Biological enlightenment learning system based on man-machine cooperative training and use method of biological enlightenment learning system

Abstract

The invention discloses a biological inspiration learning system based on man-machine cooperative training and a use method thereof.The biological inspiration learning system based on man-machine cooperative training comprises a robot learning unit and a biological inspiration teaching unit.According to the embodiment of the invention, a robot can obtain human action data in real time; key features are analyzed and extracted through a deep learning algorithm so as to quickly adapt to different motion modes; dynamic man-machine interaction is achieved in the machine learning process, the robot can not only observe human actions, but also adjust the motion strategy of the robot according to feedback of the human actions, and more natural and smooth cooperation is achieved; through human-machine collaborative learning, the robot can perform more accurate simulation and learning, the learning efficiency of the motion skill of the robot is significantly improved, and the sample dependence and sample quality problems in a traditional learning method are reduced.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of robotics technology, and in particular to a biologically inspired learning system based on human-machine collaborative training and a method for using the system. Background Art

[0002] With the rapid development of artificial intelligence and robotics, intelligent robots are increasingly used in many fields, especially in the service industry, medical assistance, and education. In order to enable robots to better interact with humans and enhance their autonomous learning and adaptability, researchers have begun to explore how to improve robots' motor skills and decision-making abilities by observing human bio-inspired behaviors. Existing research mainly focuses on using sensors, vision systems, and machine learning algorithms to analyze human actions so that robots can imitate or learn these actions.

[0003] In addition, traditional machine learning-related motion learning methods often rely on static data sets and programmed action processes, lacking real-time feedback and dynamic adaptability during the training process. In addition, existing robots often have difficulty effectively recognizing and understanding the diverse actions of humans in complex environments, resulting in low learning efficiency and insufficient adaptability. Summary of the invention

[0004] The embodiment of the present invention provides a biologically inspired learning system based on human-machine collaborative training and a method of using the same, which enables a robot to gradually improve its own movement ability by observing and analyzing human movements in real time to better simulate human movement characteristics, and has the advantages of high learning efficiency and strong adaptability.

[0005] In a first aspect, an embodiment of the present invention provides a biologically inspired learning system based on human-computer collaborative training, comprising:

[0006] A robot learning unit, comprising a robot perception module, a robot motion execution and balance control module, a robot information integration and transmission module, a robot motion learning module and a robot limb end pressure sensor, wherein the robot perception module is arranged at the head and each limb of the robot, the robot motion execution and balance control module is used to drive the robot's limbs to move, the robot information integration and transmission module and the robot motion learning module are placed in the robot's processor, and the robot limb end pressure sensors are distributed at the ends of each limb of the robot;

[0007] The biologically inspired teaching unit includes a human motion information acquisition module, a human limb extremity pressure acquisition module, a human information integration and transmission module, and a robot state feedback module. The human motion information acquisition module is used to acquire skeletal point change information of a human instructor during human motion, and the human limb extremity pressure acquisition module is used to acquire bipedal multi-point pressure information of a human instructor during human motion; the skeletal point change information and the bipedal multi-point pressure information are transmitted to the robot motion learning module and the robot motion execution and balance control module via the human information integration and transmission module to generate a robot following motion gait, and the robot following motion gait is fed back to the human instructor via the robot state feedback module.

[0008] In one embodiment, the robot perception module is used to obtain perception information, and the perception information includes optical perception information, ground three-dimensional point cloud perception information and limb end pressure perception information, wherein the optical perception information and the ground three-dimensional point cloud perception information collect environmental information with the robot as the first perspective, and map the three-dimensional point cloud to the environmental coordinate system.

[0009] In one embodiment, the robot state feedback module is used to feed back state information, and the state information includes learning accuracy value information, feedback guidance information and robot scene environment information.

[0010] In one embodiment, the logic for generating the feedback guidance information is: if the movement amplitude of the previous step of the human instructor exceeds the robot's movement threshold, the instructor is prompted to adjust the gait to within the robot's movement amplitude; if the consistency of the human instructor's movement gait is lower than the threshold, the instructor is prompted to perform bio-inspired teaching with a consistent gait; if the consistency of the human instructor's gait exceeds the threshold, the instructor is prompted to move with differentiated gait parameters.

[0011] In one embodiment, the human motion information acquisition module uses a plurality of visual depth sensors placed in the scene, and the plurality of visual depth sensors are combined with a skeleton point analysis algorithm to reconstruct a plurality of skeleton motion sequences of a human teacher.

[0012] In one embodiment, the actual motion state of the robot is based on the robot's own zero-point torque Z, and the gait M generated by the robot motion learning module robot If restrictions are imposed, the robot's real-time motion state is M physical =f(M robot ,Z), where f is the boundary function that ensures the robot does not fall.

[0013] In a second aspect, an embodiment of the present invention provides a method for using a biologically inspired learning system based on human-computer collaborative training, which is applied to the biologically inspired learning system based on human-computer collaborative training as described in the first aspect, and the method for using includes:

[0014] Step 1: The robot learning unit and the biologically inspired teaching unit are respectively arranged in area A and area B with the same area size and the same environmental configuration;

[0015] Step 2: A human teacher configures the motion name of the robot learning unit;

[0016] Step 3: The human teacher performs an action in area B;

[0017] Step 4: The human body motion information acquisition module acquires the bone point change information during the human body motion process, and the human body limb end pressure acquisition module acquires the bipedal multi-point pressure information during the human body motion process. The bone point change information and the bipedal multi-point pressure information are transmitted to the robot motion learning module via the human body information integration transmission module. The robot motion learning module combines with the motion execution and balance control module to generate the robot following motion gait;

[0018] Step 5: The robot follows the motion gait and feeds back to the human instructor via the robot information integration and transmission module;

[0019] Step 6: The human teacher performs the next action in area B and repeats steps 4 and 5;

[0020] Step 7: The robot saves the training data collected from step 4 to step 7, and trains the robot motion learning module based on the training data;

[0021] Step 8: Based on the subsequent actions of the human teacher, repeat steps 3 to 7.

[0022] In one embodiment, the human teacher performs an action in area B, including:

[0023] The human teacher performs an action in area B, and the human body motion information acquisition module and the human limb end pressure acquisition module collect the time series t 1 ~t n During the period, human motion data, real-time environmental data and plantar pressure data are collected, and the real-time environmental data is obtained through the three-dimensional point cloud E∈R T×P×3 It is represented by, where T is the number of time sampling points and P is the number of point clouds; the human body motion data includes a sequence of bone points moving over time, and the movement of all bones is represented by a multidimensional matrix M human ∈R T×J×KIt is represented by, where T is the number of time sampling points, J is the number of bone points, and K is the spatial description of each segment of bone points in space. The spatial description is adjusted to the three-dimensional coordinates of the head and end of the bone, or the coordinates of one end of the bone + the quaternion posture of the bone in three-dimensional space according to the characteristics of the machine learning algorithm; the plantar pressure data is expressed by a multidimensional matrix P∈p T×S It is represented by, where T is the number of time sampling points and S is the number of plantar sensors.

[0024] In one embodiment, the method of use further includes:

[0025] The three-dimensional point cloud matrix E, the multi-dimensional matrix M and the multi-dimensional matrix P are used as training information and synthesized into tensor inputs to the robot machine learning model. During the training process, humans first repeat the action c times, c>3, and then the robot generates its own skeleton point motion reference sequence M based on the repeated action and the network model. robot ∈R T×J×K The self-skeletal point motion reference sequence generates the rotation matrix sequence J of each joint through the geometric posture inverse solution algorithm. The self-skeletal point motion reference sequence is input into the robot motion execution and balance control module. The actual motion state of the robot is based on the robot's own zero-point torque Z. The gait M generated by machine learning is robot If restrictions are imposed, the robot's real-time motion state is M physical =f(M robot , Z), where f is the boundary function to ensure that the robot does not fall. The robot motion execution and balance control module first generates the robot's physical motion gait M based on the rotation matrix sequence J and combines its torque position or model prediction to adjust the joints. physical ; M human With M physical The difference between them is used as the evaluation function to evaluate the robot's motion process, and the robot's model parameters are trained through gradient optimization.

[0026] In one embodiment, the method of use further includes:

[0027] The robot generates feedback information, which contains a scalar M human With M physical The difference between them, and judge M human Whether the skeletal posture exceeds the robot's physical motion threshold, if it exceeds the robot's physical motion threshold, the feedback information prompts to adjust the human motion limit; and judging in different training processes of the same motion human If the consistency is lower than a preset value, random motion prompt information is generated in the feedback information.

[0028] The embodiment of the present invention includes: a biologically inspired learning system based on human-machine collaborative training and a method for using the same, wherein the biologically inspired learning system based on human-machine collaborative training includes a robot learning unit and a biologically inspired teaching unit, the robot learning unit includes a robot perception module, a robot motion execution and balance control module, a robot information integration and transmission module, a robot motion learning module and a robot limb end pressure sensor, the robot perception module is arranged on the head and each limb of the robot, the robot motion execution and balance control module is used to drive the robot's limbs to move, the robot information integration and transmission module and the robot motion learning module are arranged in the robot's processor, and the robot limb end pressure sensors are distributed in the robot The extremities of the robot; the biologically inspired teaching unit includes a human motion information acquisition module, a human limb extremity pressure acquisition module, a human information integration transmission module and a robot state feedback module. The human motion information acquisition module is used to collect the bone point change information of the human teacher during the human motion process, and the human limb extremity pressure acquisition module is used to collect the bipedal multi-point pressure information of the human teacher during the human motion process; the bone point change information and the bipedal multi-point pressure information are transmitted to the robot motion learning module and the robot motion execution and balance control module via the human information integration transmission module to generate the robot following motion gait, and the robot following motion gait is fed back to the human teacher through the robot state feedback module. Based on this, the embodiment of the present invention enables the robot to gradually improve its own motion ability by real-time observation and analysis of human movements, so as to better simulate the motion characteristics of humans, and it has the advantages of high learning efficiency and strong adaptability.

[0029] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The accompanying drawings are used to provide a further understanding of the technical solution of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the technical solution of the present invention and do not constitute a limitation on the technical solution of the present invention.

[0031] Figure 1 A physics distribution block diagram of a biologically inspired learning system based on human-computer collaborative training provided by one embodiment of the present invention;

[0032] Figure 2 A flow chart of a method for using a biologically inspired learning system based on human-computer collaborative training provided by one embodiment of the present invention;

[0033] Figure 3A schematic diagram of a feasible example of logical position configuration of a robot sensor under a training platform of the present invention provided in one embodiment of the present invention;

[0034] Figure 4 A schematic diagram of a feasible robot leg joint configuration case under the training platform of the present invention is provided as an embodiment of the present invention. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0036] It should be understood that in the description of the embodiments of the present invention, the meaning of multiple (or multiple) is more than two, greater than, less than, and exceeding are understood to exclude the number itself, and above, below, and within are understood to include the number itself. If there is a description of "first", "second", etc., it is only used for the purpose of distinguishing technical features, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features.

[0037] With the rapid development of artificial intelligence and robotics, intelligent robots are increasingly used in many fields, especially in the service industry, medical assistance, and education. In order to enable robots to better interact with humans and enhance their autonomous learning and adaptability, researchers have begun to explore how to improve robots' motor skills and decision-making abilities by observing human bio-inspired behaviors. Existing research mainly focuses on using sensors, vision systems, and machine learning algorithms to analyze human actions so that robots can imitate or learn these actions.

[0038] In addition, traditional machine learning-related motion learning methods often rely on static data sets and programmed action processes, lacking real-time feedback and dynamic adaptability during the training process. In addition, existing robots often have difficulty effectively recognizing and understanding the diverse actions of humans in complex environments, resulting in low learning efficiency and insufficient adaptability.

[0039] In view of the above-mentioned problems in the prior art, an embodiment of the present invention provides a bio-inspired learning system based on human-machine collaborative training and a method for using the same, wherein the bio-inspired learning system based on human-machine collaborative training includes a robot learning unit and a bio-inspired teaching unit, the robot learning unit includes a robot perception module, a robot motion execution and balance control module, a robot information integration and transmission module, a robot motion learning module and a robot limb end pressure sensor, the robot perception module is arranged on the head and each limb of the robot, the robot motion execution and balance control module is used to drive the robot's limbs to move, the robot information integration and transmission module and the robot motion learning module are placed in the robot's processor, and the robot limb end pressure sensor Pressure sensors are distributed at the ends of each limb of the robot; the biologically inspired teaching unit includes a human motion information acquisition module, a human limb end pressure acquisition module, a human information integration transmission module and a robot state feedback module, the human motion information acquisition module is used to collect the bone point change information of the human teacher during the human motion process, and the human limb end pressure acquisition module is used to collect the double-foot multi-point pressure information of the human teacher during the human motion process; the bone point change information and the double-foot multi-point pressure information are transmitted to the robot motion learning module and the robot motion execution and balance control module via the human information integration transmission module to generate the robot following motion gait, and the robot following motion gait is fed back to the human teacher through the robot state feedback module. Based on this, the embodiment of the present invention enables the robot to gradually improve its own motion ability by real-time observation and analysis of human movements to better simulate human motion characteristics, and it has the advantages of high learning efficiency and strong adaptability.

[0040] like Figure 1 As shown, Figure 1 It is a physics distribution block diagram of a biologically inspired learning system based on human-computer collaborative training provided by an embodiment of the present invention.

[0041] 1-A is a robot learning unit, which may include:

[0042] 1-A1 is the robot perception module, which is in the form of optical and pressure sensors and is distributed in fixed scenes, the robot head and each limb;

[0043] 1-A2 is the robot motion and balance control module, which is used to drive the robot's limbs to move;

[0044] 1-A3 is the robot information integration and transmission module, which is placed in the robot processor in the form of software;

[0045] 1-A4 is the robot motion learning module, which is placed in the robot processor in the form of an algorithm;

[0046] 1-A5 is the pressure sensor at the end of the robot's limbs, which is distributed at the end of each limb of the robot.

[0047] 1-B is a biologically inspired teaching unit, which may include:

[0048] 1-B1 is for human teachers;

[0049] 1-B2 is a human motion information acquisition module and multiple three-dimensional sensing devices placed in a fixed area of ​​the scene;

[0050] 1-B3 is a human limb end pressure collection module, which is distributed in the palms and soles of human teachers in the form of a sensor array;

[0051] 1-B4 is the robot status feedback module, which feeds back the environment, robot status and prompt information through VR / AR devices;

[0052] The human body information integration and transmission module (not shown) can transmit the bone point change information and the bipedal multi-point pressure information to the robot motion learning module.

[0053] It can be understood that the robot perception module 1-A1 is used to obtain perception information, which includes optical perception information, ground three-dimensional point cloud perception information and limb end pressure perception information. Among them, the optical perception information and ground three-dimensional point cloud perception information collect environmental information with the robot as the first perspective, and map the three-dimensional point cloud to the environmental coordinate system.

[0054] It can be understood that the human limb extremity pressure collection module 1-B3 has the following characteristics: the pressure sensor used in the human limb extremity pressure collection module 1-B3 has multiple pressure detection points, and the pressure sensor used by the robot perception module 1-A1 to obtain limb extremity pressure perception information has the same size, pressure detection point position and sampling rate as the pressure sensor used in the human limb extremity pressure collection module.

[0055] It can be understood that the robot state feedback has three attributes, namely: 1) the learning accuracy value, which is fed back to the human instructor by the display screen, reflecting the degree of fit between the robot's current motion state generated based on learning and the human body's teaching motion state; 2) feedback guidance information: This information is fed back to the human instructor by the display screen, including guidance information for instructors such as "increase / decrease movement speed", "increase / decrease leg lifting height", and "try to keep it the same as the last movement". This guidance information is generated by the robot motion learning module 1-A4; 3) robot scene environment information, which is obtained by digitally modeling the robot environment or collecting it with a fixed camera placed in the scene.

[0056] It can be understood that the logic of generating feedback guidance information is: if the movement amplitude of the human teacher's previous step exceeds the robot's movement threshold, the teacher is prompted to adjust his gait to within the robot's movement amplitude. If the consistency of the human teacher's movement gait is lower than the threshold, the teacher is prompted to conduct bio-inspired teaching with a consistent gait (i.e., try to keep each teaching movement consistent). If the consistency of the human teacher's movement gait exceeds the threshold, the teacher is prompted to exercise with differentiated gait parameters.

[0057] It can be understood that the human motion information acquisition module 1-B2 is characterized by using multiple visual depth sensors placed in the scene and combining a skeleton point analysis algorithm to reconstruct multiple skeleton motion sequences of a human teacher.

[0058] It can be understood that the robot motion and balance control module 1-A2 has the following characteristics: the actual motion state of the robot is based on the robot's own zero-point torque Z, and the gait M generated by the robot motion learning module 1-A4 is robot If restrictions are imposed, the robot's real-time motion state is M physical =f(M robot ,Z), where f is the boundary function that ensures the robot does not fall.

[0059] Among them, E is the boundary range of the robot's supporting feet.

[0060] It is understandable that before the human instructor lifts and lowers his legs, the robot state feedback module outputs teaching prompts, and the human instructor adjusts the movement based on the prompts, gait type, and actual environment.

[0061] like Figure 2 As shown, Figure 2 This is a flow chart of a method for using a biologically inspired learning system based on human-computer collaborative training provided by an embodiment of the present invention, and the method for using the system is as follows:

[0062] Step 1: Set the robot learning unit and the bio-inspired teaching unit in area A and area B with the same area size and environment configuration respectively;

[0063] Step 2: The human teacher configures the motion name of the robot learning unit;

[0064] Step 3: The human teacher performs an action in area B;

[0065] Step 4: The human motion information acquisition module acquires the bone point change information during the human motion process, and the human limb end pressure acquisition module acquires the bipedal multi-point pressure information during the human motion process. The bone point change information and bipedal multi-point pressure information are transmitted to the robot motion learning module via the human information integration transmission module. The robot motion learning module combines the motion execution and balance control modules to generate the robot following motion gait;

[0066] Step 5: The robot follows the movement gait and feeds back to the human instructor through the robot information integration transmission module;

[0067] Step 6: The human teacher performs the next action in area B and repeats steps 4 and 5;

[0068] Step 7: The robot saves the training data collected from step 4 to step 7, and trains the robot motion learning module based on the training data;

[0069] Step 8: Based on the subsequent actions of the human teacher, repeat steps 3 to 7.

[0070] like Figure 3 As shown, Figure 3 This is a feasible example of the logical position configuration of robot sensors under the training platform described in the present invention, wherein: 3-1 is the optical sensor on the robot head, 3-2 is the internal posture sensor of the robot, 3-3 is the pressure sensor at the end of the robot limbs, 3-4 is the upper limb of the multi-degree-of-freedom robot, and 3-5 is the lower limb of the multi-degree-of-freedom robot.

[0071] Figure 4 This is a feasible robot leg joint configuration example under the training platform described in the present invention, wherein 4-1 to 4-6 are the robot's 6-DOF drivers respectively, and 4-7 is the robot's plantar pressure sensor array.

[0072] The training platform is used as follows:

[0073] 1. Divide the area into two units of the same size and environment configuration. The robot learning unit (2-A) is where the robot is placed, and the biologically inspired teaching unit (2-B) is the activity area for human teachers.

[0074] 2. Before learning begins, the human teacher and the robot maintain the same posture (can be any initial posture). First, the human teacher defines a movement name (such as "single leg lift") and inputs the gait into the robot, and then the human and the robot start bio-inspired teaching.

[0075] 3. When a human performs an action, the human motion information collection module and the human foot pressure collection module collect the time series t 1 ~t nThe human motion data and real-time environmental data during the period. Environmental data can be obtained through the three-dimensional point cloud E∈R T×P×3 It is represented by , where T is the number of time sampling points and P is the number of point clouds. Human motion data includes a sequence of bone points moving over time. The motion of all bones can be represented by a multidimensional matrix M human ∈R T×J×K Where T is the number of time sampling points, J is the number of bone points, and K is the spatial description of each segment of bone points in space. This description can be adjusted to the three-dimensional coordinates of the first and last ends of the bone, or the coordinates of one end of the bone + the quaternion posture of the bone in three-dimensional space according to the characteristics of the machine learning algorithm. The plantar pressure data can be used to calculate the multidimensional matrix P∈p T×S Describe it as follows, where T is the number of time sampling points and S is the number of plantar sensors.

[0076] 4. The M, P, and E matrices are used as training information and are synthesized into tensor inputs and passed to the robot machine learning model. The machine learning model can be built using a model framework such as a deep neural network (DNN), a long-short term memory network (LSTM) combined with a generative adversarial network model (GAN). During the training process, humans first repeat the action c times (c>3), and then the robot generates its own skeleton point motion reference sequence M based on the repeated action and the network model. robot ∈R T×J×K The sequence is used to generate the rotation matrix sequence J of each joint through the geometric posture inverse solution algorithm (such as spatial matrix transformation combined with the cosine theorem). The sequence is input into the robot balance control module. The balance control module can be implemented using algorithms such as the zero-point moment algorithm (ZMP) and the model prediction algorithm (MPC). The actual motion state of the robot is based on the robot's own zero-point moment Z, and the gait M generated by machine learning is robot If restrictions are imposed, the robot's real-time motion state is M physical =f(M robot ,Z), where f is the boundary function to ensure that the robot does not fall. The balance control module first adjusts the joints based on J, combined with its torque position or model prediction, to generate the robot's physical motion gait M physical .

[0077] 5. M human With M physical The difference between them is used as the evaluation function to evaluate the robot's motion process. The robot's model parameters are trained through gradient optimization. The gradient optimization algorithm can be adjusted according to the specific model.

[0078] 6. The robot generates feedback information, which contains a scalar M human With M physical The difference between the two can be calculated as norm(M human -Mphysical ). At the same time, judge: 1) M human Whether the skeletal posture exceeds the robot's physical motion threshold. If it exceeds the threshold, the feedback information prompts to adjust the human motion limit. 2) During different training sessions of the same exercise, M human If the consistency is lower than the preset value, a random exercise prompt will be generated in the feedback information (such as "increase / decrease exercise speed", "increase / decrease leg lift height"), otherwise a prompt will be given to "try to keep it the same as the last exercise".

[0079] 7. Repeat steps 3-5.

[0080] An implementation example of the present invention is given below:

[0081] The robot learning unit (1-A) and the teaching inspiration unit (1-B) are both 5m*5m flat venues, and the ground is green plastic.

[0082] The robot movement is driven by a bus driver with angle feedback.

[0083] At the algorithm and software level, the balance control algorithm can use the zero-point torque algorithm. The robot's machine learning model can be implemented using the long short-term memory / adversarial network (LSTM / GAN). Information integration and transmission are implemented using WIFI as the physical layer medium. The balance algorithm, information integration and transmission software, and balance control algorithm are placed inside the robot in the form of software for processing. The processor can be an Intel series processor.

[0084] At the sensor level, robot perception and human motion acquisition modules can be implemented using point cloud-vision sensors. A feasible sensor model is realSense-D435i. Carbon resistor film sensor arrays can be used for limb end pressure sensors. Robot status feedback can be implemented using various wearable devices such as HoloLens.

[0085] In terms of the use of bio-inspired learning systems based on human-computer collaborative training, one implementation case is as follows:

[0086] 1) The robot has 24 degrees of freedom, of which the two arms have a total of 12 degrees of freedom and the two legs have a total of 12 degrees of freedom.

[0087] 2) The human instructor defines the exercise type as "squat".

[0088] 3) The human instructor repeats the squat action 10 times. The platform generates a human skeletal motion sequence based on AlphaPose and point cloud clustering to form a training set.

[0089] 4) The robot uses the training set and the GAN model to generate a feasible movement sequence of "squatting".

[0090] 5) The human instructor performs one squat.

[0091] 6) The robot's internal GAN ​​algorithm combines the current motion data to generate a set of motion sequences

[0092] 7) The robot uses the zero-point torque algorithm (ZMP) to adjust the generated motion sequence and generate a physical motion sequence.

[0093] 8) The physical motion sequence generates 24 (number of joints) * 3 (target position, target speed, maximum load-bearing torque) dimensional joint information through the inverse kinematics algorithm (IK), which is input into the robot joint driver through the CAN bus protocol. The driver will drive the robot to repeat the action once.

[0094] 9) The robot compares the difference between human motion and machine physical motion through difference calculation, takes the difference as the evaluation value, and uses the gradient descent method to optimize the model parameters.

[0095] 10) Repeat steps 4-9.

[0096] Based on this, the present invention proposes a bio-inspired learning system based on human-machine collaborative training and its use method. By acquiring human motion sequences, the robot can gradually observe and learn human movements in a real-time environment. The present invention has at least the following significant advantages:

[0097] Real-time learning and interactive mimicry: This system enables the robot to obtain human motion data in real time, analyze and extract key features through deep learning algorithms, and quickly adapt to different movement patterns.

[0098] Dynamic interaction: Dynamic interaction between humans and machines is achieved during machine learning. Robots can not only observe human movements, but also adjust their own movement strategies based on their feedback, achieving more natural and smooth collaboration.

[0099] Improve learning efficiency: Through human-machine collaborative learning, robots can imitate and learn more accurately, significantly improve the learning efficiency of their motor skills, and reduce sample dependence and sample quality problems in traditional learning methods.

[0100] The above is a specific description of the preferred implementation of the present invention, but the present invention is not limited to the above-mentioned implementation mode. Technical personnel familiar with the field can also make various equivalent deformations or substitutions under the shared conditions without violating the spirit of the present invention. These equivalent deformations or substitutions are all included in the scope defined by the claims of the present invention.

Claims

1. A biologically inspired learning system based on human-computer collaborative training, characterized in that: include: A robot learning unit, comprising a robot perception module, a robot motion execution and balance control module, a robot information integration and transmission module, a robot motion learning module and a robot limb end pressure sensor, wherein the robot perception module is arranged at the head and each limb of the robot, the robot motion execution and balance control module is used to drive the robot's limbs to move, the robot information integration and transmission module and the robot motion learning module are placed in the robot's processor, and the robot limb end pressure sensors are distributed at the ends of each limb of the robot; The biologically inspired teaching unit includes a human motion information acquisition module, a human limb extremity pressure acquisition module, a human information integration and transmission module, and a robot state feedback module. The human motion information acquisition module is used to acquire skeletal point change information of a human instructor during human motion, and the human limb extremity pressure acquisition module is used to acquire bipedal multi-point pressure information of a human instructor during human motion; the skeletal point change information and the bipedal multi-point pressure information are transmitted to the robot motion learning module and the robot motion execution and balance control module via the human information integration and transmission module to generate a robot following motion gait, and the robot following motion gait is fed back to the human instructor via the robot state feedback module.

2. The biologically inspired learning system based on human-computer collaborative training according to claim 1, characterized in that: The robot perception module is used to obtain perception information, which includes optical perception information, ground three-dimensional point cloud perception information and limb end pressure perception information, wherein the optical perception information and the ground three-dimensional point cloud perception information collect environmental information with the robot as the first perspective, and map the three-dimensional point cloud to the environmental coordinate system.

3. The biologically inspired learning system based on human-computer collaborative training according to claim 1, characterized in that: The robot state feedback module is used to feed back state information, and the state information includes learning accuracy numerical information, feedback guidance information and robot scene environment information.

4. The biologically inspired learning system based on human-computer collaborative training according to claim 3, characterized in that: The logic for generating the feedback guidance information is as follows: if the movement amplitude of the previous step of the human teacher exceeds the robot's movement threshold, the teacher is prompted to adjust his gait to within the robot's movement amplitude; if the consistency of the human teacher's movement gait is lower than the threshold, the teacher is prompted to perform bio-inspired teaching with a consistent gait; If the consistency of the human instructor's gait exceeds a threshold, prompts are given to guide the instructor to move with differentiated gait parameters.

5. The biologically inspired learning system based on human-computer collaborative training according to claim 1, characterized in that: The human motion information acquisition module uses a plurality of visual depth sensors placed in the scene, and the plurality of visual depth sensors are combined with a skeleton point analysis algorithm to reconstruct a plurality of skeleton motion sequences of a human teacher.

6. The biologically inspired learning system based on human-computer collaborative training according to claim 1, characterized in that: The actual motion state of the robot is based on the robot's own zero-point torque Z, and the gait M generated by the robot motion learning module robot If restrictions are imposed, the real-time motion state of the robot is M physical =f(M robot ,Z), where f is the boundary function that ensures the robot does not fall.

7. A method for using a biologically inspired learning system based on human-computer collaborative training, characterized in that: Applied to the bio-inspired learning system based on human-computer collaborative training as claimed in any one of claims 1 to 6, the method of use comprising: Step 1: The robot learning unit and the biologically inspired teaching unit are respectively arranged in area A and area B with the same area size and the same environment configuration; Step 2: A human teacher configures the motion name of the robot learning unit; Step 3: The human teacher performs an action in area B; Step 4: The human body motion information acquisition module acquires the bone point change information during the human body motion process, and the human body limb end pressure acquisition module acquires the bipedal multi-point pressure information during the human body motion process. The bone point change information and the bipedal multi-point pressure information are transmitted to the robot motion learning module via the human body information integration transmission module. The robot motion learning module combines with the motion execution and balance control module to generate the robot following motion gait; Step 5: The robot follows the movement gait and feeds back to the human instructor via the robot information integration and transmission module; Step 6: The human teacher performs the next action in area B and repeats steps 4 and 5; Step 7: The robot saves the training data collected from step 4 to step 7, and trains the robot motion learning module based on the training data; Step 8: Based on the subsequent actions of the human teacher, repeat steps 3 to 7.

8. The method of use according to claim 7, characterized in that: The human teacher performs an action in area B, including: The human teacher performs an action in area B, and the human body motion information acquisition module and the human body limb end pressure acquisition module collect the time series t1~t n During the period, human motion data, real-time environmental data and plantar pressure data are collected, and the real-time environmental data is obtained through the three-dimensional point cloud E∈R T×P×3 It is represented by, where T is the number of time sampling points and P is the number of point clouds; the human body motion data includes a sequence of bone points moving over time, and the movement of all bones is represented by a multidimensional matrix M human ∈R T ×J×K It is represented by , where T is the number of time sampling points, J is the number of bone points, and K is the spatial description of each segment of bone points in space. The spatial description is adjusted to the three-dimensional coordinates of the head and end of the bone, or the coordinates of one end of the bone + the quaternion posture of the bone in three-dimensional space according to the characteristics of the machine learning algorithm; the plantar pressure data is expressed by a multidimensional matrix P∈p T×S It is represented by, where T is the number of time sampling points and S is the number of plantar sensors.

9. The method of use according to claim 8, characterized in that: The method of use also includes: The three-dimensional point cloud matrix E, the multi-dimensional matrix M and the multi-dimensional matrix P are used as training information and synthesized into tensor inputs to the robot machine learning model. During the training process, humans first repeat the action c times, c>3, and then the robot generates its own skeleton point motion reference sequence M based on the repeated action and the network model. robot ∈R T×J×K The self-skeletal point motion reference sequence generates the rotation matrix sequence J of each joint through the geometric posture inverse solution algorithm. The self-skeletal point motion reference sequence is input into the robot motion execution and balance control module. The actual motion state of the robot is based on the robot's own zero-point torque Z. The gait M generated by machine learning is robot If restrictions are imposed, the real-time motion state of the robot is M physical =f(M robot , Z), where f is the boundary function to ensure that the robot does not fall. The robot motion execution and balance control module first generates the robot's physical motion gait M based on the rotation matrix sequence J and combines its torque position or model prediction to adjust the joints. physical ; M human With M physical The difference between them is used as the evaluation function to evaluate the robot's motion process, and the robot's model parameters are trained through gradient optimization.

10. The method of use according to claim 9, characterized in that: The method of use also includes: The robot generates feedback information, which contains a scalar M human With M physical The difference between them, and judge M human Whether the skeletal posture exceeds the robot's physical motion threshold, if it exceeds the robot's physical motion threshold, the feedback information prompts to adjust the human motion limit; and judging in different training processes of the same motion human If the consistency is lower than a preset value, random motion prompt information is generated in the feedback information.

Images