Robot control method, apparatus, device, storage medium, and computer program product

Abstract

The application discloses a robot control method, device, equipment, storage medium and computer program product, relates to the technical field of artificial intelligence, and the method comprises the following steps: in response to a control instruction, a robot is controlled through a pre-constructed environment dynamics adaptive model, the environment dynamics adaptive model is obtained through training based on an adversarial continuous dynamics domain alignment reinforcement learning framework, and the application considers that the dynamics parameter is a continuous value, and the robot is controlled by extracting the law in continuous change of the dynamics parameter.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a robot control method, apparatus, device, storage medium, and computer program product. Background Technology

[0002] Because simulator training is more efficient and safer, robot control tasks are typically trained in physical simulators rather than on real hardware when performing deep reinforcement learning on robots. However, physical simulators struggle to simulate real physical environments and complex dynamic changes, and the resulting domain differences can negatively impact the performance of robot control decisions in the real world.

[0003] Existing research only defines the target environment as an environment with specific dynamic parameters, but fails to consider that the dynamic parameters are continuous values ​​and does not attempt to extract the patterns in the continuous changes of dynamic parameters.

[0004] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this application is to provide a robot control method, device, equipment, storage medium, and computer program product, aiming to solve the technical problem that existing research only defines the target environment as an environment with specific dynamic parameters, but does not take into account that the dynamic parameters are continuous values ​​and does not attempt to extract the regularity of the continuous change of dynamic parameters.

[0006] To achieve the above objectives, this application proposes a robot control method, the method comprising:

[0007] In response to control commands, the robot is controlled by a pre-built environmental dynamics adaptation model, which is trained based on an adversarial continuous dynamics domain alignment reinforcement learning framework.

[0008] In one embodiment, the step of controlling the robot via a pre-built environmental dynamics adaptation model in response to control commands includes:

[0009] Trajectories of robots controlled by pre-built teacher reinforcement learning models in the source and target domains are collected separately and denoted as the source domain trajectory dataset and the target domain trajectory dataset.

[0010] Trajectory data are selected as support sets from the source domain trajectory dataset and the target domain trajectory dataset, respectively;

[0011] Select trajectory data from the source domain trajectory dataset as a query set;

[0012] The discriminator loss is obtained by extracting domain-invariant features from the support set using a pre-constructed adversarial dynamics domain alignment model.

[0013] Instantiate the encoder in the teacher reinforcement learning model as a meta-initial parameter;

[0014] The discriminator loss is subjected to gradient descent using a meta-optimization method, and the meta-initial parameters are updated to first task-specific parameters.

[0015] Update the pre-built student reinforcement learning model based on the specific parameters of the first task;

[0016] The similarity loss is calculated by applying the updated student reinforcement learning model to the query set.

[0017] Based on the similarity loss and the discriminator loss, backpropagation is performed to update the encoder, policy network, and value network in the student reinforcement learning model again. The updated student reinforcement learning model is denoted as the environmental dynamics adaptation model.

[0018] In one embodiment, the step of collecting the trajectories of the robot performing tasks in the source and target domains using a pre-built teacher reinforcement learning model includes:

[0019] The teacher reinforcement learning model was constructed based on the proximal policy optimization algorithm.

[0020] The student reinforcement learning model is constructed based on the teacher reinforcement learning model, and the policy network and value network of the student reinforcement learning model are aligned with the policy network and value network of the teacher reinforcement learning model.

[0021] In one embodiment, the adversarial dynamics domain alignment model uses a policy network and a value network from a reinforcement learning algorithm as task regressors. The adversarial dynamics domain alignment model further includes an encoder corresponding to the policy network, an encoder corresponding to the value network, and a continuous domain discriminator.

[0022] The step of extracting domain-invariant features from the support set using a pre-built adversarial dynamics domain alignment model to obtain the discriminator loss includes:

[0023] The support set is extracted by the encoder corresponding to the policy network and the encoder corresponding to the value network in the adversarial dynamics domain alignment model to obtain the first intermediate feature and the second intermediate feature.

[0024] The first intermediate feature and the second intermediate feature are concatenated and then input into the continuous domain discriminator of the adversarial dynamics domain alignment model;

[0025] Adversarial training is performed between the encoder corresponding to the policy network, the encoder corresponding to the value network, and the discriminator in the adversarial dynamics domain alignment model to minimize the distribution difference between the source domain trajectory dataset and the target domain trajectory dataset in the support set, so as to obtain the discriminator loss.

[0026] In one embodiment, the step of performing adversarial training between the encoder corresponding to the policy network, the encoder corresponding to the value network, and the continuous domain discriminator includes:

[0027] The loss of the task regressor is minimized by the encoder corresponding to the policy network and the encoder corresponding to the value network, the loss of the continuous domain discriminator is maximized, and the loss of the continuous domain discriminator is minimized by the continuous domain discriminator, thereby extracting domain-invariant features.

[0028] In one embodiment, the step of backpropagating based on the similarity loss and the discriminator loss and updating the encoder, policy network, and value network in the student reinforcement learning model again includes:

[0029] The initial parameters are updated to second task-specific parameters based on the similarity loss and the discriminator loss.

[0030] The encoder, policy network, and value network in the student reinforcement learning model are updated based on the specific parameters of the second task.

[0031] Furthermore, to achieve the above objectives, this application also proposes a robot control device, which includes:

[0032] A robot control module is used to control the robot in response to control commands by using a pre-built environmental dynamics adaptation model, which is trained based on an adversarial continuous dynamics domain alignment reinforcement learning framework.

[0033] In addition, to achieve the above objectives, this application also proposes a robot control device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the robot control method as described above.

[0034] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and which, when executed by a processor, implements the steps of the robot control method described above.

[0035] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the robot control method described above.

[0036] One or more technical solutions proposed in this application have at least the following technical effects:

[0037] This application responds to control commands by controlling a robot through a pre-built environmental dynamics adaptation model. The environmental dynamics adaptation model is trained based on an adversarial continuous dynamics domain alignment reinforcement learning framework. The adversarial continuous dynamics domain alignment reinforcement learning framework is a model-independent reinforcement learning framework used for dynamic adaptation robot control. It can mitigate the impact of various continuous dynamic changes (such as camera rotation, mass, robot links, and surface friction) on robot performance. Attached Figure Description

[0038] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0039] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart illustrating an embodiment of the robot control method of this application.

[0041] Figure 2 This is a schematic diagram illustrating the process of training an environmental dynamics adaptation model based on an adversarial continuous dynamics domain alignment reinforcement learning framework for the robot control method of this application.

[0042] Figure 3 A diagram showing the English and Chinese names of the algorithms used in the experimental verification of this application;

[0043] Figure 4 This represents the average total reward value for each algorithm under the Pendulum task;

[0044] Figure 5 This represents the average total reward value for each algorithm under the semi-cheetah task.

[0045] Figure 6 The average total reward value for each algorithm under the limping half-cheetah task;

[0046] Figure 7This represents the average total reward value of each algorithm under the Ant task;

[0047] Figure 8 The figures show experimental data from experiments conducted in a real environment on the robot control method of this application.

[0048] Figure 9 This is a schematic diagram of the module structure of the robot control device according to an embodiment of this application;

[0049] Figure 10 This is a schematic diagram of the device structure of the hardware operating environment involved in the robot control method in the embodiments of this application.

[0050] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0051] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0052] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0053] The main solution of this application embodiment is: in response to control commands, the robot is controlled by a pre-built environmental dynamics adaptation model, which is trained based on an adversarial continuous dynamics domain alignment reinforcement learning framework.

[0054] Technical terms involved in the embodiments of this invention:

[0055] The Pendulum task, also known as the Inverted Pendulum task, involves an agent controlling a pendulum rod to maintain its balance in a vertical position. The agent achieves this by applying force or torque to keep the rod upright for as long as possible.

[0056] Half-Cheetah Task: A common motion control task, typically involving a robot in the shape of a half-cheetah in a simulated environment. In this task, the agent's goal is to move the robot at maximum speed by controlling joint angles or forces, so as to travel as far as possible within a given time.

[0057] The Limping Half-Cheetah Mission: This is a variation of the Half-Cheetah Mission in which the robot is designed with some kind of deficiency or limitation in its mobility. This deficiency could be limited movement of certain joints or an irregular gait, thus increasing the complexity and challenge of the mission.

[0058] Ant Task: This is another classic reinforcement learning control task. In this task, an agent needs to control a multi-jointed simulated ant robot to move and navigate smoothly in a simulated environment.

[0059] In this embodiment, for ease of description, the robot control device will be used as the execution subject in the following description.

[0060] Existing research defines the target environment only as an environment specific to dynamic parameters, but does not take into account that dynamic parameters are continuous values, and does not attempt to extract the patterns in the continuous changes of dynamic parameters.

[0061] This application provides a solution that controls a robot in response to control commands by using a pre-built environmental dynamics adaptation model. The environmental dynamics adaptation model is trained based on an adversarial continuous dynamics domain alignment reinforcement learning framework. This adversarial continuous dynamics domain alignment reinforcement learning framework is a model-independent reinforcement learning framework used for dynamic adaptation robot control. It can mitigate the impact of various continuous dynamic changes (such as camera rotation, mass, robot links, and surface friction) on robot performance.

[0062] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device or robot control device capable of performing the above functions. The following description uses a robot control device as an example to illustrate this embodiment and the subsequent embodiments.

[0063] Based on this, embodiments of this application provide a robot control method, referring to... Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the robot control method of this application.

[0064] In this embodiment, the robot control method includes step S100:

[0065] In step S100, in response to control commands, the robot is controlled by a pre-built environmental dynamics adaptation model, which is trained based on an adversarial continuous dynamics domain alignment reinforcement learning framework.

[0066] It should be noted that this application proposes a novel adversarial continuous dynamics domain aligned reinforcement learning (CDA-RL) framework, which combines adversarial domain adaptation with reinforcement learning to eliminate the influence of dynamic changes.

[0067] Furthermore, training the environmental dynamics adaptation model using the adversarial continuous dynamics domain alignment reinforcement learning framework may include steps S10–S90:

[0068] Step S10: Collect the trajectories of the robot performing tasks in the source domain and the target domain using a pre-built teacher reinforcement learning model, and denot them as the source domain trajectory dataset and the target domain trajectory dataset.

[0069] It should be noted that the teacher reinforcement learning model is constructed based on the proximal policy optimization algorithm (PPO algorithm).

[0070] In this context, the virtual environment used for training the robot is considered the source domain, and the real environment used for testing is considered the target domain. Each environment has its own set of dynamic parameter values ​​(the dynamic parameter values ​​are a set of dynamic parameters such as mass and friction, denoted by U, representing the set of different values ​​of these dynamic parameters), which is denoted as u. s ,u t ∈U, and use it as the domain label for that environment.

[0071] Each dynamic parameter has its own continuous range of values. Assume the dynamic parameters of the source domain are restricted to a specific range, denoted as u. s The dynamic parameter ut in the target domain has another range of values, which is different from u. s They do not contain each other, that is

[0072] By using a teacher-controlled reinforcement learning model, the robot is trained in both the source and target domains to obtain different trajectories.

[0073] In this method, the teacher reinforcement learning model is responsible for collecting multiple trajectories τ, each containing a state s, an action a, and a reward r. The trajectory datasets sampled in the source and target domains are denoted as the source domain trajectory dataset τ, respectively. s and target domain trajectory dataset τ t .

[0074] In addition, the source domain trajectory dataset and the target domain trajectory dataset are stored in a buffer.

[0075] Step S20: Select trajectory data as support sets from the source domain trajectory dataset and the target domain trajectory dataset, respectively;

[0076] Step S30: Select trajectory data from the source domain trajectory dataset as a query set;

[0077] In this application, an unsupervised domain adaptive framework is used in the source domain trajectory dataset τ. s Train a reinforcement learning model on a small target domain trajectory dataset τt The reinforcement learning model is tuned by using only domain labels.

[0078] Step S40: Extract domain-invariant features from the support set using the pre-constructed adversarial dynamics domain alignment model to obtain the discriminator loss;

[0079] It should be noted that the adversarial dynamics domain alignment model proposed in this invention uses an adversarial domain adaptive algorithm to achieve cross-domain reinforcement learning.

[0080] Adversarial domain adaptation is a classic domain transfer algorithm that minimizes the distributional differences between the source and target domains by having the encoder and discriminator compete against each other, thereby obtaining domain-invariant features. Essentially, it is a bi-objective model consisting of an encoder, a discriminator, and a task regressor.

[0081] The encoder aims to extract discriminative features and minimize the classifier loss. The discriminator then aims to minimize the domain classifier loss. The optimization process of the two models is similar to a minimax game between two players; the encoder continuously tries to extract features that maximize the ability to confuse the discriminator, while the discriminator continuously tries to improve its domain-discriminative ability based on the features extracted by the encoder, ultimately achieving a state where the encoder extracts domain-invariant features.

[0082] Additionally, it should be noted that the support set consists of training environment trajectory data and test environment trajectory data. Features of the training environment trajectory data and test environment trajectory data are extracted through the adversarial dynamics domain alignment model. Then, the adversarial dynamics domain alignment model is used to make the features of the training environment trajectory data and the features of the test environment trajectory data look the same (i.e., the discriminator cannot distinguish whether these features are from the training environment or the test environment. At this point, it can be considered that these features have eliminated the dynamic influence), which are domain-invariant features.

[0083] It should be noted that the adversarial dynamics domain alignment model proposed in this invention employs the policy network π in reinforcement learning algorithms. θ Value Network V π As a task regressor, two encoders E were designed for the policy network and the value network. a E c To perform domain-invariant feature extraction, a continuous-domain discriminator D was designed. The optimization objectives are summarized as follows:

[0084]

[0085] Among them, L a and L c L represents the policy network loss and value network loss in reinforcement learning. disc λ represents the discriminator loss.d It controls the weight of the discriminator loss in the total loss, so that the model achieves a balance between the two tasks it needs to accomplish.

[0086] In real-world robotic scenarios, dynamics naturally change continuously within a continuous space. However, most existing work assumes that dynamic parameters are discrete within a finite range.

[0087] To address this issue, the adversarial dynamics alignment model employs CIDA and designs a continuous domain discriminator. This method treats the domain classification task as a regression problem rather than a discrete one, thereby discovering the relationship between domain indices and data distribution. To achieve this, the discriminator computes the mean D of the embeddings separately. μ and variance D σ2 Both are aligned into the same embedding. The discriminator loss is shown in Equation (3-2):

[0088]

[0089] Where s is the state, u is the domain label (dynamic parameters of the environment), and E(s,u) is the feature extracted after inputting these two parameters into the encoder.

[0090] In summary, the mini-max optimization of the adversarial dynamics alignment model mentioned in (3-1) can lead to fluctuating behavior, which may hinder the convergence of reinforcement learning. Therefore, the proposed method divides the adversarial dynamics alignment model into two parts and proposes teacher-student reinforcement learning and meta-optimization to ensure a stable training process, as described below.

[0091] During training, reinforcement learning exhibits high sensitivity to both simulated and real-world environments. Adversarial learning modifies both the actor and commentator encoders, and the resulting instability can lead to the accumulation of inaccurate behavior in the reinforcement learning model. To address this challenge, CDA-RL proposes a teacher-student reinforcement learning framework that enables the adversarial domain adaptation model to interact stably with the simulator. As shown in (3-1), a teacher reinforcement learning model interacts with the environment, while a student reinforcement learning model collaborates with the adversarial domain adaptation model to achieve domain adaptation.

[0092] Specifically, teacher reinforcement learning in teacher-student reinforcement learning consists of a policy network π and a value network V. π It consists of an encoder and an encoder, used to interact with the simulation environment and generate the optimal action for each observation.

[0093] For robot control, the PPO method is chosen as the reinforcement learning algorithm. Policy gradient and generalized advantage estimation (GAE) are used to optimize the actor and commentator. The optimization objectives for the actor and commentator in teacher reinforcement learning are summarized as follows:

[0094]

[0095] Where, π θ It is a policy network, V π It is a value network, where s is the state, a is the action, and u is the action. s E is the dynamic parameter of the source domain. a (s,u s ) is to input these two parameters into encoder E a The features extracted later, E c (s,u s ) is to input these two parameters into encoder E c The features extracted later, A π This represents the advantage function. Unlike the original formula, and compared to the standard PPO, the executor and commentator used the value from the encoder E during computation. a E c Extract the features Eactor and Ecritic, instead of using the state s.

[0096] Teacher reinforcement learning samples from training and testing environments and generates trajectories τ s τ t The data, including state s, action a, and reward r, are stored in a buffer to train students in reinforcement learning.

[0097] Student reinforcement learning has a similar structure and is designed to approximate teacher-student reinforcement learning through similarity loss. Given an observation s, the policy network π of the teacher reinforcement learning... θ Output Gaussian distribution N(μ) θ ,σ θ ) = π θ (E a (s,u)) as an action, value network V πθ Output reward r θ =V πθ (E c (s,u)) is used as the estimated return.

[0098] Student reinforcement learning strategy network π φ Value Network V πφ Designed to align with the policy and value networks of teacher-student reinforcement learning, so that the similarity of actions and estimated rewards is close, i.e., when given the same s and u, N(μ) t ,σ t )=N(μθ ,σ θ ) and r t =r θ Where N is a normal distribution, μ and σ are the mean and variance, r is the output of the value network, the subscript t indicates that this parameter is the output of the teacher reinforcement learning network, and the subscript θ indicates that this parameter is the output of the student reinforcement learning network.

[0099] KL divergence and L1 loss are used to measure the distance between actions and estimated rewards, respectively. The similarity loss function is summarized below:

[0100]

[0101] Where μ and σ are the mean and variance, r is the output of the value network, the subscript t indicates that this parameter is the output of the teacher reinforcement learning network, and the subscript θ indicates that this parameter is the output of the student reinforcement learning network. E′ represents the actor and commenter encoders E′a and E′c of the student reinforcement learning. This teacher-student reinforcement learning design avoids the interaction between teacher-student reinforcement learning and the adversarial domain adaptation model, thus preventing drastic changes in predictor loss during min-max adversarial learning. It ensures that teacher-student reinforcement learning finds the optimal action and enables student reinforcement learning to learn domain-independent actor and commenter encoders for collaborative learning.

[0102] Therefore, the action and reward evaluation losses of the adversarial domain adaptation model are replaced by the similarity loss L. sim Alternative. The optimization objective (3-1) of the adversarial domain adaptation model evolves as follows:

[0103]

[0104] Where E′ represents the enforcer and evaluator encoders E′a and E′c of the student reinforcement learning, D is the continuous domain discriminator, and λ d It controls the weight of the discriminator loss in the total loss, balancing the model between the two tasks it needs to accomplish. π θ It is a strategy network for teacher reinforcement learning, π φ It is a strategy network for students to reinforce their learning.

[0105] Specifically, step S40 involves extracting domain-invariant features from the support set using a pre-constructed adversarial dynamics domain alignment model to obtain the discriminator loss, including steps S41 to S43:

[0106] Step S41: Extract the support set by the encoder corresponding to the policy network and the encoder corresponding to the value network in the adversarial dynamics domain alignment model to obtain the first intermediate feature and the second intermediate feature.

[0107] Unlike the standard adversarial domain adaptation framework, this paper designs two encoders E a E c To extract features from a state or observation s, where E a Zactor is responsible for extracting intermediate features for the policy network π, while E c Responsible for value network V π Extract intermediate features Zcritic.

[0108] The reason for using two encoders instead of one is that during gradient backpropagation, using the same encoder to extract intermediate features would cause the policy network and the value network to share parameters, which would negatively impact the performance of reinforcement learning.

[0109] Step S42: Input the first intermediate feature and the second intermediate feature into the policy network and value network of the adversarial dynamics domain alignment model, respectively; or, concatenate the first intermediate feature and the second intermediate feature and input them into the continuous domain discriminator of the adversarial dynamics domain alignment model.

[0110] The extracted intermediate features Zactor and Zcritic will be concatenated into a single intermediate feature and then input into the continuous domain discriminator.

[0111] Step S43: Adversarial training is performed between the encoder corresponding to the policy network, the encoder corresponding to the value network, and the discriminator in the adversarial dynamics domain alignment model to minimize the distribution difference between the source domain trajectory dataset and the target domain trajectory dataset in the support set, so as to obtain the discriminator loss.

[0112] Further, step S43 includes step S431: minimizing the loss of the task regressor through the encoder corresponding to the policy network and the encoder corresponding to the value network, maximizing the loss of the continuous domain discriminator, and minimizing the loss of the continuous domain discriminator through the continuous domain discriminator, thereby extracting domain-invariant features.

[0113] It should be noted that steps S50 to S90 are optimization methods used in this application to solve the problem of insufficient target domain data in virtual-to-real migration. The optimization methods are explained below:

[0114] 1) By constructing a training framework similar to the MAML model-independent meta-learning method, we can find the initial parameters that can adapt to various tasks the fastest. Even when there is not much data in the target domain, we can achieve the maximum performance improvement through a single gradient update.

[0115] 2) Based on this training framework, modify the loss function used in the inner loop update (i.e., the discriminator loss in the adversarial dynamics domain alignment model) and the loss function used in the outer loop update (i.e., the similarity loss in the student reinforcement learning model) to ensure that the gradient directions generated by the outer loop update for the two tasks of extracting invariant features in the ascending domain and improving the total reward value are similar, so as to improve the performance of both at the same time, rather than generating two mutually exclusive gradients that cause performance fluctuations.

[0116] Specifically, the goal of the meta-optimization method borrows from the idea of ​​model-independent meta-learning methods, which is to find a set of initial parameters φ. e (denoted as the meta-initialization parameter), used for encoder E a E c And the parameters φ of the strategy network and value network of student reinforcement learning. φ (Also denoted as the meta-initialization parameter), which enables the CDA-RL model to achieve optimal performance with extremely small amounts of data and a small number of gradient update iterations.

[0117] Step S50: Instantiate the encoder in the teacher reinforcement learning model as meta-initial parameters;

[0118] It should be noted that this application divides the dataset into a support set R. sup and query set R que The support set is significantly smaller than the query set. The meta-optimization method performs initial gradient updates on the support set in the inner loop and fine-tuning on the query set in the outer loop.

[0119] Within the inner loop, the meta-optimization method incorporates the encoder E in the teacher reinforcement learning model. a E c Instantiated as the meta-initialization parameter φ e .

[0120] Step S60: Apply gradient descent to the discriminator loss using a meta-optimization method to update the meta-initial parameters to first task-specific parameters;

[0121] Among them, from supporting a batch of tasks for centralized sampling, the discriminator loss L is used. disc Perform gradient descent, setting the initial parameters φ e Updated to first task-specific parameters The details are as follows:

[0122]

[0123] Where D is the continuous-domain discriminator, and α is the learning rate of the inner loop. The inner loop is used to learn a domain-independent encoder, which is not involved in the policy network and value network. In the outer loop, optimization objective 3-6 is used as the loss function of the outer loop, as shown below:

[0124]

[0125] Where, π θ It is a strategy network for teacher reinforcement learning, π φ It is a strategy network for students' reinforcement learning, V πφ It is a value network that reinforces students' learning, φ e These are the initial parameters. For the first task-specific parameter, λ d It controls the weight of the discriminator loss in the total loss, so that the model achieves a balance between the two tasks it needs to accomplish.

[0126] Step S70: Update the pre-built student reinforcement learning model based on the specific parameters of the first task;

[0127] Specifically, the pre-built student reinforcement learning model is updated based on the first task-specific parameters, that is, the encoder in the student reinforcement learning model is updated with the first task-specific parameters. The corresponding encoder, and then, through the updated encoder and policy network π in the student reinforcement learning model. φ Value Network V πφ To calculate the similarity loss L sim This helps to make L sim and L disc The goal remains consistent. To verify this, we can substitute (3-7) into (3-8) and perform a first-order Taylor expansion on the first term of (3-8), obtaining the following result:

[0128]

[0129] Where α is the learning rate, π θ It is a strategy network for teacher reinforcement learning, π φ It is a strategy network for students to reinforce their learning. φ e V is the initial parameter. πφ It is the value network for students' reinforcement learning, and D is the discriminator for the continuous domain.

[0130] The second item contains L disc and L sim The dot product. This means that when the gradients generated by the encoder and the actor-critic reinforcement learning model are in opposite directions during backpropagation, the second term in (3-9) will be negative, leading to an increase in the total loss.

[0131] Step S80: Calculate the similarity loss by applying the updated student reinforcement learning model to the query set;

[0132] Step S90: Based on the similarity loss and the discriminator loss, backpropagate and update the encoder, policy network and value network in the student reinforcement learning model again, and denote the updated student reinforcement learning model as the environment dynamics adaptation model.

[0133] Specifically, similarity loss and discriminator loss are used for backpropagation to update the parameters φ in the student reinforcement learning model again. e and φ φ .

[0134] Further, step S90 includes steps S91 to S92:

[0135] Step S91: Update the meta-initial parameters to second task-specific parameters based on the similarity loss and the discriminator loss;

[0136] Among them, based on the similarity loss L, formula (3-8) is used. sim and discriminator loss L disc The total loss L was calculated. total Based on the total loss L total The initial parameter φ e and φ φ Updated to the second task-specific parameter φ' e and φ' φ Specifically, as shown in formula (3-10):

[0137]

[0138] Where, φ e and φ φ φ' is the initial parameter. e and φ' φ For the second task-specific parameter, α is the learning rate.

[0139] Step S92: Update the encoder in the student reinforcement learning model based on the second task-specific parameters.

[0140] Among them, the second task-specific parameter φ' e and φ' φ Update the encoder in the student reinforcement learning model.

[0141] The updated student reinforcement learning model is then applied to the robot control scenario to mitigate the impact of various continuous dynamic changes (such as camera rotation, mass, robot links, and surface friction) on robot performance.

[0142] In another feasible implementation, step S10, which involves collecting the trajectories of the robot performing tasks in the source and target domains using a pre-built teacher reinforcement learning model, may include steps A11 to A12:

[0143] Step A11: The teacher reinforcement learning model is constructed based on the proximal policy optimization algorithm;

[0144] Step A12: Based on the teacher reinforcement learning model, the student reinforcement learning model is constructed, and the policy network and value network of the student reinforcement learning model are aligned with the policy network and value network of the teacher reinforcement learning model.

[0145] In summary, this invention constructs an environment dynamics adaptation method based on adversarial domain adaptation and meta-reinforcement learning. This method draws on the idea of ​​extracting domain-invariant features in traditional adversarial domain adaptation methods and improves upon its classic network structure.

[0146] First, to prevent gradient conflicts between the value and policy networks, separate encoders are built for the adversarial domain adaptation module and the meta-reinforcement learning module.

[0147] Subsequently, in order to eliminate the impact of continuous changes in dynamic parameters on robot control, this paper uses an adversarial domain adaptation method under continuous indexing, which enables the encoder of the adversarial domain adaptation module to capture the patterns under continuous changes in dynamic parameters.

[0148] In addition, to better extract domain-invariant features, the meta-reinforcement learning module is divided into two parts: teacher and student. The student reinforcement learning model is responsible for extracting domain-invariant features and imitating the actions of the teacher reinforcement learning model, which makes feature extraction more stable.

[0149] Finally, to address the issue of insufficient target domain data, this paper employs a meta-optimization method to improve the virtual-to-real transfer effect under limited sample size and optimizes the gradient direction to accelerate training.

[0150] These four optimization points solve the problem of virtual-real migration under continuous dynamic mismatch.

[0151] This embodiment provides a robot control method. In response to control commands, the robot is controlled by a pre-constructed environmental dynamics adaptation model. The environmental dynamics adaptation model is trained based on an adversarial continuous dynamics domain alignment reinforcement learning framework. The adversarial continuous dynamics domain alignment reinforcement learning framework is a model-independent reinforcement learning framework used for dynamic adaptation robot control. It can reduce the impact of various continuous dynamic changes (such as camera rotation, mass, robot links and surface friction) on robot performance.

[0152] For example, to help understand the implementation flow of the robot control method obtained by combining this embodiment with the above embodiment one, please refer to... Figure 2 , Figure 2 This is a schematic diagram illustrating the process of training an environmental dynamics adaptation model using an adversarial continuous dynamics domain alignment reinforcement learning framework for robot control, as described in this application.

[0153] like Figure 2 As shown, this invention proposes an adversarial continuous dynamics domain aligned reinforcement learning framework that combines adversarial domain adaptation with reinforcement learning to eliminate the influence of dynamic changes.

[0154] To achieve this goal, the framework has two key components: an adversarial dynamics domain alignment model and teacher-student reinforcement learning. The top left corner shows the classic adversarial domain adaptation model and reinforcement learning model.

[0155] In addition, the framework also includes a meta-optimization section.

[0156] Among them, the adversarial domain adaptation model combines a policy network and a value network encoder with a discriminator to extract domain-invariant features.

[0157] An adversarial dynamics domain alignment model is proposed based on the adversarial domain adaptation model.

[0158] Teachers and students reinforce learning by interacting with the environment and storing their trajectories in a buffer.

[0159] Student reinforcement learning uses a dynamic invariant encoder to approximate teacher-student reinforcement learning and attempts to generate optimal actions.

[0160] Meta-optimization aligns with the optimization loss of previous methods and supports dynamic adaptation with limited information.

[0161] Finally, the student reinforcement learning policy network and value network, along with their encoder, are deployed in a real-world environment.

[0162] The following are the specific steps for training an environmental dynamics adaptation model using an adversarial continuous dynamics domain alignment reinforcement learning framework:

[0163] S300: A teacher reinforcement learning model is constructed based on the proximal policy optimization algorithm, and a student reinforcement learning model is constructed based on the teacher reinforcement learning model.

[0164] S301 uses a teacher reinforcement learning model to control the robot to perform tasks in the training and testing environments, thereby acquiring trajectory data for both environments.

[0165] S302 stores the training environment trajectory data and the test environment trajectory data in a buffer for easy retrieval later.

[0166] S303: A portion of the training environment trajectory data and a portion of the test environment trajectory data are extracted from the buffer as a support set. Domain-invariant features are extracted from the support set using an adversarial dynamics domain alignment model proposed based on the adversarial domain adaptation model. The adversarial dynamics domain alignment model uses the policy network and value network from the reinforcement learning algorithm as task regressors. The model also includes an encoder corresponding to the policy network, an encoder corresponding to the value network, and a continuous domain discriminator.

[0167] Specifically, the support set is extracted by the encoder corresponding to the policy network and the encoder corresponding to the value network in the adversarial dynamics domain alignment model to obtain the first intermediate feature and the second intermediate feature;

[0168] The first intermediate feature and the second intermediate feature are concatenated and then input into the continuous domain discriminator of the adversarial dynamics domain alignment model;

[0169] Adversarial training is performed between the encoder corresponding to the policy network, the encoder corresponding to the value network, and the discriminator in the adversarial dynamics domain alignment model to minimize the distribution difference between the source domain trajectory dataset and the target domain trajectory dataset in the support set, so as to obtain the discriminator loss.

[0170] Step S304, the encoder E in the teacher reinforcement learning model a E c Instantiated as the initial parameter φ e and φ φ The discriminator loss is subjected to gradient descent using a meta-optimization method, and the meta-initial parameters are updated to the first task-specific parameters.

[0171] Step S305: Update the encoder in the pre-built student reinforcement learning model based on the first task-specific parameters.

[0172] Step S306: Calculate the similarity loss L using the updated student reinforcement learning model on the query set. sim .

[0173] Step S307: Based on the similarity loss and the discriminator loss, backpropagate and update the encoder, policy network and value network in the student reinforcement learning model again, and denote the updated student reinforcement learning model as the environment dynamics adaptation model.

[0174] This invention was experimentally verified in a virtual environment. The experiments demonstrated that the environmental dynamics adaptation model outperforms existing advanced algorithms. A comparison of the English and Chinese names of existing advanced algorithms is shown in the figure below. Figure 3 As shown.

[0175] Reference Figure 4 , Figure 4 This represents the average total reward value for each algorithm under the Pendulum task.

[0176] Among them, according to Figure 4 The numerical results show that the environmental dynamics adaptation model proposed in this invention outperforms other algorithms in both the ordinary target domain and the extreme target domain.

[0177] Reference Figure 5 , Figure 5 This represents the average total reward value for each algorithm under the semi-cheetah task.

[0178] Among them, according to Figure 5 The numerical results show that the environmental dynamics adaptation model proposed in this invention outperforms other algorithms in the source domain, ordinary target domain, and extreme target domain.

[0179] Reference Figure 6 , Figure 6 This represents the average total reward value for each algorithm under the limping half-cheetah task.

[0180] Among them, according to Figure 6 The numerical results show that the environmental dynamics adaptation model proposed in this invention outperforms other algorithms in both the ordinary target domain and the extreme target domain.

[0181] Reference Figure 7 , Figure 7 This represents the average total reward value for each algorithm under the Ant task.

[0182] Among them, according to Figure 7 The numerical results show that the environmental dynamics adaptation model proposed in this invention outperforms other algorithms in the source domain, ordinary target domain, and extreme target domain.

[0183] Meanwhile, the environmental dynamics adaptation model proposed in this invention was also tested in a real environment. In the real environment, the camera position differed from that in the virtual environment to verify the adaptability of the environmental dynamics adaptation model when the settings of the virtual and real environments were inconsistent. The results are as follows: Figure 8 As shown, under different camera offset angles, the environmental dynamics adaptation model proposed in this invention requires fewer average steps to complete the task and has a higher overall success rate.

[0184] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the robot control method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0185] This application also provides a robot control device, please refer to... Figure 9 The robot control device includes:

[0186] The robot control module 10 is used to control the robot in response to control commands by means of a pre-built environmental dynamics adaptation model, which is trained based on an adversarial continuous dynamics domain alignment reinforcement learning framework.

[0187] The robot control device provided in this application, employing the robot control method described in the above embodiments, can solve the technical problem that existing research has not considered that dynamic parameters are continuous values ​​and has not attempted to extract the regularity of continuous changes in dynamic parameters. Compared with the prior art, the beneficial effects of the robot control device provided in this application are the same as those of the robot control method provided in the above embodiments, and other technical features in the robot control device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0188] This application provides a robot control device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the robot control method in Embodiment 1 above.

[0189] The following is for reference. Figure 10 The diagram illustrates a structural schematic suitable for implementing the robot control device of the embodiments of this application. The robot control device in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 10 The robot control device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this application.

[0190] like Figure 10As shown, the robot control device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the robot control device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to the I / O interface 1006: input devices 1007 including, for example, a touchscreen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. Communication device 1009 allows the robot control device to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows robot control devices with various systems, it should be understood that implementation or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0191] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0192] The robot control device provided in this application, employing the robot control method described in the above embodiments, can solve the technical problem that existing research has not considered that dynamic parameters are continuous values ​​and has not attempted to extract the regularity of continuous changes in dynamic parameters. Compared with the prior art, the beneficial effects of the robot control device provided in this application are the same as those of the robot control method provided in the above embodiments, and other technical features of this robot control device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0193] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0194] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0195] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the robot control method in the above embodiments.

[0196] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0197] The aforementioned computer-readable storage medium may be included in the robot control device; or it may exist independently and not be assembled into the robot control device.

[0198] The aforementioned computer-readable storage medium carries one or more programs that, when executed by a robot control device, cause the robot control device to: control the robot through a pre-built environmental dynamics adaptation model in response to control commands, wherein the environmental dynamics adaptation model is trained based on an adversarial continuous dynamics domain alignment reinforcement learning framework.

[0199] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0200] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0201] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0202] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described robot control method. This addresses the technical problem in existing research that it does not consider the continuous nature of dynamic parameters and fails to extract the patterns in the continuous changes of dynamic parameters. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the robot control method provided in the above embodiments, and will not be elaborated upon here.

[0203] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the robot control method described above.

[0204] The computer program product provided in this application can solve the technical problem that existing research has not considered that dynamic parameters are continuous values ​​and has not attempted to extract the regularity of continuous changes in dynamic parameters. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as the beneficial effects of the robot control method provided in the above embodiments, and will not be repeated here.

[0205] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A robot control method, characterized in that, The method includes: A teacher reinforcement learning model was constructed based on the proximal policy optimization algorithm; A student reinforcement learning model is constructed based on the teacher reinforcement learning model, and the policy network and value network of the student reinforcement learning model are aligned with the policy network and value network of the teacher reinforcement learning model. Trajectories of robots controlled by pre-built teacher reinforcement learning models in the source and target domains are collected separately and denoted as the source domain trajectory dataset and the target domain trajectory dataset. Trajectory data are selected as support sets from the source domain trajectory dataset and the target domain trajectory dataset, respectively; Select trajectory data from the source domain trajectory dataset as a query set; The support set is subjected to domain-invariant feature extraction by a pre-constructed adversarial dynamics domain alignment model to obtain the discriminator loss. The adversarial dynamics domain alignment model uses the policy network and value network in the reinforcement learning algorithm as the task regressor. The adversarial dynamics domain alignment model also includes an encoder corresponding to the policy network, an encoder corresponding to the value network, and a continuous domain discriminator. Instantiate the encoder in the teacher reinforcement learning model as a meta-initial parameter; The step of extracting domain-invariant features from the support set using a pre-constructed adversarial dynamics domain alignment model to obtain the discriminator loss includes: extracting the support set using the encoder corresponding to the policy network and the encoder corresponding to the value network in the adversarial dynamics domain alignment model to obtain a first intermediate feature and a second intermediate feature. The first intermediate feature and the second intermediate feature are respectively input into the policy network and value network of the adversarial dynamics domain alignment model, or the first intermediate feature and the second intermediate feature are concatenated and then input into the continuous domain discriminator of the adversarial dynamics domain alignment model. Adversarial training is performed between the encoder corresponding to the policy network, the encoder corresponding to the value network, and the discriminator in the adversarial dynamics domain alignment model to minimize the distribution difference between the source domain trajectory dataset and the target domain trajectory dataset in the support set, so as to obtain the discriminator loss. The discriminator loss is subjected to gradient descent using a meta-optimization method, and the meta-initial parameters are updated to first task-specific parameters. Update the pre-built student reinforcement learning model based on the specific parameters of the first task; The similarity loss is calculated by applying the updated student reinforcement learning model to the query set. Based on the similarity loss and the discriminator loss, backpropagation is performed and the encoder, policy network and value network in the student reinforcement learning model are updated again. The updated student reinforcement learning model is denoted as the environmental dynamics adaptation model. In response to control commands, the robot is controlled using the environmental dynamics adaptation model.

2. The method as described in claim 1, characterized in that, The step of performing adversarial training between the encoder corresponding to the policy network, the encoder corresponding to the value network, and the continuous domain discriminator includes: The loss of the task regressor is minimized by the encoder corresponding to the policy network and the encoder corresponding to the value network, the loss of the continuous domain discriminator is maximized, and the loss of the continuous domain discriminator is minimized by the continuous domain discriminator, thereby extracting domain-invariant features.

3. The method as described in claim 1, characterized in that, The step of backpropagating based on the similarity loss and the discriminator loss and updating the encoder, policy network, and value network in the student reinforcement learning model again includes: The initial parameters are updated to second task-specific parameters based on the similarity loss and the discriminator loss. The encoder in the student reinforcement learning model is updated based on the specific parameters of the second task.

4. A robot control device, characterized in that, The device includes: The robot control module is used to construct a teacher reinforcement learning model based on the proximal policy optimization algorithm. A student reinforcement learning model is constructed based on the teacher reinforcement learning model, and the policy network and value network of the student reinforcement learning model are aligned with the policy network and value network of the teacher reinforcement learning model. Trajectories of robots controlled by pre-built teacher reinforcement learning models in the source and target domains are collected separately and denoted as the source domain trajectory dataset and the target domain trajectory dataset. Trajectory data are selected as support sets from the source domain trajectory dataset and the target domain trajectory dataset, respectively; Select trajectory data from the source domain trajectory dataset as a query set; The support set is subjected to domain-invariant feature extraction by a pre-constructed adversarial dynamics domain alignment model to obtain the discriminator loss. The adversarial dynamics domain alignment model uses the policy network and value network in the reinforcement learning algorithm as the task regressor. The adversarial dynamics domain alignment model also includes an encoder corresponding to the policy network, an encoder corresponding to the value network, and a continuous domain discriminator. Instantiate the encoder in the teacher reinforcement learning model as a meta-initial parameter; The step of extracting domain-invariant features from the support set using a pre-constructed adversarial dynamics domain alignment model to obtain the discriminator loss includes: extracting the support set using the encoder corresponding to the policy network and the encoder corresponding to the value network in the adversarial dynamics domain alignment model to obtain a first intermediate feature and a second intermediate feature. The first intermediate feature and the second intermediate feature are respectively input into the policy network and value network of the adversarial dynamics domain alignment model, or the first intermediate feature and the second intermediate feature are concatenated and then input into the continuous domain discriminator of the adversarial dynamics domain alignment model. Adversarial training is performed between the encoder corresponding to the policy network, the encoder corresponding to the value network, and the discriminator in the adversarial dynamics domain alignment model to minimize the distribution difference between the source domain trajectory dataset and the target domain trajectory dataset in the support set, so as to obtain the discriminator loss. The discriminator loss is subjected to gradient descent using a meta-optimization method, and the meta-initial parameters are updated to first task-specific parameters. Update the pre-built student reinforcement learning model based on the specific parameters of the first task; The similarity loss is calculated by applying the updated student reinforcement learning model to the query set. Based on the similarity loss and the discriminator loss, backpropagation is performed and the encoder, policy network and value network in the student reinforcement learning model are updated again. The updated student reinforcement learning model is denoted as the environmental dynamics adaptation model. In response to control commands, the robot is controlled using the environmental dynamics adaptation model.

5. A robot control device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the robot control method as described in any one of claims 1 to 3.

6. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the robot control method as described in any one of claims 1 to 3.

7. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the robot control method as described in any one of claims 1 to 3.

Images