A method and system for embodied intelligent decision control based on mechanical arm joint state

Abstract

The application provides a body intelligence decision control method and system based on a mechanical arm joint state, and belongs to the technical field of body intelligence, wherein the method comprises the following steps: acquiring mechanical arm state data and a first feature in real time; for the real-time acquired mechanical arm state data, an improved Kalman filtering algorithm is used for dynamic filtering to obtain filtered data, wherein the improved Kalman filtering algorithm uses a bidirectional LSTM network to dynamically adjust a state transition matrix in real time; a hidden Markov model is used to extract the features of the filtered data to obtain a second feature; the first feature and the second feature are input into a pre-trained decision tree model to obtain a mechanical arm joint state decision result; and the mechanical arm joint state decision result is used for body intelligence decision control of the mechanical arm joint. The application meets the real-time requirement, improves the accuracy of mechanical arm joint state judgment and production efficiency.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of embodied intelligence, and particularly relates to an embodied intelligence decision control method and system based on joint states of a mechanical arm. BACKGROUND

[0002] In today's era of rapid technological development, embodied intelligence, as a highly potential and challenging research direction in the field of artificial intelligence, is attracting widespread attention from global researchers. Embodied intelligence aims to endow intelligent agents with the ability to perceive, decide, and perform actions in real physical environments. This requires intelligent agents not only to understand complex and variable environmental information, but also to make reasonable decisions based on this information and accurately perform corresponding actions, thereby achieving effective interaction with the environment.

[0003] Among the numerous technical branches covered by embodied intelligence, imitation learning occupies an important position. Imitation learning is committed to enabling intelligent agents to learn and perform specific tasks by observing human demonstrations or other excellent examples. However, the current field faces a significant challenge, namely, the accuracy of imitation learning cannot reach the ideal 100% precision, and errors occur from time to time. For example, in the industrial manufacturing scenario, when a mechanical arm performs complex assembly tasks through imitation learning, it may encounter problems such as inaccurate part assembly position and incorrect assembly sequence due to understanding bias of operation details or environmental interference; in the service robot field, when a robot imitates humans to deliver goods, it may misjudge the target position or slip when grabbing the goods. These errors not only affect the reliability and stability of embodied intelligence systems, but also limit their widespread application in real-world scenarios.

[0004] Taking the application of a mechanical arm in industrial production as an example, on the production line, the mechanical arm needs to complete a series of complex operations such as material handling and part processing. Existing models often cannot accurately determine whether the operation is successfully completed when controlling the mechanical arm to perform these operations. For example, after installing a part in a specified position, the model has difficulty determining whether the part has been correctly and firmly installed, which may lead to quality problems in subsequent production links, or even cause equipment failure. This uncertainty is continuously amplified in large-scale production environments, seriously affecting production efficiency and product quality. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application proposes an embodied intelligence decision control method and system based on joint states of a mechanical arm.

[0006] In a first aspect, the present application proposes an embodied intelligence decision control method based on joint states of a mechanical arm, comprising:

[0007] real-time acquisition of mechanical arm state data and first features;

[0008] For the real-time acquired mechanical arm state data, an improved Kalman filtering algorithm is used for dynamic filtering to obtain filtered data, and the improved Kalman filtering algorithm uses a bidirectional LSTM network to dynamically adjust a state transition matrix in real time;

[0009] A hidden Markov model is used to extract features of the filtered data to obtain second features;

[0010] The first features and the second features are input into a pre-trained decision tree model to obtain a mechanical arm joint state decision result;

[0011] The mechanical arm joint state decision result is used for embodied intelligent decision control of the mechanical arm joint.

[0012] The mechanical arm state data includes joint positions of the mechanical arm, angular velocities of the mechanical arm, angular accelerations of the mechanical arm, and jerk of the mechanical arm.

[0013] The first features include energy fluctuation values of the mechanical arm, JMI values of the mechanical arm, and potential energy fields of the mechanical arm.

[0014] The bidirectional LSTM network includes an input layer, a hidden layer, and an output layer; the input layer is connected to the hidden layer, and the hidden layer is connected to the output layer;

[0015] The input layer takes a historical mechanical arm state observation sequence at a previous time and mechanical arm state data at the previous time as input data of the input layer;

[0016] The hidden layer is two layers of bidirectional LSTM units, each layer having 64 bidirectional LSTM units, and a Tanh function being used as an activation function;

[0017] The output layer is a fully connected layer.

[0018] The bidirectional LSTM network has the following calculation formula:

[0019] ;

[0020] wherein, is the historical mechanical arm state observation sequence at the previous time, is the mechanical arm state data at the previous time, is a state transition matrix in the Kalman filtering algorithm at the previous time, is a state transition matrix in the Kalman filtering algorithm at the current time.

[0021] The hidden Markov model extracts features of the filtered data to obtain second features, including:

[0022] The filtered data is taken as a state sequence of a hidden Markov model, and a historical robot arm state observation sequence is taken as an observation sequence of the hidden Markov model, each observation value in the observation sequence corresponding to a joint position of the filtered robot arm, an angular velocity of the robot arm, an angular acceleration of the robot arm, and a jerk of the robot arm at a certain moment.

[0023] The hidden Markov model is used to extract features of the filtered data, to obtain second features, the second features including: robot arm action coherence and a trajectory matching degree between the filtered data and the historical robot arm state observation sequence.

[0024] The pre-trained decision tree model includes:

[0025] The first features and the second features are taken as inputs of the pre-trained decision tree model, and corresponding threshold values of the first features and the second features are taken to form the pre-trained decision tree model, wherein the corresponding threshold values are obtained by training the historical robot arm state observation sequence and corresponding decision labels.

[0026] In a second aspect, the present application provides a body-aware intelligent decision control system based on robot arm joint states, including:

[0027] A data acquisition module is configured to acquire robot arm state data and first features in real time.

[0028] A data filtering module is configured to perform dynamic filtering on the acquired robot arm state data in real time by using an improved Kalman filtering algorithm, to obtain filtered data, and the improved Kalman filtering algorithm uses a bidirectional LSTM network to dynamically adjust a state transition matrix.

[0029] A feature extraction module is configured to extract features of the filtered data by using a hidden Markov model, to obtain second features.

[0030] A state decision module is configured to input the first features and the second features into a pre-trained decision tree model, to obtain robot arm joint state decision results.

[0031] A decision control module is configured to perform body-aware intelligent decision control on robot arm joints by using the robot arm joint state decision results.

[0032] In a third aspect, the present application provides an electronic device, including one or more processors and a memory, the memory being configured to store instructions, when the instructions are executed by the one or more processors, causing the one or more processors to perform the body-aware intelligent decision control method based on robot arm joint states.

[0033] In a fourth aspect, the present application provides a computer readable storage medium storing executable instructions, which, when executed, cause a processor to perform the method for embodied intelligent decision control based on joint state of a mechanical arm.

[0034] In a fifth aspect, the present application provides a computer program product comprising a computer program or instructions, which, when executed by a processor, implement the method for embodied intelligent decision control based on joint state of a mechanical arm.

[0035] Advantages:

[0036] The present application provides a method and system for embodied intelligent decision control based on joint state of a mechanical arm. The improved Kalman filtering algorithm is used to dynamically filter the real-time acquired mechanical arm state data. The hidden Markov model is used to extract the features of the filtered data to obtain second features. The fusion features are input into the pre-trained decision tree model to obtain the joint state decision result of the mechanical arm. The method of the present application fundamentally improves the intelligent level and operation efficiency of the embodied intelligent system. Through accurate task completion judgment, the problem of excessive operation or insufficient operation caused by the inability of the previous model to accurately identify the task endpoint is effectively avoided, the energy consumption and equipment wear are significantly reduced, and the service life of the robot is prolonged. At the same time, the fast task switching capability enables the robot to handle more tasks in a unit of time, greatly improves the production efficiency, and has great application potential in fields such as industrial manufacturing, logistics and warehousing, medical care, and other fields with high requirements for real-time performance, efficiency and accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 The method for embodied intelligent decision control based on joint state of a mechanical arm according to an embodiment of the present application is shown in the flowchart.

[0038] Figure 2 The specific example decision control method flowchart of the embodiment of the present application is shown in the flowchart.

[0039] Figure 3 The principle block diagram of the system for embodied intelligent decision control based on joint state of a mechanical arm according to an embodiment of the present application is shown in the block diagram. DETAILED DESCRIPTION

[0040] The specific embodiments of the present application are described in detail below in conjunction with the drawings and examples.

[0041] In view of the current state of the art, the present application focuses on the "brain" part of the embodied intelligence system and conducts in-depth research, aiming to propose an innovative mechanical arm state monitoring method to accurately determine whether the large model has successfully completed the action, fill the technical gap in this aspect of the current embodied intelligence field, and improve the overall performance and reliability of the embodied intelligence system, laying a solid foundation for its wide application in many fields such as industry, medicine, and service.

[0042] By continuously observing the behavior of the model, for example, in the reasoning process of the ACT (Action Chunking with Transformers) model, when the model performs the task of finding target objects and carrying out a series of actions, it will continuously drive joint movement. When the model completes the preset series of tasks, it will return to a waiting point. At this time, although the reasoning program of the model is still running, it no longer drives joint movement.

[0043] The present application takes the pause time of the model at the completion point as the key condition for determining whether the task is completed. Specifically, when the joint is in a stable state and continuously maintained for a duration of t, or only performs micro-movement within a very small range, it can be determined that the current reasoning task has been completed. Based on this judgment, even if the reasoning program of the model is still running, it can gracefully pause the current task. This technical solution enables the robot to autonomously and accurately determine whether the task is completed and be ready to start the next task at any time. In specific reasoning, it needs to be noted that the real-time nature of mechanical arm detection and judgment is considered. If a complex and large model is used for reasoning, it can indeed ensure accurate reasoning and judgment, but it cannot meet the real-time requirements. Therefore, both accuracy and real-time requirements are required to achieve decision control of the mechanical arm joint state in the field of embodied intelligence.

[0044] Embodiment 1:

[0045] The present embodiment proposes an embodied intelligence decision control method based on the state of the mechanical arm joint, as shown in Figure 1 , which includes:

[0046] Step S1: Real-time acquisition of mechanical arm state data and first features;

[0047] In the present embodiment, the mechanical arm state data includes the joint position of the mechanical arm, the angular velocity of the mechanical arm, the angular acceleration of the mechanical arm, and the jerk of the mechanical arm. The first features include the energy fluctuation value of the mechanical arm, the JMI (Joint Movement Index) value of the mechanical arm, and the potential energy field of the mechanical arm. The JMI value is an index for quantifying the smoothness and efficiency of joint movement. It is usually used to evaluate the quality of mechanical arm movement, predict potential failures, or optimize control algorithms to reduce unnecessary vibration and energy consumption.

[0048] Generally, to determine the joint state of the robot arm, only the joint position of the robot arm and the angular velocity of the robot arm are needed. In order to accurately obtain the joint state of the robot arm, the angular acceleration of the robot arm and the jerk of the robot arm are also obtained in real time. Meanwhile, the first feature is obtained, that is, the energy fluctuation value of the robot arm is obtained through the analysis of the energy consumption curve of the robot arm, the JMI value of the robot arm is obtained through the JMI calculation, and the potential energy field of the robot arm is obtained through the calculation of the target potential energy field. The first feature obtained here is fused with the second feature calculated later to determine the joint state of the robot arm.

[0049] In a specific implementation, the joint position of the robot arm, the angular velocity of the robot arm, the angular acceleration of the robot arm and the jerk of the robot arm are collected in real time and accurately by means of the high-precision encoder of the robot arm joint. Anti-interference protocol is adopted in the transmission process to ensure low delay and distortion of the data, and to provide a reliable data basis for subsequent analysis.

[0050] Meanwhile, a distributed time series database is built by using a time series data storage architecture, and the robot arm state data is stored according to the nanosecond timestamp, supporting millisecond-level fast read-write and backtracking query. The distributed architecture can stably bear large-scale data and provide data support with strong scalability for historical state analysis and model training.

[0051] Step S2: For the real-time acquired robot arm state data, a dynamic filtering is performed by using an improved Kalman filtering algorithm to obtain filtered data, and the improved Kalman filtering algorithm uses a bidirectional LSTM network (Bidirectional Long Short-Term Memory Network) to dynamically adjust the state transition matrix in real time;

[0052] In this embodiment, the key behavior parameters of the focus model during task execution are focused, especially the joint motion state and action sequence of the robot arm, and a task completion determination system is constructed. The proposed robot arm state determination system creatively combines the improved Kalman filtering, the hidden Markov model, the decision tree algorithm, and the newly introduced features of the energy fluctuation value of the robot arm, the JMI value of the robot arm, and the potential energy field of the robot arm to form a progressive determination system. The system not only meets the real-time requirement but also efficiently and accurately determines the joint state of the robot arm.

[0053] In this embodiment, the high-precision encoder is used to obtain the joint position of the robot arm, the angular velocity of the robot arm, the angular acceleration of the robot arm and the jerk of the robot arm, and the improved Kalman filtering algorithm is used to optimize the noisy data.

[0054] The definition state vector of the Kalman filtering algorithm is: ; wherein, θ is the joint position of the robot arm, is the angular velocity of the robot arm, is the angular acceleration of the robot arm, is the jerk, is the current robot arm state data.

[0055] The state transition matrix of the Kalman filter algorithm is a state transition matrix constructed by using a fourth-order Taylor expansion:

[0056] F= ;

[0057] The state transition equation uses a high-order motion model: ;

[0058] wherein, F is the state transition matrix, is the process noise, Δt is the sampling interval, is the current robot arm state data, is the robot arm state data at the next time.

[0059] In this embodiment, the improved Kalman filter algorithm can capture transient response, and the introduction of jerk can accurately describe the sudden change in high-speed motion and reduce the lag error of the traditional third-order model under nonlinear motion. At the same time, it has dynamic adaptability, and the fourth-order Taylor expansion covers the motion characteristics of the robot arm in the full frequency band, and is especially suitable for high dynamic tasks (such as rapid grabbing and emergency braking).

[0060] In this embodiment, the improved Kalman filter algorithm uses a bidirectional LSTM network to dynamically adjust the state transition matrix in real time, and the bidirectional LSTM network comprises an input layer, a hidden layer and an output layer; the input layer is connected with the hidden layer, and the hidden layer is connected with the output layer.

[0061] The input layer takes the historical robot arm state observation sequence (length k=10) at the last time and the robot arm state data at the last time as the input data of the input layer, is the historical robot arm state observation sequence at the last time;

[0062] The hidden layer is two layers of bidirectional LSTM units, each layer has 64 bidirectional LSTM units, and the Tanh function is used as the activation function.

[0063] The output layer is a fully connected layer.

[0064] The bidirectional LSTM network has the following calculation formula:

[0065] ;

[0066] wherein, is a historical robot arm state observation sequence of the previous time, is robot arm state data of the previous time, is a state transition matrix in the Kalman filtering algorithm of the previous time, is a state transition matrix in the Kalman filtering algorithm of the current time.

[0067] In this embodiment, a bidirectional LSTM network is used to dynamically perceive environmental interference (such as vibration and electromagnetic noise), and the covariance matrix is adjusted in real time. Compared with fixed parameter filtering, the noise suppression capability is improved by 40%. Lightweight network design (parameter quantity <10k) ensures that the single inference time consumption is <0.5ms, meeting the real-time requirements of industry.

[0068] Step S3: extracting features of the filtered data by using a hidden Markov model to obtain second features;

[0069] The hidden Markov model extracts features of the filtered data, and includes:

[0070] The filtered data is taken as a state sequence of the hidden Markov model, and the historical robot arm state observation sequence is taken as an observation sequence of the hidden Markov model. Each observation value in the observation sequence corresponds to a joint position of the filtered robot arm, an angular velocity of the robot arm, an angular acceleration of the robot arm, and jerk of the robot arm at a certain time.

[0071] The hidden Markov model extracts features of the filtered data to obtain second features, and the second features include: motion coherence of the robot arm, and trajectory matching degree of the filtered data and the historical robot arm state observation sequence.

[0072] This embodiment proposes a task completion state direct calculation framework based on a hidden Markov model (HMM). By modeling the time sequence features and state transition rules of robot arm motion, the task execution process is accurately quantified. The following is a specific implementation scheme:

[0073] The hidden Markov model analyzes the processed data, excavates the time sequence features and state transition rules of joint motion, and extracts key features of the task execution process, such as motion coherence and trajectory matching degree.

[0074] To further quantify the feature indicators of the process, define a state sequence , an observation sequence , and each observation value in the observation sequence corresponds to a joint angle, a velocity, and a fusion feature at a certain time. There are a hidden state transition matrix , an observation probability matrix , which describes the mapping relationship from the hidden state to the observation value, and an initial state probability vector , then the probability of the observation sequence under given model parameters can be expressed as:

[0075] ;

[0076] in, For the model parameters The probability of the occurrence of the observation sequence O is, From state S i Transfer to S j The probability of , characterizing the transition rules between states; In state S j Observed The probability of , which measures the degree of match between observation and state; At the initial moment, the system is in state S i The probability of , T is the total length or number of steps of the observation sequence, S is the set of all possible hidden state paths, is the probability that the system is in state S1 at the initial moment, is observed in state S1 The probability of From state S t-1 Transfer to S t The probability of To indicate that the system is in the hidden state S at time t t When , the observation value is generated The probability of S t is the hidden state at time t (e.g. some internal state of the robotic arm), is the observation value at time t (such as sensor data such as joint angle and speed).

[0077] when When the value is high, it means that the current observed trajectory has a high degree of match with the standard trajectory, which can be used as a metric for action coherence and trajectory consistency and input into the subsequent decision model.

[0078] Step S4: inputting the first feature and the second feature into the pre-trained decision tree model to obtain the robot arm joint state decision result;

[0079] The pre-trained decision tree model includes:

[0080] The first feature and the second feature are used as inputs of a pre-trained decision tree model, and the thresholds corresponding to the first feature and the second feature are used to form a pre-trained decision tree model, wherein the corresponding thresholds are obtained by training using a historical robotic arm state observation sequence and a corresponding decision label.

[0081] In this embodiment, based on the extracted features (i.e., second features), and the fusion of the first features, a decision model is constructed by means of a decision tree algorithm. By training the decision tree, a mapping relationship between joint motion features and task states is established, for example, according to the matching features of the joint trajectory and the preset standard trajectory, it is accurately determined whether the mechanical arm completes the task actions such as grabbing and placing.

[0082] In this embodiment, the task completion determination problem can be formalized as a binary classification problem, judging whether the task state is "completed" or "not completed", and the input feature vector is set as:

[0083]

[0084] wherein each represents a feature, for example: JMI value, trajectory matching score, end velocity convergence index, ΔE energy fluctuation value, etc. The output is a class label y∈{0,1}, wherein 0 represents "not completed", and 1 represents "completed".

[0085] The decision tree determination process is as follows:

[0086] Each node performs feature division, and the judgment condition is: wherein, is the jth feature judged by the node; is the division threshold value of the feature. If the condition is met, the sample enters the left subtree, otherwise it enters the right subtree, until it reaches the leaf node, and outputs the classification result: =label(L).

[0087] Step S5: adopting the decision result of the mechanical arm joint state to make embodied intelligent decision control of the mechanical arm joint.

[0088] In order to describe an embodied intelligent decision control method based on the joint state of a mechanical arm in more detail, an example is listed as shown in Figure 2 .

[0089] (1) Start stage: start the entire task flow, and the robot enters the state of preparing to execute the task.

[0090] (2) Robot action execution: the robot starts to execute the action according to the preset instruction or model reasoning result. In this process, the robot constantly tries to complete the set atomic task. The atomic task is the smallest, indivisible basic task unit that constitutes a complex task, for example, the mechanical arm grabbing a specific object, moving to a specified position, etc.

[0091] ​(3) Determine whether the atomic task is completed: Steps S1-S5 are used to determine whether the atomic task is completed. During the execution of the robot action, the system monitors the relevant state parameters (such as position, attitude, force feedback, etc., depending on the type of task) in real time, and determines whether the atomic task is completed. If not (the result of the determination is no), the robot continues to maintain the action execution state and keeps trying until the atomic task is completed; if it is completed (the result of the determination is yes), the next step is entered.

[0092] (4) Robot is stationary: When it is determined that the robot has completed the atomic task, the robot stops the current action and enters a stationary state. At this time, the robot no longer actively moves or changes its position or attitude.

[0093] (5) Determine the stationary time: The system starts timing and monitors the stationary time of the robot. Determine whether the stationary time reaches the pre-set time t. If not (the result of the determination is no), continue to monitor and accumulate time; if it is reached (the result of the determination is yes), the next step is entered.

[0094] Atomic action execution completion determination: When the robot stationary time reaches t, the system determines that the atomic action execution is completed. At this time, according to the specific task requirements, it can be decided whether to continue to execute the next atomic task or to complete the entire task flow.

[0095] Through the above steps, the example can use the action state and stationary time of the robot and other information to accurately determine whether the atomic action is executed and completed, providing a reliable determination method for efficient and accurate execution of tasks in embodied intelligent systems. On the other hand, the example also provides an intelligent task management strategy based on the above determination mechanism. When the model is determined to complete the current task, the strategy can automatically trigger a series of subsequent operations, including but not limited to gracefully pausing the reasoning program of the model, releasing relevant computing resources, and quickly completing the reset and initialization of the system state, so that the robot can be ready to accept the next task instruction at any time. And the strategy has good scalability and compatibility, can be easily integrated into existing various embodied intelligent system architectures, without the need for large-scale redesign and modification of the system.

[0096] The embodiment proposes a body-possessed intelligent decision control method based on the joint state of a mechanical arm, comprising: acquiring mechanical arm state data and first features in real time; performing dynamic filtering on the real-time acquired mechanical arm state data by using an improved Kalman filtering algorithm to obtain filtered data, wherein the improved Kalman filtering algorithm uses a bidirectional LSTM network to dynamically adjust a state transition matrix; extracting features of the filtered data by using a hidden Markov model to obtain second features; fusing the first features and the second features and inputting them into a pre-trained decision tree model to obtain a mechanical arm joint state decision result; and using the mechanical arm joint state decision result to perform body-possessed intelligent decision control on the mechanical arm joints. The method of the embodiment fundamentally improves the intelligent level and operation efficiency of the body-possessed intelligent system. Through accurate task completion judgment, the problem of excessive operation or insufficient operation caused by the inability of the previous model to accurately identify the task endpoint is effectively avoided, the energy consumption and equipment wear are significantly reduced, and the service life of the robot is prolonged. At the same time, the rapid task switching capability enables the robot to handle more tasks in a unit of time, greatly improves the production efficiency, and has great application potential in fields such as industrial manufacturing, logistics and warehousing, medical care and the like which have extremely high requirements on efficiency and accuracy. In addition, good compatibility and scalability provide strong support for the popularization and promotion of body-possessed intelligent technology, and help to accelerate the application of related technologies in different industries.

[0097] Embodiment 2

[0098] The embodiment proposes a body-possessed intelligent decision control system based on the joint state of a mechanical arm, as shown in Figure 3 The embodiment proposes a body-possessed intelligent decision control system based on the joint state of a mechanical arm, as shown in

[0099] The data acquisition module is used for acquiring mechanical arm state data and first features in real time.

[0100] The data filtering module is used for performing dynamic filtering on the real-time acquired mechanical arm state data by using an improved Kalman filtering algorithm to obtain filtered data, wherein the improved Kalman filtering algorithm uses a bidirectional LSTM network to dynamically adjust a state transition matrix.

[0101] The feature extraction module is used for extracting features of the filtered data by using a hidden Markov model to obtain second features.

[0102] The state decision module is used for inputting the first features and the second features into a pre-trained decision tree model to obtain a mechanical arm joint state decision result.

[0103] A decision control module is configured to make embodied intelligent decision control of the joints of the robot arm based on the joint state decision result of the robot arm.

[0104] Embodiment 3

[0105] The embodiment provides an electronic device, which comprises one or more processors and a memory. The memory is configured to store instructions. When the instructions are executed by the one or more processors, the one or more processors are caused to perform the embodied intelligent decision control method based on the joint state of a robot arm.

[0106] The electronic device can be a mobile phone, a computer, a tablet computer or the like, and comprises a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the embodied intelligent decision control method based on the joint state of a robot arm is implemented. It can be understood that the electronic device can further comprise an input / output (I / O) interface and a communication component.

[0107] The processor is configured to perform all or part of the embodied intelligent decision control method based on the joint state of a robot arm. The memory is configured to store various types of data. The data can comprise instructions of any application program or method in the electronic device and application program related data.

[0108] The processor can be an Application Specific Integrated Cricuit (ASIC), a Digital Signal Processor (DSP), a Programmable Logic Device (PLD), a Field Programmable Gate Array (FPGA), a controller, a microcontroller, a microprocessor or other electronic elements, and is configured to execute the embodied intelligent decision control method based on the joint state of a robot arm.

[0109] Embodiment 4

[0110] The embodiment provides a computer readable storage medium, which stores executable instructions. When the instructions are executed, if the instructions are implemented in the form of a software function unit and sold or used as an independent product, the instructions can be stored in a computer readable storage medium.

[0111] The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of an embodied intelligent decision-making and control method based on the joint state of a robotic arm as described in various embodiments of the present application.

[0112] The aforementioned storage media include: flash memory, hard disk, multimedia card, card-type memory (for example, SD (Secure Digital Memory Card) or DX (Memory Data Register, MDR) memory, etc.), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, server, APP (Application, abbreviation of application software) application store and other media that can store program verification codes, on which a computer program is stored. When the computer program is executed by the processor, it can implement the various steps of the above-mentioned embodied intelligent decision-making and control method based on the joint state of the robotic arm.

[0113] Example 5:

[0114] This embodiment proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the embodied intelligent decision-making and control method based on the joint state of a robotic arm.

[0115] Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution can be embodied in the form of a computer program product.

[0116] The various embodiments in this application are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

[0117] The scope of the present application is not limited to the above-described embodiments, and it will be apparent to those skilled in the art that various changes and modifications can be made to the present disclosure without departing from the scope and spirit of the present disclosure. The present disclosure is intended to include such changes and modifications within the scope and spirit of the present disclosure.

Claims

1. An embodied intelligent decision-making and control method based on the joint state of a robotic arm, characterized in that: include: Acquire the robot arm status data and the first feature in real time; For the real-time acquired robotic arm state data, an improved Kalman filter algorithm is used to perform dynamic filtering to obtain filtered data. The improved Kalman filter algorithm uses a bidirectional LSTM network to dynamically adjust the state transfer matrix in real time; The hidden Markov model is used to extract the features of the filtered data to obtain the second feature; Input the first feature and the second feature into the pre-trained decision tree model to obtain the robot arm joint state decision result; Using the robot arm joint state decision result to perform embodied intelligent decision control on the robot arm joint; The first feature includes: the energy consumption fluctuation value of the robot arm, the JMI value of the robot arm, and the potential energy field of the robot arm; The hidden Markov model extracts features of the filtered data to obtain a second feature, including: The filtered data is used as the state sequence of the hidden Markov model, and the historical robot state observation sequence is used as the observation sequence of the hidden Markov model. Each observation value in the observation sequence corresponds to the filtered joint position, angular velocity, angular acceleration and acceleration of the robot at a certain moment. A hidden Markov model is used to extract features of the filtered data to obtain second features, which include: continuity of robot arm movements and trajectory matching between the filtered data and a historical robot arm state observation sequence.

2. The embodied intelligent decision-making and control method based on the joint state of a robotic arm according to claim 1, characterized in that: The state data of the robotic arm include: joint positions of the robotic arm, angular velocity of the robotic arm, angular acceleration of the robotic arm, and acceleration of the robotic arm.

3. The embodied intelligent decision-making and control method based on the joint state of a robotic arm according to claim 1, characterized in that: The bidirectional LSTM network includes: an input layer, a hidden layer, and an output layer; the input layer is connected to the hidden layer, and the hidden layer is connected to the output layer; The input layer uses the historical robot arm state observation sequence at the previous moment and the robot arm state data at the previous moment as input data of the input layer; The hidden layer is a two-layer bidirectional LSTM unit with 64 bidirectional LSTM units in each layer, using the Tanh function as the activation function; The output layer is a fully connected layer.

4. The embodied intelligent decision-making and control method based on the joint state of a robotic arm according to claim 1, characterized in that: The bidirectional LSTM network is calculated as follows: ； in, is the historical robot arm state observation sequence at the previous moment, is the state data of the robot arm at the last moment, is the state transfer matrix in the Kalman filter algorithm at the previous moment, is the state transfer matrix in the Kalman filter algorithm at the current moment.

5. The embodied intelligent decision-making and control method based on the joint state of a robotic arm according to claim 1, characterized in that: The pre-trained decision tree model includes: The first feature and the second feature are used as inputs of a pre-trained decision tree model, and the thresholds corresponding to the first feature and the second feature are used to form a pre-trained decision tree model, wherein the corresponding thresholds are obtained by training using a historical robotic arm state observation sequence and a corresponding decision label.

6. An embodied intelligent decision-making control system based on the joint state of a robotic arm, used to implement the embodied intelligent decision-making control method based on the joint state of a robotic arm according to any one of claims 1 to 5, characterized in that: include: A data acquisition module, used for acquiring the robot arm state data and the first feature in real time; A data filtering module is used to dynamically filter the real-time acquired robotic arm state data using an improved Kalman filter algorithm to obtain filtered data. The improved Kalman filter algorithm uses a bidirectional LSTM network to dynamically adjust the state transfer matrix. A feature extraction module is used to extract features of the filtered data using a hidden Markov model to obtain a second feature; A state decision module is used to input the first feature and the second feature into a pre-trained decision tree model to obtain a robot arm joint state decision result; A decision control module is used to perform embodied intelligent decision control on the joints of the robotic arm using the decision results of the robotic arm joint states.

7. An electronic device, characterized in that: include: One or more processors, and a memory, wherein the memory is used to store instructions, and when the instructions are executed by the one or more processors, the one or more processors execute the embodied intelligent decision-making and control method based on the joint state of a robotic arm as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that It stores executable instructions, which, when executed, enable the processor to execute the embodied intelligent decision-making and control method based on the joint state of the robotic arm as described in any one of claims 1 to 5.

9. A computer program product comprising a computer program or instructions, characterized in that When the computer program or instruction is executed by a processor, an embodied intelligent decision-making and control method based on the joint state of a robotic arm as described in any one of claims 1 to 5 is implemented.

Images