An attractor-based data augmentation method and system

Abstract

This application provides an attractor-based data augmentation method and system, belonging to the field of artificial intelligence technology. The invention acquires action data; based on the action data, it acquires pattern actions; based on the pattern actions, it determines equilibrium points or limit cycles, wherein for non-repetitive actions, a series of equilibrium points are determined, and for repetitive pattern actions, limit cycles are determined; based on the equilibrium points or limit cycles, it constructs dynamic equations, wherein the dynamic equations are divided into tangential and normal directions, representing the evolution of augmented data sampling points along the tangential and normal directions of the pattern actions, respectively; based on the dynamic equations, it completes data augmentation. This invention can effectively perform data augmentation, providing more diverse data support for imitation learning and effectively improving the efficiency and effectiveness of the pre-training stage.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an attractor-based data augmentation method and system. Background Technology

[0002] In robot reinforcement learning, imitation learning can be used for pre-training to accelerate policy convergence. However, existing imitation learning techniques have the following limitations in terms of data augmentation:

[0003] Limited dataset size: The high cost of acquiring expert examples results in a limited dataset size, which cannot effectively cover the wide range of state spaces that the robot may encounter.

[0004] Insufficient data diversity: Existing imitation learning methods tend to rely on a single expert path, resulting in weak model generalization ability and difficulty in adapting to dynamic or unseen environments.

[0005] Overfitting problem: With only a few expert examples, the model is prone to overfitting to specific action sequences, which affects its performance in new environments.

[0006] Insufficient generalization ability: Existing imitation learning methods often can only adapt to specific environments and are difficult to effectively transfer to changing environments or multi-task scenarios.

[0007] Low learning efficiency: A large number of expert examples are needed to effectively learn complex strategies, especially in high-dimensional perceptual tasks (such as vision and speech), where the efficiency of imitation learning drops significantly.

[0008] Difficulty in acquiring expert data: Robot imitation learning relies on expert demonstrations, but acquiring large amounts of high-quality expert data is costly.

[0009] Insufficient exploration of the state space: The pre-training stage of traditional imitation learning often only covers a local area of ​​the state space, resulting in poor model performance when faced with new tasks.

[0010] Sample size limits generalization: The lack of sufficient and diverse example data limits the model's generalization ability. Summary of the Invention

[0011] The main objective of this application is to provide a data augmentation method and system based on attractors.

[0012] The technical solution adopted in this invention is:

[0013] On one hand, embodiments of the present invention provide an attractor-based data augmentation method, the method comprising the following steps:

[0014] Acquire motion data;

[0015] Based on the action data, obtain the pattern action;

[0016] Based on the described pattern of action, determine the equilibrium point or limit cycle;

[0017] Based on the equilibrium point or the limiting cycle, construct the dynamic equations;

[0018] Based on the aforementioned kinetic equations, data augmentation is performed.

[0019] Furthermore, the acquisition of motion data includes the following steps:

[0020] Motion capture equipment is used to collect the motion trajectory of a target object to obtain motion data; the motion capture equipment includes an optical tracking system, an inertial sensor or a depth camera.

[0021] Furthermore, obtaining the pattern action based on the action data includes the following steps:

[0022] From the motion data, multiple motion modes are selected based on robot kinematics and task requirements.

[0023] Further, determining the equilibrium point or limit cycle based on the said pattern action includes:

[0024] By integrating multiple described pattern actions, a target action sequence is generated, forming an equilibrium point or limit cycle.

[0025] Furthermore, constructing the dynamic equations based on the equilibrium point or the limiting cycle includes the following steps:

[0026] Determine the number of sampling points;

[0027] Determine the dimensions of the action space;

[0028] Based on the action space dimension and the number of augmented data sampling points, several current state points are obtained;

[0029] Calculate the distance between each current state point and each original data point on the limit cycle to obtain the original data point on the limit cycle that is closest to each current state point;

[0030] Based on the current state point and the nearest original data point on the limit cycle, the dynamic equation is obtained. Further, the data augmentation based on the dynamic equation includes the following steps:

[0031] The iteration step size is set based on the distance between the current state point and the nearest original data point on the distance limit cycle;

[0032] The current state point is updated based on the dynamic equation and the iteration step size to complete data augmentation.

[0033] On the other hand, embodiments of the present invention also provide an attractor-based data augmentation system, the system comprising:

[0034] The first module is used to acquire motion data;

[0035] The second module is used to obtain the pattern action based on the action data;

[0036] The third module is used to determine the equilibrium point or limit cycle based on the said pattern action;

[0037] The fourth module is used to construct the dynamic equations based on the equilibrium point or the limit cycle;

[0038] The fifth module is used to perform data augmentation based on the aforementioned kinetic equations.

[0039] Furthermore, the attractor-based data augmentation system described in this embodiment of the invention further includes:

[0040] An attractor construction module is used to construct the attractor based on the motion data;

[0041] The data augmentation module is used to sample in the action space and generate corresponding state trajectories;

[0042] The multi-attractor parallel mechanism module is used to generate corresponding state trajectories and action sequences by utilizing multiple attractor regions.

[0043] On the other hand, embodiments of the present invention also provide an attractor-based data augmentation apparatus, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the attractor-based data augmentation method as described above.

[0044] On the other hand, embodiments of the present invention also provide a computer-readable storage medium storing computer-executable instructions for causing a computer to perform the attractor-based data augmentation method as described above.

[0045] The embodiments of this application include at least the following beneficial effects: This application provides a data augmentation method and system based on attractors. The invention acquires action data; acquires pattern actions based on the action data; determines equilibrium points or limit cycles based on the pattern actions, wherein for non-repetitive actions, a series of equilibrium points are determined, and for repetitive pattern actions, limit cycles are determined; constructs dynamic equations based on the equilibrium points or limit cycles, wherein the dynamic equations are divided into tangential and normal directions, representing the evolution of action sampling points in the tangential and normal directions of the pattern actions, respectively; and completes data augmentation based on the dynamic equations. This invention can effectively perform data augmentation, providing more diverse data support for imitation learning and greatly improving the efficiency and effectiveness of the pre-training stage. Attached Figure Description

[0046] Figure 1 This is a flowchart of the attractor-based data augmentation method provided in an embodiment of the present invention;

[0047] Figure 2 This is a schematic diagram of the attractor construction process provided in an embodiment of the present invention;

[0048] Figure 3 This is an illustration of the attractor effect for each leg of the quadruped robot provided in this embodiment of the invention;

[0049] Figure 4 This is a schematic diagram of the data augmentation module process provided in an embodiment of the present invention;

[0050] Figure 5 This is a diagram illustrating the effect of iterating through 20 times the current state point of each leg of the quadruped robot provided in this embodiment of the invention.

[0051] Figure 6 This is a diagram showing the effect of iterating 100 times over 20 current state points of one leg of a quadruped robot according to an embodiment of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0053] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various concepts, but unless otherwise stated, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the words “if,” “when,” or “in response to a determination” as used herein may be interpreted as “when…” or “when…” or “in response to a determination.”

[0054] As used in this application, the terms "at least one", "multiple", "each", "any", etc., "at least one" includes one, two or more, "multiple" includes two or more, "each" refers to each of the corresponding multiples, and "any" refers to any one of the multiples.

[0055] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0056] This invention takes into account the widespread application of attractor theory in dynamical and chaotic systems, describing the behavior of system states tending towards a specific region over time. Attractor-based augmentation methods can effectively generate diverse samples, thereby providing more pre-training data for imitation learning and improving the model's generalization and learning efficiency.

[0057] This invention aims to address the shortcomings of existing imitation learning methods, such as insufficient data diversity, inadequate state space exploration, and insufficient generalization ability, through an attractor-based data augmentation method. By generating diverse samples through an attractor system, this method provides a richer expert dataset for the pre-training stage of robot imitation learning, thereby improving the model's generalization ability and learning efficiency.

[0058] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0059] On one hand, embodiments of the present invention provide an attractor-based data augmentation method, referring to... Figure 1 The method includes the following steps:

[0060] S100, Obtain motion data;

[0061] S200: Obtain the pattern action based on the action data;

[0062] S300: Determine the balance point or limit cycle based on the pattern action;

[0063] S400. Construct the dynamic equations based on the equilibrium point or limit cycle;

[0064] S500, Based on the dynamic equations, complete the data augmentation.

[0065] This invention discloses step S100 for acquiring motion data, including the following steps:

[0066] S110. Use motion capture equipment to collect the motion trajectory of the target object and obtain motion data; motion capture equipment includes optical tracking system, inertial sensor or depth camera.

[0067] This invention discloses step S200, which involves obtaining a pattern action based on action data, including the following steps:

[0068] S210. From the motion data, select multiple motion modes based on the robot's kinematics and task requirements.

[0069] This invention discloses step S300, which determines the equilibrium point or limit cycle based on the pattern action, including:

[0070] S310. Integrate multiple mode actions to generate a target action sequence, forming an equilibrium point or limit cycle.

[0071] This invention discloses step S400, which involves constructing a dynamic equation based on an equilibrium point or a limit cycle, including the following steps:

[0072] S410. Determine the number of sampling points;

[0073] S420, Determine the dimensions of the motion space;

[0074] S430. Based on the dimension of the action space and the number of action sampling points, several current state points are obtained;

[0075] S440. Calculate the distance between each current state point and each original data point on the limit cycle, and obtain the original data point on the limit cycle that is closest to each current state point.

[0076] S450. Based on the current state point and the nearest original data point on the limit cycle, the dynamic equation is obtained.

[0077] This invention discloses step S500, which involves data augmentation based on the kinetic equations, including the following steps:

[0078] S510. Set the iteration step size based on the distance between the current state point and the nearest original data point on the distance limit ring;

[0079] S520. Update the current state point according to the dynamic equation and iteration step size to complete data augmentation.

[0080] As an optional implementation, the present invention includes:

[0081] 1. Attractor building block

[0082] Acquiring motion data: Using motion capture devices (such as optical tracking systems, inertial sensors, or depth cameras) to collect the motion trajectory of the target object and record relevant data.

[0083] Determine the mode action: From the captured data, select multiple mode actions based on the robot's kinematics and task requirements to form an equilibrium point or limit cycle.

[0084] Determine the equilibrium points or limit cycles: For non-repetitive actions, determine a series of equilibrium points; for repetitive pattern actions, determine the limit cycles as the robot's final behavioral trajectory.

[0085] 2. Data Enlargement Module

[0086] Determine the parameters: set the number of action sampling points m and the maximum number of iterations t.

[0087] Randomly generate action space sampling points:

[0088] Randomly select action sampling point a1 from the entire action space, and the number of samples is m1.

[0089] Randomly select action sampling point a2 within three times the range of action data, and the number of samples is m2.

[0090] The original motion data selects motion sampling point a3, and the number of samples is m-m1-m2.

[0091] Find the nearest original data point on the limit cycle: Given n original data points, calculate the distance l between each current state point and each original data point on the limit cycle. i (i = 1, 2, ..., n), determine the nearest original data point p on the limit cycle.

[0092] Determine the dynamic equations:

[0093] Calculate the normal velocity q, with the current state point pointing in the direction of the nearest original data point on the limit cycle.

[0094] Calculate the tangential velocity t and the direction of the nearest original data point on the limit cycle.

[0095] The dynamic equation is the sum of the normal vector and the tangent vector.

[0096] Determine the step size: based on the distance l between the current state point and the nearest original data point on the limit cycle. min Set the iteration step size s.

[0097] Data augmentation generation:

[0098] The sampling points are updated based on the dynamic equation and the step size s.

[0099] Repeat the iterative process until the maximum number of iterations t is reached.

[0100] 3. Multi-attractor parallel mechanism

[0101] Simultaneously, multiple attractor regions are utilized to generate corresponding state trajectories and action sequences, ensuring coverage of various complex scenarios and enhancing the model's generalization ability.

[0102] On the other hand, embodiments of the present invention also provide an attractor-based data augmentation system, the system comprising:

[0103] The first module is used to acquire motion data;

[0104] The second module is used to obtain the pattern action based on the action data;

[0105] The third module is used to determine the equilibrium point or limit cycle based on the pattern action;

[0106] The fourth module is used to construct the dynamic equations based on the equilibrium point or limit cycle;

[0107] The fifth module is used to augment the data based on the dynamic equations.

[0108] The attractor-based data augmentation system provided in this embodiment of the invention further includes:

[0109] The attractor building module is used to build attractors based on motion data;

[0110] The data augmentation module is used to sample in the action space and generate corresponding state trajectories;

[0111] The multi-attractor parallel mechanism module is used to generate corresponding state trajectories and action sequences by utilizing multiple attractor regions.

[0112] As an optional implementation, this invention addresses the problems of poor generalization ability, low learning efficiency, and limited data scale in existing imitation learning techniques by proposing a data augmentation method based on attractor theory, suitable for the pre-training stage of robot imitation learning. By utilizing the dynamic characteristics of attractors to generate diverse state trajectories and action sequences, this invention can provide rich training data for imitation learning, thereby improving the model's performance in complex environments.

[0113] This invention first extracts representative data from expert demonstration data and transforms it into attractor regions. The attractors not only contain the expert's original trajectory but also generate more diverse approximate expert behaviors through perturbations or random sampling within the attractor boundaries. This attractor generation mechanism based on expert behavior enables broader state space exploration.

[0114] This system includes the following main modules:

[0115] Attractor building module: Builds attractors based on action data.

[0116] Data augmentation module: Samples some data in the action space and generates corresponding state trajectories.

[0117] Multi-attractor parallel mechanism: This invention utilizes multiple attractor systems to generate data in parallel, with each attractor region corresponding to a different robot behavior strategy. The multi-attractor parallel mechanism ensures that the generated data can cover a variety of complex scenarios that the robot may face, avoiding excessive reliance on a single trajectory in model pre-training.

[0118] refer to Figure 2 Attractor module building process:

[0119] Attractors are stable regions in a dynamical system to which the system state converges. This invention utilizes attractors to generate diverse samples, mainly including the following steps:

[0120] 1. Acquire motion data: Use motion capture devices (such as optical tracking systems, inertial sensors, or depth cameras) to capture the motion trajectory of the target object and record relevant data.

[0121] 2. Determine the pattern action: Select the pattern action from the motion data based on the robot's kinematics and task requirements.

[0122] 3. Determine equilibrium points or limit cycles: For non-repetitive actions, determine a series of equilibrium points; for repetitive pattern actions, determine limit cycles as the robot's final trajectory. (Reference) Figure 3 The attractor effect for each leg of the quadruped robot.

[0123] refer to Figure 4 Data augmentation module process:

[0124] 1. Determine the number of action sampling points and the maximum number of iterations.

[0125] 2. Randomly generate motion space sampling points: Randomly select motion sampling points in the entire motion space, and randomly select motion sampling points within three times the range of motion capture data. The remaining motion sampling points are the original motion data points, which are used as samples for this batch.

[0126] 3. Obtain the nearest original data point on the limit cycle: After obtaining the action sampling points, calculate the distance between each current state point and each original data point on the limit cycle to determine the nearest original data point on the limit cycle.

[0127] 4. Determine the dynamic equation of the current state point: The dynamic equation consists of two parts. The first part is the normal velocity, which is the velocity of the current state point toward the nearest original data point on the limit cycle. The second part is the tangential velocity, which is the velocity of the nearest original data point on the limit cycle itself.

[0128] 5. Determine the step size for each iteration: Determine the iteration step size based on the distance to the nearest original data point on the limit cycle.

[0129] 6. Attractor-based data augmentation generation: Based on the dynamic equations and iteration step size, perform one iteration to obtain a data augmentation, i.e., update the current state point. Repeat steps 4 and 5 until the maximum number of iterations is reached. (Reference) Figure 5 The image shows the effect of iterating through the current state point of each leg of a quadruped robot 20 times. Figure 6 The effect of iterating through 20 current state points of one leg of a quadruped robot 100 times.

[0130] As an optional implementation, the data augmentation module process of this embodiment of the invention is as follows:

[0131] 1. Determine the number of action sampling points m and the maximum number of iterations t.

[0132] 2. Randomly select initial sampling points in the action space: Randomly select action sampling point a1 from the entire action space, with a sampling count of m1. Randomly select initial sampling point a2 within three times the range of action data, with a sampling count of m2. The remaining action sampling points are the original action data a3, with a sampling count of m-m1-m2. In one embodiment, m1 and m2 can be m / 2 and m / 4, respectively. a1, a2, a3, etc., are used as samples in this batch.

[0133] 3. Obtain the nearest original data point on the limit cycle: The number of original data points is n. After obtaining the action sampling points, calculate the distance l between each current state point a (including a1, a2, a3, etc.) and each original data point on the limit cycle. i (i = 1, 2, ..., n), where the current state point is p, which is the closest original data point on the limit cycle, and the distance between the current state point and p is l. p .

[0134] 4. Determine the velocity of the current state point according to the dynamic equation: The dynamic equation consists of two parts. The first part is the normal velocity q, which is the velocity of the current state point toward the nearest original data point on the limit cycle. The second part is the tangential velocity t, which is the direction of the nearest original data point on the limit cycle itself.

[0135] 5. Calculate the new state point: Based on the dynamic equation and the iteration step size s, perform one iteration to obtain a data augmentation point a′, a′=a+(q+t)s, and update the current state point.

[0136] 6. Attractor-based data augmentation generation: Repeat steps 4 and 5 until the maximum number of iterations t is reached.

[0137] Advantages of this invention:

[0138] Significantly enhanced data diversity: The diverse samples generated based on attractors greatly improve the richness of the dataset and avoid overfitting to a single expert's behavior.

[0139] Improving the efficiency of imitation learning: Attractor-enhanced datasets enable imitation learning models to be exposed to more diverse behavioral patterns during the pre-training phase, thereby accelerating policy learning and optimization.

[0140] Enhanced generalization ability: By attracting the diversified generation of sub-regions, this invention effectively improves the generalization ability of the model in real-world tasks, enabling it to cope with changes in different environments and tasks.

[0141] Reduced data dependency: It reduces reliance on large-scale expert data and expands the coverage of the dataset through an attractor generation mechanism based on a small number of expert examples.

[0142] Key technologies of this invention:

[0143] Determine the normal velocity q and tangential velocity t in the dynamic equations.

[0144] The number of original data points is a scalar n, and the displacement of the current state point a from each original data point on the limit cycle is described as a vector d. i (i = 1, 2, ..., n), the distance from the current state point a to each original data point on the limit cycle is described by the scalar l. i (i = 1, 2, ..., n), where p is the closest original data point on the limit cycle to the current state point a, and the displacement of the current state point a from p is the vector d. p The distance is a scalar l p The velocity of each raw data point on the limit cycle is described by the vector v. i (i = 1, 2, ..., n), ||v i || represents the vector v iThe size of the 2-norm, with a maximum value of ||v|| max The velocity of p is a vector v p ,||v p || represents the vector v p The size of the 2-norm, v p ′ is v p The unit vector, i.e., v p ′=v p / ||v p In one embodiment, the distance threshold l t For scalars, l t =60||v p ||, where constant c is a scalar, c = 0.001, constant factor μ is a scalar, μ = 2000, coefficient k is a scalar, k = 10, and weight ω i ω is a scalar i =||v i || / ||v|| max (i = 1, 2, ..., n).

[0145] The normal velocity q is the normal speed u multiplied by the normal unit vector n′, i.e., q = un′. Where:

[0146] The normal velocity u is a constant factor μ divided by the ω corresponding to each equilibrium point. i / l i The sum, divided by the number of equilibrium points n, is...

[0147] The method for calculating the normal unit vector n′ is the same as l. p It depends on the size.

[0148] When l p Greater than or equal to l t hour:

[0149] First, calculate the attraction g of the i-th equilibrium point on the current state point a. i The steps are weights ω i Multiply by the displacement d from the i-th equilibrium point to the next equilibrium point i Then divide by the distance l from the current state point to the i-th equilibrium point. i The square of g i =ω i d i / l i 2 The normal unit vector n′ is the normalized sum of the attractive forces at each point, i.e.

[0150] When l p Less than l t hour:

[0151] d p In v p The projection on is j p That is, j p =d p v p The normal unit vector n′ is d. p -j p Normalization after that, i.e., n′=(d p -j p ) / ||d p -j p ||.

[0152] The calculation of tangential velocity q needs to take into account the distance of the current state point a from each equilibrium point and the velocity of each equilibrium point.

[0153] drag force τ i The velocity v at the i-th equilibrium point i Divide by the square of the distance from the current state point to the equilibrium point, and then divide by the kl corresponding to the equilibrium point. i +1, i.e., τ i =v i / (l i 2 (kl i +1)). The total drag force τ is the sum of the drag components, i.e. The equivalent distance r is 1 / l corresponding to each equilibrium point. i 2 The sum of Drag speed v t It is the total drag force divided by the equivalent distance, i.e., v t =τ / r.

[0154] When l p Greater than c and greater than l t hour:

[0155] The tangential velocity is the drag velocity, i.e., t = v. t

[0156] When l p Greater than c and l p Less than l t hour:

[0157] Tangential velocity is the drag velocity v t The speed v at the nearest point p The component of velocity. Dragging force divided by distance in v p The size of the projection on the surface is b, that is, b = v. t v p The tangential velocity t is b multiplied by v. p unit vector v p ′, that is, t=bvp ′.

[0158] When l p When less than c:

[0159] The tangential velocity is the velocity v at the closest point. p t = v p .

[0160] Protected points of this invention:

[0161] Pattern action extraction and equilibrium point or limit cycle generation: Extract pattern actions from action data and construct equilibrium points or limit cycles.

[0162] Multi-attractor parallel generation mechanism: Data is generated in parallel in multiple attractor regions, which improves data generation efficiency and the coverage of the state space.

[0163] Modular Data Augmentation System: A modular data augmentation system was designed, including sampling point generation and data augmentation modules, to ensure efficient system operation and enhance learning effects and the pre-training dataset for imitation learning.

[0164] Calculation of dynamic equations: including the calculation of the normal vector direction and the tangent vector direction.

[0165] This invention proposes a data augmentation method based on attractor theory, which is particularly suitable for the pre-training stage of robot imitation learning. By generating diverse samples, the model can learn policies more effectively, avoid overfitting, and improve generalization ability. This method can effectively augment data, providing more diverse data support for imitation learning and significantly improving the efficiency and effectiveness of the pre-training stage.

[0166] On the other hand, embodiments of the present invention also provide an attractor-based data augmentation apparatus, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the attractor-based data augmentation method as described above.

[0167] The attractor-based data augmentation device according to embodiments of the present invention includes a memory and a processor.

[0168] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0169] Memory can include various types of storage units, such as system memory, read-only memory (ROM), and permanent storage devices. ROM can store static data or instructions required by the processor or other modules of the computer. Permanent storage devices can be read-write storage devices. Permanent storage devices can be non-volatile storage devices that retain stored instructions and data even when the computer is powered off. In some embodiments, permanent storage devices use high-capacity storage devices (e.g., magnetic or optical disks, flash memory) as permanent storage devices. In other embodiments, permanent storage devices can be removable storage devices (e.g., floppy disks, optical drives). System memory can be a read-write storage device or a volatile read-write storage device, such as dynamic random access memory. System memory can store some or all of the instructions and data required by the processor during operation. Furthermore, memory can include any combination of computer-readable storage media, including various types of semiconductor memory chips (DRAM, SRAM, SDRAM, flash memory, programmable read-only memory), and disks and / or optical disks can also be used. In some implementations, the memory may include removable storage devices that are readable and / or writable, such as laser discs (CDs), read-only digital versatile optical discs (e.g., DVD-ROMs, dual-layer DVD-ROMs), read-only Blu-ray discs, ultra-high density optical discs, flash memory cards (e.g., SD cards, mini SD cards, Micro-SD cards, etc.), magnetic floppy disks, etc. Computer-readable storage media do not contain carrier waves or transient electronic signals transmitted wirelessly or via wired connections.

[0170] The memory stores executable code, which, when processed by the processor, can cause the processor to execute some or all of the methods described above.

[0171] On the other hand, embodiments of the present invention also provide a computer-readable storage medium storing computer-executable instructions for causing a computer to perform the attractor-based data augmentation method described above.

[0172] It will be understood by those skilled in the art that all or some of the steps and systems in the methods disclosed above can be implemented as software, firmware, hardware, and suitable combinations thereof. Some or all of the physical components can be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, which can include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, as is known to those skilled in the art, communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0173] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. An attractor-based data augmentation method, characterized in that, The method includes the following steps: Acquire motion data; Based on the action data, obtain the pattern action; Based on the described pattern of action, determine the equilibrium point or limit cycle in the action space; Based on the equilibrium point or the limit cycle, construct a dynamic equation in the action space to attract the state point in the space to approach the equilibrium point or the limit cycle; while attracting the state point to approach the limit cycle, the dynamic equation drags the state point so that its velocity is close to the velocity of the state point on the limit cycle. Based on the aforementioned kinetic equations, data augmentation is performed.

2. The method according to claim 1, characterized in that, The acquisition of action data includes the following steps: Motion capture equipment is used to collect the motion trajectory of a target object to obtain motion data; the motion capture equipment includes an optical tracking system, an inertial sensor or a depth camera.

3. The method according to claim 1, characterized in that, The step of obtaining the pattern action based on the action data includes the following steps: From the motion data, multiple motion modes are selected based on robot kinematics and task requirements.

4. The method according to claim 1, characterized in that, The step of determining the equilibrium point or limit cycle based on the described action pattern includes: By integrating multiple described pattern actions, a target action sequence is generated, forming an equilibrium point or limit cycle.

5. The method according to claim 1, characterized in that, The process of constructing the dynamic equations based on the equilibrium point or the limiting cycle includes the following steps: Determine the number of action sampling points; Determine the dimensions of the action space; Based on the action space dimension and the number of action sampling points, several current state points are obtained; Calculate the distance between each current state point and each original data point on the limit cycle to obtain the original data point on the limit cycle that is closest to each current state point; The dynamic equations are obtained based on the current state point and the nearest original data point on the limit cycle.

6. The method according to claim 5, characterized in that, The data augmentation based on the dynamic equation includes the following steps: The iteration step size is set based on the distance between the current state point and the nearest original data point on the distance limit cycle; The current state point is updated based on the dynamic equation and the iteration step size to complete data augmentation.

7. An attractor-based data augmentation system for implementing the attractor-based data augmentation method as described in any one of claims 1 to 6, characterized in that, The system includes: The first module is used to acquire motion data; The second module is used to obtain the pattern action based on the action data; The third module is used to determine the equilibrium point or limit cycle based on the said pattern action; The fourth module is used to construct the dynamic equations based on the equilibrium point or the limit cycle; The fifth module is used to perform data augmentation based on the aforementioned kinetic equations.

8. The system according to claim 7, characterized in that, The system also includes: An attractor construction module is used to construct the attractor based on the motion data; The data augmentation module is used to sample in the action space and generate corresponding state trajectories; The multi-attractor parallel mechanism module is used to generate corresponding state trajectories and action sequences by utilizing multiple attractor regions.

9. An attractor-based data augmentation device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the attractor-based data augmentation method as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions for causing a computer to perform the attractor-based data augmentation method as described in any one of claims 1 to 6.

Images