, a step adjustment network, and an action generation network; the step adjustment network comprises a task step adjustment subnetwork and a state step adjustment subnetwork.

[0074] In step S103, the task analysis network is used to perform action task analysis on the action control instruction to obtain a target action task.

[0075] In step S104, the task step adjustment subnetwork is used to assign a basic task step to the target action task.

[0076] In step S105, the state step adjustment subnetwork, the target image, and the proprioceptive data are used to adjust the basic task step to obtain an adjusted task step.

[0077] In step S106, the basic task step and the adjusted task step are weighted and fused to obtain a fused task step.

[0078] In step S107, action generation is performed based on the action generation network and the fused task step to obtain a target action.

[0079] The steps S101 to S107 shown in the embodiments of the present application are helpful to the robot to fully perceive the task demand and the environment state and provide rich data basis for subsequent action generation by obtaining the action control instruction for controlling the target robot, the proprioceptive data, and the target image; then, the control data of the target robot are subjected to action task analysis, basic task step assignment, task step adjustment, and action generation by the robot action generation model, which can perform differentiated task step control for different task demands, assigns more task steps for high-precision tasks to ensure the accuracy of the generated action, reduces the task steps for low-precision tasks, improves the efficiency of the generated action while ensuring the accuracy of the generated action, and thus improves the action control effect of the robot.

[0080] In step S101 of some embodiments, specifically, the action control instruction is a control instruction issued by a user to guide the robot to complete a task, and the action control instruction can be a text instruction in the form of natural language.

[0081] For example, in a financial application scenario, the action control instruction can be "please deliver the account opening file to the customer"; in a medical application scenario, the action control instruction can be "please deliver the medicine to the patient room".

[0082] Specifically, the proprioceptive data refers to the motion state data of the robot, including but not limited to the joint position, joint speed, and gripper state of the robot.

[0083] Specifically, the target image is an environment image obtained by a robot vision sensor, and is used to perceive the position and state of the target object.

[0084] For example, in a financial application scenario, when the robot receives an action control instruction of "please deliver the account opening file to the customer", the robot can obtain the position and speed of its own joints through the joint sensor, and obtain the position and state image of the file through the vision sensor; in a medical application scenario, when the robot receives an action control instruction of "please deliver the medicine to the patient room", the robot can obtain the position and speed of its own joints through the joint sensor, and obtain the position and state image of the medicine through the vision sensor.

[0085] In this embodiment, by obtaining the action control instruction and obtaining the body perception data and the target image of the target robot according to the action control instruction, the robot can comprehensively perceive the task demand and the environment state, and ensure the accuracy and continuity of subsequent action generation.

[0086] In step S102 of some embodiments, specifically, the robot action generation model refers to a neural network model composed of a task analysis network, a step adjustment network, and an action generation network. The step adjustment network is further divided into a task step adjustment sub-network and a state step adjustment sub-network.

[0087] Specifically, the task analysis network can be a large language model (LLM, Large Language Model) for analyzing the action task information of the action control instruction; the task step adjustment sub-network can be a task adapter (Task-Based Adaptor) for allocating a basic task step according to the task demand accuracy; the state step adjustment sub-network can be a state adapter (State-Based Adaptor) for further adjusting the basic task step in combination with the target image and the body perception data; and the action generation network can be a diffusion model-based policy (Diffusion Policy) network for generating the final robot action according to the fused task step and the adjusted step.

[0088] Please refer to Figure 2 In some embodiments, step S103 includes but is not limited to steps S201 to S205:

[0089] In step S201, a task analysis network is used to extract semantic features of the action control instruction to obtain control instruction features.

[0090] In step S202, the control instruction features are subjected to initial action task analysis to obtain initial action tasks.

[0091] Step S203, task decomposition is performed on the initial action task to obtain an action subtask.

[0092] Step S204, task precision requirement identification is performed on the action subtask to obtain a task precision requirement.

[0093] Step S205, a target action task is determined based on the action subtask and the task precision requirement.

[0094] In step S201 of some embodiments, specifically, the control instruction feature refers to an action control semantic vector extracted from the action control instruction.

[0095] Specifically, the action control instruction can be encoded into a control instruction token through a text encoding layer (such as a Bert model).

[0096] For example, in a financial application scenario, if the robot receives an instruction to "deliver a file to a customer", the task parsing network can extract key semantic feature vectors such as "customer", "delivery", and "file"; in a medical application scenario, for the instruction "send medicine to the ward", semantic feature vectors such as "medicine" and "ward" are extracted.

[0097] In step S202 of some embodiments, specifically, the initial action task can be obtained by performing initial action task analysis on the control instruction feature.

[0098] For example, in a financial application scenario, the control instruction feature is a key semantic vector such as "customer", "delivery", and "file", and through initial action task analysis, these features are converted into a specific action task such as "move to the file location and grab the file". In a medical application scenario, the control instruction feature is a semantic vector such as "medicine" and "ward", and the analyzed initial action task can be "move to the medicine cabinet and grab the medicine".

[0099] In this embodiment, the initial action task is obtained by performing initial action task analysis on the control instruction feature, which ensures that the robot can extract a specific executable action task from an abstract instruction.

[0100] In step S203 of some embodiments, specifically, task decomposition is performed by analyzing the semantics and context information of the initial action task to decompose the initial action task into multiple specific subtasks.

[0101] For example, in a financial application scenario, when the robot receives an initial action task of "delivering a file to a customer", the task of "delivering a file to a customer" is split into multiple action sub-tasks: moving to the file location, grabbing the file, moving to the customer location, and delivering the file. In a medical application scenario, for an initial action task of "delivering medicine to a patient room", the task of "delivering medicine to a patient room" is split into multiple action sub-tasks: moving to the medicine cabinet, grabbing the medicine, moving to the patient room, and delivering the medicine.

[0102] In step S204 of some embodiments, specifically, the task demand accuracy refers to the accuracy that the robot needs to achieve when performing the action sub-tasks.

[0103] Specifically, the task demand accuracy can be determined based on whether the robot interacts with the environment. If the robot needs to interact with the environment, it indicates that the task demand accuracy is high. If only the robot itself moves, it indicates that the task demand accuracy is low.

[0104] For example, in a financial application scenario, for the "grabbing a file" action sub-task, the robot needs to interact with the file in the environment, indicating that this task is a high-precision task demand. For the "moving to the file location" action sub-task, only the robot itself moves, indicating that this task is a low-precision task demand. In a medical application scenario, for the "grabbing medicine" action sub-task, the robot needs to interact with the medicine in the environment, and the shape and size of the medicine may be irregular, indicating that this task is a high-precision task demand.

[0105] In this embodiment, by identifying the task accuracy demand of the action sub-tasks, the robot can adjust the fineness of the action when performing the task to improve the success rate of task execution.

[0106] In step S205 of some embodiments, specifically, the target action task refers to the specific and decomposed action sequence that the robot needs to complete when executing the action control instruction.

[0107] For example, in a financial application scenario, the target action task can be "grabbing a file with high-precision task demand and quickly moving to the customer location with low-precision task demand"; in a medical application scenario, the target action task can be "grabbing medicine with high-precision task demand and safely moving to the patient room with low-precision task demand".

[0108] Through steps S201 to S205, the action control instruction is deeply analyzed by the task parsing network, the robot can accurately understand the task demand and decompose it into specific sub-tasks, and the accuracy of the robot action control is improved through the accurate task parsing and task decomposition mechanism.

[0109] In step S104 of some embodiments, specifically, the basic task step refers to an initial sampling step that the robot needs to take when performing the action task.

[0110] Specifically, the higher the accuracy required by the task, the larger the task step required by the action task; the lower the accuracy required by the task, the smaller the task step required by the action task, so as to improve the efficiency of generating actions subsequently.

[0111] For example, in a financial application scenario, for the target action task of "grabbing a file", the task step adjustment sub-network allocates a basic task step (such as 80 steps) according to the step of the historical grabbing action, so as to ensure the fineness of the action. In a medical application scenario, for the target action task of "grabbing a medicine", the task step adjustment sub-network also allocates a basic task step (such as 80 steps) according to the step of the historical grabbing action.

[0112] In this embodiment, by using the task step adjustment sub-network to allocate a basic task step for the target action task, an explicit initial sampling step is provided for the robot, which helps the robot to generate corresponding actions according to the requirements of the task.

[0113] Please refer to Figure 3 In some embodiments, step S105 includes but is not limited to steps S301 to S304:

[0114] Step S301: using a state step adjustment sub-network to perform visual coding on the target image to obtain visual features.

[0115] Step S302: performing feature extraction on the proprioceptive data to obtain proprioceptive data features.

[0116] Step S303: performing motion state recognition on the target robot based on the visual features and the proprioceptive data features to obtain a target motion state.

[0117] Step S304: adjusting the basic task step according to the target motion state to obtain an adjusted task step.

[0118] In step S301 of some embodiments, specifically, the visual features refer to vector representations of key information such as object position, shape, size, and state extracted from the target image.

[0119] Specifically, the target image can be visually coded by an image coding layer to convert the target image into a set of numerical vectors, i.e., visual tokens, which capture the features of object position, shape, size, and state in the target image.

[0120] For example, in a financial application scenario, the target image shows a file located 0.5 meters in front of the robot, and the image encoding layer extracts this position information as a visual feature. In a medical application scenario, the target image shows a medicine located 0.6 meters in front of the robot, and the image encoding layer extracts this position information as a visual feature.

[0121] In step S302 of some embodiments, specifically, the body perception data features refer to features extracted from the body perception data.

[0122] For example, in a financial application scenario, when the robot performs a task in a bank hall, it can extract joint angle and joint speed numerical features through joint sensors, gripper angle numerical features through gripper sensors, torque and motor current numerical features through torque sensors, and motion speed, acceleration, and angular speed numerical features through speed sensors. In a medical application scenario, when the robot delivers medicine in a hospital, it can also extract joint angle and joint speed numerical features through joint sensors, gripper angle numerical features through gripper sensors, and motion speed, acceleration, and angular speed numerical features through speed sensors.

[0123] In step S303 of some embodiments, specifically, the target motion state comprehensive visual features and the body perception data features determine the dynamic behavior mode of the robot, including but not limited to the robot's movement state, speed, and posture.

[0124] For example, in a financial application scenario, when the bank service robot performs the task of "delivering a file to a customer", the visual features can be the location of the file cabinet and the obstacles on the path, and the body perception data features can be the current joint angle and motion speed of the robot. By integrating this information, it can be determined whether the robot is in the stage of approaching the target object, or whether the motion speed needs to be adjusted to avoid obstacles. In a medical application scenario, when the nursing robot performs the task of "delivering medicine to a patient room", the visual features can be the location of the medicine cabinet and the path to the patient room, and the body perception data features can be the current balance state and load condition of the robot. By integrating this information, it can be determined whether the robot is in the stage of grabbing the medicine, or whether the gripper angle or posture needs to be adjusted to grab the medicine.

[0125] In this embodiment, the target robot is identified based on visual features and body perception data features, and the robot can dynamically adjust its behavior according to real-time visual and body perception information to ensure the accuracy of subsequent action generation.

[0126] Please refer to Figure 4In some embodiments, the target motion state includes an interactive motion state and a non-interactive motion state, and step S304 includes but is not limited to steps S401-S402:

[0127] In step S401, if the target motion state is an interactive motion state, the basic task step is incrementally adjusted to obtain an adjusted task step.

[0128] In step S402, if the target motion state is a non-interactive motion state, the basic task step is decrementally adjusted to obtain an adjusted task step.

[0129] In step S401 of some embodiments, specifically, the target motion state can include an interactive motion state and a non-interactive motion state.

[0130] Specifically, the interactive motion state refers to a task state in which the robot needs to directly contact or interact with a person or other object.

[0131] For example, in a financial application scenario, the motion state of a bank service robot grabbing text is an interactive motion state; in a medical application scenario, the motion state of a nursing robot grabbing medicine is also an interactive motion state.

[0132] Specifically, the non-interactive motion state refers to a state in which the robot completes the task independently without the need for direct contact or interaction with a person or other object.

[0133] For example, in a financial application scenario, the motion state of a robot autonomously patrolling in a bank hall is a non-interactive motion state; in a medical application scenario, the motion state of a robot moving to a medicine is also a non-interactive motion state.

[0134] Specifically, when the target motion state is an interactive motion state, the robot needs to be incrementally adjusted, i.e., the basic task step is increased, to ensure that the task can be completed more accurately and stably during the interaction.

[0135] For example, in a financial application scenario, when a bank service robot is grabbing a file, the target motion state is an interactive motion state, and the task demand accuracy is high accuracy, the robot needs to adjust the basic task step from 80 steps to 100 steps to ensure that the file can be accurately delivered when interacting with the customer. In a medical application scenario, when a nursing robot is grabbing medicine, the target motion state is also an interactive motion state, and the task demand accuracy is also high accuracy, but due to the irregular shape of the medicine, more delicate grabbing control is needed, and the basic task step needs to be adjusted from 80 steps to 120 steps.

[0136] In step S402 of some embodiments, specifically, when the target motion state is a non-interactive motion state, the robot needs to reduce the task step length to improve the efficiency of task execution.

[0137] For example, in a financial application scenario, when a bank service robot autonomously patrols in a bank hall, the target motion state is a non-interactive motion state, and the basic task step length can be adjusted from 80 steps to 60 steps to improve the patrol efficiency and reduce unnecessary energy consumption. In a medical application scenario, when a nursing robot moves to a medicine, the target motion state is a non-interactive motion state, and since the moving distance is short, the basic task step length can be adjusted from 80 steps to 50 steps.

[0138] Through steps S401 to S402, by dynamically adjusting the basic task step length according to the target motion state, the robot can balance precision and efficiency in different task scenarios. Specifically, in an interactive motion state, the step length is automatically adjusted incrementally, which helps to generate higher precision and stability of the action, ensuring the precision of the robot's action when interacting with people or objects; in a non-interactive motion state, the step length is automatically adjusted decrementally, which can improve the motion efficiency of the robot with lower precision requirements, reduce the energy consumption of the robot and shorten the task execution time, realizing a mechanism of flexibly adjusting the step length according to the task requirements, ensuring high-precision execution of interactive tasks, and improving the efficiency of non-interactive tasks.

[0139] Through steps S301 to S304, by combining visual features and proprioceptive data features, the motion state of the robot can be comprehensively reflected, realizing accurate identification of the motion state of the robot, and incrementally sampling the step length for interactive target motion states to ensure action precision, and decrementally sampling the step length for non-interactive target motion states to improve action execution efficiency, realizing adaptive sampling step length control, and solving the problem of poor robot action control effect caused by generating robot actions with fixed sampling step length.

[0140] Please refer to Figure 5 In some embodiments, step S106 includes but is not limited to steps S501 to S502:

[0141] Step S501, obtaining a basic step length weight of a basic task step length, and obtaining an adjustment step length weight for adjusting the task step length.

[0142] Step S502, step length weighted fusion of the basic task step length and the adjustment task step length based on the basic step length weight and the adjustment step length weight to obtain a fused task step length.

[0143] In step S501 of some embodiments, specifically, the basic step length weight is used to adjust the proportion of the basic task step length in the robot action generation process.

[0144] Specifically, the adjustment step weight is used to adjust the proportion of the adjustment task step in the robot action generation process.

[0145] In step S502 of some embodiments, specifically, the fused task step refers to the sampling step obtained by weighting adjustment of the basic task step and the adjustment task step.

[0146] For example, if the basic task step is 80 steps, the adjustment task step is 120 steps, the basic step weight is 0.4, and the adjustment step weight is 0.6, then the fused task step = (80 x 0.4) + (120 x 0.6) = 104 steps.

[0147] In this embodiment, the fused task step takes into account both the requirements of the basic task and the requirements of real-time adjustment, enabling the robot to better complete tasks in complex environments.

[0148] Through steps S501 to S502, by weighting and fusing the basic task step and the adjustment task step, the robot can generate a more optimized fused task step, so that the robot can dynamically adjust the sampling step according to the task requirement accuracy and the complexity of the environment, ensuring the safety and reliability of the task while ensuring the efficiency of the task completion.

[0149] Please refer to Figure 6 In some embodiments, the robot action generation method further includes but is not limited to steps S601 to S604:

[0150] Step S601: Obtain the initial window size and window weight of a preset smoothing window filter.

[0151] Step S602: Adjust the window size according to the fused task step to obtain an adjusted window size.

[0152] Step S603: Adjust the window weight according to the fused task step to obtain an adjusted window weight.

[0153] Step S604: Obtain a historical task step, and perform smoothing processing on the fused task step based on the historical task step, the adjusted window size, and the adjusted window weight to obtain a smoothed fused task step.

[0154] In step S601 of some embodiments, specifically, the smoothing window filter is an algorithm tool for processing a sequence of sampling steps, which smoothes the fluctuations of the data by applying a sliding window on the sampling steps and calculating the weighted average of the sampling steps in the window, to reduce the influence of noise.

[0155] For example, in a financial application scenario, when a bank service robot is performing a task, the task step length may fluctuate sharply due to environmental changes or task switching. At this time, the smoothing window filter can smooth the step length change to ensure the stability of the action.

[0156] Specifically, the initial window size and the window weight are two key parameters of the smoothing window filter. The initial window size determines the number of data points contained in the window, and the window weight determines the contribution degree of each data point in calculating the weighted average value.

[0157] For example, in a financial application scenario, when a bank service robot is performing a file delivery task, the initial window size may be set to contain data points of the last 5 task steps, and the window weight may decrease according to the time sequence to pay more attention to the recent data. In a medical application scenario, when a nursing robot is performing a medicine delivery task, the initial window size may be set to contain data points of the last 10 task steps, and the window weight may be adjusted according to the urgency and importance of the task to ensure that the smoothed data can better reflect the current task requirements.

[0158] In step S602 of some embodiments, specifically, the adjustment of the window size is used to adjust the number of data points contained in the window.

[0159] For example, in a financial application scenario, if the bank service robot needs to quickly move to the customer location, the fusion task step length may be short (such as 50 steps), and the initial window size is adjusted from 5 to 3 to reduce the smoothing degree and retain more action details. In a medical application scenario, if the nursing robot needs to accurately grasp the medicine, the fusion task step length may be long (such as 150 steps), and the initial window size is adjusted from 15 to 20 to increase the smoothing degree and ensure the stability of the action.

[0160] In step S603 of some embodiments, specifically, the window weight is dynamically adjusted according to the characteristics of the fusion task step length, which can better reflect the importance of different data points.

[0161] For example, in a financial application scenario, if the fusion task step length is small, the weight of the data point with a closer time sequence can be reduced to avoid over-reliance on a single data point. In a medical application scenario, if the fusion task step length is large, more attention may be paid to the stability of the current task, and the weight of all data points in the window can be increased to smooth the action jitter caused by task switching.

[0162] In step S604 of some embodiments, the smoothed fusion task step length refers to the task step length determined by combining the historical task step length data and the adjusted window parameters to meet the current task requirements.

[0163] For example, if the current fusion task step size is 80 steps, the historical task step size is [95, 100, 105, 110, 115], the adjusted window size is 10, and the window weight is Gaussian distribution, the smoothed fusion task step size is 91 by calculating the weighted average value.

[0164] By steps S601 to S604, the fusion task step size is smoothed by combining the historical task step size, the adjusted window size, and the window weight, which can reduce the drastic change of the fusion task step size when the task is switched or the interaction state is changed, thereby improving the smoothness and continuity of the robot motion control, and significantly improving the robot motion control effect.

[0165] Please refer to Figure 7 In some embodiments, step S107 includes but is not limited to steps S701 to S704:

[0166] Step S701, determining the diffusion step number based on the action generation network and the fusion task step size.

[0167] Step S702, obtaining initial noise data.

[0168] Step S703, gradually diffusing the initial noise data into the action generation network based on the diffusion step number to obtain a noise action sequence.

[0169] Step S704, gradually denoising the noise action sequence based on the diffusion step number to obtain a target action.

[0170] In step S701 of some embodiments, specifically, the diffusion step number is a key parameter in the action generation network, which is used to determine the number of iterations in the process of gradually generating a target action sequence from initial noise data.

[0171] Specifically, the determination of the diffusion step number is based on the fusion task step size, and the diffusion step number integrates the accuracy requirement and real-time state of the task to ensure the efficiency and quality of action generation.

[0172] In step S702 of some embodiments, specifically, the initial noise data refers to a noise vector randomly generated from a Gaussian distribution in the action generation process, which is used as the initial input of the action generation network to provide randomness for the generation of the action sequence.

[0173] In the specific implementation process, the initial noise data is obtained by a random number generator.

[0174] For example, a Gaussian noise vector with a dimension of D can be generated by a random number generator, where each element zi follows a standard normal distribution N(0, 1).

[0175] In this embodiment, the initial noise data is obtained to provide a random starting point for the action generation network, so that the robot can generate diverse action sequences under different task requirements.

[0176] In step S703 of some embodiments, specifically, the noise action sequence refers to an action sequence containing random noise generated by gradually adding the initial noise data to the action generation network during the action generation process.

[0177] Specifically, the diffusion model can be used to gradually construct a target action sequence containing noise by increasing noise for each fusion task step.

[0178] For example, in a financial application scenario, if a bank service robot needs to quickly move to a customer location and deliver a file, the fusion task step is 80 steps. According to this step, the action generation network determines that the diffusion step is 40 steps. Starting from the initial noise data, the noise data is gradually added to the action generation network through 40 steps of forward diffusion to generate an action sequence containing noise.

[0179] In step S704 of some embodiments, specifically, the target action refers to the action sequence generated by gradually removing the random noise data.

[0180] Specifically, the diffusion model can be used to gradually remove the random noise data in the noise action sequence through reverse diffusion to recover the target action sequence.

[0181] For example, in a financial application scenario, starting from an action sequence containing noise, denoising processing is performed through 40 diffusion steps to gradually reduce noise, and finally a smooth and coherent quick file grabbing action is generated. In a medical application scenario, starting from an action sequence containing noise, denoising processing can be performed through 60 diffusion steps to generate a precise target action of grabbing medicine.

[0182] Through steps S701 to S704, the initial noise data is gradually diffused into the action generation network through the diffusion steps, and the noise action sequence is gradually denoised. Through more diffusion steps, the robot can generate more precise action sequences, or through fewer diffusion steps, the action sequence generation efficiency is improved, so that the robot can dynamically adjust the generated action under different task requirements, and ensure the efficiency and accuracy of the generated action.

[0183] The embodiment of the application helps the robot to comprehensively perceive the task demand and the environment state by acquiring the action control instruction for controlling the target robot, the body perception data and the target image, and provides rich data basis for subsequent action generation; then, the control data of the target robot are subjected to action task analysis, basic task step allocation, task step adjustment and action generation by the robot action generation model, which can perform differentiated task step control for different task demands, realizes allocating more task steps for high-precision tasks to ensure the accuracy of the generated action, reduces the task steps for low-precision tasks, improves the efficiency of the generated action while ensuring the accuracy of the generated action, thereby improving the action control effect of the robot.

[0184] Please refer to Figure 8 The embodiment of the application also provides a robot action generation device, which can realize the robot action generation method described above, and the device comprises:

[0185] An action control data acquisition module is configured to acquire an action control instruction for controlling a target robot, and acquire body perception data and a target image of the target robot according to the action control instruction;

[0186] An action generation model acquisition module is configured to acquire a pre-trained robot action generation model; wherein the robot action generation model comprises a task analysis network, a step adjustment network and an action generation network; the step adjustment network comprises a task step adjustment subnetwork and a state step adjustment subnetwork;

[0187] An action task analysis module is configured to perform action task analysis on the action control instruction by using the task analysis network to obtain a target action task;

[0188] A basic step matching module is configured to allocate a basic task step for the target action task by using the task step adjustment subnetwork;

[0189] A step adjustment module is configured to adjust the basic task step by using the state step adjustment subnetwork, the target image and the body perception data to obtain an adjusted task step;

[0190] A step fusion module is configured to perform weighted fusion on the basic task step and the adjusted task step to obtain a fused task step;

[0191] An action generation module is configured to perform action generation based on the action generation network and the fused task step to obtain a target action.

[0192] The specific implementation of the robot action generation device is basically the same as the specific embodiment of the robot action generation method described above, and will not be repeated here.

[0193] The embodiment of the present application further provides an electronic device, which comprises a memory and a processor. The memory stores a computer program, and the processor implements the robot action generation method when executing the computer program. The electronic device can be any intelligent terminal, such as a tablet computer or a vehicle-mounted computer.

[0194] Please refer to Figure 9 , Figure 9 The hardware structure of the electronic device of another embodiment is illustrated, which comprises:

[0195] The processor 901 can be implemented in the form of a general-purpose CPU (Central Processing Unit), a microprocessor, an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits, and is used to execute related programs to implement the technical solutions provided by the embodiments of the present application.

[0196] The memory 902 can be implemented in the form of a ROM (Read Only Memory), a static storage device, a dynamic storage device, or a RAM (Random Access Memory). The memory 902 can store a processing system and other application programs. When the technical solutions provided by the embodiments of the present application are implemented by software or firmware, the related program codes are stored in the memory 902 and are called and executed by the processor 901 to implement the robot action generation method of the embodiments of the present application.

[0197] The input / output interface 903 is used to realize information input and output.

[0198] The communication interface 904 is used to realize the communication interaction between the device and other devices. The communication can be realized by a wired manner (for example, a USB, a network cable, etc.) or a wireless manner (for example, a mobile network, WIFI, Bluetooth, etc.).

[0199] The bus 905 is used to transmit information between various components (for example, the processor 901, the memory 902, the input / output interface 903, and the communication interface 904) of the device.

[0200] The processor 901, the memory 902, the input / output interface 903, and the communication interface 904 are connected to each other through the bus 905 to realize the communication connection between them in the device.

[0201] The embodiment of the present application further provides a computer readable storage medium, which stores a computer program. The computer program is executed by a processor to implement the robot action generation method.

[0202] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. In addition, the memory can include a high-speed random access memory and can also include a non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state memory device. In some embodiments, the memory can optionally include a memory that is remotely disposed relative to the processor, and these remote memories can be connected to the processor through a network. Examples of the above network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0203] The embodiments described in the embodiments of the present application are used to more clearly illustrate the technical solutions of the embodiments of the present application, and do not constitute a limitation on the technical solutions provided by the embodiments of the present application. Those skilled in the art can know that, with the evolution of technology and the appearance of new application scenarios, the technical solutions provided by the embodiments of the present application are also applicable to similar technical problems.

[0204] Those skilled in the art can understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present application, and can include more or fewer steps than shown in the figures, or combine certain steps, or different steps.

[0205] The device embodiments described above are only schematic, and the units described as separate components can or can not be physically separate, that is, they can be located in one place, or can be distributed on multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the embodiments of the present application.

[0206] Those skilled in the art can understand that all or some of the steps in the above disclosed method, the functional modules / units in the system and the device can be implemented as software, firmware, hardware and their appropriate combinations.

[0207] The terms "first", "second", "third", "fourth" and the like used in the specification of the present application and the above-described drawings (if any) are used to distinguish similar objects, and do not necessarily have to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0208] It should be understood that, in the application, "at least one" refers to one or more, and "multiple" refers to two or more. "And / or" is used to describe the association relationship of the associated objects, which means that there can be three relationships, for example, "A and / or B" can represent three cases of only A, only B and A and B existing at the same time, wherein A and B can be singular or plural. The character " / " generally represents an "or" relationship between the associated objects before and after it. "At least one of the following" or similar expressions means any combination of these items, including any combination of single or multiple items. For example, at least one of a, b or c can represent a, b, c, "a and b", "a and c", "b and c", or "a and b and c", wherein a, b and c can be single or multiple.

[0209] In several embodiments provided in the application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only illustrative, for example, the division of the above units is only a logical function division, and actual implementation can have another division manner, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. The coupling or direct coupling or communication connection between the displayed or discussed each other can be through some interface, indirect coupling or communication connection between devices or units, which can be electrical, mechanical or other forms.

[0210] The units described above as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place, or they can be distributed on multiple network units. According to actual needs, part or all of the units can be selected to achieve the purpose of the embodiment scheme.

[0211] In addition, each functional unit in each embodiment of the application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0212] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application, essentially or in other words, the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes multiple instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various program storage media.

[0213] The preferred embodiments of the embodiments of the present application are described above with reference to the accompanying drawings, and are not limited to the scope of the embodiments of the present application. Any modifications, equivalent replacements and improvements made by those skilled in the art without departing from the scope and essence of the embodiments of the present application shall be within the scope of the embodiments of the present application.

Claims

1. A method for generating robot actions, characterized in that, The method includes: Obtain motion control commands for controlling the target robot, and obtain the target robot's body perception data and target image based on the motion control commands; Obtain a pre-trained robot motion generation model; wherein the robot motion generation model includes a task parsing network, a step size adjustment network, and a motion generation network; the step size adjustment network includes a task step size adjustment sub-network and a state step size adjustment sub-network; The task parsing network is used to parse the action control instructions into action tasks to obtain the target action task; The task step size adjustment subnetwork is used to assign a basic task step size to the target action task; The basic task step size is adjusted using the state step size adjustment subnetwork, the target image, and the ontology perception data to obtain the adjusted task step size; The basic task step size and the adjusted task step size are weighted and fused to obtain the fused task step size; Based on the action generation network and the fusion task step size, action generation is performed to obtain the target action; The step of adjusting the basic task step size using the state step size adjustment subnetwork, the target image, and the ontology perception data to obtain the adjusted task step size includes: The target image is visually encoded using the state step size adjustment subnetwork to obtain visual features; Feature extraction is performed on the ontology-aware data to obtain ontology-aware data features; Based on the visual features and the ontological perception data features, the motion state of the target robot is identified to obtain the target motion state; The basic task step size is adjusted according to the target motion state to obtain the adjusted task step size; wherein, the target motion state includes interactive motion state and non-interactive motion state; The step of adjusting the basic task step size according to the target motion state to obtain the adjusted task step size includes: If the target motion state is an interactive motion state, then the basic task step size is incrementally adjusted to obtain the adjusted task step size. If the target motion state is a non-interactive motion state, then the basic task step size is reduced to obtain the adjusted task step size.

2. The method according to claim 1, characterized in that, The action generation based on the action generation network and the fusion task step size to obtain the target action includes: The number of diffusion steps is determined based on the action generation network and the fusion task step size; Obtain initial noise data; Based on the number of diffusion steps, the initial noise data is gradually diffused into the action generation network to obtain a noise action sequence; The target action is obtained by progressively denoising the noisy action sequence based on the number of diffusion steps.

3. The method according to claim 2, characterized in that, Before determining the diffusion steps based on the action generation network and the fusion task step size, the method further includes: Obtain the initial window size and window weights of the preset smoothing window filter; The initial window size is adjusted according to the fusion task step size to obtain the adjusted window size; The window weights are adjusted according to the fusion task step size to obtain the adjusted window weights. Obtain the historical task step size, and smooth the fusion task step size based on the historical task step size, the adjustment window size, and the adjustment window weight to obtain the smoothed fusion task step size.

4. The method according to claim 1, characterized in that, The step of parsing the action control command using the task parsing network to obtain the target action task includes: The semantic features of the action control command are extracted using the task parsing network to obtain the control command features; The control command features are parsed to obtain the initial action task; The initial action task is decomposed into action subtasks; The task accuracy requirement is identified for the action sub-tasks to obtain the task accuracy requirement. The target action task is determined based on the action sub-task and the task accuracy requirements.

5. The method according to any one of claims 1 to 4, characterized in that, The step of weightedly fusing the basic task step size and the adjusted task step size to obtain the fused task step size includes: Obtain the basic step size weight of the basic task step size, and obtain the adjustment step size weight of the adjusted task step size; The basic task step size and the adjusted task step size are weighted and fused based on the basic step size weight and the adjusted step size weight to obtain the fused task step size.

6. A robot motion generation device, characterized in that, The device includes: The motion control data acquisition module is used to acquire motion control commands for controlling the target robot, and to acquire the body perception data and target image of the target robot based on the motion control commands. The motion generation model acquisition module is used to acquire a pre-trained robot motion generation model; wherein, the robot motion generation model includes a task parsing network, a step size adjustment network, and a motion generation network; the step size adjustment network includes a task step size adjustment sub-network and a state step size adjustment sub-network; The action task parsing module is used to parse the action control instructions into action tasks using the task parsing network to obtain the target action task. A basic step size matching module is used to allocate a basic task step size to the target action task using the task step size adjustment sub-network. The step size adjustment module is used to adjust the basic task step size using the state step size adjustment sub-network, the target image and the body perception data to obtain the adjusted task step size; The step size fusion module is used to perform a weighted fusion of the basic task step size and the adjusted task step size to obtain a fused task step size. The action generation module is used to generate actions based on the action generation network and the fusion task step size to obtain the target action; The step size adjustment module also includes: The target image is visually encoded using the state step size adjustment subnetwork to obtain visual features; Feature extraction is performed on the ontology-aware data to obtain ontology-aware data features; Based on the visual features and the ontological perception data features, the motion state of the target robot is identified to obtain the target motion state; The basic task step size is adjusted according to the target motion state to obtain the adjusted task step size; wherein, the target motion state includes interactive motion state and non-interactive motion state; The step size adjustment module also includes: If the target motion state is an interactive motion state, then the basic task step size is incrementally adjusted to obtain the adjusted task step size. If the target motion state is a non-interactive motion state, then the basic task step size is reduced to obtain the adjusted task step size.

7. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the robot motion generation method according to any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the robot motion generation method according to any one of claims 1 to 5.

Images