Finger capturing device and method for humanoid robot reinforcement learning

Abstract

The invention relates to the technical field of robots, and discloses a finger capturing device and method for reinforcement learning of a humanoid robot, and the device comprises finger modules which are disposed on each finger of the humanoid robot, and each finger module comprises an IMU sensor, an LED lamp, a touch sensor, and a processing module; the IMU sensor is used for collecting motion state data of fingers; the LED lamps are used for carrying out ID identification on each finger through a flicker sequence of the LED lamps; the processing module is used for processing data acquired by the IMU sensor and controlling the LED lamp; the touch sensor is used for sensing the contact state of the finger and an external object; the camera is used for acquiring two-dimensional image data of the hand position of the humanoid robot in multiple directions; and the analysis center is used for identifying and obtaining the spatial position coordinates of each finger according to the two-dimensional image data in the multiple directions, and judging the hand posture according to the spatial position coordinates of each finger and the contact state and motion state data of the finger and the external object, so that the anti-interference performance and the accuracy of optical motion capture are improved.

Description

Technical Field

[0001] This application relates to the technical field of robot finger motion capture, and in particular, to a finger capture device and method for humanoid robot reinforcement learning. Background Art

[0002] With the development of robot technology, humanoid robots are increasingly widely used in the field of reinforcement learning. In the learning process of humanoid robots, the accurate capture and control of fingers are the key to realizing various complex tasks. By obtaining the motion postures, force states, and motion data of fingers, the humanoid robot can be better controlled and managed.

[0003] The existing finger capture methods for humanoid robots include inertial motion capture, optical motion capture, and tactile and force feedback, etc. Among them, the inertial motion capture system records finger motion data through micro sensors, combines AI technology for robot motion training to realize the motion simulation process, and also integrates fingertip tracking sensors to real-time capture finger bending, stretching and other motions, and then combines tactile feedback and machine learning algorithms to optimize and improve the fidelity of robot hand motions. The optical motion capture collects the whole body and hand motions by tracking feature points with multiple cameras.

[0004] In the existing finger capture methods, there are problems such as low accuracy and poor stability in single acquisition methods. Among them, optical motion capture is easily affected by the outside world, resulting in large errors in the finger joint tracking process. Therefore, how to improve the anti-interference ability and accuracy of optical motion capture is the fundamental problem to be solved in the present invention. Summary of the Invention

[0005] In order to improve the anti-interference ability and accuracy of optical motion capture, this application provides a finger capture device and method for humanoid robot reinforcement learning.

[0006] In the first aspect, this application provides a finger capture device for humanoid robot reinforcement learning, adopting the following technical solutions:

[0007] A finger capture device for humanoid robot reinforcement learning, comprising:

[0008] A finger module, arranged on each finger of the humanoid robot, including an IMU sensor, an LED lamp, a tactile sensor, and a processing module; the IMU sensor is used to collect the motion state data of the finger; the LED lamp is used to identify each finger through its flashing sequence; the processing module is used for processing the data obtained by the IMU sensor and controlling the LED lamp; the tactile sensor is used to sense the contact state between the finger and an external object;

[0009] A microcontroller, which is used to receive and process IMU data from a finger module and transmit it to an analysis center;

[0010] A camera, which is provided with several groups and is used to obtain two-dimensional image data of multiple orientations of the hand position of a humanoid robot;

[0011] An analysis center, which is used to identify the spatial position coordinates of each finger based on two-dimensional image data of multiple orientations, and judge the hand posture according to the spatial position coordinates of each finger, the contact state with an external object, and the motion state data.

[0012] By adopting the above technical solution, the finger where the LED lamp is located is determined through the flashing sequence of the LED lamp, which avoids the influence of external interference factors on the judgment of the position point, and further improves the accuracy of the acquisition of the position point of the LED, etc. The spatial position coordinates of each finger, the contact state with an external object, and the motion state data are used to comprehensively judge the hand posture, which avoids problems such as a large error rate and inaccurate acquisition results in a single acquisition method.

[0013] Optionally, the process of controlling the flashing sequences of several groups of LED lamps to perform ID recognition on each finger includes:

[0014] A unique flashing sequence is preset for different fingers, and the ID of each finger is parsed through continuous two-dimensional image data.

[0015] By adopting the above technical solution, each finger can have a unique flashing sequence, and the serial number of the finger can be identified through this flashing sequence, thereby ensuring the anti-interference and accuracy of the acquisition of the spatial coordinate position of the finger.

[0016] Optionally, the process of the analysis center judging the hand posture includes:

[0017] Taking the position of the LED lamp as a feature point, extracting the feature points in the two-dimensional image data, and calculating the three-dimensional coordinates of the feature points according to the parallax of the feature points in multiple orientations;

[0018] Performing stereo matching and screening on the three-dimensional coordinates of the feature points with a preset posture set in a database to obtain a three-dimensional hand model;

[0019] Judging the hand posture according to the three-dimensional hand model and the contact state and motion state data of each finger with an external object.

[0020] By adopting the above technical solution, the hand posture is judged according to the three-dimensional hand model and the contact state and motion state data of each finger with an external object, thereby improving the anti-interference of finger state acquisition and obtaining more accurate finger position and posture information.

[0021] Optionally, each finger module is provided with two groups of LED lights, and the two groups of LED lights are arranged on different finger segments. The process of stereo matching and screening includes:

[0022] Obtain the spatial distances of the LED lights on different finger segments of each finger to form a first distance sequence [x1, x2,..., xm]; compare the first distance sequence [x1, x2,..., xm] with the first distance sequence [xi1, xi2,..., xim] corresponding to each preset posture in the preset posture set, and sort each preset posture in ascending order, and extract the preset postures according to the preset screening rules to obtain the first screening set; where m is the number of fingers, k is a positive integer and k ∈ [1, m], is the length influence coefficient of the kth finger;

[0023] Obtain the midpoints of the connections of the LED lights on different finger segments of each finger, and obtain the distances between the midpoints of the connections of adjacent fingers according to the finger order to form a second distance sequence [y1, y2,..., y(m - 1)]; compare the second distance sequence [y1, y2,..., y(m - 1)] with the second distance sequence [yi1, yi2,..., yi(m - 1)] corresponding to each preset posture in the first screening set, and sort them in ascending order of numerical value, select the first Q names for the first vector comparison process, and obtain the corresponding preset posture as the stereo matching screening result according to the first vector comparison process, and obtain the corresponding three-dimensional hand model, is the proportional adjustment coefficient, j ∈ [1, m - 1].

[0024] By adopting the above technical solution, on the premise of ensuring the accuracy of the screening result, the calculation amount in the screening process is reduced, and the screening efficiency is improved.

[0025] Optionally, the preset screening rules include:

[0026] Judge the corresponding quantity g. If the quantity g exceeds the preset value gt, then select the first gt names in the sorting as the first screening set; otherwise, select the first g names in the sorting as the first screening set, where Xt is the preset threshold.

[0027] By adopting the above technical solution, obtaining the preset postures in the first screening set through the preset screening rules can ensure the accuracy and comprehensiveness of the obtained results.

[0028] Optionally, the first vector comparison process includes:

[0029] Form a first vector sequence by connecting the midpoints of the LED lights on different finger segments of the first finger with the midpoints of the LED lights on different finger segments of other fingers Compare the first vector sequence with the first vector sequence of the first Q preset postures before sorting, and select the preset posture corresponding to the minimum value as the stereo matching screening result, where s ∈ [1, m - 2], is the s-th vector in the first vector sequence and the included angle between the first vector sequence corresponding to the preset posture, is the bias error coefficient.

[0030] By adopting the above technical solution, the corresponding preset posture can be accurately matched.

[0031] Optionally, each finger module is provided with a group of LED lights, and the process of stereo matching screening includes:

[0032] Obtain the distances between adjacent LED lights according to the finger order to form a third distance sequence [z1, z2,..., z(m - 1)]; compare the third distance sequence [z1, z2,..., z(m - 1)] with the third distance sequence [zi1, zi2,..., zi(m - 1)] corresponding to each preset posture in the first screening set, and sort them in ascending order of values. Select the first Q before sorting for the second vector comparison process, and obtain the corresponding preset posture as the stereo matching screening result according to the second vector comparison process, and obtain the corresponding three-dimensional hand model. By adopting the above technical solution, the corresponding preset posture can be screened out faster, and the efficiency of the screening process can be improved.

[0033] By adopting the above technical solution, the corresponding preset posture can be screened out faster, and the efficiency of the screening process can be improved.

[0034] Optionally, the second vector comparison process includes:

[0035] Form a second vector sequence with the LED light coordinates on the first finger of the serial number and the LED light coordinates on other fingers , and compare the second vector sequence with the first vector sequence of the first Q preset postures before sorting, and select the preset posture corresponding to the minimum value as the stereo matching screening result, where p ∈ [1, m - 2], is the p-th vector in the second vector sequence and the included angle between the first vector sequence corresponding to the preset posture, is the bias error coefficient.

[0036] By adopting the above technical solution, the corresponding preset posture can be accurately matched.

[0037] In a second aspect, the present application provides a finger capture method for humanoid robot reinforcement learning, adopting the following technical solution:

[0038] A finger capture method for reinforcement learning of humanoid robots, the method using a finger capture device for reinforcement learning of humanoid robots according to any one of the above, including:

[0039] Collecting the motion state data of the fingers through an IMU sensor; identifying the ID of each finger through the LED light flashing sequence; processing the data obtained by the IMU sensor and controlling the LED lights through a processing module; sensing the contact state between the finger and an external object through a tactile sensor;

[0040] Receiving and processing the IMU data from the finger module through a microcontroller and transmitting it to the analysis center;

[0041] Obtaining two-dimensional image data of multiple orientations of the hand position of the humanoid robot through a plurality of groups of cameras;

[0042] The analysis center identifies the spatial position coordinates of each finger based on the two-dimensional image data of multiple orientations, and judges the hand posture according to the spatial position coordinates of each finger, the contact state with the external object, and the motion state data.

[0043] In summary, the present application includes at least one of the following beneficial technical effects:

[0044] According to the present invention, the finger where the LED light is located can be determined through the LED light flashing sequence, avoiding the influence of external interference factors on the judgment of the position point, thereby improving the accuracy of the LED position point collection. The spatial position coordinates of each finger, the contact state with the external object, and the motion state data are used to comprehensively judge the hand posture of the humanoid robot, avoiding problems such as large error rates and inaccurate collection results in a single collection method. Description of the Drawings

[0045] Figure 1 It is a logic block diagram of the finger capture device for reinforcement learning of humanoid robots according to the present invention.

[0046] Figure 2 It is the finger capture effect when two groups of LED lights are set for each finger module.

[0047] Figure 3 It is the finger capture effect when one group of LED lights is set for each finger module.

[0048] Figure 4 It is a flowchart of the finger capture method for reinforcement learning of humanoid robots according to the present invention. Detailed Embodiments

[0049] The following details the embodiments of the present application, and the examples of the embodiments are shown in the drawings.

[0050] In the description of this specification, the description referring to terms such as "certain embodiments", "one embodiment", "some embodiments", "schematic embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiments or examples are included in at least one embodiment or example of this application. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiments or examples. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0051] An embodiment of this application discloses a finger capture device for the reinforcement learning of humanoid robots. Referring to Figure 1 , it includes a finger module, a microcontroller, several groups of cameras, and an analysis center. Among them, the finger module is set on each finger of the humanoid robot and includes an IMU sensor, an LED light, a tactile sensor, and a processing module; the IMU sensor is used to collect the motion state data of the finger, and the motion state data includes inertial data such as the acceleration and angular velocity of the finger. The LED light is used to perform ID identification on each finger through its flashing sequence. The flashing sequence of the LED light is pre-set according to the numbers of different fingers. The color and brightness of the LED light can be selected according to actual needs. Through the flashing sequence of the LED light, the finger where the LED light is located can be determined, avoiding the influence of external interference factors on the judgment of the position point, and thus improving the accuracy of the acquisition of the position point of the LED, etc. At the same time, the processing module is used for the processing of the data obtained by the IMU sensor and the control of the LED light; the tactile sensor is used to sense the contact state between the finger and an external object. The tactile sensor is installed at the end of each finger and can sense the contact force between the finger and an external object. The tactile sensor can be selected from types such as piezoelectric, capacitive, or resistive, and a suitable sensor is selected according to the specific application scenario; in addition, the microcontroller is used to receive and process the IMU data from the finger module and transmit it to the analysis center; several groups of cameras are provided, which can capture the two-dimensional image information of the positions of the hands of the humanoid robot in different directions. Then, through the analysis center, the spatial position coordinates of each finger are obtained by identifying the two-dimensional image data in multiple directions. Furthermore, based on the spatial position coordinates of each finger, the contact state with an external object, and the motion state data, the hand posture is judged. By using the above data to comprehensively judge the hand posture of the humanoid robot, problems such as a large error rate and inaccurate acquisition results existing in a single acquisition method are avoided.

[0052] In one embodiment, a comprehensive judgment process is provided. By filtering the three-dimensional coordinate data to remove noise points, screening and matching according to the obtained three-dimensional coordinate data and the preset information in the database, a corresponding three-dimensional finger model is obtained; calibrating and filtering the IMU data to eliminate errors and improve the accuracy of attitude calculation; making a threshold judgment on the tactile sensor data to determine whether an effective contact occurs; performing visual-inertial fusion on the preprocessed three-dimensional coordinate data and the IMU data, using the Kalman filter algorithm, taking the state transition model provided by the three-dimensional model and the IMU data as the system model, performing state estimation, and obtaining more accurate finger position and attitude information through continuous iterative updates.

[0053] In one embodiment, a process for identifying the ID of each finger by controlling the flashing sequences of several groups of LED lights is provided, including: presetting corresponding unique flashing sequences for different fingers, parsing out the ID of each finger from the continuous two-dimensional image data. Through this process, each finger can have a unique flashing sequence, and the serial number of the finger can be identified through this flashing sequence, thereby ensuring the anti-interference and accuracy of the acquisition of the finger spatial coordinate position.

[0054] In one embodiment, a process for the analysis center to judge the hand posture is provided, including: taking the positions of the LED lights as feature points, extracting the feature points in the two-dimensional image data, calculating the three-dimensional coordinates of the feature points based on the parallax of multiple azimuth feature points. The realization of this process is based on the three-dimensional reconstruction technology of multiple cameras, which will not be further described here. Performing stereo matching and screening on the three-dimensional coordinates of the feature points and the preset posture set in the database to obtain a three-dimensional hand model; judging the hand posture according to the three-dimensional hand model and the contact state and motion state data of each finger with the external object, thereby improving the anti-interference of finger state acquisition and obtaining more accurate finger position and attitude information.

[0055] In one embodiment, please refer to the attached Figure 2 ., each finger module is provided with two groups of LED lights, and the two groups of LED lights are arranged on different finger segments. The process of stereo matching and screening in this setting includes: obtaining the spatial distances of the LED lights on different finger segments of each finger to form a first distance sequence [x1, x2,..., xm]; comparing the first distance sequence [x1, x2,..., xm] with the first distance sequence [xi1, xi2,..., xim] corresponding to each preset posture in the preset posture set, and sorting each preset posture in ascending order, where Sort each preset posture in ascending order, where is the length influence coefficient of the k-th finger, which is related to the finger length of the humanoid robot. Since the deviation amount generated is larger when the finger length is longer, and the influence degree of the deviation amount of the shorter finger on the overall result is higher than that of the longer finger, the length influence coefficient is negatively correlated with the finger length of the humanoid robot. When the finger lengths of all humanoid robots are the same, the length influence coefficients are also the same. The first screening set is obtained by extracting the preset postures according to the preset screening rules; where m is the number of fingers, k is a positive integer and k ∈ [1, m]. Through the first comparison process, the preliminary screening process of the preset postures in the database can be realized. Then, the midpoints of the LED lamp connections of different finger segments on each finger are obtained, and the distances between the midpoints of the adjacent finger connections are obtained in the order of the fingers to form the second distance sequence [y1, y2,..., y(m - 1)]; The second distance sequence [y1, y2,..., y(m - 1)] is compared with the second distance sequence [yi1, yi2,..., yi(m - 1)] corresponding to each preset posture in the first screening set, and sorted in ascending order of the numerical values, where is the proportional adjustment coefficient, which is obtained by fitting according to the empirical data. In addition, j ∈ [1, m - 1]. The first Q names before sorting are selected for the first vector comparison process, and then the secondary screening process is completed. The corresponding preset posture is obtained according to the first vector comparison process as the stereo matching screening result, and the corresponding three-dimensional hand model is obtained. Through the above multi-level screening process, the calculation amount in the screening process can be reduced while ensuring the accuracy of the screening result, and the screening efficiency is improved.

[0056] In one embodiment, a preset screening rule is given, including: judging the corresponding quantity g, Xt is the preset threshold, which is set according to the empirical data. The quantity g is compared with the preset value gt, and the preset value gt is set according to the empirical data. If the quantity g exceeds the preset value gt, it means that there are more matching items. Therefore, the first gt names before sorting are selected as the first screening set to ensure the quantity of the matching preset postures in the first screening set. Otherwise, the first g names before sorting are selected as the first screening set to ensure the effectiveness and matching of the matching preset postures in the first screening set. By obtaining the preset postures in the first screening set through the preset screening rule, the accuracy and comprehensiveness of the obtained result can be ensured.

[0057] In one embodiment, a first vector comparison process is given, including: forming a first vector sequence from the midpoints of the LED lamp connections of different finger segments on the first finger of the serial number and the midpoints of the LED lamp connections of different finger segments on other fingers , and comparing the first vector sequence with the first vector sequences of the first Q preset postures before sorting, and selecting the preset posture corresponding to the minimum value as the stereo matching screening result, where s ∈ [1, m - 2]. is the s-th vector in the first vector sequence the included angle between the first vector sequence corresponding to the preset posture is the bias error coefficient, and the bias error coefficient is set according to the distance between other fingers and the first finger with the serial number. Since the angle deviation amount is larger when the distance between two fingers is farther, and the influence degree of the angle deviation amount of the shorter finger on the overall result is higher when the distance is closer, the bias error coefficient is negatively correlated with the finger distance of the humanoid robot. Through the above first vector comparison process, the corresponding preset posture can be accurately matched.

[0058] In one embodiment, please refer to the appendix Figure 3 , each finger module is provided with a group of LED lights, and the process of stereo matching and screening includes: obtaining the distances between adjacent LED lights in the finger order to form a third distance sequence [z1, z2,..., z(m - 1)]; comparing the third distance sequence [z1, z2,..., z(m - 1)] with the third distance sequence [zi1, zi2,..., zi(m - 1)] corresponding to each preset posture in the first screening set, and arranging them in ascending order of the numerical values, selecting the top Q before sorting for the second vector comparison process, and obtaining the corresponding preset posture as the stereo matching and screening result according to the second vector comparison process, and obtaining the corresponding three-dimensional hand model. In the state where each finger module is provided with a group of LED lights, through the third comparison process and the second vector comparison process, the corresponding preset posture can be screened faster, improving the efficiency of the screening process.

[0059] In one embodiment, a second vector comparison process is given, including: forming a second vector sequence from the LED light coordinates on the first finger with the serial number and the LED light coordinates on other fingers , and comparing the second vector sequence with the first vector sequences of the top Q preset postures before sorting, and selecting the preset posture corresponding to the minimum value as the stereo matching and screening result, where p ∈ [1, m - 2], is the p-th vector in the second vector sequence the included angle between the first vector sequence corresponding to the preset posture is the bias error coefficient. Through the above second vector comparison process, the corresponding preset posture can be accurately matched.

[0060] The embodiment of the present application also discloses a finger capture method for humanoid robot reinforcement learning. Refer to Figure 4, including: using an IMU sensor to collect the motion state data of the fingers; using the LED light flashing sequence to identify the ID of each finger; using a processing module to process the data obtained by the IMU sensor and control the LED lights; using a tactile sensor to sense the contact state between the fingers and external objects; using a microcontroller to receive and process the IMU data from the finger module and transmit it to the analysis center; using several groups of cameras to obtain two-dimensional image data of multiple orientations of the hand position of the humanoid robot; using the analysis center to identify the spatial position coordinates of each finger based on the two-dimensional image data of multiple orientations, and judge the hand posture according to the spatial position coordinates of each finger, the contact state with external objects, and the motion state data. In this embodiment, the finger where the LED light is located can be determined through the LED light flashing sequence, which avoids the influence of external interference factors on the judgment of the position point, and thus improves the accuracy of the LED position point collection. The hand posture of the humanoid robot is comprehensively judged by using the spatial position coordinates of each finger, the contact state with external objects, and the motion state data, which avoids problems such as a large error rate and inaccurate collection results in a single collection method.

[0061] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present application. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present application.

Claims

1. A finger capture device for reinforcement learning of humanoid robots, characterized in that, Comprising: Finger modules, each set on a finger of the humanoid robot, including an IMU sensor, an LED light, a tactile sensor, and a processing module; the IMU sensor is used to collect the motion state data of the finger; the LED light is used to perform ID recognition on each finger through its blinking sequence; the processing module is used for processing the data obtained by the IMU sensor and controlling the LED light; the tactile sensor is used to sense the contact state between the finger and an external object; A microcontroller, used to receive and process the IMU data from the finger module and transmit it to the analysis center; Cameras, provided in several groups, used to obtain two-dimensional image data of multiple orientations of the hand position of the humanoid robot; An analysis center, used to identify the spatial position coordinates of each finger based on the two-dimensional image data of multiple orientations, and judge the hand posture according to the spatial position coordinates of each finger, the contact state with an external object, and the motion state data.

2. The finger capture device for reinforcement learning of a humanoid robot according to claim 1, characterized in that, The process of controlling the blinking sequences of several groups of LED lights to perform ID recognition on each finger includes: Pre-setting corresponding unique blinking sequences for different fingers, and parsing the ID of each finger through continuous two-dimensional image data.

3. The finger capture device for reinforcement learning of a humanoid robot according to claim 2, characterized in that, The process of the analysis center judging the hand posture includes: Taking the position of the LED light as a feature point, extracting the feature points in the two-dimensional image data, and calculating the three-dimensional coordinates of the feature points according to the parallax of the feature points in multiple orientations; Performing stereo matching and screening on the three-dimensional coordinates of the feature points and the preset posture set in the database to obtain a three-dimensional hand model; Judging the hand posture according to the three-dimensional hand model, the contact state between each finger and an external object, and the motion state data.

4. A finger capture device for reinforcement learning of a humanoid robot according to claim 3, characterized in that, Each finger module is provided with two groups of LED lights, and the two groups of LED lights are arranged on different finger segments. The process of the stereo matching and screening includes: Obtain the spatial distances of the LED lights on different finger segments of each finger to form a first distance sequence [x1, x2, …, xm]; perform a first comparison between the first distance sequence [x1, x2, …, xm] and the first distance sequence [xi1, xi2, …, xim] corresponding to each preset posture in the preset posture set, and sort each preset posture in ascending order according to Sort each preset posture in ascending order, and extract the preset postures according to the preset screening rules to obtain a first screening set; where m is the number of fingers, k is a positive integer and k ∈ [1, m], is the length influence coefficient of the kth finger; Obtain the midpoints of the connections of the LED lights on different finger segments of each finger, and obtain the distances between the midpoints of the connections of adjacent fingers in the order of the fingers, forming a second distance sequence [y1, y2, …, y(m-1)]; compare the second distance sequence [y1, y2, …, y(m-1)] with the second distance sequence [yi1, yi2, …, yi(m-1)] corresponding to each preset posture in the first screening set, and sort them in ascending order of values, select the top Q before sorting for the first vector comparison process, obtain the corresponding preset posture as the stereo matching screening result according to the first vector comparison process, and obtain the corresponding three-dimensional hand model. is the proportional adjustment coefficient, j ∈ [1, m-1].

5. The finger capture device for reinforcement learning of a humanoid robot according to claim 4, characterized in that, The preset screening rules include: Judge For the corresponding quantity g, if the quantity g exceeds the preset value gt, then select the top gt before sorting as the first screening set; otherwise, select the top g before sorting as the first screening set, where Xt is the preset threshold value.

6. The finger capturing device for reinforcement learning of a humanoid robot according to claim 4, wherein The first vector comparison process includes: Connect the midpoints of the LED lamp connections on different finger segments of the first finger with the midpoints of the LED lamp connections on different finger segments of other fingers to form a first vector sequence , and compare the first vector sequence with the first vector sequences of the first Q preset postures before sorting, and select the preset posture corresponding to the minimum value as the stereo matching screening result, where s ∈ [1, m - 2], is the s-th vector in the first vector sequence and the included angle of the first vector sequence corresponding to the preset posture, is the offset error coefficient 7. The finger capture device for reinforcement learning of a humanoid robot according to claim 3, characterized in that, Each finger module is provided with a group of LED lights. The process of the stereo matching and screening includes: Obtain the distances between adjacent LED lights in the order of fingers to form a third distance sequence [z1, z2, …, z(m-1)]; perform a third comparison between the third distance sequence [z1, z2, …, z(m-1)] and the third distance sequence [zi1, zi2, …, zi(m-1)] corresponding to each preset posture in the first screening set, and sort them in ascending order of numerical values, select the top Q before sorting for the second vector comparison process, obtain the corresponding preset posture as the stereo matching screening result according to the second vector comparison process, and obtain the corresponding three-dimensional hand model.

8. A finger capture device for reinforcement learning of a humanoid robot according to claim 7, characterized in that, The second vector comparison process includes: Form a second vector sequence from the LED lamp coordinates on the first finger with the serial number and the LED lamp coordinates on other fingers , and compare the second vector sequence with the first vector sequence of the first Q preset postures before sorting, and select the preset posture corresponding to the minimum value as the stereo matching screening result, where p ∈ [1, m - 2], is the p-th vector in the second vector sequence the included angle with the first vector sequence corresponding to the preset posture, is the offset error coefficient.

9. A finger capture method for reinforcement learning of humanoid robots, characterized in that, The method adopts a finger capture device for reinforcement learning of a humanoid robot according to any one of claims 1-8, including: Collecting the motion state data of the finger through the IMU sensor; performing ID recognition on each finger through the blinking sequence of the LED light; processing the data obtained by the IMU sensor and controlling the LED light through the processing module; sensing the contact state between the finger and an external object through the tactile sensor; Receiving and processing the IMU data from the finger module through the microcontroller and transmitting it to the analysis center; Obtaining two-dimensional image data of multiple orientations of the hand position of the humanoid robot through several groups of cameras; Identifying the spatial position coordinates of each finger based on the two-dimensional image data of multiple orientations through the analysis center, and judging the hand posture according to the spatial position coordinates of each finger, the contact state with an external object, and the motion state data.

Images