A method and system for evaluating the motion ability of a virtual intelligent body based on terrain generation

Abstract

The present invention provides a method and system for evaluating the motion ability of a virtual intelligent body based on terrain generation, wherein the method includes: generating a target motion trajectory based on configured initial parameters and a motion trajectory generation model; generating a test terrain based on the target motion trajectory and the terrain generation model; driving the virtual intelligent body to move based on the test terrain to obtain an actual motion trajectory; and generating virtual intelligent body motion ability evaluation data based on the difference between the actual motion trajectory and the target motion trajectory. The present invention can provide a variety of test scenarios for the virtual intelligent body through the terrain generation method, and more comprehensively reflect the motion ability performance of the virtual intelligent body under various conditions.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and system for evaluating the movement ability of a virtual intelligent body based on terrain generation. Background Art

[0002] The evaluation of the motor ability of virtual agents is of great significance in the fields of artificial intelligence and robotics. With the development of virtual reality, augmented reality, and robotic simulation technology, the performance of virtual agents in virtual environments has been widely used in various scenarios such as games, training simulations, and autonomous driving tests. In order to ensure that these virtual agents have sufficient adaptability and motor flexibility in a variety of environments and tasks, it is particularly important to conduct a systematic motor ability evaluation. Motor ability evaluation can not only help researchers understand the motor skill performance of virtual agents, but also provide data support for optimizing and improving the motion algorithms of virtual agents. Most existing technologies focus on functional testing of virtual agents or performance testing under specific working conditions, and fail to comprehensively evaluate the performance of virtual agents under various working environment conditions. Summary of the invention

[0003] One of the purposes of the present invention is to provide a method and system for evaluating the athletic ability of a virtual intelligent body based on terrain generation. Through the terrain generation method, a variety of test scenarios can be provided for the virtual intelligent body, which can more comprehensively reflect the athletic ability performance of the virtual intelligent body under various terrain conditions.

[0004] An embodiment of the present invention provides a method for evaluating the motion ability of a virtual intelligent body based on terrain generation, comprising:

[0005] Generate the target motion trajectory according to the configured initial parameters and motion trajectory generation model;

[0006] Generate test terrain based on target motion trajectory and terrain generation model;

[0007] Drive the virtual agent to move based on the test terrain and obtain the actual movement trajectory;

[0008] Based on the difference between the actual motion trajectory and the target motion trajectory, the virtual agent's motion ability evaluation data is generated.

[0009] Preferably, the initial parameters of the configuration include: setting the initial difficulty , initial position , end point position , Termination judgment conditions , maximum number of test rounds max_epoch, difficulty update step and , The map width corresponding to the preset test terrain.

[0010] Preferably, the training steps of the motion trajectory generation model are as follows:

[0011] Random sampling training initial position and the end position ,in is a fixed normal value; and the length of random sampling The target sampling trajectory ;

[0012] Calculate target sampling trajectory The mean curvature , and define it as the difficulty of the sampling trajectory;

[0013] Building a trajectory generation neural network , enter the initial position , end point position and difficulty , the output is of length Trajectory , with the target sampling trajectory And the actual generated trajectory The error between them is the loss function, and the trajectory is optimized to generate the neural network ; Obtain trajectory generation model ;

[0014] The loss function is , where To actually generate the trajectory exist The coordinates of the point; Sample trajectory for the target exist The coordinates of the point.

[0015] Preferably, the average curvature The calculation process is:

[0016] Compute the vector between each pair of consecutive points: ;

[0017] Compute the angle between adjacent vectors: ;

[0018] The calculation trajectory is located at The local curvature of the point; ;

[0019] Calculate the overall average curvature of the trajectory, .

[0020] Preferably, the training steps of the terrain generation model are as follows:

[0021] Random sampling training initial position and the end position ,in is a fixed normal value; and the length of random sampling The target sampling trajectory ;

[0022] Building a Neural Network , the input is the optimal trajectory of the target , the output is The matrix ,in is the preset map width;

[0023] Matrix Each element is the height of the corresponding position point, and the generated length and width are , and calculate the optimal trajectory from the initial position to the final position under this map ,The calculation process can use the traditional path planning algorithm;

[0024] Calculate the optimal trajectory under the current map And the target sampling trajectory The error between , optimize the terrain generation network , where Optimal trajectory exist The coordinates of the point; Sample trajectory for the target exist The coordinates of the point;

[0025] Repeat the above steps for a preset number of repetitions (the value configured by max_epoch) to obtain the converged terrain generation model .

[0026] Preferably, the difference between the actual motion trajectory and the target motion trajectory includes: a normalized error accumulation sum; the calculation formula of the normalized error accumulation sum is as follows:

[0027] ;

[0028] In the formula, represents the normalized error cumulative sum; Indicates the actual motion trajectory The position corresponding to the time, Represents the target motion trajectory The position corresponding to the moment.

[0029] Preferably, after driving the virtual intelligent body to move based on the test terrain and obtaining the actual movement trajectory, the initial parameters are updated and the virtual intelligent body is driven again; wherein updating the initial parameters includes:

[0030] renew , , and randomly sample the initial position of the letter To the end position To complete the update;

[0031] The termination conditions for terminating re-driving include:

[0032] The total number of rounds exceeds the setting value of the parameter max_epoch or .

[0033] Preferably, the virtual agent athletic ability rating data includes a virtual agent athletic ability evaluation index, and the virtual agent athletic ability evaluation index is .

[0034] Preferably, the method for evaluating the motion ability of a virtual agent based on terrain generation further includes:

[0035] Simulate the test terrain based on the terrain simulation sandbox to obtain the simulated terrain;

[0036] loading the virtual agent onto a test model pre-configured to a point on the simulated terrain corresponding to the initial position;

[0037] The test model is driven by a virtual agent to run the trajectory and obtain a simulated trajectory;

[0038] The normalized cumulative error sum of the simulation trajectory and the target motion trajectory is calculated, and the normalized cumulative error sum of the simulation trajectory and the target motion trajectory is processed to obtain simulation evaluation data.

[0039] The present invention also provides a virtual intelligent body movement ability evaluation system based on terrain generation, comprising:

[0040] Target motion trajectory generation module, test terrain generation module, driving module and evaluation module;

[0041] Among them, the target motion trajectory generation module generates the target motion trajectory according to the configured initial parameters and the motion trajectory generation model; the test terrain generation module generates the test terrain based on the target motion trajectory and the terrain generation model; the driving module drives the virtual intelligent body to move based on the test terrain to obtain the actual motion trajectory; the evaluation module generates the virtual intelligent body's motion ability evaluation data based on the difference between the actual motion trajectory and the target motion trajectory.

[0042] The present invention has the following advantages:

[0043] The present invention can automatically generate the target motion trajectory of the virtual agent and the corresponding test terrain, and iteratively update the test scene according to the actual performance of the virtual agent on the test terrain. The motion ability of the virtual agent is measured by the final convergent motion performance of the virtual agent, providing guidance for the development of the motion ability of the virtual agent.

[0044] 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 written description and the accompanying drawings.

[0045] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The accompanying drawings are used to provide a further understanding 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 present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0047] Figure 1 A schematic diagram of a method for evaluating the motion ability of a virtual intelligent body based on terrain generation in an embodiment of the present invention;

[0048] Figure 2 This is a diagram of the evaluation steps in the embodiment of the present invention;

[0049] Figure 3 Schematic diagram of a virtual intelligent body motion ability evaluation system based on terrain generation in an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0051] The embodiment of the present invention provides a method for evaluating the motion ability of a virtual intelligent body based on terrain generation, such as Figure 1 As shown, including:

[0052] Generate the target motion trajectory according to the configured initial parameters and motion trajectory generation model;

[0053] Generate test terrain based on target motion trajectory and terrain generation model;

[0054] Drive the virtual agent to move based on the test terrain and obtain the actual movement trajectory;

[0055] Based on the difference between the actual motion trajectory and the target motion trajectory, the virtual agent's motion ability evaluation data is generated.

[0056] like Figure 2 As shown, a specific implementation step of the present invention is as follows:

[0057] Step 1: Load the virtual agent, motion trajectory generation model and terrain generation model, and configure the initial parameters;

[0058] Among them, the motion trajectory generation model and terrain generation module are trained and converged in advance. The initial parameters are configured as follows: Setting the initial difficulty , initial position , end point position , Termination judgment conditions , maximum number of test rounds max_epoch, difficulty update step and .

[0059] Step 2: Generate a target motion trajectory based on the initial parameters and motion trajectory generation model;

[0060] According to difficulty , initial position , end point position , using trajectory generation model Generate a virtual agent from the initial position To the end position Target motion trajectory ;

[0061] Step 3: Generate test terrain based on the target motion trajectory and terrain generation model;

[0062] Generate models using terrain Based on the target motion trajectory Generates a test terrain, which is represented as a matrix ; With matrix Build test terrain;

[0063] Step 4: Drive the virtual agent to obtain the actual motion trajectory based on the test terrain;

[0064] The virtual agent uses its own strategy to complete the To the end position movement, get its Actual movement trajectory of time length ;

[0065] Step 5: Calculate the normalized error accumulation sum between the target motion trajectory and the actual motion trajectory;

[0066] Calculate target trajectory The actual motion trajectory The normalized error accumulation sum is calculated as follows:

[0067] ;

[0068] In the formula, Indicates the actual motion trajectory The position corresponding to the time, Represents the target motion trajectory The position corresponding to the moment;

[0069] Step 6: Update the initial parameters and repeat steps 2 to 5 until the preset conditions are met;

[0070] renew , , and randomly sample the initial position of the letter To the end position To complete the update; after the update, repeat steps 2 to 5 until the total number of rounds exceeds the set value of the parameter max_epoch or .

[0071] Step 7: Process the normalized error accumulation and obtain the virtual agent's motion ability evaluation data.

[0072] The virtual agent motion ability rating data includes the virtual agent motion ability evaluation index, which is ; The higher the index is (closer to 1), the stronger the motor ability of the intelligent agent is; conversely, the lower the index is (closer to 0), the weaker the motor ability of the intelligent agent is.

[0073] In one embodiment, the trajectory generation module used in the present invention generates trajectories based on a given target difficulty, and the training steps of the motion trajectory generation model are as follows:

[0074] Random sampling training initial position and the end position ,in is a fixed normal value. And randomly sample the length The target sampling trajectory ;

[0075] Calculate target sampling trajectory The mean curvature , and define it as the difficulty of the sampling trajectory. The specific process is: calculate the vector between each pair of consecutive points: ; Calculate the angle between adjacent vectors: ; The calculation trajectory is located at The local curvature of the point; ; Calculate the overall average curvature of the trajectory, ;

[0076] Construct a motion trajectory generation neural network (you can use a 3-hidden-layer MLP network structure) , enter the initial position , end point position and difficulty , the output is of length The trajectory , with the target sampling trajectory And the actual generated trajectory The error between them is the loss function, and the trajectory is optimized to generate the neural network ; Obtain trajectory generation model ;

[0077] The loss function is , where To actually generate the trajectory exist The coordinates of the point; Sample trajectory for the target exist The coordinates of the point.

[0078] In one embodiment, the terrain generation model generates terrain based on the optimal estimate of a given target; the training steps of the terrain generation model are as follows:

[0079] Random sampling training initial position and the end position ,in is a fixed normal value. And randomly sample the length The target sampling trajectory ;

[0080] Construct a neural network (you can use an MLP network structure with 3 hidden layers) , the input is the optimal trajectory of the target , the output is The matrix ,in The preset map width.

[0081] Matrix Each element is the height of the corresponding position point, and the generated length and width are , and calculate the optimal trajectory from the initial position to the final position under this map The calculation process can use the traditional path planning algorithm, which includes: Dijkstra algorithm, A* (A-Star) algorithm, D* (Dynamic A*) algorithm, RRT (Rapidly-exploring Random Tree) algorithm, artificial potential field method.

[0082] Calculate the optimal trajectory under the current map And the target sampling trajectory The error between , optimize the terrain generation network , where Optimal trajectory exist The coordinates of the point; Sample trajectory for the target exist The coordinates of the point;

[0083] Repeat the above steps for a preset number of repetitions (the value configured by max_epoch) to obtain the converged terrain generation model .

[0084] In order to achieve a more comprehensive athletic ability assessment, in one embodiment, the method for assessing the athletic ability of a virtual agent based on terrain generation further includes:

[0085] Simulate the test terrain based on the terrain simulation sandbox to obtain the simulated terrain;

[0086] loading the virtual agent onto a test model pre-configured to a point on the simulated terrain corresponding to the initial position;

[0087] The test model is driven by a virtual agent to run the trajectory and obtain a simulated trajectory;

[0088] The normalized cumulative error sum of the simulation trajectory and the target motion trajectory is calculated, and the normalized cumulative error sum of the simulation trajectory and the target motion trajectory is processed to obtain simulation evaluation data.

[0089] In this embodiment, the test terrain generated by the terrain generation model is simulated through a terrain simulation sandbox, and the virtual intelligent body drives the test model to perform actual simulation, and the simulated trajectory and the target motion trajectory are used to perform actual simulation evaluation; the processing method for the normalized error accumulation sum of the simulation evaluation data and the virtual intelligent body motion ability evaluation data is consistent, which will not be explained in detail here.

[0090] The present invention also provides a virtual intelligent body movement ability evaluation system based on terrain generation, such as Figure 3 As shown, it includes: a target motion trajectory generation module 1, a test terrain generation module 2, a driving module 3 and an evaluation module 4;

[0091] Among them, the driving module 3 first loads the virtual intelligent body, the motion trajectory generation model and the terrain generation model, and configures the initial parameters; the target motion trajectory generation module 1 generates the target motion trajectory according to the configured initial parameters and the motion trajectory generation model; the test terrain generation module 2 generates the test terrain based on the target motion trajectory and the terrain generation model; the driving module 3 drives the virtual intelligent body to obtain the actual motion trajectory based on the test terrain; after driving the virtual intelligent body to move based on the test terrain and obtaining the actual motion trajectory, the initial parameters are updated and the driving is re-performed; wherein updating the initial parameters includes:

[0092] renew , , and randomly sample the initial position of the letter To the end position To complete the update;

[0093] The termination conditions for terminating re-driving include:

[0094] The total number of rounds exceeds the setting value of the parameter max_epoch or Repeat steps 2 to 5 until the preset conditions are met; the evaluation module 4 generates virtual intelligent body motion ability evaluation data based on the difference between the actual motion trajectory and the target motion trajectory; the difference between the actual motion trajectory and the target motion trajectory includes: the normalized error accumulation sum; based on the error between the actual motion trajectory, the target motion trajectory and each position on the target motion trajectory, calculate the normalized error accumulation sum.

[0095] In order to achieve a more comprehensive evaluation of athletic ability, in one embodiment, the system for evaluating the athletic ability of a virtual agent based on terrain generation further includes: a real-time simulation module;

[0096] Among them, the live simulation module performs the following operations:

[0097] Simulate the test terrain based on the terrain simulation sandbox to obtain the simulated terrain;

[0098] loading the virtual agent onto a test model pre-configured to a point on the simulated terrain corresponding to the initial position;

[0099] The test model is driven by a virtual agent to run the trajectory and obtain a simulated trajectory;

[0100] The normalized cumulative error sum of the simulation trajectory and the target motion trajectory is calculated, and the normalized cumulative error sum of the simulation trajectory and the target motion trajectory is processed to obtain simulation evaluation data.

[0101] Among them, the terrain simulation sand table includes: a sand table main body, a gantry arranged above the sand table main body, an image acquisition module and a supplementary material suction and distribution mechanism arranged on the gantry, and a mechanical claw arranged on the gantry; the supplementary material suction and distribution mechanism includes: a discharge port for placing supplementary materials (fine sand) on the sand table main body and a vacuum suction port for sucking supplementary materials from the sand table main body.

[0102] The test model includes: an automatic navigation vehicle in communication connection with the live simulation module;

[0103] The main body of the sand table includes: array-arranged and mutually interlocked lattice bodies, the lattice bodies can slide up and down with each other, and a lifting mechanism is arranged under the lattice bodies; the lifting mechanism controls the rise and fall of the lattice bodies to simulate the height of each point on the test terrain; after being lifted and lowered into place, the test terrain is screened for points to determine the uniform distribution points and uniform distribution areas of the supplementary materials; according to the parameters of each point in the uniform distribution area, an analysis parameter set is constructed according to the regional position, and then the control parameters of the supplementary material absorption uniform distribution mechanism are retrieved from the pre-configured uniform distribution control library according to the analysis parameter set, and the control parameters are used to control the supplementary material absorption uniform distribution mechanism to distribute the supplementary materials uniformly at the supplementary material uniform distribution points;

[0104] Among them, the uniform distribution points of the supplementary materials are points in the test terrain that are higher than the surrounding points, and the uniform distribution area is a closed area composed of points whose heights continue to decrease after the uniform distribution points of the supplementary materials radiate outwards; the uniform distribution control library is a pre-configured storage, in which the analysis parameter set and the control parameter set are associated one by one;

[0105] Since the uniform distribution area is not necessarily a regular shape (such as a circular area), the effect cannot be achieved by a single uniform distribution. The image after the single uniform distribution is collected by the image acquisition module (infrared laser scanning can also be used), and modeling is performed based on the image to obtain a three-dimensional model; and the three-dimensional model is compared with the model corresponding to the test terrain to determine the secondary uniform distribution point and the secondary uniform distribution area; then a secondary analysis set is constructed based on the secondary uniform distribution area, and the control parameters of the supplementary material absorption uniform distribution mechanism are retrieved from the pre-configured secondary uniform distribution control library, and then a secondary uniform distribution operation is performed. Among them, the three-dimensional model is compared with the model corresponding to the test terrain to determine the secondary uniform distribution point and the secondary uniform distribution area, including: sampling the points of the three-dimensional model and the model corresponding to the test terrain respectively, determining the height difference of each corresponding sampling point, and taking the sampling point with the largest height difference as the secondary uniform distribution point; the secondary uniform distribution area is a closed area composed of points with continuously reduced height after traversing outward from the secondary uniform distribution point; in order to further improve the accuracy, three times of uniform distribution can be continued until the determined uniform distribution area is less than or equal to the preset area threshold.

[0106] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for evaluating the motion ability of a virtual agent based on terrain generation, characterized in that: include: Generate the target motion trajectory according to the configured initial parameters and motion trajectory generation model; Generate test terrain based on target motion trajectory and terrain generation model; Drive the virtual agent to move based on the test terrain and obtain the actual movement trajectory; Generate virtual agent motion ability evaluation data based on the difference between the actual motion trajectory and the target motion trajectory; Among them, the training steps of the motion trajectory generation model are as follows: Random sampling training initial position and the end position ,in is a fixed normal value; and the length of random sampling The target sampling trajectory ; Calculate target sampling trajectory The mean curvature , and define it as the difficulty of the sampling trajectory; Constructing a Neural Network for Motion Trajectory Generation , enter the initial position , end point position and difficulty , the output is of length The trajectory , with the target sampling trajectory And the actual generated trajectory The error and loss function are used to optimize the motion trajectory to generate a neural network. ; Obtain trajectory generation model ; The loss function is , where To actually generate the trajectory exist The coordinates of the point; Sample trajectory for the target exist The coordinates of the point.

2. The method for evaluating the motion ability of a virtual agent based on terrain generation according to claim 1, characterized in that: The initial parameters of the configuration include: setting the initial difficulty , initial position , end point position , Termination judgment conditions , maximum number of test rounds max_epoch, difficulty update step and , The preset map width.

3. The method for evaluating the motion ability of a virtual agent based on terrain generation according to claim 1, characterized in that: Mean curvature The calculation process is: Compute the vector between each pair of consecutive points; Calculate the angle between adjacent vectors; The calculation trajectory is located at The local curvature of the point; Calculate the overall mean curvature of the trajectory.

4. The method for evaluating the motion ability of a virtual agent based on terrain generation according to claim 1, characterized in that: The training steps of the terrain generation model are as follows: Random sampling training initial position and the end position ,in is a fixed normal value; and the length of random sampling The target sampling trajectory ; Building a terrain generation network , the input is the optimal trajectory of the target , the output is The matrix ,in is the preset map width; Matrix Each element is the height of the corresponding position point, and the generated length and width are , and calculate the optimal trajectory from the initial position to the final position under this map ,The calculation process uses the traditional path planning algorithm; Calculate the optimal trajectory under the current map And the target sampling trajectory The error between , optimize the terrain generation network , where Optimal trajectory exist The coordinates of the point; Sample trajectory for the target exist The coordinates of the point; Repeat the above training steps for a preset number of times to obtain a converged terrain generation model. .

5. The method for evaluating the motion ability of a virtual agent based on terrain generation according to claim 1, characterized in that: The difference between the actual motion trajectory and the target motion trajectory includes: a normalized error accumulation sum; and calculating the normalized error accumulation sum according to the error between each position on the actual motion trajectory and the target motion trajectory.

6. The method for evaluating the motion ability of a virtual agent based on terrain generation according to claim 5, characterized in that: After driving the virtual agent to move based on the test terrain and obtaining the actual movement trajectory, the initial parameters are updated and the agent is driven again; The updating of initial parameters includes: renew , , and randomly sample the initial position of the letter To the end position To complete the update; The termination conditions for terminating re-driving include: The total number of rounds exceeds the setting value of the parameter max_epoch or .

7. The method for evaluating the motion ability of a virtual agent based on terrain generation according to claim 6, characterized in that: The virtual agent motion ability rating data includes virtual agent motion ability evaluation indicators. The virtual agent motion ability evaluation indicators are: .

8. The method for evaluating the motion ability of a virtual agent based on terrain generation according to claim 1, characterized in that: Also includes: Simulate the test terrain based on the terrain simulation sandbox to obtain the simulated terrain; loading the virtual agent onto a test model pre-configured to a point on the simulated terrain corresponding to the initial position; The test model is driven by a virtual agent to run the trajectory and obtain a simulated trajectory; The normalized cumulative error sum of the simulation trajectory and the target motion trajectory is calculated, and the normalized cumulative error sum of the simulation trajectory and the target motion trajectory is processed to obtain simulation evaluation data.

9. A virtual agent motion ability evaluation system based on terrain generation, applied to the virtual agent motion ability evaluation method based on terrain generation as claimed in any one of claims 1 to 8, characterized in that: include: Target motion trajectory generation module, test terrain generation module, driving module and evaluation module; Among them, the target motion trajectory generation module generates the target motion trajectory according to the configured initial parameters and the motion trajectory generation model; the test terrain generation module generates the test terrain based on the target motion trajectory and the terrain generation model; the driving module drives the virtual intelligent body to move based on the test terrain to obtain the actual motion trajectory; the evaluation module generates the virtual intelligent body's motion ability evaluation data based on the difference between the actual motion trajectory and the target motion trajectory.

Images