Teleoperation intent recognition and shared control method based on imitation learning

Abstract

This invention discloses a method and apparatus for teleoperation intent recognition and shared control based on imitation learning. The method includes: acquiring the real-time position of the robot's slave hand and the operator's master hand control commands; responding to receiving actual control commands, determining a demonstration task corresponding to the actual control commands, and acquiring the imitation learning results corresponding to the demonstration task; obtaining a preliminary intent recognition result based on the slave hand's real-time position, the master hand control commands, and the imitation learning results; and obtaining a position control command based on the preliminary intent recognition result and the master hand control commands through a shared control arbitration rule, and transmitting the position control command to the lower-level machine to achieve human-machine shared control. This invention completes demonstration learning through virtual simulation or real demonstration, and obtains imitation learning results based on the demonstration learning results using an imitation learning method, thereby making this invention adaptable to complex unstructured environments and improving the applicability of teleoperation shared control.

Description

Technical Field

[0001] This invention relates to the field of teleoperated robot technology, and in particular to a method and apparatus for teleoperation intent recognition and shared control based on imitation learning. Background Technology

[0002] In industrial applications, robotic arms play an irreplaceable role in achieving fast, precise, and repetitive tasks. However, in some complex, unstructured work environments, such as nuclear power plant emergency rescue, underwater exploration and inspection, and space maintenance, tasks still rely on human remote control. This means that a human operator issues control commands to the master arm device, and the robotic arm follows the master arm's movements. However, in this control process, issues such as long-distance delays, human hand tremors, and operator unfamiliarity can lead to lower control accuracy and efficiency at the slave end of the master-slave remote control compared to autonomous control. Therefore, shared controllers, which allow human operators and robotic autonomous controllers to share remote control, have emerged. These controllers can reduce the operator's workload and / or improve performance during task execution.

[0003] In existing technologies, the relationship between humans and robots is transformed from a master-slave model to a collaborative one. Specifically, when there is a significant difference between the human's instructions and the robot's reference trajectory, the robot will actively adjust its own reference trajectory locally, incorporating the human's virtual interaction forces. When the differences between the human and robot intentions are small, the instructions from both sides are comprehensively considered, and the human-robot hybrid cost function is dynamically adjusted based on system safety assessment indicators to calculate the optimal control quantity, achieving shared human-robot control. However, the above method is based on single-path planning of robot control instructions. Although its operator intention recognition based on force feedback interaction can determine the operational intention, its reference trajectory instructions may not be suitable for complex, unstructured environments, potentially leading to inappropriate trajectory planning and incorrect intention classification. Summary of the Invention

[0004] The present invention aims to at least partially solve one of the technical problems in the related art.

[0005] To address this, this invention proposes a teleoperation intent recognition and shared control method based on imitation learning. This method obtains a preliminary intent recognition result based on the slave hand's real-time position, the master hand's control commands, and the imitation learning results. Then, based on the preliminary intent recognition result and the master hand's control commands, a position control command is obtained through a shared control arbitration rule and transmitted to the lower-level machine to achieve human-machine shared control. This invention utilizes virtual simulation or real-world demonstration methods for demonstration learning, and uses imitation learning methods to obtain imitation learning results based on the demonstration learning results. This makes the invention adaptable to complex, unstructured environments, improving the applicability of teleoperation shared control.

[0006] Another objective of this invention is to propose a teleoperation intent recognition and sharing control device based on imitation learning.

[0007] To achieve the above objectives, this invention proposes a teleoperation intent recognition and shared control method based on imitation learning, wherein the method is applied to a shared controller and includes:

[0008] Acquire the robot's slave hand's real-time position and the operator's master hand control commands;

[0009] In response to receiving an actual control command, determine the demonstration task corresponding to the actual control command, and obtain the imitation learning result corresponding to the demonstration task;

[0010] Based on the real-time position of the slave hand, the control commands of the master hand, and the imitation learning results, a preliminary intention recognition result is obtained;

[0011] Based on the preliminary intent recognition result and the master control command, a position control command is obtained through the shared control arbitration rule, and the position control command is transmitted to the lower-level machine to realize human-machine shared control.

[0012] The teleoperation intent recognition and shared control method based on imitation learning in this invention may also have the following additional technical features:

[0013] In one embodiment of the present invention, the step of obtaining a position control command based on the preliminary intent identification result and the master control command through a shared control arbitration rule includes:

[0014] If the preliminary intention identification result is to execute the imitation learning trajectory, then based on the slave hand's real-time position, the master hand's control command, and the imitation learning result, a predicted intention trajectory is obtained, and based on the predicted intention trajectory and the master hand's control command, a position control command is obtained through a shared control arbitration rule;

[0015] If the preliminary identification result of the intent is not to execute the imitation learning trajectory, then the master hand control command is determined to be a position control command.

[0016] In one embodiment of the present invention, obtaining the predicted intention trajectory based on the slave hand's real-time position, the master hand's control command, and the imitation learning result includes:

[0017] Based on the real-time position of the slave hand and the control command of the master hand, two probability values ​​are obtained as the preliminary recognition result of the intention and the predicted intention trajectory at the next moment through preset rules;

[0018] The two probability values ​​are fuzzy fused to obtain the preliminary intent probability, and the Markov process transition probability matrix is ​​iteratively updated to obtain the target intent probability.

[0019] Consciousness recognition is performed based on the stated intention probability to obtain the predicted intention trajectory.

[0020] In one embodiment of the present invention, before obtaining the preliminary intent recognition result based on the slave hand's real-time position, the master hand's control command, and the imitation learning result, the method further includes:

[0021] In response to receiving a demonstration instruction, obtain the expert demonstration trajectory information corresponding to the demonstration task;

[0022] The expert demonstration trajectory information is segmented to extract the action sub-trajectory of the demonstration task;

[0023] Imitation learning is performed based on the action sub-trajectories to obtain the imitation learning results of the demonstration task.

[0024] In one embodiment of the present invention, the expert demonstration trajectory information is segmented to extract the action sub-trajectory of the demonstration task, including:

[0025] The position and velocity information of the demonstration trajectory are collected, and the velocity information is normalized to obtain directional velocity information;

[0026] The demonstration trajectory is encoded by introducing coded action primitives;

[0027] Based on the directional velocity information, the demonstration trajectory is segmented using a clustering method to obtain the action segment primitives of the demonstration trajectory;

[0028] Extract features from the action fragment primitives to obtain the feature action fragment primitives;

[0029] Based on the aforementioned feature action fragment primitives, the action sub-trajectories of the demonstration task are extracted.

[0030] In one embodiment of the present invention, the imitation learning based on the action sub-trajectory to obtain the imitation learning result of the demonstration task includes:

[0031] The action sub-trajectory is learned by imitation learning method to obtain the imitation learning trajectory;

[0032] The imitation learning trajectory is modulated to obtain the imitation learning result of the demonstration task.

[0033] To achieve the above objectives, another aspect of the present invention proposes a teleoperation intent recognition and shared control device based on imitation learning, wherein the device is applied to a shared controller and includes:

[0034] The acquisition module is used to acquire the real-time position of the robot's slave hand and the operator's master hand control commands;

[0035] The determination module is used to, in response to receiving an actual control command, determine the demonstration task corresponding to the actual control command and obtain the imitation learning result corresponding to the demonstration task;

[0036] The recognition module is used to obtain a preliminary intention recognition result based on the real-time position of the slave hand, the control command of the master hand, and the imitation learning result;

[0037] The control module is used to obtain a position control command based on the preliminary intention recognition result and the master control command through a shared control arbitration rule, and transmit the position control command to the lower-level machine to realize human-machine shared control.

[0038] In one embodiment of the present invention, the control module is specifically used for:

[0039] If the preliminary intention identification result is to execute the imitation learning trajectory, then based on the slave hand's real-time position, the master hand's control command, and the imitation learning result, a predicted intention trajectory is obtained, and based on the predicted intention trajectory and the master hand's control command, a position control command is obtained through a shared control arbitration rule;

[0040] If the preliminary identification result of the intent is not to execute the imitation learning trajectory, then the master hand control command is determined to be a position control command.

[0041] In one embodiment of the present invention, the control module is further configured to:

[0042] Based on the real-time position of the slave hand and the control command of the master hand, two probability values ​​are obtained as the preliminary recognition result of the intention and the predicted intention trajectory at the next moment through preset rules;

[0043] The two probability values ​​are fuzzy fused to obtain the preliminary intent probability, and the Markov process transition probability matrix is ​​iteratively updated to obtain the target intent probability.

[0044] Consciousness recognition is performed based on the stated intention probability to obtain the predicted intention trajectory.

[0045] In one embodiment of the present invention, the device is further configured to:

[0046] In response to receiving a demonstration instruction, obtain the expert demonstration trajectory information corresponding to the demonstration task;

[0047] The expert demonstration trajectory information is segmented to extract the action sub-trajectory of the demonstration task;

[0048] Imitation learning is performed based on the action sub-trajectories to obtain the imitation learning results of the demonstration task.

[0049] The teleoperation intent recognition and shared control method based on imitation learning in this invention obtains a preliminary intent recognition result based on the slave hand's real-time position, the master hand's control command, and the imitation learning result. Based on the preliminary intent recognition result and the master hand's control command, a position control command is obtained through a shared control arbitration rule and transmitted to the lower-level machine to achieve human-machine shared control. This invention completes demonstration learning through virtual simulation or real-world demonstration, and obtains imitation learning results based on the demonstration learning results. This makes the invention adaptable to complex, unstructured environments and improves the applicability of teleoperation shared control.

[0050] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0051] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0052] Figure 1 This is a flowchart of a teleoperation intent recognition and shared control method based on imitation learning according to an embodiment of the present invention;

[0053] Figure 2 This is a schematic diagram of fuzzy fusion according to an embodiment of the present invention;

[0054] Figure 3 This is a structural diagram of a teleoperation intent recognition and sharing control device based on imitation learning according to an embodiment of the present invention. Detailed Implementation

[0055] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0056] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0057] The following description, with reference to the accompanying drawings, describes a method and apparatus for teleoperation intent recognition and shared control based on imitation learning, according to embodiments of the present invention.

[0058] Figure 1 This is a flowchart of a teleoperation intent recognition and shared control method based on imitation learning, according to an embodiment of the present invention.

[0059] like Figure 1 As shown, this method is applied to a shared controller and includes:

[0060] S1, acquire the real-time position of the robot's slave hand and the operator's master hand control commands;

[0061] In one embodiment of the present invention, the real-time position of the robot's slave hand can be obtained using... This indicates that the operator's main hand control commands can be used express.

[0062] S2, in response to receiving the actual control command, determines the demonstration task corresponding to the actual control command and obtains the imitation learning result corresponding to the demonstration task;

[0063] In one embodiment of the present invention, upon receiving an actual control command, it is necessary to determine a demonstration task corresponding to the actual control command and obtain the imitation learning result corresponding to the demonstration task. This allows for the preliminary identification of intent based on the imitation learning result, the real-time position of the slave hand, and the master hand control command. Furthermore, in one embodiment of the present invention, before determining the demonstration task corresponding to the actual control command and obtaining the imitation learning result corresponding to the demonstration task, the above method may further include the following steps:

[0064] Step 1: Upon receiving the demonstration instruction, obtain the expert demonstration trajectory information corresponding to the demonstration task;

[0065] Step 2: Perform demonstration segmentation on the expert demonstration trajectory information and extract the action sub-trajectories of the demonstration task;

[0066] Step 3: Perform imitation learning based on action sub-trajectories to obtain the imitation learning results of the demonstration task.

[0067] In one embodiment of the present invention, the method for segmenting expert demonstration trajectory information and extracting action sub-trajectories of the demonstration task may include the following steps:

[0068] Step a: Collect the position and velocity information of the demonstration trajectory, and normalize the velocity information to obtain the directional velocity information;

[0069] Step b: Encode the demonstration trajectory by introducing coded action primitives;

[0070] Step c: Based on the direction and velocity information, the demonstration trajectory is segmented using a clustering method to obtain the action segment primitives of the demonstration trajectory;

[0071] Step d: Extract features from action fragment primitives to obtain feature action fragment primitives;

[0072] Step e: Extract the action sub-trajectories of the demonstration task based on the feature action fragment primitives.

[0073] Specifically, in one embodiment of the present invention, the location information of the demonstration trajectory is collected. Speed ​​information , , For the number of data points, the speed information Mean filtering is applied across all dimensions, and... Normalization is performed to obtain direction and velocity information. .

[0074] In one embodiment of the invention, coded action primitives are introduced. The encoding of the demonstration trajectory was implemented. Among other things, Used to mark the n The category to which the action primitive belongs; It is based on Re-encode each action primitive in sequence The index value.

[0075] In one embodiment of the present invention, the demonstration trajectory is segmented based on directional velocity information using a clustering method; that is, the demonstration trajectory is segmented according to the velocity direction. part, The larger the value, the higher the trajectory segmentation accuracy, but the higher the sensitivity to noise. Specifically, in one embodiment of the present invention, when a segment of motion primitive velocity direction... When they are similar, they are grouped into one category and considered to conform to the Gaussian model. ,Right now:

[0076] (1)

[0077] In one embodiment of the present invention, the K-means clustering algorithm is used to obtain the average value. Initial values ​​are used to obtain the parameters of the GMM model. The angle between two unit velocities is measured using the arc distance (i.e., cosine distance).

[0078] (2)

[0079] Furthermore, in one embodiment of the present invention, the probability density function in the GMM model... During the calculation process, it is also necessary to use the arc distance (i.e., cosine distance) to measure the distance from the action primitive in space. To cluster average distance :

[0080] (3)

[0081] Specifically, the Expectation-Maximization (EM) method is used to update the parameters of the GMM model. And can be updated :

[0082] (4)

[0083] And, based on Updates can be made recursively :

[0084] (5)

[0085] Furthermore, according to The data can be used to set action primitives Divided into Action segments Among them, the introduction To define these fragments, where , and For fragments All included In and , The total length of the trajectory, It should be noted that, All in the set They are all equal. To obtain new feature fragments after further feature extraction for subsequent action segments. The index value.

[0086] Furthermore, regarding The vector properties are used to complete feature extraction, and obtain... :

[0087] (6)

[0088] in, ,according to The data can be used to analyze action segments. Divided into A characteristic action segment And, the introduction To define these fragments, where , and For fragments All included In and , .

[0089] Furthermore, in one embodiment of the present invention, non-featured action primitives in the feature action segments are excluded to extract the action sub-trajectories of the demonstration task. Specifically, they are excluded one by one. and Angular deviation greater than The action primitives, and exclude two action primitives with a distance less than 1. The action primitives are updated. , , Furthermore, in order to obtain sub-trajectories, the feature segments are compared. Curve difference less than They are grouped into a single sub-trajectory.

[0090] (7)

[0091] Furthermore, in one embodiment of the present invention, according to It can identify characteristic action fragments Divided into Each action sub-trajectory, among which, utilizing To define the sub-trajectory, where , for All sets The total number of action primitives in the text. Among them, and For sub-trajectories All included , In and , for The position of each action primitive in the entire trajectory: Increment the number of feature segments in each sub-trajectory by 1.

[0092] (8)

[0093] In one embodiment of the present invention, the above method can be used to complete the demonstration in a real environment through teaching demonstration, or in a virtual environment through simulation. That is, the above demonstration segmentation method is insensitive to real / virtual environments and different operation mapping ratios, and can achieve "real-virtual compatible" demonstration segmentation and learning.

[0094] Furthermore, in one embodiment of the present invention, after obtaining the action sub-trajectory through the above steps, imitation learning can be performed based on the action sub-trajectory to obtain the imitation learning result of the demonstration task. Specifically, in one embodiment of the present invention, the method for obtaining the imitation learning result of the demonstration task based on the action sub-trajectory may include: performing imitation learning on the action sub-trajectory using an imitation learning method to obtain an imitation learning trajectory; and modulating the imitation learning trajectory to obtain the imitation learning result of the demonstration task.

[0095] In one embodiment of the present invention, each of the above sub-trajectories Training sets can be built. , For For the input data of the element, For This refers to the output data for the elements. Furthermore, it utilizes a type-2 fuzzy imitation learning method to achieve data processing. This method employs imitation learning, is computationally simple, can modulate trajectories from arbitrary positions, and does not prioritize shape preservation. Furthermore, the training set is set as follows:

[0096] (9)

[0097] in, To output the fuzzy range of the data. As the number of clusters, the Type 2 fuzzy model consists of several "IF-THEN" fuzzy rules. Assuming there are J rules in total, rule j can be described as:

[0098] (10)

[0099] in, For input data, These represent the type 2 fuzzy set, local linear model parameters, and local linear model output in rule j of the type 2 fuzzy model, respectively. for The upper and lower boundaries, for The upper and lower bounds of . Then the first The trajectory of the strip, the first The output of the fuzzy model of the type II model in one dimension for:

[0100] (11)

[0101] in, and It is to satisfy The lower and upper bounds of the type II fuzzy membership degree.

[0102] (12)

[0103] in, To define the range of fuzzy membership degree variation, Type 1 fuzzy membership degree:

[0104] (13)

[0105] in, , This is the norm-inducing matrix. As cluster center, and Let m be the input and output cluster centers of the j-th fuzzy cluster. Here, m>1 controls the fuzziness of the fuzzy clustering; the larger m is, the more fuzzy the division. m=2 is a widely used choice.

[0106] Furthermore, in one embodiment of the present invention, the above-mentioned recognition based on antecedent recognition and result recognition realizes the parameters of the type II fuzzy model. .

[0107] Furthermore, in one embodiment of the present invention, when an operator reproduces the trajectory, there may be instances where they return along the original path. Based on this, at most, it is necessary to generate... Imitation learning trajectory , ,defined by , and It can generate imitation learning trajectories Therefore, a maximum of [number] can be generated. Intent trajectories available for reference. And, a set of design intent trajectories. To further determine the most likely intention trajectory, let ,in For trajectory Total length, For based on The direction of velocity at the next moment , Current hand speed and The included angle, let for Distance sub-trajectory closest distance:

[0108] (14)

[0109] in, And, based on , , Information generation based on Imitation learning trajectory and The method may include the following steps:

[0110] Step 1: Input Data Update: Current End-effector Position of the Robot The input data are the constraints for the modulation trajectory. , The position closest to this sub-trajectory

[0111] (15)

[0112] In one embodiment of the present invention, based on the modulation trajectory, one can travel from the current position along... Direction, that is arrive Direction, or along Direction, that is arrive For each of the two directions, set two input data points. and :

[0113] (16)

[0114] Step 2: Calculation : Output data At this point, the insertion constraint of the modulation trajectory is obtained: That is, the modulation trajectory must pass through this position. If insertion constraints are present... The nearest fuzzy cluster center ( distance ( ) at a predefined threshold Within, then for the first If the clustering parameters are not adjusted, then a new cluster category is added, as follows:

[0115] (17)

[0116] in, Update the parameters for the parameter numbers that need updating:

[0117] (18)

[0118] And, will and According to the revised parameters And formulas (18)(19)(20)(21) can obtain the... Modulation trajectory after imitation learning .

[0119] Step 3: Obtain the output: Repeat step 2 above to integrate the results and obtain the output. Modulation trajectory after learning: ,as well as and based on Information Update , Other related information.

[0120] It should be noted that, in one embodiment of the present invention, the above-mentioned imitation learning method can also be any one of GMM, KMP, and DMP.

[0121] S3, based on the real-time position of the slave hand, the master hand control commands, and the imitation learning results, obtains the preliminary intention recognition result;

[0122] In one embodiment of the present invention, after obtaining the imitation learning result through the above steps, a preliminary intention recognition result can be obtained based on the real-time position of the slave hand, the master hand control command, and the imitation learning result.

[0123] Furthermore, in one embodiment of the present invention, in order to reduce the amount of real-time calculation during actual control, it is not necessary to calculate all... A learning trajectory based on imitation. Specifically, in one embodiment of the present invention, all are calculated using the above formula (14). , and when ( When distance is a constraint, the operator's intention will not be along... When the movement is initiated, the initial intention recognition result is that the imitation learning trajectory will not be executed; otherwise, the initial intention recognition result is that the imitation learning trajectory will be executed, and the imitation learning trajectory will be modulated to obtain... The various data.

[0124] S4, based on the preliminary intent recognition result and the master control command, obtains the position control command through the shared control arbitration rule, and transmits the position control command to the lower computer to realize human-machine shared control.

[0125] In one embodiment of the present invention, the method for obtaining a position control command based on the preliminary intent identification result and the master hand control command through a shared control arbitration rule may include: if the preliminary intent identification result is to execute the imitation learning trajectory, then based on the slave hand's real-time position, the master hand control command, and the imitation learning result, a predicted intent trajectory is obtained, and based on the predicted intent trajectory and the master hand control command, a position control command is obtained through a shared control arbitration rule; if the preliminary intent identification result is not to execute the imitation learning trajectory, then the master hand control command is determined as the position control command.

[0126] Furthermore, in one embodiment of the present invention, the method for obtaining the predicted intention trajectory based on the real-time position of the slave hand, the master hand control command, and the imitation learning result may include the following steps:

[0127] S401, based on the real-time position of the slave hand and the control commands of the master hand, obtains two probability values ​​for the preliminary intention recognition result as the prediction of the intention trajectory at the next moment through preset rules;

[0128] S402, perform fuzzy fusion processing on the two probability values ​​to obtain the preliminary intent probability, and use Markov process to iteratively update the transition probability matrix to obtain the target intent probability;

[0129] S403, based on the probability of purpose and intent, consciousness recognition is performed to obtain the predicted intent trajectory.

[0130] In one embodiment of the present invention, the following is defined: and Used to describe respectively based on and Information, the intention to operate in the next moment is the trajectory. The probability of [the outcome]. Furthermore, in one embodiment of the present invention, two probability values ​​are obtained as the predicted intention trajectory for the next moment, based on the preliminary intention recognition result, through the following preset rules.

[0131] Specifically, in one embodiment of the present invention, when At that time, even if the operator has followed The movement trend is there, but the trajectory is about to be completed and should not be considered as an alternative to the intended trajectory. .

[0132] When satisfied hour, The smaller, the more it means With intention trajectory The closer the distance, the better. The larger; in order to handle when In this case, the probability is described using a soft maximization layer method:

[0133] (27)

[0134] When satisfied , The smaller, the more it means With intention trajectory The smaller the included angle, the better. The larger; It can be described as:

[0135] (28)

[0136] And, in obtaining and After considering two probability values, a fuzzy method is introduced to determine... and The non-linear relationship between them. Specifically, with and The input is the operation intention at the next moment, which is the trajectory. Initial intention probability For output. The range of input and output variables is defined as follows: , , Use the standard Gaussian membership function. Where, Figure 2 This is a schematic diagram of fuzzy fusion.

[0137] Furthermore, the operator's predicted intention trajectory may follow a certain... It may proceed, but it's also possible that at the next moment, the operator will expect the movement to proceed according to their own intentions, meaning there are a total of Possible intentions, assuming The target state of the operation intention, where For the number of iterations or operation loops:

[0138] (29)

[0139] Among them, state variables It can be seen as having A Markov chain with states This represents the state transition matrix. Represents the first This operation starts from the initial state. To the target state The probability of . Let The initial state is hour, The state with the highest probability:

[0140] (30)

[0141] Wherein, the transformation matrix All elements have the same probability 1 / Initialize. The state transition matrix is ​​from... arrive The update and iteration, i.e., obtaining the probability of the intended purpose. A matrix is ​​defined as follows:

[0142] (31)

[0143] in, Let be the learning rate, and the necessary and sufficient condition for convergence is: Appropriate learning rate The learning rate setting is crucial. A higher learning rate leads to faster convergence but also increases the impact of inappropriate updates due to erroneous operations. A low learning rate results in slower convergence, requiring more iteration cycles and experience. The unique solution to the iterative equation is:

[0144] (32)

[0145] Furthermore, in one embodiment of the present invention, during the control process, firstly, = And can be obtained through formula (29) ,when = ,Right now In other words, the probability of the operator executing any intentional trajectory is not high. Therefore, the operator control mode should be entered at this time. To get ,in To ensure a larger learning rate, thereby guaranteeing Conversely, use To get ,in A smaller learning rate could potentially lead to If the above situation occurs, further assessment is needed to determine whether... If the condition is met, then let = Otherwise, it can be assumed that the current predicted trajectory does not conform to the operator's intention, and then let = and reuse To get Through the above process, the final result is... This is the predicted trajectory of the current operation.

[0146] In one embodiment of the present invention, the above steps are used to obtain... , , Subsequently, location control instructions must be obtained through the shared control arbitration rules. These shared control arbitration rules can be: if... In this case, it is assumed that the robot will not intervene in the operator's actions, meaning that the generated intention trajectory is considered unhelpful to the operator. Control is then simply transferred to the operator, similar to traditional master-slave control. If... Then the operator and robot enter a shared control state.

[0147] Furthermore, in one embodiment of the present invention, in order to solve the problem that the operating speed in the virtual demonstration does not match the actual speed, the generated robot operation instructions... There is only the direction of velocity, not the magnitude of velocity. Control speed for the robot based on intent recognition:

[0148] (33)

[0149] Furthermore, in one embodiment of the invention, the robot is controlled only by the operator. Output speed for intervention for:

[0150] (34)

[0151] Furthermore, in one embodiment of the invention, the weights of the robot's participation in control... for:

[0152] (35)

[0153] Furthermore, in one embodiment of the present invention, the position control command is:

[0154] (36)

[0155] According to an embodiment of the present invention, a teleoperation intent recognition and shared control method based on imitation learning obtains a preliminary intent recognition result based on the slave hand's real-time position, the master hand's control command, and the imitation learning result. Based on the preliminary intent recognition result and the master hand's control command, a position control command is obtained through a shared control arbitration rule, and the position control command is transmitted to the lower-level machine to achieve human-machine shared control. In this invention, demonstration learning is completed through virtual simulation or real demonstration, and the imitation learning result is obtained using an imitation learning method based on the demonstration learning result. This makes the invention adaptable to complex unstructured environments and improves the applicability of teleoperation shared control.

[0156] To achieve the above embodiments, such as Figure 3As shown, this embodiment also provides a teleoperation intent recognition and sharing control device 10 based on imitation learning. This device is applied to a sharing controller and includes an acquisition module 100, a prediction module 200, and a control module 300.

[0157] The acquisition module 100 is used to acquire the real-time position of the robot's slave hand and the operator's master hand control commands;

[0158] The determination module 200 is used to respond to the received actual control command, determine the demonstration task corresponding to the actual control command, and obtain the imitation learning result corresponding to the demonstration task.

[0159] The recognition module 300 is used to obtain preliminary intention recognition results based on the real-time position of the slave hand, the master hand control commands, and the imitation learning results;

[0160] The control module 400 is used to obtain position control commands based on the preliminary intent recognition results and the master control commands through shared control arbitration rules, and transmit the position control commands to the lower-level machine to realize human-machine shared control.

[0161] Furthermore, the aforementioned control module 400 is specifically used for:

[0162] If the initial intention identification result is to execute the imitation learning trajectory, then based on the slave hand's real-time position, the master hand's control command, and the imitation learning result, the predicted intention trajectory is obtained, and based on the predicted intention trajectory and the master hand's control command, the position control command is obtained through the shared control arbitration rule;

[0163] If the initial identification result indicates that the imitation learning trajectory will not be executed, then the master hand control command will be determined as a position control command.

[0164] Furthermore, the aforementioned control module 400 is also used for:

[0165] Based on the real-time position of the slave hand and the control commands of the master hand, two probability values ​​are obtained through preset rules to predict the trajectory of the intention at the next moment, which are the preliminary recognition results of the intention.

[0166] The two probability values ​​are fuzzily fused to obtain the initial intent probability, and the Markov process transition probability matrix is ​​iteratively updated to obtain the target intent probability.

[0167] Consciousness recognition is performed based on the probability of purpose and intent to obtain the predicted intent trajectory.

[0168] Furthermore, the aforementioned device 10 is also used for:

[0169] In response to receiving a demonstration instruction, obtain the expert demonstration trajectory information corresponding to the demonstration task;

[0170] Perform demonstration segmentation on the expert demonstration trajectory information and extract the action sub-trajectories of the demonstration task;

[0171] Imitation learning is performed based on action sub-trajectories to obtain the imitation learning results for the demonstration task.

[0172] According to an embodiment of the present invention, a teleoperation intent recognition and shared control device based on imitation learning obtains a preliminary intent recognition result based on the slave hand's real-time position, the master hand's control command, and the imitation learning result. Based on the preliminary intent recognition result and the master hand's control command, a position control command is obtained through a shared control arbitration rule, and the position control command is transmitted to the lower-level machine to achieve human-machine shared control. In this invention, demonstration learning is completed through virtual simulation or real demonstration methods, and imitation learning results are obtained based on the demonstration learning results. This makes the invention adaptable to complex unstructured environments and improves the applicability of teleoperation shared control.

[0173] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0174] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for teleoperation intent recognition and shared control based on imitation learning, characterized in that, The method is applied to a shared controller and includes: Acquire the robot's slave hand's real-time position and the operator's master hand control commands; In response to receiving an actual control command, determine the demonstration task corresponding to the actual control command, and obtain the imitation learning result corresponding to the demonstration task; Based on the real-time position of the slave hand, the control commands of the master hand, and the imitation learning results, a preliminary intention recognition result is obtained; Based on the preliminary intent recognition result and the master control command, a position control command is obtained through the shared control arbitration rule, and the position control command is transmitted to the lower-level machine to realize human-machine shared control.

2. The method according to claim 1, characterized in that, The process of obtaining position control commands based on the preliminary intent identification result and the master control command through shared control arbitration rules includes: If the preliminary intention identification result is to execute the imitation learning trajectory, then based on the slave hand's real-time position, the master hand's control command, and the imitation learning result, a predicted intention trajectory is obtained, and based on the predicted intention trajectory and the master hand's control command, a position control command is obtained through a shared control arbitration rule; If the preliminary identification result of the intent is not to execute the imitation learning trajectory, then the master hand control command is determined to be a position control command.

3. The method according to claim 2, characterized in that, The process of obtaining the predicted intention trajectory based on the slave hand's real-time position, the master hand's control commands, and the imitation learning results includes: Based on the real-time position of the slave hand and the control command of the master hand, two probability values ​​are obtained as the preliminary recognition result of the intention and the predicted intention trajectory at the next moment through preset rules; The two probability values ​​are fuzzy fused to obtain the preliminary intent probability, and the Markov process transition probability matrix is ​​iteratively updated to obtain the target intent probability. Consciousness recognition is performed based on the stated intention probability to obtain the predicted intention trajectory.

4. The method according to claim 1, characterized in that, Before obtaining the preliminary intent recognition result based on the slave hand's real-time position, the master hand's control commands, and the imitation learning results, the method further includes: In response to receiving a demonstration instruction, obtain the expert demonstration trajectory information corresponding to the demonstration task; The expert demonstration trajectory information is segmented to extract the action sub-trajectory of the demonstration task; Imitation learning is performed based on the action sub-trajectories to obtain the imitation learning results of the demonstration task.

5. The method according to claim 4, characterized in that, The step of segmenting the expert demonstration trajectory information and extracting the action sub-trajectories of the demonstration task includes: The position and velocity information of the demonstration trajectory are collected, and the velocity information is normalized to obtain directional velocity information; The demonstration trajectory is encoded by introducing coded action primitives; Based on the directional velocity information, the demonstration trajectory is segmented using a clustering method to obtain the action segment primitives of the demonstration trajectory; Extract features from the action fragment primitives to obtain the feature action fragment primitives; Based on the aforementioned feature action fragment primitives, the action sub-trajectories of the demonstration task are extracted.

6. The method according to claim 5, characterized in that, The imitation learning based on the action sub-trajectory to obtain the imitation learning result of the demonstration task includes: The action sub-trajectory is learned by imitation learning method to obtain the imitation learning trajectory; The imitation learning trajectory is modulated to obtain the imitation learning result of the demonstration task.

7. A teleoperation intent recognition and sharing control device based on imitation learning, characterized in that, The device is applied to a shared controller and includes: The acquisition module is used to acquire the real-time position of the robot's slave hand and the operator's master hand control commands; The determination module is used to, in response to receiving an actual control command, determine the demonstration task corresponding to the actual control command and obtain the imitation learning result corresponding to the demonstration task; The recognition module is used to obtain a preliminary intention recognition result based on the real-time position of the slave hand, the control command of the master hand, and the imitation learning result; The control module is used to obtain a position control command based on the preliminary intention recognition result and the master control command through a shared control arbitration rule, and transmit the position control command to the lower-level machine to realize human-machine shared control.

8. The apparatus according to claim 7, characterized in that, The control module is specifically used for: If the preliminary intention identification result is to execute the imitation learning trajectory, then based on the slave hand's real-time position, the master hand's control command, and the imitation learning result, a predicted intention trajectory is obtained, and based on the predicted intention trajectory and the master hand's control command, a position control command is obtained through a shared control arbitration rule; If the preliminary identification result of the intent is not to execute the imitation learning trajectory, then the master hand control command is determined to be a position control command.

9. The apparatus according to claim 8, characterized in that, The control module is also used for: Based on the real-time position of the slave hand and the control command of the master hand, two probability values ​​are obtained as the preliminary recognition result of the intention and the predicted intention trajectory at the next moment through preset rules; The two probability values ​​are fuzzy fused to obtain the preliminary intent probability, and the Markov process transition probability matrix is ​​iteratively updated to obtain the target intent probability. Consciousness recognition is performed based on the stated intention probability to obtain the predicted intention trajectory.

10. The apparatus according to claim 7, characterized in that, The device is also used for: In response to receiving a demonstration instruction, obtain the expert demonstration trajectory information corresponding to the demonstration task; The expert demonstration trajectory information is segmented to extract the action sub-trajectory of the demonstration task; Imitation learning is performed based on the action sub-trajectories to obtain the imitation learning results of the demonstration task.

Images