Robot dog intelligent gait planning method based on deep learning

Abstract

The invention discloses a robot dog intelligent gait planning method based on deep learning, which comprises a gait acquisition module, a scene acquisition module, a virtual human positioning module, a path planning module, an execution control module and a deep learning module, the virtual human positioning module is used for determining and controlling the position of a virtual human in a three-dimensional scene, and the scene acquisition module is used for acquiring three-dimensional model information between the virtual human and a target location. The gait collection system can collect the motion state of a real organism in real time so as to convert and plan the motion posture suitable for walking of the robot dog, and continuous correction learning can be carried out according to continuously collected different data through the deep learning system, so that the walking flexibility of the robot dog is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of quadruped robot gait planning, and particularly to an intelligent gait planning method for a quadruped robot based on deep learning. Background Technique

[0002] The walking gait of a quadruped robot is achieved by collecting the pace movements of a biological (dog), converting them into information flows, and then controlling them by various corresponding module systems. The traditional gait of a quadruped robot requires artificial input of various data streams to the system side as a basic database, and then controlling the quadruped robot to walk according to the information stored in this database. However, this method has great limitations. It does not have the function of collecting the surrounding environment and the moving postures of organisms during walking, making the database information of the quadruped robot unable to be updated in real time. The quadruped robot can only move according to the pre-input information settings, which may cause the quadruped robot to be blocked by obstacles and difficult to pass through, greatly limiting the activity range of the quadruped robot and reducing its environmental adaptability.

[0003] Secondly, some iteratively updated quadruped robots have the function of collecting environmental information, but they cannot directly complete the storage and conversion of information by themselves. They still need to manually store and convert the received information data to complete the upgrade of the quadruped robot. This process is too cumbersome and has high limitations. Moreover, it is difficult to upgrade in a timely manner on-site due to the limitations of the task environment, resulting in the failure of the quadruped robot to execute tasks and reducing the flexibility of use of the quadruped robot.

[0004] In view of the above problems, the present invention is improved. Summary of the Invention

[0005] The present invention provides an intelligent gait planning method for a quadruped robot based on deep learning, which solves the above problems existing in the prior art during use.

[0006] The technical solution of the present invention is implemented as follows: An intelligent gait planning method for a quadruped robot based on deep learning includes a gait acquisition module, a scene acquisition module, a virtual human positioning module, a path planning module, an execution control module, and a deep learning module. The gait acquisition module is used to collect the pace movements of a biological (dog), the virtual human positioning module is used to determine and control the position of a virtual human in a three-dimensional scene, the scene acquisition module is used to obtain the three-dimensional model information between the virtual human and the target location, the path planning module determines multiple paths connecting the virtual human and the target location according to the three-dimensional model information, and the real dog interacts with the system by using the execution control module;

[0007] S31. Set a base line segment according to the position and orientation of the virtual human;

[0008] S32. Select a bifurcation point on the base line segment;

[0009] S33. Connect the bifurcation point and the target location to obtain a quasi-path segment;

[0010] S34. Determine whether the quasi-path segment meets the slope requirement. If it meets, use the connection line between the virtual person and the bifurcation point and the quasi-path segment as the travel path. If it does not meet, jump to step S35;

[0011] S35. Find a change point on the quasi-path segment, draw a perpendicular line to the quasi-path segment through the change point, and find two temporary points A and B that meet the terrain height requirement on the perpendicular line. The temporary points are located on both sides of the quasi-path segment;

[0012] S36. Calculate the energy consumption index according to the corresponding formula;

[0013] S36. Select the point with a lower energy consumption index as the path waypoint, and use the connection line between the virtual person and the bifurcation point, the connection line between the bifurcation point and the waypoint, and the connection line between the waypoint and the target location as the travel path, so as to form a moving gait motion process;

[0014] When the steering switch is turned on, the gait acquisition module only acquires the steering data of the real person and determines the direction of the virtual person. The path planning module regenerates several travel paths. After one of the paths is selected, the steering switch automatically turns off. The gait acquisition module only acquires the forward gait data of the real person. The virtual person positioning module controls the virtual person to move forward on the selected path according to the forward gait data. The deep learning module includes being configured to be accessible via a communication network for enabling a first artificial neural network model to learn;

[0015] The data classification and storage module is configured to receive and store source learning data;

[0016] The learning module is configured to use the source learning data to enable a second artificial neural network model in the system to learn relevant information;

[0017] The adjustment information transmission module is configured to transmit adjustment information including learning relevant information and parameter information of the second artificial neural network model to an external deep learning system to request adjustment of the first artificial neural network model, wherein the learning relevant information and parameter information are modified according to the learning result;

[0018] The adjustment module is configured to update the parameter information of the second artificial neural network model according to the adjustment information transmitted from the external deep learning system.

[0019] The machine dog intelligent gait planning method based on deep learning as described above in the present invention further includes: the deep learning module further includes a diagnosis module, which is configured to determine whether it has walked along the optimal established route based on the first artificial neural network model optimized by the repeated operations of the learning module and the adjustment module.

[0020] The machine dog intelligent gait planning method based on deep learning as described above in the present invention further includes: the adjustment information transmission module preferentially transmits hyperparameters required to synchronize the initial learning state of the first artificial neural network model as the adjustment information.

[0021] The machine dog intelligent gait planning method based on deep learning as described above in the present invention further includes: the learning-related information includes any one of the ratio of learning data to source learning data to be learned and the number of learned source learning data segments.

[0022] The machine dog intelligent gait planning method based on deep learning as described above in the present invention further includes: the first-level classification criterion of the data classification unit includes color values for identifying one or more in a normal image as classification criterion values.

[0023] The machine dog intelligent gait planning method based on deep learning as described above in the present invention further includes: the adjustment information transmission module includes at least two or more of the first-level classification criterion, the second-level classification criterion, and the third-level classification criterion in the adjustment information.

[0024] The machine dog intelligent gait planning method based on deep learning as described above in the present invention further includes: the adjustment information transmission module transmits the adjustment information to the external deep learning system.

[0025] The machine dog intelligent gait planning method based on deep learning as described above in the present invention further includes: in step S32, multiple fork points are selected, and the path planning module generates multiple travel paths for selection.

[0026] The machine dog intelligent gait planning method based on deep learning as described above in the present invention further includes: the execution control module includes three basic functions: selecting a target location in the three-dimensional scene, selecting an execution path from the paths generated by the path planning module, and turning on the steering switch.

[0027] In summary, the beneficial effects of the present invention are as follows:

[0028] 1. In the structural design of the present invention that combines an optimized gait acquisition system and an optimized deep learning system, the gait acquisition system can collect the motion state of real organisms in real time and thus transform and plan a motion posture suitable for the walking of a robot dog. Moreover, through the deep learning system, continuous correction learning can be carried out according to continuously collected different data, thereby improving the walking flexibility of the robot dog.

[0029] 2. This system divides the gait acquisition data of real people into turning data and forward data. The turning data is used to change the orientation of the virtual person and re-collect and generate a path, and the forward data is used to control the virtual person to move on the selected path. Therefore, a real person can turn freely during forward movement without being collected turning data, realizing the control of a virtual person to move in a larger virtual space in a smaller real space. And the path planning module provides multiple suitable paths by analyzing the changes in the terrain height in the three-dimensional scene, making the provided paths closer to the paths that real people want to move along.

[0030] 3. Since the artificial neural network model performs partial self-optimization using parameters obtained from the learning results generated by an external system, it is actually impossible to restore or infer the source learning data, thereby minimizing the damage caused by information exposure.

[0031] 4. In particular, without sharing the source learning data, the present invention can achieve the same effects as those obtained by learning the same learning data. And thus, even when the artificial neural network model is in an environment where it can only learn a small amount of source data, the artificial neural network model can be changed into an artificial neural network model that has learned a large amount of source learning data. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0033] Figure 1 It is a schematic diagram of the overall structural framework of the present invention;

[0034] Figure 2 It is a schematic diagram of the relationship between the change points and the temporary points of the present invention; DETAILED DESCRIPTION OF THE EMBODIMENTS

[0035] The following will be combined with the drawings in the embodiments of the present invention Figure 1-2, the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0036] Embodiment

[0037] A machine dog intelligent gait planning method based on deep learning, including a gait acquisition module, a scene acquisition module, a virtual human positioning module, a path planning module, an execution control module, and a deep learning module. The gait acquisition module is used to collect the pace actions of a biological (dog). The virtual human positioning module is used to determine and control the position of the virtual human in a three-dimensional scene. The scene acquisition module is used to obtain the three-dimensional model information between the virtual human and the target location. The path planning module determines multiple paths connecting the virtual human and the target location according to the three-dimensional model information. The real dog interacts with the system by using the execution control module;

[0038] S31. Set a base line segment according to the position and orientation of the virtual human;

[0039] S32. Select a fork point on the base line segment;

[0040] S33. Connect the fork point and the target location to obtain a quasi-path line segment;

[0041] S34. Determine whether the quasi-path line segment meets the slope requirement. If it meets, use the connection line between the virtual human and the fork point and the quasi-path line segment as the traveling path. If it does not meet, jump to step S35;

[0042] S35. Find a change point on the quasi-path line segment, draw a perpendicular line to the quasi-path line segment through the change point, and find two temporary points A and B that meet the terrain height requirement on the perpendicular line. The temporary points are located on both sides of the quasi-path line segment;

[0043] S36. Calculate the energy consumption index according to the corresponding formula;

[0044] Let the path length from the fork point to point A be L_{1}, and the energy consumption coefficient per unit length be k_{1};

[0045] The path length from point A to the target location is L_{2}, and the energy consumption coefficient per unit length is k_{2};

[0046] The path length from the fork point to point B is L_{3}, and the energy consumption coefficient per unit length is k_{3};

[0047] The path length from point B to the target location is L_{4}, and the energy consumption coefficient per unit length is k_{4};

[0048] The path length from the bifurcation point to the change point is \(L_{5}\), and the energy consumption coefficient per unit length is \(k_{5}\);

[0049] The path length from the change point to the target location is \(L_{6}\), and the energy consumption coefficient per unit length is \(k_{6}\);

[0050] The path length from the change point to point A is \(L_{7}\), and the energy consumption coefficient per unit length is \(k_{7}\);

[0051] The path length from the change point to point B is \(L_{8}\), and the energy consumption coefficient per unit length is \(k_{8}\).

[0052] Calculation formula

[0053] If going from the bifurcation point through point A to the target location: the total energy consumption \(E_{A}=k_{1}L_{1}+k_{2}L_{2}\)

[0054] If going from the bifurcation point through point B to the target location: the total energy consumption \(E_{B}=k_{3}L_{3}+k_{4}L_{4}\)

[0055] If going from the bifurcation point through the change point to the target location: the total energy consumption \(E_{change}=k_{5}L_{5}+k_{6}L_{6}\)

[0056] If going from the bifurcation point through the change point and then through point A to the target location: the total energy consumption \(E_{change

[0057] A}=k_{5}L_{5}+k_{7}L_{7}+k_{2}L_{2}\)

[0058] If going from the bifurcation point through the change point and then through point B to the target location: the total energy consumption \(E_{change

[0059] B}=k_{5}L_{5}+k_{8}L_{8}+k_{4}L_{4}\)

[0060] In practical applications, the \(k\) values can be determined according to specific path attributes, transportation tools (or energy consumption entities), etc., and then combined with the actually measured or given \(L\) values to calculate the energy consumption;

[0061] S36. Select the point with a lower energy consumption index as the waypoint of the path, and use the connection line between the virtual person and the bifurcation point, the connection line between the bifurcation point and the waypoint, and the connection line between the waypoint and the target location as the traveling path, thereby forming the motion process of the dynamic gait;

[0062] The execution control module includes three basic functions: selecting a target location in the three-dimensional scene, selecting an execution path from the paths generated by the path planning module, and turning on the steering switch;

[0063] When the steering switch is turned on, the gait acquisition module only acquires the steering data of a real person and determines the direction of the virtual person. The path planning module regenerates several travel paths. After one of the paths is selected, the steering switch automatically turns off, and the gait acquisition module only acquires the forward gait data of the real person. The virtual person positioning module controls the virtual person to move forward on the selected path according to the forward gait data;

[0064] The deep learning module includes being configured to be accessible via a communication network for enabling a first artificial neural network model to learn;

[0065] The data classification and storage module is configured to receive and store source learning data;

[0066] The learning module is configured to use the source learning data to enable a second artificial neural network model in the system to learn relevant information;

[0067] The adjustment information transmission module is configured to transmit adjustment information including learning relevant information and parameter information of the second artificial neural network model to an external deep learning system to request adjustment of the first artificial neural network model, wherein the learning relevant information and parameter information are modified according to the learning result;

[0068] The adjustment module is configured to update the parameter information of the second artificial neural network model according to the adjustment information transmitted from the external deep learning system;

[0069] Therefore, the system of the present invention classifies the gait acquisition data of a real person into steering data and forward data. The steering data is used to change the orientation of the virtual person and regenerate a path, and the forward data is used to control the virtual person to move on the selected path. Therefore, a real person can turn freely during forward movement without being collected steering data, realizing controlling a virtual person to move in a larger virtual space in a smaller real space. The path planning module provides multiple suitable paths by analyzing the changes in terrain height in a three-dimensional scene, so that the provided paths can be closer to the paths that a real person wants to move;

[0070] Since the artificial neural network model performs partial self-optimization using parameters obtained from the learning results generated by an external system, it is actually impossible to recover or infer the source learning data, thereby minimizing the damage caused by information exposure;

[0071] Among them, without sharing the source learning data, the present invention can achieve the same effect as that obtained by learning the same learning data. Therefore, even if the artificial neural network model is in an environment where it can only learn a small amount of source data, the artificial neural network model can be changed into an artificial neural network model that has learned a large amount of source learning data.

[0072] The deep learning module further includes a diagnosis module configured to determine whether the optimal established route has been followed based on the first artificial neural network model optimized by the repeated operations of the learning module and the adjustment module.

[0073] The adjustment information transmission module preferentially transmits hyperparameters required to synchronize the initial learning state of the first artificial neural network model as the adjustment information.

[0074] The learning-related information includes either the ratio of learning data to source learning data to be learned or the number of learned source learning data segments, so as to improve the speed and efficiency of information learning, enabling the robotic dog to quickly respond to emergencies during use.

[0075] The first-level classification criterion of the data classification unit includes color values for identifying one or more in a normal image as classification criterion values, improving the recognition accuracy of the robotic dog for different environments and multiple color gamuts, expanding the information collection range of the robotic dog, and improving the learning efficiency of the robotic dog.

[0076] The adjustment information transmission module includes at least two or more of the first-level classification criterion, the second-level classification criterion, and the third-level classification criterion in the adjustment information, improving the classification criterion level of the robotic dog, so that the robotic dog can classify and organize the collected information data more meticulously. At the same time, it is also convenient for users to use and consult, and convenient for users to quickly find problems during the maintenance of the robotic dog, providing convenience for the user's maintenance work.

[0077] The adjustment information transmission module transmits the adjustment information to the external deep learning system, improving the autonomous learning ability of the robotic dog and the speed of new information collection and conversion applications, so as to ensure that the robotic dog can adapt to more and more complex working environments and provide convenience for the user's work use.

[0078] In step S32, multiple fork points are selected, and the path planning module generates multiple travel paths for selection.

[0079] It should be noted that the functions to be realized by each hardware in the present invention are supported by a large number of mature technologies and belong to the prior art. The essence of the present invention lies in optimizing the combination of existing hardware and its connection methods for a specific application scenario to meet the adaptation requirements in a specific application scenario and solve the problems raised in the background art (without involving improvements to the software inside the hardware).

[0080] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. An intelligent gait planning method for a robotic dog based on deep learning, characterized in that: It includes a gait acquisition module, a scene acquisition module, a virtual human positioning module, a path planning module, an execution control module, and a deep learning module. The gait acquisition module is used to acquire the pace actions of a living being (dog). The virtual human positioning module is used to determine and control the position of the virtual human in a three-dimensional scene. The scene acquisition module is used to obtain the three-dimensional model information between the virtual human and the target location. The path planning module determines multiple paths connecting the virtual human and the target location based on the three-dimensional model information. The real dog interacts with the system by using the execution control module; S31. Set a base line segment according to the position and orientation of the virtual human; S32. Select a bifurcation point on the base line segment; S33. Connect the bifurcation point and the target location to obtain a quasi-path line segment; S34. Determine whether the quasi-path line segment meets the slope requirement. If it meets, use the connection line between the virtual human and the bifurcation point and the quasi-path line segment as the travel path. If it does not meet, jump to step S35; S35. Find a change point on the quasi-path line segment, draw a perpendicular line to the quasi-path line segment through the change point, and find two temporary points A and B that meet the terrain height requirement on the perpendicular line. The temporary points are located on both sides of the quasi-path line segment; S36. Calculate the energy consumption index according to the corresponding formula; S36. Select the point with a lower energy consumption index as the path waypoint, and use the connection line between the virtual human and the bifurcation point, the connection line between the bifurcation point and the waypoint, and the connection line between the waypoint and the target location as the travel path, so as to form a motion process of the dynamic gait; When the steering switch is turned on, the gait acquisition module only acquires the steering data of the real person and determines the direction of the virtual human. The path planning module regenerates several travel paths. After one of the paths is selected, the steering switch automatically turns off. The gait acquisition module only acquires the forward gait data of the real person. The virtual human positioning module controls the virtual human to move forward on the selected path according to the forward gait data. The deep learning module includes being configured to be accessible through a communication network for enabling a first artificial neural network model to learn; A data classification and storage module, which is configured to receive and store source learning data; A learning module, which is configured to use the source learning data to enable a second artificial neural network model in the system to learn relevant information; An adjustment information transmission module, which is configured to transmit adjustment information including learning relevant information and parameter information of the second artificial neural network model to an external deep learning system to request adjustment of the first artificial neural network model, wherein the learning relevant information and parameter information are modified according to the learning result; An adjustment module, which is configured to update the parameter information of the second artificial neural network model according to the adjustment information transmitted from the external deep learning system.

2. The intelligent gait planning method for a robotic dog based on deep learning according to claim 1, characterized in that: The deep learning module further includes a diagnosis module, which is configured to judge whether it has walked according to the optimal established route based on the first artificial neural network model optimized by the repeated operations of the learning module and the adjustment module.

3. A method for intelligent gait planning of a robotic dog based on deep learning according to claim 1, characterized in that: The adjustment information transmission module preferentially transmits, as the adjustment information, hyperparameters required to synchronize the initial learning state of the first artificial neural network model.

4. A method for intelligent gait planning of a robotic dog based on deep learning according to claim 1, characterized in that: The learning-related information includes any one of the ratio of learning data to source learning data to be learned and the number of learned source learning data segments.

5. A method for intelligent gait planning of a robotic dog based on deep learning according to claim 1, characterized in that: The first-level classification criterion of the data classification unit includes color values for identifying one or more of normal images as classification criterion values.

6. A method for intelligent gait planning of a robotic dog based on deep learning according to claim 1, characterized in that: The adjustment information transmission module includes at least two or more of the first-level classification criterion, the second-level classification criterion, and the third-level classification criterion in the adjustment information.

7. A method for intelligent gait planning of a robotic dog based on deep learning according to claim 6, characterized in that: The adjustment information transmission module transmits the adjustment information to the external deep learning system.

8. A method for intelligent gait planning of a robotic dog based on deep learning according to claim 1, characterized in that: In step S32, multiple bifurcation points are selected, and the path planning module generates multiple travel paths for selection.

9. A method for intelligent gait planning of a robotic dog based on deep learning according to claim 1, characterized in that: The execution control module includes three basic functions: selecting a target location in the three-dimensional scene, selecting an execution path from the paths generated by the path planning module, and turning on the steering switch.

Images