A dexterous manipulator operation skill transfer method based on relational geometric prior

Abstract

The application discloses a dexterous manipulator operation skill transfer method based on relational geometric prior, comprising the following steps: extracting relational operation vectors from the demonstration trajectory of a human hand and a target object, and constructing a unified hand geometric feature representation; learning a low-dimensional latent geometric prior representing human hand coordination rules by using a variational auto-encoding network; in the reinforcement learning training stage, introducing the latent geometric prior into the strategy state space, and combining a phased curriculum learning mechanism to optimize the control strategy; constructing a state pool based on the latent space reconstruction error and the task effectiveness, which is used for environment initialization in the free exploration stage, so as to constrain the strategy to perform efficient exploration within the action manifold conforming to the human coordination rules and physically feasible. The application is verified in various dexterous manipulator platforms and simulation and physical operation tasks, can reduce the physical conflicts caused by explicit redirection, and improve the physical feasibility, training stability and efficiency of the dexterous manipulator operation skill transfer.

Description

Technical Field

[0001] This invention relates to the field of robotics, and more particularly to a method for transferring dexterous manipulator operation skills based on relational geometric priors. Background Technology

[0002] Dexterous robotic hands, with their multiple fingers, multiple degrees of freedom, and complex contact capabilities, are considered an important vehicle for achieving precise manipulation and embodied intelligence in robots, and have broad application prospects in service robots, industrial automation, medical assistance, and human-computer interaction. To enable robots to possess human-like dexterous manipulation capabilities, utilizing large-scale human demonstration data for skill transfer to robotic hands has become an important technical approach in dexterous manipulation research in recent years.

[0003] In existing technologies, one type of method describes human hand demonstrations using keypoints and joint angles, and combines this with target-conditional reinforcement learning for policy training. While this type of method can utilize demonstration information to narrow the exploration space, most of them use keypoint representations in absolute space, which makes it difficult to effectively characterize the fine-grained, structured collaborative relationships between fingers. When there are scale differences and coupling variations in different parts of the hand, absolute keypoint representations often fail to stably describe universally applicable operational structures.

[0004] Meanwhile, they typically use explicit redirection to map the human hand demonstration to the joint trajectories of a specific robotic hand, which are then used as training references or initialization sources. This type of method can accelerate policy convergence to some extent, but due to significant differences between human and robotic hands in terms of the number of degrees of freedom, link ratios, kinematic constraints, and physical properties, explicit redirection often introduces physical conflicts such as self-collision, penetration, and mismatched contact, resulting in unreasonable initialization states and further misleading the reinforcement learning exploration process.

[0005] Therefore, how to extract geometric features that are consistent and transferable across different dexterous manipulators from human demonstrations, and on this basis, improve the physical feasibility and training stability of the skill transfer process, remains a pressing technical problem to be solved in this field. Based on this, it is necessary to propose a dexterous manipulator skill transfer method based on relational geometric priors. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for transferring dexterous manipulator operation skills based on relational geometric priors.

[0007] The objective of this invention is achieved through the following technical solution: a method for transferring dexterous manipulator manipulation skills based on relational geometry priors, the method comprising:

[0008] Acquire demonstration trajectory data of the human hand and the target object, and construct a unified representation of the geometric features of the hand;

[0009] The hand geometric feature representation is modeled using a pre-trained latent representation learning network to obtain a low-dimensional latent geometric prior that characterizes the rules of human hand collaborative operation;

[0010] To construct a reinforcement learning environment for dexterous manipulator operation tasks, the task state information, the latent representation corresponding to the robot's current hand state, and the latent representation corresponding to the reference demonstration state are used as policy inputs to establish a state space that integrates latent geometric priors.

[0011] Based on the state space, a reinforcement learning algorithm is used to train and optimize the control strategy of the dexterous manipulator. The training process combines a phased learning mechanism and filters candidate states based on the latent space reconstruction error and task effectiveness, so as to realize the transfer of human hand demonstration to the dexterous manipulator's operation skills.

[0012] Furthermore, the construction of a unified hand geometric feature representation includes:

[0013] A local coordinate system is established with the human wrist as a reference, and the positions of each key point in the global coordinate system are uniformly transformed to the local coordinate system of the wrist. Under the local coordinate system of the wrist, a set of relational operation vectors representing the collaborative structure of the hand is constructed, including the fingertip to wrist vector, the PIP joint to wrist vector, the fingertip to thumb fingertip vector, and the vector between adjacent fingertips. For each relational operation vector, it is further decomposed into direction information and length information. The direction component and the length component are concatenated to obtain a four-dimensional representation of a single relational operation vector. The representations of all relational operation vectors are stacked in a preset order to obtain the geometric feature representation of the hand.

[0014] Furthermore, the latent representation learning network is a variational autoencoder, including an encoder and a decoder. The encoder takes as input the geometric feature representation of the hand and outputs the latent geometric prior that represents the rules of human hand coordination. The decoder reconstructs the geometric feature representation of the hand based on the latent geometric prior.

[0015] Furthermore, the latent representation learning network is pre-trained using human hand demonstration trajectory data; during the pre-training process, the total reconstruction loss and KL divergence regularization term are used for joint training;

[0016] The total reconstruction loss is represented by a weighted average of the homoscedastic uncertainty of the directional reconstruction loss and the length reconstruction loss of each relation operation vector.

[0017] Furthermore, the reinforcement learning environment for constructing the dexterous manipulator operation task is to model the dexterous manipulator operation task as a Markov decision process. The policy network input includes task state information, the latent representation corresponding to the robot's current hand state, the latent representation corresponding to the reference demonstration state, and the similarity measure between the latent representations of the robot's current hand state and the reference demonstration. The similarity measure is represented by cosine similarity. The policy network outputs the target position command for each drive joint of the dexterous manipulator.

[0018] The task status information includes the target object's pose, target object's velocity, object's geometric features, and the relative distance information from the mechanical fingertip to the target object;

[0019] The latent representation is obtained through encoding via a latent representation learning network.

[0020] Furthermore, in the process of training and optimizing the dexterous manipulator control strategy using reinforcement learning algorithms, the total reward function includes a task reward for constraining the error between the target object state and the reference trajectory, an imitation reward for constraining hand deviations, and a smoothing reward for constraining the continuity and stability of robot joint movements.

[0021] Furthermore, the phased learning mechanism includes:

[0022] Phase 1: The state of the target object is directly set to the state of the object in the reference demonstration, so that the target object is not affected by free dynamics and contact effects; at the same time, task rewards are turned off and only imitation rewards and smooth rewards are used to optimize the strategy; when the imitation success rate in the first phase of training reaches the preset threshold, or the number of training steps reaches the preset limit, the second phase of training is switched to.

[0023] Phase Two: Release the target object, subjecting it to realistic physical dynamics and contact effects, while simultaneously activating mission rewards to guide the strategy to complete the realistic control task while maintaining human-like collaborative principles.

[0024] Furthermore, in the second stage of training, a state pool is constructed to store high-quality intermediate states obtained during historical training. The corresponding state is added to the state pool only when the reconstruction error of the geometric features of the robot's hand relationship is less than a preset threshold and the target object is in a stable and operable state. Each time the environment is reset, intermediate states are sampled from the state pool as the initial state.

[0025] According to another aspect of the specification, a dexterous manipulator skill transfer device based on relational geometric priors is also provided, including a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it implements the aforementioned dexterous manipulator skill transfer method based on relational geometric priors.

[0026] According to another aspect of the specification, a computer-readable storage medium is also provided, on which a program is stored, which, when executed by a processor, implements the aforementioned method for transferring dexterous manipulator manipulation skills based on relational geometry priors.

[0027] The beneficial effects of this invention are:

[0028] 1. A method for skill transfer of dexterous manipulators based on relational geometric priors is proposed, which can extract stable internal geometric relation representations from human hand demonstrations, providing a unified geometric description basis for skill transfer of dexterous manipulators.

[0029] 2. A method is proposed to introduce latent geometric priors into the state space of reinforcement learning, so that the control policy can maintain the cooperative operation features consistent with human hand demonstration while learning the target object operation task, thereby reducing the generation of abnormal actions and unnatural postures.

[0030] 3. A latent geometric prior-guided learning mechanism for skill transfer in dexterous robotic hands is proposed. This mechanism first guides the robotic hand to learn hand alignment and finger coordination consistent with human hand demonstrations through a kinematic anchoring phase. Then, through a free exploration phase, it learns object manipulation processes under real contact conditions, reducing the training difficulty caused by high-dimensional motion space and complex contact dynamics. Furthermore, this invention selects candidate states based on latent space reconstruction error and task effectiveness, allowing strategies to be trained starting from physically feasible states that conform to human hand manipulation patterns, thereby improving training efficiency, operational stability, and generalization ability.

[0031] 4. This invention constructs a unified geometric feature representation of the hand through relational operation vectors and utilizes a latent representation learning network to extract the collaborative operation rules of the human hand, enabling policy learning to be independent of explicit joint mapping between the human hand and the robotic hand. This invention achieves skill transfer from human hand demonstrations to dexterous robotic hands with different degrees of freedom and kinematic structures without explicit kinematic retargeting, demonstrating excellent cross-configuration adaptability and practical deployment value. Attached Figure Description

[0032] Figure 1 This is a schematic diagram illustrating the extraction of hand relationship operation vectors in an embodiment of the present invention;

[0033] Figure 2 This is an architecture diagram of a dexterous manipulator skill transfer method based on relational geometry priors in an embodiment of the present invention;

[0034] Figure 3 This is a schematic diagram illustrating the effect of different dexterous robotic arm platforms performing operational skill transfer tasks in a real environment, as described in the embodiments of the present invention.

[0035] Figure 4This is a schematic diagram of a dexterous manipulator skill transfer device based on relational geometry priors in an embodiment of the present invention. Detailed Implementation

[0036] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0037] It should be understood that the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0038] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. The singular forms “a,” “the,” and “the” used in the embodiments of this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0039] The present invention will be further described below with reference to the accompanying drawings and specific examples.

[0040] like Figure 1 As shown, this embodiment provides a method for transferring dexterous manipulator skills based on relational geometric priors. In this embodiment, the method is implemented based on the PyTorch deep learning framework, and the reinforcement learning environment is constructed using the IsaacGym simulation platform. The demonstration trajectory of the human hand and the target object is derived from the hand-object interaction dataset GRAB, and the target dexterous hand is selected as the Shadow Hand. After completing the policy training, the acquired manipulator skills are transferred to dexterous manipulators of different configurations, such as XHand, Wuji Hand, and Inspire Hand, thereby verifying the transferability and applicability of the method on heterogeneous dexterous hand platforms.

[0041] Step 1: Obtain demonstration trajectory data of the human hand and the target object, and construct a unified geometric feature representation of the hand.

[0042] (1.1) Obtain the demonstration trajectory data of the human hand and the target object: denoted as ,in, Indicates the status of manpower. Indicates the state of the target object. This indicates the total duration of the entire demonstration trajectory.

[0043] (1.2) Establish a local coordinate system for the wrist: In order to eliminate the influence of global spatial position and orientation differences on representation learning, a local coordinate system is established with the human wrist as a reference, and the positions of each key point in the global coordinate system are uniformly transformed to the local coordinate system of the wrist.

[0044] (1.3) Extracting relational operation vectors: Under the local coordinate system of the wrist, construct a set of relational operation vectors representing the hand cooperative structure. Preferably, such as Figure 2 As shown, this example contains 17 relation operation vectors, which are divided into the following four categories:

[0045] (1.3.1) Five vectors from fingertip to wrist are used to characterize the full hand opening scale and overall coverage state;

[0046] (1.3.2) Five vectors from the PIP joint to the wrist are used to characterize the proximal flexion state of the fingers and the shape of the palmar arch;

[0047] (1.3.3) There are 4 vectors from each fingertip to the thumb fingertip, which are used to represent the thumb orientation relationship and the fine grasping structure;

[0048] (1.3.4) There are 3 vectors between adjacent fingertips, which are used to characterize the distance between fingers and fine contact relationships.

[0049] (1.4) Construct a unified geometric feature representation of the hand:

[0050] (1.4.1) For each relation operation vector Further decompose it into directional information and length information The definition is as follows:

[0051]

[0052] (1.4.2) Concatenate the direction component and the length component to obtain a four-dimensional representation of a single relation operation vector:

[0053] (1.4.3) Stack the representations of all 17 relation operation vectors in a preset order to form a hand geometric feature tensor of size 17×4. , which serves as the input for subsequent latent representation learning networks.

[0054] Step 2: Model the geometric feature representation of the hand using a latent representation learning network to obtain a low-dimensional latent geometric prior.

[0055] (2.1) Constructing a latent representation learning network: In this embodiment, the latent representation learning network is preferably a variational autoencoder, including an encoder. and decoder .

[0056] The encoding process is represented as , These are low-dimensional latent variables used to characterize the latent geometric priors corresponding to the patterns of human-hand coordination. The decoding process is represented as follows: , Reconstructed geometric features of the hand.

[0057] (2.2) Constructing the direction reconstruction loss: In order to ensure that the direction of the reconstructed vector is consistent with the direction of the original vector, a direction loss is constructed for each relation operation vector. , This represents the directional component of the reconstructed vector.

[0058] (2.3) Constructing the length reconstruction loss: In order to ensure that the length of the reconstructed vector is consistent with the length of the original vector, a length loss is constructed for each relation operation vector. , This represents the length component of the reconstructed vector.

[0059] (2.4) Constructing grouped weighted reconstruction loss: Considering the differences in scale and function of the relation operation vectors of different categories, automatic weighting is performed based on the four groups of classification in (1.3) and homoscedastic uncertainty modeling.

[0060]

[0061] In the formula, For the first The set of indices corresponding to the class relation operation vector; The number of categories in the relation operation vector; and These are the base weights for directional loss and length loss, respectively; and The first The learnable log-variance parameters of the direction loss and length loss corresponding to the class relation operation vector.

[0062] (2.5) Constructing the latent distribution regularization term: In order to make the latent space smooth and continuous, a KL divergence regularization term is introduced.

[0063]

[0064] In the formula, Indicates encoder The output posterior distribution, This represents the prior distribution.

[0065] (2.6) Constructing the overall training objective: The overall loss function of the variational autoencoder is defined as... , These are the balancing coefficients. After training is complete, the encoder parameters are frozen, and the latent variables of the encoder output are set to... This serves as a low-dimensional latent geometric prior corresponding to the rules governing human-manual collaborative operations.

[0066] Step 3: Construct a reinforcement learning environment for the dexterous robotic arm's manipulation tasks and establish a policy input that incorporates latent geometric priors.

[0067] (3.1) Constructing a reinforcement learning environment: Modeling the dexterous manipulator operation task as a Markov decision process.

[0068]

[0069] In the formula, Representing the state space, Represents the action space. Represents the state transition function. Show the reward function, This represents the discount factor.

[0070] (3.2) Constructing task status information: Task-related status is denoted as This includes the target object's pose, velocity, geometric features, and the relative distance between the mechanical fingertip and the target object.

[0071] (3.3) Constructing the latent prior state: Representing the relational geometric features of the robot's hand at the current moment. And the geometric features of the hand-person relationship in the next moment, as shown in the reference demonstration. Encode the robot's current latent representation. and reference potential representation Further calculate the cosine similarity between the two. .

[0072] (3.4) Establishing the policy input: The task state, the current robot latent representation, the reference latent representation, and the similarity are concatenated to obtain the policy input state.

[0073] (3.5) Define the motion space: The policy network outputs the target position commands for each driven joint of the dexterous manipulator. , This represents the number of driven joints in the dexterous robotic hand. In this embodiment, the pose of the robotic wrist is directly given by a reference demonstration trajectory, and strategy learning is performed only on the finger joints.

[0074] (3.6) Define the policy optimization objective: In the reinforcement learning environment, the policy The optimization objective is to maximize the cumulative discount return. .

[0075] Step 4: Optimize the dexterous manipulator control strategy under the guidance of latent geometric priors: In this embodiment, the PPO algorithm is used to iteratively optimize the dexterous manipulator control strategy. The training process includes a kinematic anchoring phase and a free exploration phase, with the two phases corresponding to different object state settings and reward activation methods.

[0076] (4.1) Constructing the reward function: Time step The total reward function for reinforcement learning is defined as follows:

[0077] ,

[0078] In the formula, As a task reward, it is used to constrain the error between the target object's state and the reference trajectory; To mimic rewards and constrain hand deviations; To smooth out rewards, it is used to constrain the continuity and stability of robot joint movements; , and These are the weighting coefficients for the corresponding reward items.

[0079] (4.2) Perform the first phase of training: kinematic anchoring phase. In the first phase, the state of the target object is directly set to the state of the object in the reference demonstration, so that the target object is not affected by free dynamics and contact action; at the same time, task rewards are turned off. The optimization strategy utilizes only imitation rewards and smooth rewards. It guides the robotic arm to prioritize learning stable hand coordination patterns and basic operational postures.

[0080] (4.3) When the imitation success rate in the first stage of training reaches the preset threshold, or the number of training steps reaches the preset upper limit, switch to the second stage of training.

[0081] (4.4) Conduct the second phase of training: free exploration phase. In the second phase, release the target object and subject it to real physical dynamics and contact effects, while activating task rewards to guide the strategy to complete the real control task while maintaining human-like collaborative rules.

[0082] (4.5) Constructing a state pool based on latent geometric priors: During the second stage of training, a state pool is constructed. This is used to store high-quality intermediate states obtained during historical training. For candidate states, relational geometric features are calculated based on the robot's current joint states. The reconstruction results are obtained through a latent representation learning network. When the reconstruction error satisfies:

[0083] ,

[0084] Furthermore, the corresponding state is added to the state pool only when the target object is in a stable and operable state. . This is a preset threshold.

[0085] (4.6) Environment initialization using a state pool: In the second phase, the environment is reset when a training round ends, the target object falls, an operation fails, or the maximum number of steps is reached. Each time the environment is reset, a probability is used to initialize the environment. From the state pool Sample intermediate states as initial states, with probability Initialize starting from the reference trajectory start frame.

[0086] (4.7) Output control strategy: After training, the strategy outputs the target position command of each drive joint according to the current environmental state, thereby realizing stable operation of the target object. Figure 3 This invention demonstrates the results of different configurations of dexterous manipulators performing operational skill transfer tasks in a real environment, including the grasping, moving, and posture adjustment processes of target objects such as cups, wine glasses, stamps, handles, and glasses. It shows that the control strategy can be transferred to real dexterous manipulator platforms such as XHand and WujiHand, and achieve stable object manipulation.

[0087] Corresponding to the aforementioned embodiment of a dexterous manipulator skill transfer method based on relational geometric priors, the present invention also provides an embodiment of a dexterous manipulator skill transfer device based on relational geometric priors.

[0088] See Figure 4 The present invention provides a dexterous manipulator skill transfer device based on relational geometric priors, comprising a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it is used to implement a dexterous manipulator skill transfer method based on relational geometric priors as described in the above embodiments.

[0089] The present invention provides an embodiment of a dexterous robotic arm skill transfer device based on relational geometric priors. This device can be applied to any device with data processing capabilities, such as a computer. The device embodiment can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 4The diagram shown is a hardware structure diagram of any data processing-capable device, in which a dexterous robotic arm skill transfer device based on relational geometry priors provided by the present invention is located. (Except for...) Figure 4 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0090] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0091] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0092] This invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements a method for transferring dexterous manipulator operation skills based on relational geometric priors as described in the above embodiments.

[0093] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device of any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices of any data processing device. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0094] The present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements the aforementioned method for transferring dexterous manipulator operation skills based on relational geometry priors.

[0095] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this application are indicated by the claims.

[0096] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this application. This application is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for transferring manipulator skills based on relational geometry priors, characterized in that, The method includes: Acquire demonstration trajectory data of a human hand and a target object, and construct a unified geometric feature representation of the hand; the construction of the unified geometric feature representation of the hand includes: A local coordinate system is established with the human wrist as a reference, and the positions of each key point in the global coordinate system are uniformly transformed to the local coordinate system of the wrist. Under the local coordinate system of the wrist, a set of relational operation vectors representing the collaborative structure of the hand is constructed, including the fingertip to wrist vector, the PIP joint to wrist vector, the fingertip to thumb fingertip vector, and the vector between adjacent fingertips. For each relational operation vector, it is further decomposed into direction information and length information. The direction component and the length component are concatenated to obtain a four-dimensional representation of a single relational operation vector. The representations of all relational operation vectors are stacked in a preset order to obtain the geometric feature representation of the hand. The hand geometric feature representation is modeled using a pre-trained latent representation learning network to obtain a low-dimensional latent geometric prior that characterizes the rules of human hand collaborative operation; To construct a reinforcement learning environment for dexterous manipulator operation tasks, the task state information, the latent representation corresponding to the robot's current hand state, and the latent representation corresponding to the reference demonstration state are used as policy inputs to establish a state space that integrates latent geometric priors. Based on the state space, a reinforcement learning algorithm is used to train and optimize the control strategy of the dexterous manipulator. The training process combines a phased learning mechanism and filters candidate states based on the latent space reconstruction error and task effectiveness, so as to realize the transfer of human hand demonstration to the dexterous manipulator's operation skills. In the process of training and optimizing the control strategy of the dexterous manipulator using reinforcement learning algorithm, the total reward function includes a task reward for constraining the error between the target object state and the reference trajectory, an imitation reward for constraining hand deviation, and a smoothing reward for constraining the continuity and stability of robot joint movements. The phased learning mechanism includes: Phase 1: The state of the target object is directly set to the state of the object in the reference demonstration, so that the target object is not affected by free dynamics and contact effects; at the same time, task rewards are turned off and only imitation rewards and smooth rewards are used to optimize the strategy; when the imitation success rate in the first phase of training reaches the preset threshold, or the number of training steps reaches the preset limit, the second phase of training is switched to. Phase Two: Release the target object, subjecting it to realistic physical dynamics and contact effects, while simultaneously activating mission rewards to guide the strategy to complete the realistic control task while maintaining human-like collaborative principles. In the second stage of training, a state pool is constructed to store high-quality intermediate states obtained during historical training. The corresponding state is added to the state pool only when the reconstruction error of the geometric features of the robot's hand relationship is less than a preset threshold and the target object is in a stable and operable state. When a training round ends, or the target object falls, the operation fails, or the maximum number of steps is reached, the environment is reset. Each time the environment is reset, an intermediate state is sampled from the state pool as the initial state.

2. The method for skill transfer of a dexterous robotic arm based on relational geometry priors according to claim 1, characterized in that, The latent representation learning network is a variational autoencoder, which includes an encoder and a decoder. The encoder takes as input the geometric feature representation of the hand and outputs the latent geometric prior that represents the rules of human hand coordination. The decoder reconstructs the geometric feature representation of the hand based on the latent geometric prior.

3. The method for skill transfer of a dexterous robotic arm based on relational geometry priors according to claim 2, characterized in that, The latent representation learning network is pre-trained using human hand demonstration trajectory data; during the pre-training process, the total reconstruction loss and KL divergence regularization term are used for joint training; The total reconstruction loss is represented by a weighted average of the homoscedastic uncertainty of the directional reconstruction loss and the length reconstruction loss of each relation operation vector.

4. The method for skill transfer of a dexterous robotic arm based on relational geometry priors according to claim 1, characterized in that, The reinforcement learning environment for constructing the dexterous manipulator operation task is to model the dexterous manipulator operation task as a Markov decision process. The input of the policy network includes task state information, the latent representation corresponding to the current hand state of the robot, the latent representation corresponding to the reference demonstration state, and the similarity measure between the latent representations of the current hand state of the robot and the reference demonstration. The similarity measure is represented by cosine similarity. The policy network outputs the target position command of each drive joint of the dexterous manipulator. The task status information includes the target object's pose, target object's velocity, object's geometric features, and the relative distance information from the mechanical fingertip to the target object; The latent representation is obtained through encoding via a latent representation learning network.

5. A dexterous manipulator skill transfer device based on relational geometry priors, comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that, When the processor executes the executable code, it implements a method for transferring dexterous manipulator operation skills based on relational geometric priors as described in any one of claims 1-4.

6. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements a method for transferring dexterous manipulator operation skills based on relational geometric priors as described in any one of claims 1-4.

Images