Mobile mechanical arm control method and system based on deep meta-imitation learning

Abstract

The invention belongs to the technical field of mechanical arm control, and discloses a mobile mechanical arm control method based on depth element imitation learning, which introduces a depth element imitation learning (DMIL) algorithm, and the algorithm combines depth element learning and imitation learning (IL) technologies to improve the adaptability and efficiency of a mobile mechanical arm in such tasks. The expert trajectory is elaborately made and carefully checked in a AppeliaSim simulation environment, and the interaction of the real world is evaluated at the same time. The position and the direction of the button lock are determined by adopting a deep learning 6D attitude estimation method, and key operation reference points are visually identified. In the whole IL stage, through integration of resistance inverse reinforcement learning (AIRL) and variable impedance control (VIC) technologies, a strategy for controlling operation of the mechanical arm is enhanced. The expert-guided trajectory and tangible force feedback data from real-world settings enhance this.

Description

Technical Field

[0001] The present invention belongs to the technical field of robot arm control, and in particular, relates to a mobile robot arm control method and system based on deep meta-imitation learning. Background Art

[0002] In recent years, deep learning technology has become a catalyst for industrial applications, especially in the automation system of Industry 5.0. The rapid development of robotics and automation technology has led to its increasing application in power equipment inspection scenarios, facilitating the autonomous maintenance and inspection of essential power equipment. However, these intelligent operations require improved operational capabilities of inspection robots. Therefore, integrating devices with operational functions, such as robotic arms, has become a key focus of smart substation inspection. These solutions aim to reduce reliance on human maintenance personnel, quickly manage emergency failures to reduce downtime, maintain power systems, and replace personnel in dangerous conditions to ensure their safety. Robots face considerable challenges in complex and dynamic power equipment inspection environments, such as managing large amounts of data, limited task generalization, and strict safety and stability requirements.

[0003] In complex environments, robots need to accurately identify the position and orientation of the surrounding environment and objects in order to accurately perform tasks. However, in dynamically changing environments, traditional perception methods often fail to provide sufficient accuracy and reliability. Deep learning methods, such as the YOLO (You Only Look Once) family, have improved the accuracy and reliability of perception in challenging scenarios through real-time object detection and network modifications tailored to industrial needs. Hao et al. proposed a streamlined multi-target detection model CaSnLi-YOLO based on YOLOv5 for real-time detection of personnel and machinery on construction sites. Using GhostNet, coordinated attention mechanism, DWConv and C3Ghost technology, the model improves detection accuracy while reducing parameter count, making it very suitable for deployment on small edge devices. Wang et al. developed an enhanced YOLOv7 model based on visual attention fusion to analyze the state of spring locks in substations, which improved accuracy and efficiency. For pose estimation, Wang et al. proposed DenseFusion, which improves the accuracy and speed of 6D target pose estimation by iteratively improving pixel-level features of RGB-D images. Hu et al. proposed a new 6D pose estimation method through 3D to 2D correspondence, which is convenient for end-to-end training and further improves the performance.

[0004] Only with such detailed task analysis and careful path planning can precise robotic manipulation be performed through effective motion control. In practical applications, calibration and modeling will inevitably involve certain errors, which will cause deviations in the trajectory of the robot end effector. These errors may not affect work in free space, but even small deviations can generate large contact forces in direct contact with the environment - which may damage the equipment and increase danger during operation. Therefore, according to the literature, compliant control technology is of primary significance. Compliance control can be passive or active. Passive compliance control relies entirely on the physical state of the machine, while active compliance control dynamically adjusts control parameters according to changes in the environment. Anand et al. explored ways to improve the interaction of robotic systems with real-world environments by incorporating compliance. They proposed a deep model-predictive variable impedance control (VIC) that combines VIC with model predictive control. The controller obtains flexible operation by learning a general Cartesian impedance model and modifying the impedance parameters in the model through a model predictive control framework. Another method is DEVILC, which utilizes a model-based reinforcement learning framework. It also uses Gaussian process regression in iterative learning of system dynamics and optimizes a neural network policy to adapt to the robot's impedance. Both methods were experimentally validated on a Franka Emika Panda robotic arm, showing the ability to perform complex manipulation tasks.

[0005] Imitation learning (IL) has been shown to be one of the strongest methods for robots to learn complex tasks based on observing and imitating expert behaviors, including motion trajectory planning and decision-making in dynamic environments. In fact, robots can acquire necessary skills without explicit programming. However, these traditional imitation learning methods, such as behavior cloning (BC), usually suffer from high data requirements and poor generalization when the number of expert demonstration data is limited or there is a change in the distribution of tasks. The methods proposed by researchers include generative adversarial imitation learning (GAIL) and adversarial inverse reinforcement learning (AIRL). Li et al. proposed a safety-aware version of AIRL (S-AIRL), which not only encodes a safety critic to evaluate the safety of a behavior, but also adds additional penalty terms in the loss function to prevent high-risk behaviors. Wang et al. proposed a new framework Meta-AIRL, which can simultaneously learn policies and reward functions from existing training tasks through adversarial learning and inverse reinforcement learning mechanisms, and adapt to new tasks using a meta-learning framework. However, the computational complexity and dependence on high-quality training data are not considered. In practical applications, only a small part of the input data is labeled and is usually limited; therefore, the trained model always needs to adapt to new tasks with few samples.

[0006] The application of robots in complex environments requires strong adaptability and generalization. First, the amount of data in complex environments usually exceeds the ability of robots to process high-dimensional data, resulting in increasing difficulties in perception and decision-making. Second, in the structure of advanced robotics, existing robotics cannot guarantee good performance in achieving new tasks or in dynamically changing environments. In addition, the execution of tasks requires high security and stability; any operational errors may lead to serious consequences. In recent years, imitation learning has achieved some results in such problems, but it still has limitations. Traditional methods usually rely on a large amount of demonstration data from experts, which is time-consuming, costly, and poorly adaptable, making it difficult to cover all scenarios. It also requires high-quality data, and noisy data may bring inaccurate policy learning processes; therefore, it increases operational risks. Therefore, in order to improve the adaptability and generalization of robots in complex environments, some improvements to existing methods are needed.

[0007] Through the above analysis, the problems and defects of the prior art are as follows:

[0008] Traditional methods usually rely on a large amount of demonstration data from experts, which is time-consuming, costly, and poorly adaptable, making it difficult to cover all scenarios. It also requires high-quality data, while noisy data may lead to inaccurate policy learning processes; therefore, it increases operational risks. Summary of the invention

[0009] In view of the problems existing in the prior art, the present invention provides a mobile robotic arm control method based on deep meta-imitation learning.

[0010] The present invention is implemented in this way: a mobile robot arm control method based on deep meta-imitation learning includes:

[0011] Step 1: Introduce the deep meta-imitation learning DIL algorithm to combine deep meta-learning with imitation learning IL technology;

[0012] Step 2: Use a deep learning 6D posture estimation method to determine the position and orientation of the button lock and intuitively identify key operating reference points;

[0013] Step 3, in the whole IL stage, the adversarial inverse reinforcement learning AIRL and variable impedance control VIC technology are integrated;

[0014] Step 4, introduce latent embedding optimization (LEO) into the deep meta-learning architecture.

[0015] Furthermore, the 6D posture estimation method using deep learning:

[0016] (1) Visually guided trajectory based on deep learning;

[0017] (2) IL combined with VIC;

[0018] (3)DMIL.

[0019] Furthermore, the deep learning-based visual guidance trajectory:

[0020] a. Mobile robot end position and posture analysis

[0021] The electric cabinet door includes a door frame and a door leaf, and the button door lock includes a button and a handle; the coordinate system is defined as follows: {W} is the world coordinate system, fixed at the lower corner of the door frame; {O1} represents the lock head coordinate system, and {O2} represents the handle coordinate system, both of which are associated with the button door lock; all coordinate systems follow the right-hand rule;

[0022] {O_1} and {O_2} are attached to the button door lock, and their relative position relationship |O_1O_2|=O_12 remains unchanged; the homogeneous transformation matrix from {O_1} to {O_2} is expressed as:

[0023]

[0024] {W} is fixed on the door frame, and {O_1} is fixed on the cabinet door; their relative posture changes with the cabinet door rotation angle θ_o; assuming that the origin O_1 of the {O_1} coordinate system to the door hinge is w_o and the height is z_o, then the coordinates ^W O_1 (x_o, y_o, z_o) are:

[0025] W O1(x o ,y o ,z o )= W O1(w o sinθ o ,w o cosθ o ,z o ) (2)

[0026] The homogeneous transformation matrix of the button lock's pressed position relative to {O_1} relative to {W} is:

[0027]

[0028] The unlocking trajectory includes pressing the button to lock and rotating the handle; after pressing the button to a preset distance along the x-axis, release the handle; the transformation matrix of {O2} in {W} is:

[0029]

[0030] In order to model the motion of the handle, a feature point P_s at the end of the handle is selected, whose motion follows a cylindrical arc with O_2 as the vertex and the handle length l as the generatrix; to depict this arc, assume that the bounce angle of the handle is The coordinates of P_s in {O_2} are:

[0031]

[0032] The unlocking trajectory "T"_1 rotating around the x-axis is represented by the arc in {O_2}}:

[0033]

[0034] To open the door, you need to rotate the door around the door hinge and unlock it ( Figure 1 ); Assume that {W} is located at the bottom of the door hinge and is axially aligned with {O1} and {O2}; when the handle is pulled, the cabinet door rotates clockwise along the {W}z-axis; the opening motion is represented by the trajectory of the handle end point P_s;

[0035] b. Visually guided trajectory generation

[0036] The visual guidance system uses the YOLOv8 instance segmentation algorithm to accurately identify target objects and perform real-time image segmentation in the complex environment of the substation; the system is compatible with the capture of high-precision depth and image data. Realsense TM The D435 depth camera is used in conjunction with the robot to accurately determine the position and posture of key components that the robot arm needs to operate.

[0037] Furthermore, the IL is combined with the VIC:

[0038] a. Contact force model between the mobile robot and the door lock environment

[0039] The impedance model formula of the mobile robot in the task space is:

[0040]

[0041] As shown in Equation 8, x_e=x-x_d represents the position error between the actual position x and the desired position x_d; Mr is the inertia matrix, Dr is the damping matrix, and Kr is the stiffness matrix; for an n-degree-of-freedom robot, these are n×n positive definite symmetric matrices, which are usually assumed to be positive definite diagonal matrices; F_e is the contact force of the door lock on the end effector of the robot; the one-dimensional impedance model formula in the joint space is expressed as:

[0042]

[0043] Where q, q˙ and q¨ are the vectors of generalized joint coordinates, angular velocity and angular acceleration respectively, and τ_e represents the external force / torque in the joint space. In addition, using the Jacobian matrix J(q) of the robot, the correlation between τ_e and F_e is as follows:

[0044] τe=J(q)TFe (10)

[0045] The dynamic equations of the mobile manipulator describe in detail the dynamic relationship between torque and state variables in different spaces, providing a basis for the force control of the manipulator. In the presence of contact impedance, the dynamics of the joint space is usually expressed by the Lagrangian equation:

[0046]

[0047] Where C represents the coupling matrix of Coriolis and centripetal forces; F is the friction force; G is the gravity load; where is the joint driving torque;

[0048] The door lock mechanism is modeled as a first-order spring system, the lock stiffness is represented by k_i, and the contact pressure of the lock is:

[0049] f e =k i (x o -x c ) (12)

[0050] Where x_o is the natural position of the door lock contact surface when unloaded, f_e can be measured using a six-axis force sensor in the actual system; the expected tracking force is represented by f_d, and the difference between the expected tracking force and the actual contact force is expressed as:

[0051] Δf=f e -f d (13)

[0052] b. Mobile Robotic Arm VIC

[0053] VIC parameters K and D are further introduced; VIC adjusts stiffness K and damping D to find the optimal values ​​that ensure the compliance and stability of the arm in different tasks;

[0054] For the target impedance, the input is the force deviation Δf, and the output is the trajectory deviation e(t); the second-order transfer function of the position-based impedance control model is:

[0055]

[0056] The model describes the dynamic trajectory deviation e(t) and force deviation Δf as a second-order differential system of “mass-damper-spring”:

[0057]

[0058] M_r, B_r and K_r represent the mass, damping and stiffness coefficient matrices of the target impedance model, respectively;

[0059] c. AIRL combined with VIC.

[0060] Further, the AIRL is combined with VIC:

[0061] AIRL consists of a generator G that tries to learn from expert behavior and a discriminator D that tries to distinguish between trajectories obtained from generated policies and actual expert trajectories. AIRL learns from the reward functions of expert behavior and then uses these rewards to guide the optimal control strategy of the robot arm to achieve precise door opening action.

[0062] In AIRL, the reward function r θ (s,a,K,D) is summarized as follows

[0063] r θ (s,a,K,D)=f(s,a)-λ1(KK t ) 2 -λ2(DD t ) 2 (16)

[0064] f(s,a)F(s,a) represents the basic reward function based on state and action to ensure that the position and posture of the robot arm meet the expected trajectory of opening the door button lock; K_t and D_t are the target values ​​of the damping coefficient and stiffness coefficient, which are determined by the system design requirements; λ_1 and λ_2 are regularization weights used to balance different aspects of the reward;

[0065] The discriminator D_φ and the generator π_θ have objective functions that are at the heart of adversarial training and aim to ensure that the generated policy closely replicates the expert behavior; the goal of the discriminator is to distinguish between the trajectories produced by the generator π_θ and the true trajectories from the expert π_E; its optimization objective is defined as:

[0066]

[0067] It consists of the log-likelihood of the generated trajectory and the log-likelihood of the expert trajectory, which improves the discriminator's ability to correctly distinguish between the generated and expert trajectories; the generator aims to generate trajectories similar to the expert's behavior; the optimization objective of the generator is:

[0068]

[0069] This means that the generator π_θ tries to minimize the discriminator The log-likelihood of is used to simulate expert behavior, effectively “fooling” the discriminator;

[0070] VIC allows the robot arm to dynamically change its stiffness and damping according to different interactions and task requirements, which is critical when performing high-precision operations (such as opening a push-button lock); Impedance control force formula:

[0071]

[0072] Where K and D represent the stiffness matrix and damping matrix respectively, p_“D” and p˙D represent the target position and velocity, and p and p˙ represent the current position and velocity;

[0073] Incorporating VIC parameters K and D into the AIRL framework, the optimized reward function is:

[0074]

[0075] λ_p and λ_d are weight factors for position and velocity changes, which aim to balance the trajectory learning objective and dynamics requirements.

[0076] Furthermore, the DMIL:

[0077] DMIL: Integrates two meta-learning techniques: AIRL and LEO;

[0078] The essence of trajectory generalization technology is to integrate the optimization goal of AIRL with the dynamic adjustment of LEO, and obtain the following optimization formula:

[0079]

[0080] Where γ is the discount factor, r_φ represents the reward function derived from the discriminator, and λ∥θ∥^2 acts as a regularization term to reduce model complexity and promote generalization;

[0081] In addition, the dynamic adjustment of LEO is given by the regularized formula:

[0082]

[0083] Here, λ_K and λ_D are regularization parameters that regulate the adjustment of stiffness K and damping D, ensuring that the performance in new environments is not affected by previous learning;

[0084] Training loss calculation: For each task T_i, the loss function on the training set D^tr is defined as:

[0085]

[0086] where r_(θ, φ) represents the reward function derived by the discriminator, Q^π(s, a) represents the expected action value according to the current policy π; after calculating the training loss, the encoding vector z^' is updated by gradient descent:

[0087]

[0088] The modified z^' is then decoded using the decoder g_(φ_d)(z^') to extract the new parameters θ_i^':

[0089]

[0090] After completing a training cycle, calculate the loss on the validation set D^val:

[0091]

[0092] In successive rounds of training, the validation loss of each task is used to update the global parameters

[0093]

[0094] Another object of the present invention is to provide a mobile manipulator control research system based on deep meta-imitation learning, comprising:

[0095] The combination module is used to introduce the deep meta-imitation learning DIL algorithm and combine deep meta-learning with imitation learning IL technology;

[0096] A determination module is used to determine the position and orientation of the button lock using a deep learning 6D posture estimation method to intuitively identify key operating reference points;

[0097] Integration module for integration of adversarial inverse reinforcement learning (AIRL) and variable impedance control (VIC) techniques throughout the IL phase;

[0098] Introducing a module for introducing latent embedding optimization (LEO) in deep meta-learning architectures.

[0099] Another object of the present invention is to provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the mobile robot arm control method based on deep meta-imitation learning.

[0100] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to perform the steps of the mobile robotic arm control method based on deep meta-imitation learning.

[0101] Another object of the present invention is to provide an information data processing terminal, which is used to implement the mobile robotic arm control research system based on deep meta-imitation learning.

[0102] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0103] First, the present invention introduces a deep meta-imitation learning algorithm that combines deep meta-learning with imitation learning (IL) techniques to improve the adaptability and efficiency of mobile manipulators in such tasks. Expert trajectories are carefully crafted and carefully examined in the CoppeliaSim simulation environment, while real-world interactions are also evaluated. The present invention uses a deep learning 6D pose estimation method to determine the position and orientation of the button lock and intuitively identify key operating reference points. Throughout the IL phase, the strategy for controlling the operation of the manipulator is enhanced by the integration of adversarial inverse reinforcement learning (AIRL) and variable impedance control (VIC) techniques. This is enhanced by expert-guided trajectories and tangible force feedback data from real-world settings. In addition, the introduction of latent embedding optimization (LEO) in the deep meta-learning architecture promotes rapid adaptation to new tasks and significantly improves the generalization ability of the model. Experiments were conducted in the CoppeliaSim simulation environment and on a mobile manipulator platform, focusing on the identification process, trajectory formulation, and compliance management. To evaluate the effectiveness of the algorithm, three different push-button lock cabinet door opening tasks were performed, each with a different challenge level. Experimental results demonstrate that the mobile robot arm is able to skillfully locate and manipulate multiple push-button locks, highlighting its practicality and feasibility in advancing industrial automation and robotics.

[0104] Dual learning in simulation and real environment: By generating trajectories and performing initial learning in a simulation environment, and then performing fine-tuning and online learning in the actual environment, the mechanical data collected in the real environment can be effectively utilized to continuously adjust and optimize the operation strategy.

[0105] Integration of VIC and IL: This innovative approach combines VIC with IL to achieve optimal control under different physical interaction conditions by dynamically adjusting their impedance parameters (stiffness and damping).

[0106] Combining IL and LEO: Within the meta-learning framework, LEO is used to optimize embeddings in the latent space and quickly build initial parameters for new tasks. This combination enables the algorithm to quickly adapt to new button locking tasks, enhancing the generalization ability and execution speed of the robot arm.

[0107] Application of Deep Meta-Inverse Reinforcement Learning and Enhanced Multi-Task Adaptability: By introducing DMIL, we can learn and infer the reward function of complex tasks from expert operation trajectories. This method captures the key behaviors in expert operation, promotes efficient knowledge transfer in multi-task learning, and significantly improves the ability of the robot arm to select the most appropriate operation strategy and force control when operating different types of button locks.

[0108] Second, as auxiliary evidence of the inventiveness of the claims of the present invention, it is also reflected in the following important aspects:

[0109] (1) The expected benefits and commercial value of the technical solution of the present invention after transformation are:

[0110] The integration of the deep meta-imitation learning (DMIL) algorithm and the variable impedance control (VIC) technology proposed in this paper is not only innovative in the field of academic research, but also has great commercial value in industrial applications. Its expected benefits are mainly reflected in the following aspects:

[0111] Improved automated production efficiency: By dynamically adapting to button lock operation tasks under different physical conditions, the mobile robot arm can achieve more efficient and precise operations, greatly reducing the need for human intervention and improving the level of industrial automation.

[0112] Reduction of maintenance and operating costs: This solution can significantly reduce maintenance time and failure rate, and reduce the operation and maintenance costs of equipment through deep learning 6D posture estimation and inverse reinforcement learning optimization control strategy.

[0113] Wide range of industry application potential: Due to the advantages of the present invention in multi-task learning and rapid adaptation, its application scope is not limited to power substations, but can also be widely used in automobile manufacturing, medical equipment operation, precision instrument assembly and other fields with high operational requirements, and has extremely high market potential.

[0114] Economic benefits of knowledge transfer: The present invention can quickly adapt to the deployment requirements of new tasks through efficient transfer of knowledge in a multi-task environment, shorten the development cycle of new projects, and enhance the market competitiveness of enterprises.

[0115] (2) The technical solution of the present invention fills the technical gap in the industry at home and abroad:

[0116] This paper proposes a new mobile robot control scheme by integrating deep meta-learning, imitation learning, inverse reinforcement learning and variable impedance control technology, filling the gap in real-time control in complex physical environments in the existing technology. Specifically:

[0117] Most of the existing mobile robot arm technologies at home and abroad can only operate in fixed or simple environments and lack the intelligent control capabilities to adapt to complex physical conditions. The present invention overcomes the technical barriers of precise operation of the robot arm in a complex mechanical environment through real-time control of impedance parameters, solves the problem of efficient multi-task knowledge transfer, and fills the unresolved technical gap in this field.

[0118] There is relatively little research on the integrated application of inverse reinforcement learning technology and actual physical force control systems at home and abroad. The present invention innovatively combines the two and greatly improves the adaptability and control accuracy of the robotic arm in real scenarios by dynamically adapting to different interaction conditions, thus promoting technological breakthroughs in the industry.

[0119] (3) Whether the technical solution of the present invention solves the technical problems that people have been eager to solve but have not been able to solve successfully:

[0120] The present invention successfully solves several long-standing technical problems in the industry:

[0121] Efficient adaptation to complex tasks: In the field of industrial automation, how to make mobile robotic arms quickly adapt to multiple tasks, especially to operate flexibly in different mechanical environments, has always been an unresolved technical bottleneck. Through the application of DMIL algorithm and LEO, this invention successfully achieves rapid task adaptation and knowledge transfer, significantly improving the task adaptability of the robotic arm.

[0122] Intelligent control and force feedback optimization: Traditional robot arm control methods cannot dynamically adjust impedance parameters and are difficult to cope with changes in mechanical conditions in actual operations. This invention successfully solves this problem by integrating variable impedance control and imitation learning technology, and realizes intelligent control and real-time force feedback optimization of the robot arm under various physical interaction conditions, filling a gap that the industry has long been eager to solve.

[0123] (4) The present invention overcomes common technical prejudices in the industry, which is specifically reflected in the following aspects:

[0124] Limitations of traditional reinforcement learning: In the prior art, reinforcement learning technology is considered to be only suitable for training in specific tasks or environments, and it is difficult to apply it efficiently in a multi-task environment. This invention successfully breaks this limitation by introducing deep meta-imitation learning and inverse reinforcement learning, enabling the robot arm to achieve efficient transfer and reuse of knowledge in a variety of complex tasks.

[0125] Misconception of fixed impedance parameters: Traditional control theory often believes that impedance parameters should not be adjusted dynamically once set. However, the present invention achieves rapid adaptation to different physical interaction conditions by dynamically adjusting impedance parameters, successfully breaking through this mindset.

[0126] Breakthrough in the disconnection between academics and applications: In the existing technologies, many theoretical studies are difficult to apply directly to industrial environments. However, the present invention breaks the barrier between academics and practical applications through dual learning in the CoppeliaSim simulation environment and the real environment, achieving seamless connection from theory to actual industrial operations, and is extremely practical.

[0127] Third, the present invention proposes a mobile robot arm control method based on deep meta-imitation learning (DMIL), which solves many problems in the prior art of accurately controlling mobile robot arms in complex environments and has achieved significant technical progress in practical industrial applications. The specific problems solved and the progress made are as follows:

[0128] 1. Precision control issues in complex environments

[0129] Existing robot arm control methods often find it difficult to achieve high-precision operations in complex and changing environments, especially when facing dynamic targets (such as button locks, etc.), where accurate identification and precise execution of tasks are very challenging. The present invention introduces a deep meta-imitation learning algorithm, combined with a 6D posture estimation method and a visual guidance system, so that the robot arm can efficiently identify the position and direction of the operating object, greatly improving the ability to accurately control in complex environments. In addition, through the integration of AIRL and VIC technology, the robot arm's ability to respond to external force feedback is further enhanced, ensuring that it can operate stably under uncertain conditions.

[0130] 2. Insufficient adaptability and generalization of the control model

[0131] Traditional robotic arm control methods lack the ability to generalize in different tasks and environments. Even models trained with a large amount of data are difficult to adapt to new environments and operating requirements. This invention introduces latent embedding optimization (LEO) technology and deep meta-learning architecture, which can perform transfer learning between different tasks, quickly adapt to new tasks and environments, and significantly enhance the adaptability and generalization ability of the model. This dynamic optimization mechanism enables the robotic arm to efficiently handle a variety of different types of operating tasks, reducing the reliance on retraining and debugging.

[0132] 3. It is difficult to balance control accuracy and stability

[0133] While pursuing high-precision control, the existing technology often fails to take into account the stability of the system. In particular, during high-precision operations (such as unlocking a button lock), the operation may fail due to external interference or dynamic changes in the robotic arm itself. The present invention integrates variable impedance control (VIC) technology to dynamically adjust the stiffness and damping of the robotic arm, so that it can adjust its compliance and stability in real time according to task requirements. Combined with the adversarial training mechanism of AIRL, the present invention not only ensures the accuracy of the operation, but also significantly improves the stability of the system, especially showing excellent performance in fine operation tasks.

[0134] 4. Integration of imitation learning and reinforcement learning

[0135] Imitation learning technology often relies on data from expert behavior, while reinforcement learning continuously optimizes control strategies by interacting with the environment. Existing robot arm control methods make it difficult to effectively combine the two for practical applications. The present invention successfully combines imitation learning and reinforcement learning through deep integration of AIRL and VIC technologies, forming an intelligent control system that can learn from expert behavior and continuously optimize through real-time environmental feedback. This technological advancement has broad prospects in industrial applications, especially in the field of industrial automation that requires autonomous decision-making and high-precision operations, and can effectively improve production efficiency and operational accuracy.

[0136] In the field of mobile robotic arm control, the present invention solves the problem of precise operation in complex environments through an innovative deep meta-imitation learning algorithm and the integration of multiple advanced technologies, significantly improving the adaptability, stability and intelligence level of the system, and providing a solid technical foundation for the widespread application of robotic arms in industrial automation, medical equipment, service robots and other fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0137] Figure 1 It is a flow chart of a mobile robotic arm control method based on deep meta-imitation learning provided in an embodiment of the present invention.

[0138] Figure 2 It is a structural block diagram of a mobile robotic arm control research system based on deep meta-imitation learning provided by an embodiment of the present invention.

[0139] Figure 3 Schematic diagram of an unlocking button of a mobile mechanical arm provided in an embodiment of the present invention.

[0140] Figure 4 An embodiment of the present invention provides a demonstration of combining the YOLOv8 instance segmentation model with depth information in CoppeliaSim to estimate the 6D posture of the button lock and guide the robot arm trajectory generation graph.

[0141] Figure 5 It is an impedance model diagram of a button lock and a mobile mechanical arm provided in an embodiment of the present invention.

[0142] Figure 6 It is an impedance model diagram of a button lock and a mobile mechanical arm provided in an embodiment of the present invention.

[0143] Figure 7 It is an IL combined with VIC diagram provided by an embodiment of the present invention.

[0144] Figure 8 The mobile robot integration and simulation workflow provided by the embodiment of the present invention is as follows. (a) Mobile robot and components. (b) Software and integration workflow. (c) CoppeliaSim simulation. (d) Robot operation diagram supporting ROS.

[0145] Fig. 9 This is a diagram of a mobile robot arm task provided by an embodiment of the present invention: opening and locking a cabinet door.

[0146] Fig.10 A diagram of a visually guided mobile robotic arm for estimating push-lock position and posture using a calibrated depth camera provided by an embodiment of the present invention.

[0147] Fig.11 It is an example segmentation and Aruco posture estimation verification diagram of three different button locks provided by an embodiment of the present invention.

[0148] Fig.12 It is a robot arm motion control diagram using the CoppeliaSim trajectory in RViz provided by an embodiment of the present invention.

[0149] Fig.13 1 is a graph of reward scores and training iterations of three algorithms provided in embodiments of the present invention.

[0150] Fig.14 It is a three-key lock trajectory generation and force feedback data diagram based on DMIL provided in an embodiment of the present invention.

[0151] Fig.15 It is a process diagram of a mobile robot using a DMRL algorithm to open various button locks provided by an embodiment of the present invention.

[0152] Fig.16 Graph showing average success rates of different key lock control methods provided by embodiments of the present invention.

[0153] Fig.17 It is a schematic diagram of the door opening process of the robotic arm provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0154] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0155] like Figure 1 As shown, a mobile mechanical arm control method based on deep meta-imitation learning provided by an embodiment of the present invention comprises the following steps:

[0156] S101, introduces the deep meta-imitation learning DMIL algorithm, combining deep meta-learning with imitation learning IL technology;

[0157] S102, uses a deep learning 6D posture estimation method to determine the position and orientation of the button lock, and intuitively identifies key operating reference points;

[0158] S103, integration of adversarial inverse reinforcement learning AIRL and variable impedance control VIC technology throughout the IL stage;

[0159] S104, introducing latent embedding optimization (LEO) in deep meta-learning architecture.

[0160] The present invention provides a mobile robot arm control method based on deep meta-imitation learning, which optimizes the autonomous control capability of the robot arm, especially the precise operation in complex environments, by combining deep learning, imitation learning and reinforcement learning. The method first introduces the deep meta-imitation learning (DMIL) algorithm, which integrates deep meta-learning with imitation learning technology, so that the robot arm can quickly learn complex tasks from limited training data. This algorithm architecture not only improves the adaptability and learning efficiency of the system, but also provides a more intelligent decision-making basis for subsequent operations.

[0161] In step S102, the system uses a deep learning 6D posture estimation method to accurately determine the position and orientation of the button lock. This step uses a deep neural network to locate objects in space in three dimensions and identify key reference points for operations, so that the robot arm can accurately perform the operation task. This posture estimation-based approach provides important visual feedback for the precise operation of the robot arm, ensuring efficient positioning and precise control in complex environments.

[0162] Throughout the imitation learning (IL) phase, the system introduced adversarial inverse reinforcement learning (AIRL) and variable impedance control (VIC) technology. AIRL technology helps the robot arm better understand the task objectives by simulating human behavior patterns, while VIC technology provides flexible impedance adjustment to ensure that the robot arm can cope with different force feedback conditions during operation. This technology integration greatly enhances the system's ability to operate in uncertain environments, making the robot arm more stable and accurate when performing fine movements.

[0163] In addition, the introduction of latent embedding optimization (LEO) in the deep meta-learning architecture further improves the generalization ability of the model. LEO technology optimizes the latent space, allowing the robot arm to transfer learning between different tasks and reduce the dependence on large amounts of data. This optimization not only improves the training efficiency of the model, but also enhances the system's ability to quickly adapt to new tasks, thereby improving the overall control accuracy and robustness.

[0164] The 6D posture estimation method using deep learning provided by the embodiment of the present invention:

[0165] (1) Visually guided trajectory based on deep learning;

[0166] (2) IL combined with VIC;

[0167]

[0168] The deep learning-based visual guidance trajectory provided by the embodiment of the present invention:

[0169] a. Mobile robot end position and posture analysis

[0170] The electric cabinet door includes a door frame and a door leaf, and the button door lock includes a button and a handle; the coordinate system is defined as follows: {W} is the world coordinate system, fixed at the lower corner of the door frame; {O1} represents the lock head coordinate system, and {O2} represents the handle coordinate system, both of which are associated with the button door lock; all coordinate systems follow the right-hand rule;

[0171] {O_1} and {O_2} are attached to the button door lock, and their relative position relationship |O_1O_2|=O_12 remains unchanged; the homogeneous transformation matrix from {O_1} to {O_2} is expressed as:

[0172]

[0173] {W} is fixed on the door frame, and {O_1} is fixed on the cabinet door; their relative posture changes with the cabinet door rotation angle θ_o; assuming that the origin O_1 of the {O_1} coordinate system to the door hinge is w_o and the height is z_o, then the coordinates ^W O_1 (x_o, y_o, z_o) are:

[0174] W O1(x o ,y o ,z o )= W 01(w o sinθ o ,w o cosθ o ,z o ) (2)

[0175] The homogeneous transformation matrix of the button lock's pressed position relative to {O_1} relative to {W} is:

[0176]

[0177] The unlocking trajectory includes pressing the button to lock and rotating the handle; after pressing the button to a preset distance along the x-axis, release the handle; the transformation matrix of {O2} in {W} is:

[0178]

[0179] In order to model the motion of the handle, a feature point P_s at the end of the handle is selected, whose motion follows a cylindrical arc with O_2 as the vertex and the handle length l as the generatrix; to depict this arc, assume that the bounce angle of the handle is The coordinates of P_s in {O_2} are:

[0180]

[0181] The unlocking trajectory "T"_1 rotating around the x-axis is represented by the arc in {O_2}}:

[0182]

[0183]

[0184] To open the door, you need to rotate the door around the door hinge and unlock it ( Figure 1 ); Assume that {W} is located at the bottom of the door hinge and is axially aligned with {O1} and {O2}; when the handle is pulled, the cabinet door rotates clockwise along the {W}z-axis; the opening motion is represented by the trajectory of the handle end point P_s;

[0185] b. Visually guided trajectory generation

[0186] The visual guidance system uses the YOLOv8 instance segmentation algorithm to accurately identify target objects and perform real-time image segmentation in the complex environment of the substation; the system is compatible with the capture of high-precision depth and image data. Realsense TM The D435 depth camera is used in conjunction with the robot to accurately determine the position and posture of key components that the robot arm needs to operate.

[0187] The IL combined with VIC provided in the embodiment of the present invention:

[0188] a. Contact force model between the mobile robot and the door lock environment

[0189] The impedance model formula of the mobile robot in the task space is:

[0190]

[0191] As shown in Equation 8, x_e=x-x_d represents the position error between the actual position x and the desired position x_d; Mr is the inertia matrix, Dr is the damping matrix, and Kr is the stiffness matrix; for an n-degree-of-freedom robot, these are n×n positive definite symmetric matrices, which are usually assumed to be positive definite diagonal matrices; F_e is the contact force of the door lock on the end effector of the robot; the one-dimensional impedance model formula in the joint space is expressed as:

[0192]

[0193] Where q, q˙ and q¨ are the vectors of generalized joint coordinates, angular velocity and angular acceleration respectively, and τ_e represents the external force / torque in the joint space. In addition, using the Jacobian matrix J(q) of the robot, the correlation between τ_e and F_e is as follows:

[0194] τe=J(q)TFe (10)

[0195] The dynamic equations of the mobile manipulator describe in detail the dynamic relationship between torque and state variables in different spaces, providing a basis for the force control of the manipulator. In the presence of contact impedance, the dynamics of the joint space is usually expressed by the Lagrangian equation:

[0196]

[0197] Where C represents the coupling matrix of Coriolis and centripetal forces; F is the friction force; G is the gravity load; where is the joint driving torque;

[0198] The door lock mechanism is modeled as a first-order spring system, the lock stiffness is represented by k_i, and the contact pressure of the lock is:

[0199] f e =k i (x o -x c ) (12)

[0200] Where x_o is the natural position of the door lock contact surface when unloaded, f_e can be measured using a six-axis force sensor in the actual system; the expected tracking force is represented by f_d, and the difference between the expected tracking force and the actual contact force is expressed as:

[0201] Δf=f e -f d (13)

[0202] b. Mobile Robotic Arm VIC

[0203] VIC parameters K and D are further introduced; VIC adjusts stiffness K and damping D to find the optimal values ​​that ensure the compliance and stability of the arm in different tasks;

[0204] For the target impedance, the input is the force deviation Δf, and the output is the trajectory deviation e(t); the second-order transfer function of the position-based impedance control model is:

[0205]

[0206] The model describes the dynamic trajectory deviation e(t) and force deviation Δf as a second-order differential system of “mass-damper-spring”:

[0207]

[0208] M_r, B_r and K_r represent the mass, damping and stiffness coefficient matrices of the target impedance model, respectively;

[0209] c. AIRL combined with VIC.

[0210] The AIRL provided in the embodiment of the present invention is combined with the VIC:

[0211] AIRL consists of a generator G that tries to learn from expert behavior and a discriminator D that tries to distinguish between trajectories obtained from generated policies and actual expert trajectories. AIRL learns from the reward functions of expert behavior and then uses these rewards to guide the optimal control strategy of the robot arm to achieve precise door opening action.

[0212] In AIRL, the reward function r θ (s,a,K,D) is summarized as follows

[0213] r θ (s,a,K,D)=f(s,a)-λ1(K-Kt) 2 -λ2(DD t ) 2 (16)

[0214] f(s,a)F(s,a) represents the basic reward function based on state and action to ensure that the position and posture of the robot arm meet the expected trajectory of opening the door button lock; K_t and D_t are the target values ​​of the damping coefficient and stiffness coefficient, which are determined by the system design requirements; λ_1 and λ_2 are regularization weights used to balance different aspects of the reward;

[0215] The discriminator D_φ and the generator π_θ have objective functions that are at the heart of adversarial training and aim to ensure that the generated policy closely replicates the expert behavior; the goal of the discriminator is to distinguish between the trajectories produced by the generator π_θ and the true trajectories from the expert π_E; its optimization objective is defined as:

[0216]

[0217] It consists of the log-likelihood of the generated trajectory and the log-likelihood of the expert trajectory, which improves the discriminator's ability to correctly distinguish between the generated and expert trajectories; the generator aims to generate trajectories similar to the expert's behavior; the optimization objective of the generator is:

[0218]

[0219] This means that the generator π_θ tries to minimize the discriminator The log-likelihood of is used to simulate expert behavior, effectively “fooling” the discriminator;

[0220] VIC allows the robot arm to dynamically change its stiffness and damping according to different interactions and task requirements, which is critical when performing high-precision operations (such as opening a push-button lock); Impedance control force formula:

[0221]

[0222] Where K and D represent the stiffness matrix and damping matrix respectively, p_“D” and p˙D represent the target position and velocity, and p and p˙ represent the current position and velocity;

[0223] Incorporating VIC parameters K and D into the AIRL framework, the optimized reward function is:

[0224]

[0225] λ_p and λ_d are weight factors for position and velocity changes, which aim to balance the trajectory learning objective and dynamics requirements.

[0226] The DMIL provided in the embodiment of the present invention:

[0227] DMIL: Integrates two meta-learning techniques: AIRL and LEO;

[0228] The essence of trajectory generalization technology is to integrate the optimization goal of AIRL with the dynamic adjustment of LEO, and obtain the following optimization formula:

[0229]

[0230] Where γ is the discount factor, r_φ represents the reward function derived from the discriminator, and λ∥θ∥^2 acts as a regularization term to reduce model complexity and promote generalization;

[0231] In addition, the dynamic adjustment of LEO is given by the regularized formula:

[0232]

[0233] Here, λ_K and λ_D are regularization parameters that regulate the adjustment of stiffness K and damping D, ensuring that the performance in new environments is not affected by previous learning;

[0234] Training loss calculation: For each task T_i, the loss function on the training set D^tr is defined as:

[0235]

[0236] where r_(θ, φ) represents the reward function derived by the discriminator, Q^π(s, a) represents the expected action value according to the current policy π; after calculating the training loss, the encoding vector z^' is updated by gradient descent:

[0237]

[0238] The modified z^' is then decoded using the decoder g_(φ_d)(z^') to extract the new parameters θ_i^':

[0239]

[0240] After completing a training cycle, calculate the loss on the validation set D^val:

[0241]

[0242] In successive rounds of training, the validation loss of each task is used to update the global parameters

[0243]

[0244] like Figure 2 As shown, a mobile manipulator control research system based on deep meta-imitation learning provided by an embodiment of the present invention includes:

[0245] The combination module is used to introduce the deep meta-imitation learning DIL algorithm and combine deep meta-learning with imitation learning IL technology;

[0246] A determination module is used to determine the position and orientation of the button lock using a deep learning 6D posture estimation method to intuitively identify key operating reference points;

[0247] Integration module for integration of adversarial inverse reinforcement learning (AIRL) and variable impedance control (VIC) techniques throughout the IL phase;

[0248] Introducing a module for introducing latent embedding optimization (LEO) in deep meta-learning architectures.

[0249] Another object of the present invention is to provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the mobile robot arm control method based on deep meta-imitation learning.

[0250] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to perform the steps of the mobile robotic arm control method based on deep meta-imitation learning.

[0251] Another object of the present invention is to provide an information data processing terminal, which is used to implement the mobile robotic arm control research system based on deep meta-imitation learning.

[0252] The present invention is specifically implemented:

[0253] 1. This invention will propose an algorithm: DMIL combines the advantages of deep meta-learning and deep meta-learning; therefore, it allows the system to quickly learn new tasks with limited data. It implements deep learning for 6D pose estimation and utilizes the state-of-the-art YOLOv8 instance segmentation algorithm, which enables precise visual guidance. In addition, VIC technology is embedded to enhance the compliance of its robotic arm. The system utilizes various other techniques, such as Latent Embedding Optimization (LEO), which enhances the system's adaptability and generalization capabilities to new task environments.

[0254] The contributions of the present invention are as follows:

[0255] Dual learning in simulation and real environment: By generating trajectories and performing initial learning in a simulation environment, and then performing fine-tuning and online learning in the actual environment, the mechanical data collected in the real environment can be effectively utilized to continuously adjust and optimize the operation strategy.

[0256] Integration of VIC and IL: This innovative approach combines VIC with IL to achieve optimal control under different physical interaction conditions by dynamically adjusting their impedance parameters (stiffness and damping).

[0257] Combining IL and LEO: Within the meta-learning framework, LEO is used to optimize embeddings in the latent space and quickly build initial parameters for new tasks. This combination enables the algorithm to quickly adapt to new button locking tasks, enhancing the generalization ability and execution speed of the robot arm.

[0258] Application of Deep Meta-Inverse Reinforcement Learning and Enhanced Multi-Task Adaptability: By introducing DMIL, we can learn and infer the reward function of complex tasks from expert operation trajectories. This method captures the key behaviors in expert operation, promotes efficient knowledge transfer in multi-task learning, and significantly improves the ability of the robot arm to select the most appropriate operation strategy and force control when operating different types of button locks.

[0259] The structure of the present invention is organized as follows:

[0260] 2. Related Work

[0261] 2.1 Visual Guidance

[0262] The role of vision-guided systems in robotics is becoming increasingly important in the development of Industry 5.0, where vision is becoming one of the cornerstones for the development of better perception and interaction in complex interactions of robots. These robots use cameras and other imaging devices as eyes, allowing for powerful analysis, localization, navigation, and precise manipulation of the robots. Of all the adopted technologies, the YOLO algorithm stands out the most in terms of fast and accurate object detection. This forms the basis for the YOLOv8 feature extension, with advanced object segmentation, integrated depth cameras, and providing a robotic arm with 6D position and pose estimation to guide accurate manipulation. It is this integration that provides the robot with detailed 3D spatial information, which greatly improves precision and safety during operation; therefore, it ensures greater efficiency and reliability when working in complex environments.

[0263] 2.2 Inverse Reinforcement Learning

[0264] IL is one of the core areas of robotics, of which BC is the most basic form. It is common to learn from experts by imitating their action trajectories, allowing the robot to simulate the expert's decision-making process. However, it does not work when faced with a new environment. Subsequently, advanced methods such as GAIL and AIRL were proposed. GAIL uses a generator to imitate the behavior of experts and a discriminator to distinguish between different behaviors, generating high-quality decision data without additional labeling. AIRL further develops the concept of adversarial learning. While learning the strategy itself, it also learns the reward function through a specially designed discriminator. This helps AIRL generalize more across different environments because it not only learns the behavior pattern, but also understands the intention behind the behavior, in other words, the reward structure. In addition, AIRL has achieved particular success in cross-task learning and operation in complex high-dimensional action spaces.

[0265] 2.3 Variable impedance control

[0266] Compliance control is very important in the interaction between robots and their operating environment. Advances in impedance control technology have introduced VIC, allowing robots to dynamically adjust their stiffness, damping, and inertia parameters in response to changes in the operating environment. By adopting VIC technology, robots can interact with the environment safely and precisely, and are more adaptable and flexible in operation when working conditions and unexpected situations vary. In particular, VIC technology is well suited for complex and dynamic industrial environments, such as automated production lines, collaborative robot applications, and many other high-mobility, environmentally sensitive scenarios. When a robot with VIC comes into contact with the environment, its mechanical properties automatically change to effectively respond to unexpected changes in external forces, which are common in industrial operations or precision assembly. In addition, VIC technology will also enable robots to respond to collisions or collisions, unstable surfaces, and other interactions by self-adjusting their own impedance parameters, thereby reducing impact forces to protect the machine and manipulated objects from damage.

[0267] like Fig. 9 As shown, the task involves a specially equipped cabinet door with a push button lock mechanism and a mobile robotic arm equipped with the necessary tools and sensors. The cabinet door features a push button lock that is unlocked by pressing the button itself, thereby releasing the handle, which can be turned and pulled to open the door. The robotic arm, in turn, will move the clamp and tactile sensors to perform fine movements; it should be equipped with a camera system for identifying the lock button and door handle. The robotic arm completes this task through several operating steps, including: it identifies the position of the lock button by using a camera system; then, with the corresponding force, the arm moves to the correct position to press the button. After that, in order to accurately grasp and turn the handle, the arm adjusts its position and angle according to the generated trajectory. Finally, the door handle should be pulled with sufficient force to open the door, and its operation should not cause any kind of damage to the cabinet door or any other equipment.

[0268] In order to verify the generalization ability of the DMIL method, an experiment was designed using only one type of cabinet door button lock and different types of operation tasks. Taking a certain model of cabinet door button lock as an example, the robot was trained by integrating multimodal sensor data to learn the button position, pressing force, clamping angle and rotation force. Then, a generalization test was conducted on two types of button locks to record the performance of the robot arm in identifying the button position, applying appropriate pressing force, adjusting the clamping angle, and the success rate of unlocking and opening the door.

[0269] 4.2. Visual Guidance and Expert Trajectory Creation

[0270] a. Visual guidance based on YOLOv8 combined with depth camera

[0271] This experimental setup aims to verify the effectiveness of 6D pose estimation for a toggle switch by combining the YOLOv8 instance segmentation algorithm with a depth camera. RealSense TM The depth camera D435 is used to capture high-precision depth and image data. The image recognition model used is YOLOv8 for real-time image segmentation and object recognition.

[0272] like Fig.10 As shown, a calibrated depth camera D435 captures depth and color images of the button lock in various states. These images are manually annotated to identify and mark key components, such as the position and state of the button and handle on the lock. To simulate the variability encountered in actual operation, three different button locks are randomly placed in different positions, angles, and states. The annotated dataset is then used to train a YOLOv8 model, which aims to accurately identify and locate these components and determine their states. The model performs instance segmentation on the image and accurately determines the position of the button lock, thereby accurately cropping the area containing the button and handle. By processing the depth data from the camera, the pixel depth value is converted into a 3D spatial coordinate to estimate the 6D pose of each component. Finally, the trained model and pose estimation algorithm are integrated into the control system of the robotic arm and a comprehensive integration test is carried out to ensure accuracy and efficiency in real-world scenarios.

[0273] 3. Proposed Method

[0274] 3.1. Deep Learning-Based Vision-Guided Trajectory

[0275] a. Mobile robot end position and posture analysis

[0276] The electric cabinet door includes a door frame and a door leaf, and the button door lock includes a button and a handle. Figure 3 As shown in the figure, the definitions are as follows: {W} is the world coordinate system, fixed at the lower corner of the door frame; {O1} represents the lock head coordinate system, and {O2} represents the handle coordinate system, both of which are associated with the push button door lock. All coordinate systems follow the right-hand rule.

[0277] {O_1} and {O_2} are attached to the button door lock, and their relative position relationship |O_1O_2|=O_12 remains unchanged. The homogeneous transformation matrix from {O_1} to {O_2} is expressed as:

[0278]

[0279] {W} is fixed on the door frame, and {O_1} is fixed on the cabinet door; their relative posture changes with the cabinet door rotation angle θ_o. Assuming that the origin O_1 of the {O_1} coordinate system to the door hinge is w_o and the height is z_o, the coordinates ^W O_1 (x_o, y_o, z_o) are:

[0280]

[0281] The homogeneous transformation matrix of the button lock's pressed position relative to {O_1} relative to {W} is:

[0282]

[0283] The unlocking trajectory includes pressing the button to lock and rotating the handle. After pressing the button to a preset distance along the x-axis, release the handle. The transformation matrix of {O2} in {W} is:

[0284]

[0285] To model the motion of the handle, a feature point P_s at the end of the handle is selected, whose motion follows a cylindrical arc with O_2 as the vertex and the handle length l as the generatrix. To depict this arc, assume that the bounce angle of the handle is The coordinates of P_s in {O_2} are:

[0286]

[0287] The unlocking trajectory "T"_1 rotating around the x-axis is represented by the arc in {O_2}}:

[0288]

[0289] To open the door, you need to rotate the door around the door hinge and unlock it ( Figure 1 ). Assume {W} is located at the bottom of the door hinge and is axially aligned with {O1} and {O2}. When the handle is pulled, the cabinet door rotates clockwise along the {W}z axis. The opening motion is represented by the trajectory of the handle end point P_s.

[0290] b. Visually guided trajectory generation

[0291] In the field of automation and robotics, traditional trajectory teaching methods often rely on manual guidance, which is not only inefficient but also prone to inaccuracies and instabilities. To overcome these challenges, we developed a vision-based deep learning method for expert trajectory generation, using CoppeliaSim simulation software to optimize and automate the process. The method accurately replicates expert operations and can generate a variety of task scenarios and initial states, greatly enriching the dataset.

[0292] Figure 4 The YOLOv8 instance segmentation model is combined with depth information in CoppeliaSim to estimate the 6D pose of the button lock and guide the robot arm trajectory generation. The visual guidance system uses the YOLOv8 instance segmentation algorithm to accurately identify target objects and perform real-time image segmentation in the complex environment of the substation. The system is compatible with the capture of high-precision depth and image data. RealsenseTM The D435 depth camera is used in conjunction to accurately determine the position and posture of the key components that the robot arm needs to operate. Therefore, the system effectively generates the expert trajectory in CoppeliaSim. This method not only improves the accuracy of the expert trajectory, but also provides a solid foundation for the generation of the robot arm's operation trajectory.

[0293] 3.2.IL and VIC

[0294] In the task of opening the push-button lock on the electrical cabinet door of the substation, the operation of the robot arm is based on the trajectory generated in the simulation environment and must interact smoothly with the cabinet door. In order to avoid damage or improper handling during these interactions, active compliance control technology is used. The present invention adopts the VIC method to dynamically adjust the impedance parameters of the system at key interaction points to handle complex operating conditions and unforeseen external interference. The control strategy uses a six-axis force sensor installed at the end joint of the robot arm to monitor and adjust the changes in external force in real time and modify the position of the end effector.

[0295] a. Contact force model between the mobile robot and the door lock environment

[0296] The key to the interaction between the mobile robot arm and the button lock is to control the contact force. During the engagement, especially in the nonlinear transition area of ​​the collision, the key challenge comes from the inevitable collision and its transient, obvious nonlinear characteristics. For manipulation tasks such as activating the lock, the trajectory of the end of the arm and the button follows a predefined linear path corresponding to the direction of the pressure.

[0297] The assembly consisting of the mobile manipulator and the six-axis force sensor is conceptualized as a second-order "mass damper-spring" system, while the door lock unlocking mechanism is described as a first-order spring system. The corresponding impedance model is as follows Figure 5 shown.

[0298] like Figure 5 As shown, the impedance model formula of the mobile robot in the task space is:

[0299]

[0300] As shown in Equation 8, x_e = x-x_d represents the position error between the actual position x and the desired position x_d; Mr is the inertia matrix, Dr is the damping matrix, and Kr is the stiffness matrix. For an n-degree-of-freedom robot, these are n×n positive definite symmetric matrices, which are usually assumed to be positive definite diagonal matrices; F_e is the contact force of the door lock on the end effector of the robot. The one-dimensional impedance model formula in the joint space is expressed as:

[0301]

[0302] Where q, q˙ and q¨ are the vectors of generalized joint coordinates, angular velocity and angular acceleration respectively, and τ_e represents the external force / torque in the joint space. In addition, using the Jacobian matrix J(q) of the robot, the relationship between τ_e and F_e is as follows:

[0303] τe=J(q)TFe (10)

[0304] The dynamic equations of the mobile manipulator describe in detail the dynamic relationship between torque and state variables in different spaces, providing the basis for the force control of the manipulator. In the presence of contact impedance, the dynamics of the joint space is usually expressed by the Lagrangian equation:

[0305]

[0306] Where C represents the Coriolis and centripetal force coupling matrix; F is the friction force; G is the gravity load; and is the joint driving torque.

[0307] The door lock mechanism is modeled as a first-order spring system, the lock stiffness is represented by k_i, and the contact pressure of the lock is:

[0308] f e =k i (x o -x c ) (12)

[0309] Where x_o is the natural position of the door lock contact surface when unloaded, and f_e can be measured using a six-axis force sensor in the actual system. The expected tracking force is represented by f_d, and the difference between the expected tracking force and the actual contact force is expressed as:

[0310] Δf=f e -f d (13)

[0311] b. Mobile Robotic Arm VIC

[0312] In order to further improve the interactive mechanical performance of the robot arm in motion, VIC parameters K and D are further introduced. VIC adjusts the stiffness K and damping D to find the optimal values ​​that ensure the compliance and stability of the arm in different tasks.

[0313] For the target impedance, the input is the force deviation Δf, and the output is the trajectory deviation e(t). The second-order transfer function of the position-based impedance control model is:

[0314]

[0315] Figure 6 Impedance model of the push button lock and the moving robot arm.

[0316] The model describes the dynamic trajectory deviation e(t) and force deviation Δf as a second-order differential system of “mass-damper-spring”:

[0317]

[0318] Mr, B and K represent the mass, damping and stiffness coefficient matrices of the target impedance model, respectively. During the control of the mobile manipulator, VIC affects the execution of different tasks by changing K and D. The change in these values ​​helps to provide better compliance and stability for a variety of different tasks. High values ​​of K and D are suitable for tasks that require high stability and high accuracy, while lower values ​​are more suitable for tasks that rely on greater compliance and force feedback. Therefore, these parameters need to be studied in order to outline the optimal values ​​of K and D for different tasks.

[0319] c. Combination of AIRL and VIC

[0320] like Figure 7 The present invention combines the AIRL framework with VIC to train a robot arm to perform the task of opening a cabinet door button lock. Using this method, trajectory generation and flexible control strategy of the mobile robot arm are realized.

[0321] To be precise, it can be simulated in the CoppeliaSim environment first through a preliminary expert system, which prepares a series of operation trajectories, including but not limited to the position and posture data of the robotic arm, and the detailed action sequence when working with the door lock. During the execution of the task, the sensors installed on the end effector obtain force feedback data in real time. Information about contact force and torque is detected by these sensors. The vision strategy guides the robotic arm by using the on-board camera to identify the exact position and status of the button. The image processing algorithm of the vision system analyzes the status of the door, such as the position and type of the lock, to guide the robotic arm to accurately position and operate.

[0322] AIRL consists of a generator G that tries to learn from expert behavior and a discriminator D that tries to distinguish between trajectories obtained from the generated policy and actual expert trajectories. Based on the adversarial relationship between the generator and the discriminator, the performance is improved. Specifically, AIRL learns from the reward function of the expert behavior and then uses these rewards to guide the optimal control strategy of the robot arm to achieve precise door opening action.

[0323] In AIRL, the reward function r θ (s,a,K,D) is summarized as follows

[0324] r θ (s,a,K,D)=f(s,a)-λ1(K-Kt) 2 -λ2(DD t )2 (16)

[0325] f(s,a)F(s,a) represents the basic reward function based on state and action to ensure that the position and posture of the robot arm meet the expected trajectory of opening the door button lock. K_t and D_t are the target values ​​of the damping coefficient and stiffness coefficient, which are determined by the system design requirements. λ_1 and λ_2 are regularization weights used to balance different aspects of the reward.

[0326] The discriminator D_φ and the generator π_θ have objective functions that are at the heart of adversarial training and aim to ensure that the generated policy closely replicates the expert behavior. The goal of the discriminator is to distinguish between the trajectories produced by the generator π_θ and the true trajectories from the expert π_E. Its optimization objective is defined as:

[0327]

[0328] It consists of the log-likelihood of the generated trajectory and the log-likelihood of the expert trajectory, which improves the discriminator's ability to correctly distinguish between the generated and expert trajectories. The generator aims to generate trajectories similar to the expert behavior; the optimization objective of the generator is:

[0329]

[0330] This means that the generator π_θ tries to minimize the discriminator The log-likelihood of is used to simulate expert behavior, effectively “fooling” the discriminator.

[0331] VIC allows the robot arm to dynamically change its stiffness and damping according to different interactions and task requirements, which is critical when performing high-precision operations such as opening a push-button lock. Impedance control force formula:

[0332]

[0333] Among them, K and D represent the stiffness matrix and damping matrix respectively, p_“D” and p˙D represent the target position and velocity, and p and p˙ represent the current position and velocity.

[0334] Incorporating VIC parameters K and D into the AIRL framework, the optimized reward function is:

[0335]

[0336] λ_p and λ_d are weighting factors for position and velocity changes, designed to balance trajectory learning objectives and dynamics requirements. It explicitly attempts to leverage the strengths of the AIRL framework by inferring an implicit reward function from data collected from the behavior of an expert operator. We construct an adversarial network and iteratively optimize the reward function so that the derived control policy can replicate expert behavior with high accuracy. Combining AIRL with the VIC tuning strategy allows the impedance parameters on the manipulator to be modified in real time to adapt to the specific operating environment and the task at hand. The control strategy is a model-based control approach that changes the impedance parameters by approximating the interaction with the environment and the consequences of its manipulation.

[0337] 3.3.DMIL

[0338] This paper proposes a method to improve the adaptability and effectiveness of a mobile robot in various lock door tasks: DMIL. It integrates inverse reinforcement learning and meta-learning techniques: AIRL and Local Execution Optimizer (LEO). This innovative integration not only learns a deep reward function from limited expert behavior data, but also quickly adapts to new task environments through a meta-learning mechanism, optimizing and adjusting the control strategy of the robot.

[0339] The AIRL framework instantiates a generative adversarial network while inferring an implicit reward function from demonstrated expert behavior trajectories. Essentially, LEO is a meta-learning framework that focuses on quickly generalizing to new tasks with minimal data in a multi-task learning setting. This is important because it can quickly adapt to new button configurations and cabinet door types without extensive retraining of the robotic arm. LEO does this efficiently by optimizing local representations of the model parameters, which helps reduce learning time and improve generalization capabilities.

[0340] The essence of trajectory generalization technology is to integrate the optimization goal of AIRL with the dynamic adjustment of LEO, and obtain the following optimization formula:

[0341]

[0342] where γ is the discount factor, r_φ represents the reward function derived from the discriminator, and λ|θ|^2 acts as a regularization term to reduce model complexity and promote generalization.

[0343] In addition, the dynamic adjustment of LEO is given by the regularized formula:

[0344]

[0345] Here, λ_K and λ_D are regularization parameters that regulate the tuning of stiffness K and damping D, ensuring that performance in new environments is not affected by prior learning.

[0346] Training loss calculation: For each task T_i, the loss function on the training set D^tr is defined as:

[0347]

[0348] Where r_(θ, φ) represents the reward function derived by the discriminator, and Q^π(s, a) represents the expected action value according to the current policy π. After calculating the training loss, the encoding vector z^' is updated by gradient descent:

[0349]

[0350] The modified z^' is then decoded using the decoder g_(φ_d)(z^') to extract the new parameters θ_i^':

[0351]

[0352] After completing a training cycle, calculate the loss on the validation set D^val:

[0353]

[0354] In successive rounds of training, the validation loss of each task is used to update the global parameters

[0355]

[0356] By combining AIRL with LEO meta-learning technology, the present invention not only designs an effective robot control learning strategy, but also significantly improves the robot's generalization ability and ability to quickly adapt to new environments through dynamic operation adjustment. This method uses a deep reward function learned from expert behavior and an adversarial training mechanism to ensure that the behavior can be quickly adjusted to adapt to different operation requirements in unfamiliar situations. At the same time, the introduction of LEO meta-learning optimizes the learning trajectory, enabling the robot to quickly adapt to new tasks and reduce exposure to example data, which is crucial for environments where tasks change frequently in practical applications.

[0357] Table 1. DMIL algorithm

[0358]

[0359]

[0360] The DMIL algorithm combines meta-learning with IL for common structure discovery across tasks, while relying on LEO and VIC for better model adaptability and generalization. The goal of meta-learning is to enable the model to learn common initializations across multiple tasks so that it can quickly adapt to new tasks. DIL leverages the meta-learning framework by utilizing both inner and outer loops. The inner loop is used to learn task-specific policies, and the outer loop is used to update the latent space to enhance generalization across tasks.

[0361] LEO is the key to DMIL, encoding the generated task trajectories into latent space vectors and decoding them into initial policy parameters, which promotes the effective extraction of shared information between tasks. The latent space optimization of LEO is achieved by minimizing the loss function of the training task, as shown in step 26. This technique optimizes the latent representation across various tasks, achieving fast policy adjustment and enhanced adaptability. In DMIL, VIC dynamically changes the parameters K and D to change the interaction of its agent with its physical parts in various tasks. Combining VIC with the reward function in AIRL results in a model that can not only learn policies but also dynamically adjust the flexibility of physical interactions according to the environment. The discriminator is given by AIRL in step 11, while DMIL integrates VIC into the optimization of physical control policies during task execution, so that it can adapt to tasks more skillfully.

[0362] DIL embodies a technique that combines three different techniques: AIRL, VIC, and LEO. The approach is specifically designed to provide greater adaptability and improved generalization capabilities of robotic systems for specific tasks, such as door opening. The algorithm extracts task instances from a meta-training dataset, generates trajectories using the AIRL policy and reward function, then converts these trajectories into latent vectors using an encoder and a relational network, which are decoded to initialize policy parameters. In addition, the algorithm updates the VIC parameters online to achieve the best interactive performance, and iteratively optimizes the model parameters using gradient descent to minimize the difference between the generated trajectories and the expert trajectories, thereby improving the model's adaptability and efficiency for new tasks.

[0363] 4. Results

[0364] 4.1. Environment and door opening task settings

[0365] a. Environment settings

[0366] In the present invention, the computer used for the mobile robotic arm experiment is equipped with 64GB RAM, 13th generation Intel(R) Core(TM) i5-13400F CPU (2.5GHz), NVIDIA GeForce RTX 4060. The constructed simulation model will be available on GitHub: https: / / github.com / Poserdon / Deep-Meta-Imitation-Learning.

[0367] In the simulation environment, Figure 8 As shown in (b) and (c), SolidWorks was used to perform high-precision 3D modeling of the mobile robot arm, maintaining an exact 1:1 ratio with the actual device. The model was imported into CoppeliaSim to ensure the authenticity and reliability of the simulation results. The simulation model contains all moving parts and sensor locations, providing a comprehensive operating experience. Through the API interface, Python code facilitates dynamic interaction between the robot and the simulation environment. The simulation environment features various scenarios and tasks to evaluate the performance of the robot under different conditions. The simulation trajectory data is synchronized with the actual operation data in real time for comparative analysis to verify the accuracy of the simulation results. The high-performance computing capabilities of the simulation platform are used to speed up the experimental process and shorten the training cycle. Specific optimization measures include high-precision modeling of each component of the robot and interactive control of the environment through the API interface.

[0368] The experimental platform uses the Agile composite mobile robot produced by Songling Robotics Company, such as Figure 8 (a) and (d). The robot consists of multiple high-performance components (Table 2):

[0369] Table 2. System components of the mobile robotic platform

[0370]

[0371] b. Door opening task setting

[0372] Tasks such as picking and placing with robotic arms have been extensively studied, but automating complex and precise manipulations remains a significant challenge for the industry. Applications that require high precision and dexterity, such as opening a cabinet door with a push-button lock, require precise control capabilities, high adaptability, and advanced sensory skills to manage changing environments and task demands.

[0373] Fig.11 The segmentation performance of three types of push-button locks is demonstrated, where the handle and button components are evaluated separately using an ArUco marker-based vision localization system.

[0374] We use three different datasets to evaluate the position and pose of the Bounce Lock handle and button. To evaluate the pose error of the Bounce Lock in three dimensions, three metrics are used: Intersection over Union (IoU), 5mm5°, and Average 3D Distance (ADD) of the model vertices. The IoU metric evaluates the overlap between the predicted and actual segmented regions. The 5mm5° metric confirms that the pose estimation is accurate if the displacement error does not exceed 5mm and the angular error remains within 5°. For ADD, the accuracy of the pose estimation is verified by ensuring that the average distance between the point cloud and the true pose and the average distance between the estimated pose does not exceed 10% of the object diameter.

[0375] Table 3. Performance analysis of three push button locks under visual inspection indicators

[0376]

[0377] As shown in Table 3, the performance of the three button locks was analyzed using the YOLOv8 instance segmentation algorithm, achieving accurate image segmentation under the strict IoU0.9 criterion. The technique reliably determined the orientation and position of the key components of the button lock, especially performing well in the 5mm5° and ADD indicators. These results provide a solid foundation for the visual guidance of the subsequent robot arm maneuver trajectory generation.

[0378] b. Expert trajectory creation

[0379] This research focuses on deploying the trajectory data produced by CoppeliaSim in RViz using a vision-guided system to manage the motion of the robotic arm and subsequently transferring these trajectories to an actual robotic arm for the door opening task.

[0380] Fig.12 Initially, the vision system in CoppeliaSim records the specific position and status information of the door. Trajectory planning is performed based on this data, and trajectory data is generated and exported in JSON format. Then the ROS environment is configured, the necessary dependencies are installed, the ROS workspace is established, the ROS nodes are established, the trajectory data is read, and the joint status information is published. Subsequently, the driver of the robot arm is configured and activated, ready to receive control commands. Finally, the ROS core is started, the trajectory publishing node is activated, the trajectory data is realized on the actual robot arm, and the motion visualization is realized in RViz, which ensures precise motion control and real-time monitoring of simulation data. This method provides key technical support for the operation function of the mobile robot arm.

[0381] c. Assumptions and adjustments for actual implementation

[0382] In the process of translating simulation data into practical applications, several key assumptions are made and system performance metrics are adjusted to ensure that the generated expert trajectories can be effectively used in practical environments.

[0383] First, it is assumed that the environment is stable throughout the operation, including that the physical modeling of the robot in CoppeliaSim is highly consistent with the real robot. Specifically, the deviation between the simulation environment and the real environment in joint speed, maximum torque, and friction characteristics should not be greater than 4%. This also means that disturbances in the environment, such as interference or resistance, will not have a strong impact on the robot. This stability will ensure that the control strategies generated in the simulation can be directly applied to real-world tasks and promote the successful completion of these tasks.

[0384] Secondly, the accuracy of the sensor must remain constant between the simulated environment and the real environment. In CoppeliaSim, the simulated vision system should capture the position and state, assuming that the maximum error does not exceed 2mm; further assuming that the sensors on the real robot arm can achieve a similar level of accuracy. More calibration is required before application to further narrow the gap between simulation and reality, ensuring that in real-world tasks, the robot arm can accurately find the position of the door, avoid trajectory deviation, and ensure the accuracy of control.

[0385] What’s more, since there are only minor differences between simulation and reality, such as door resistance or joint friction, the system must make real-time adjustments. This also involves continuous real-time optimization of control parameters such as joint speed and force output, so that the trajectory execution error does not exceed 5% in actual operation. In this way, the robot arm can effectively respond to the uncertainty of the real environment while performing movements that are closely related to the simulation.

[0386] It is assumed that very low-probability events, such as sensor failure or sudden external interference, are excluded; it is assumed that in actual operation, the occurrence rate of these events is less than 1%. This will guarantee the overall stability and reliability of the system, so the robot arm will perform all operations smoothly in most cases and will not be affected by unexpected external events.

[0387] The above indicators ensure that the transition from simulation to reality is smooth and add a certain degree of robustness to the system. Therefore, this accurate trajectory generation, integration with ROS, and visualization in RViz guarantee the control accuracy and responsiveness of the robot arm in real operation, allowing it to reliably perform complex tasks such as opening a door.

[0388] 4.3. Experiment of using the mobile robot to open various button locks using the DMIL method

[0389] This experiment implements the DMIL strategy that combines AIRL with LEO and VIC to operate a mobile robot on different types of push button locks. The present invention first trains the A-type push button lock, and then uses meta-learning technology to quickly adapt to the new B-type and C-type push button locks to test the generalization ability of the learning strategy.

[0390] The experimental process includes two stages: pre-training and generalization testing. In pre-training, the initial training of the A-type push button lock uses cabinet door samples and integrates multimodal sensor data to master the button position, pressing force, grasping angle, and rotational force. The AIRL method is used to develop a deep reward function through adversarial training to refine the operation strategy of the robot arm. At the same time, the LEO meta-learning technique is used to enhance the learning process, so that the robot arm can quickly adapt to new tasks and reduce exposure to example data. In the generalization test phase, the B-type and C-type push button lock tasks are evaluated. The robot arm is located in front of the B-type and C-type cabinet doors, and the DIL method is applied for real-time recognition and operation adjustment, and the environment interaction is optimized together with VIC. The performance indicators recorded include the accuracy of the robot arm in identifying the button position, applying pressure, adjusting the grasping angle and trajectory, and the success rate of the door opening operation.

[0391] a. Expert policy training

[0392] A series of experiments were conducted to verify the generalization ability of the method in the robot arm control task, especially when performing the button lock opening operation. The DQN, A2C and PPO algorithms based on the reward function were evaluated, and the PPO algorithm was finally selected as the expert strategy.

[0393] Table 4. Hyperparameters of DQN, A2C, PPO, VIC, and LEO

[0394]

[0395] Table 4 details the hyperparameters of various algorithms and control methods, including key settings for reinforcement learning algorithms (DQN, A2C, PPO), such as learning rate, batch size, and discount factor, as well as stiffness, damping, and maximum force limit in VIC. In addition, the table outlines the basic hyperparameters of LEO, such as latent space dimension, meta-learning rate, and regularization coefficient.

[0396] Through these experiments, we aim to demonstrate the effectiveness and adaptability of our approach in managing various button locking tasks.

[0397] like Fig.13As shown, PPO (red line) performs more stably and slightly higher in reward scores than DQN (black line) and A2C (blue line). PPO shows higher stability during training compared to DQN and A2C. While all three algorithms show significant improvements in the initial stages, PPO has less fluctuation in the later stages and maintains a more stable and higher reward score. PPO achieves higher reward scores in the early training iterations and stabilizes faster than the other two algorithms. This rapid convergence enables PPO to provide effective policies faster. PPO maintains a balance between exploration and exploitation by limiting the step size of each update, which helps prevent drastic changes in policy during updates and avoids significant policy fluctuations.

[0398] Based on the above analysis, we choose PPO algorithm as the expert strategy because it has good stability, convergence speed and adaptability. It guides and measures the performance of other strategies.

[0399] b. Practical evaluation

[0400] To evaluate the generalization ability of the DMIL method in the flexible control of mobile manipulators and push-button locks, we conducted experiments on key locations of three different types of push-button lock cabinet doors. Trajectory generation and mechanical feedback control data show that the system exhibits excellent performance and adaptability in handling complex tasks. The following is an in-depth analysis of the experimental results:

[0401] like Fig.14 As shown in the figure, the experiment explored the interaction with various types of button locks, specifically, the pressing operation of button locks A and C, and the more complex unlocking operation of button lock B. Button locks A and C only require pressing the button to complete the task, while button lock B requires additional downward force after pressing to unlock. Throughout the experiment, the robot arm methodically completed the steps of approaching, touching, and pressing the button lock. The figure shows representative data of DMIL and impedance control (IC), marked as CH and CH-i, respectively. Under position control, we monitored the mechanical parameters of each channel (CH-i) of the six-axis force sensor over time and tested it using the DMIL algorithm (CHi-1). The x-axis of the experimental data represents the sampling point, and the y-axis represents the corresponding mechanical feedback parameter. At the critical moment of pressing the button, there is a significant fluctuation in the force feedback, indicating that the handle of the button lock was successfully released.

[0402] To elucidate the importance of each component of DMIL, this paper includes sensitivity analysis and ablation studies to measure the effectiveness of various model configurations in 50 experimental tasks. The evaluation focuses on the impact of LEO, impedance control (IC) and VIC on force feedback during task execution, especially regarding the maximum contact force output of button lock types A, B and C.

[0403] Table 5. Comparative analysis of contact force control in DMIL: evaluating the impact of LEO, impedance control and VIC on task types A, B and C

[0404]

[0405]

[0406] As shown in Table 5, Model 1 of Button Lock A has moderate performance with a maximum output force of 40.3N. Although VIC facilitates dynamic force adjustment, the lack of LEO leads to less precise control. This pattern persists in Button Lock B, where the force output is upgraded to 60.8N, indicating that the force control is not perfect. The most complex task in this experiment is Button Lock C, where the maximum force output of Model 1 is 155.4N. From experimental observations, VIC alone cannot handle tasks with high complexity without the optimization of LEO. Model 2 uses a combination of LEO and IC, but without VIC. For Button LockA, its accuracy is improved and the output force is 35.6N lower than that of Model 1. This can prove that LEO improves its adaptability. On Button LockB, the force output is 50.2N, reflecting that LEO and IC together help them maintain reasonable control. However, on Button Lock C, Model 2 performed poorly without VIC, with force output rising to 120.3N, proving that while LEO improved performance, VIC was necessary for better control in complex tasks. Model 3 performed the worst among the three tasks because it only used IC. Button Lock A had poor control and a high force output of 50.5N. For Button Lock B, force output increased further to 74.6N, reflecting its imprecision. In the complex Button Lock C task, Model 3's force output soared to 180.6N, exceeding the safe operating limit. This is also an example of how IC cannot effectively control force output in a difficult task.

[0407] Model 4 (DIL), which integrates LEO and VIC, outperforms the other models on all tasks. In Button Lock A, it provides the most precise control with a force output of 25.5N, significantly exceeding its peers. For Button Lock B, it maintains excellent precision with a force output of 42.5N. Even in the challenging Button Lock C, Model 4 demonstrates excellent control, achieving a maximum force of 45.6N. These results highlight the synergistic advantages of combining LEO and VIC, with LEO enhancing task adaptability and VIC enhancing dynamic control, thus establishing the dominant configuration of Model 4. Model 4 performs well in handling all button lock tasks, especially in complex scenarios, indicating that the combined application of LEO and VIC significantly improves precision and adaptability. Model 1 (VIC only) shows moderate performance but performs poorly in complex tasks without LEO support. Model 2 (LEO+IC) shows good adaptability but lacks the precise control provided by VIC in more demanding situations. Model 3 (IC only), relying solely on IC, proved the least effective, consistently producing high force outputs, indicating that both LEO and VIC were required for optimal performance.

[0408] These results verify that DIL generalizes well in situations where it is compatible with control, moving the robot arm, and button locks, helping to avoid using excessive forces that could damage the button or the robot arm. The robot arm is able to match its force output depending on the type of button lock it is facing. This ensures stability and safety during operation. The experiments summarize the trajectories of different button lock openings, as well as the operation steps and trajectories ( Fig.15 ).

[0409] This paper effectively proposes a learning strategy for robot control by uniformly integrating LEO and AIRL meta-learning methods. VIC dynamically adjusts the robot's operational performance, greatly enhancing the robot's generalization ability and increasing the speed of adapting to new environments. It utilizes the deep reward function learned from expert behavior and adversarial training mechanisms, so that when unknown tasks arrive, the robot can quickly shift its behavior to various operational requirements. In addition, VIC allows for more fine-grained interactions with the environment. Finally, the introduction of LEO meta-learning optimizes this learning process, enabling it to quickly adapt to new tasks with a small amount of example data. This is very important for practical applications, where changes in task environments are very typical.

[0410] c. Performance evaluation of four IL algorithms under different key locking tasks

[0411] This experiment compares the performance of four key lock algorithms, BC, GAIL, AIRL, and DMIL, from four key performance indicators: completion time, trajectory deviation, force feedback, and success rate. It reflects the performance of the algorithms in performing multiple unlocking tasks from multiple perspectives. There are three types of button locks in total: type A, type B, and type C. 50 tests were conducted for each type. The average success rate of each algorithm in different tasks, including the button lock task, can be used to discover the effectiveness of various operational controls in mobile manipulators.

[0412] Table 6. Performance evaluation of four IL algorithms in different button locking tasks

[0413]

[0414] As shown in Table 6, force feedback reflects the maximum contact force applied by the robot during operation, and excessive force feedback may cause damage to the device. The results show that DMIL effectively regulates the force feedback for all tasks, and the force feedback values ​​are kept at a low level, with force feedback values ​​of 25.5N and 45.6N for task type A and task type C, respectively. In contrast, the force feedback values ​​of the BC algorithm are higher, reaching 74.6N in task type B and rising to 180.6N in task type C, indicating weaker force control and higher risk of damage. Both GAIL and AIRL show more balanced force feedback control, although their values ​​are higher than DIL, highlighting the advantage of DIL in managing complex mechanical interactions.

[0415] Completion time represents the total time required for each algorithm to complete the task. From the results, BC has the least execution time in all these button lock tasks because its task execution mode has become simplified. BC generates control policies by directly imitating experts. Therefore, unlike GAIL and AIRL, it does not need to perform deep strategy optimization and deep environment exploration. This avoids a lot of calculations and iterations when BC performs tasks, thereby completing predefined operations faster. Since BC only relies on demonstration data from experts, no additional inference or policy updates from the environment are required: the significant reduction in overhead during execution translates into a shorter completion time. However, with the improvement in time efficiency, BC's simplified approach may face significant challenges when handling complex tasks. In contrast, more sophisticated algorithms such as DIL, GAIL, and AIRL, although they make some tasks slightly slower to complete, can more effectively optimize policies and adapt to changes in the environment, making them more stable and accurate when performing the same tasks. Therefore, although BC has an advantage in completion time, the faster running of tasks comes at the expense of accuracy in controlling and handling complex tasks. That said, a shorter completion time may not always be the best solution, as the speed at which the BC performs the task may result in suboptimal overall performance for tasks that also require higher stability and accuracy.

[0416] The trajectory deviation reflects the control accuracy due to the difference between the ideal path and the actual path of the robot's motion. Among all tasks, DMIL has the smallest deviation, with a deviation of only 5.3 mm for task C, indicating that its control accuracy is very good. BC brings higher trajectory deviations, especially in task type B, with a deviation of up to 13.2 mm, indicating that its control is unstable in more complex tasks. For task types A and C, GAIL and AIRL have smaller trajectory deviations, while for task type B, their trajectory deviation is slightly higher than DIL.

[0417] The success rate is a key indicator to measure the efficiency of task execution, such as Fig.16 As shown in Figure 1, it describes the success rates of different algorithms in different task types. The DMIL algorithm achieved the highest success rate, 96.0% for type A, 90.0% for type B, and 94.0% for type C, with small error bar fluctuations, high stability and reliability. In contrast, the GAIL algorithm achieved a success rate of 94.0% for type A, 74.0% for type B, and 84.0% for type C, with moderate error bar fluctuations and moderate stability. Similarly, the AIRL algorithm achieved a success rate of 92.0% for type A, 78.0% for type B, and 90.0% for type C, but showed obvious error bar fluctuations, especially in type B, indicating a certain instability. The BC algorithm lagged behind relatively, with a success rate of 92.0% for type A, 66.0% for type B, and 72.0% for type C, while there were large error bar fluctuations, indicating that the algorithm was unstable and had limited generalization ability. Overall, DIL achieved the highest success rate in all task types, while BC performed poorly, especially in types B and C.

[0418] In summary, the performance results of four IL algorithms tested on various button lock tasks show that DIL consistently outperforms the other algorithms in terms of control accuracy, force feedback, and success rate, especially in terms of force feedback with minimal trajectory deviation. BC has a simplified execution style and the shortest completion time, but weaker control, especially for complex tasks, as manifested by large force feedback and large trajectory deviation. GAIL and AIRL both perform moderately, maintaining a balance between execution time and control accuracy, but are still far from the stability of DIL. Quantitatively, DMIL produces the strongest and most reliable algorithm on all performance metrics; while in BC, fast execution comes at the cost of control accuracy and generalization.

[0419] 5. Discussion

[0420] DMIL’s success likely comes from the fact that it merges vision guidance, trajectory generation, and compliance control into a single framework. Recipe YOLOv8 works with a depth camera to precisely locate the button lock via pose estimation—in other words, the exact input required for trajectory generation. This is further supported by VIC, which dynamically adjusts the stiffness and damping parameters of the robot’s manipulator to achieve optimal interaction with the environment, preventing over-pressing or damaging the button lock.

[0421] A. Interpretation and Analysis of Results

[0422] These experimental findings strongly demonstrate the effectiveness of the DMIL algorithm in controlling the opening of several types of push-button lock cabinet doors by a mobile robot. As a result, the DMIL algorithm achieved the highest success rate for all tested push-button lock types: 96.0% for type A, 90.0% for type B, and 94.0% for type C. These results show the superior performance, stability, and reliability of the algorithm in different scenarios. DIL outperformed other algorithms such as BC, GAIL, and AIRL. While GAIL and AIRL also achieved higher success rates, the errors in the error bars were larger and the stability was poorer compared to DIL. On the other hand, the BC algorithm had larger value fluctuations and an overall lower success rate, especially for the challenging locks of types B and C, which again reflects the poor generalization ability. The very high success rate and small error bar fluctuations indicate that DIL is robust and effective in adapting to changes. The DMIL algorithm supports meta-learning, retains prior knowledge, and quickly adapts to new push-button lock types with minimal retraining, which again demonstrates its good generalization under different operating conditions.

[0423] These results do show that DMIL seems to be very effective for complex robotic manipulation tasks that require a high degree of precision and adaptability. Moreover, due to the ability to combine more advanced techniques with different learning strategies, this branch of research starts to look very promising in industrial automation tasks, especially in tasks with dynamic and unpredictable environments.

[0424] B. Theoretical and practical significance

[0425] 1. It theoretically contributes to the work proposed in this paper, on the novel integration of deep meta-learning, IL and VIC to control mobile robotic arms in dynamic and complex environments. By integrating AIRL and LEO within the DMIL framework, this work greatly improves the adaptability and generalization ability of the robotic system. The key insight here is that DMIL can learn a deep reward function from a small amount of expert behavior data and quickly adapt to unknown tasks using a meta-learning mechanism.

[0426] 2. This has important practical implications for industrial automation and robotics. DIL is particularly suitable for applications in industrial environments, such as inspection and maintenance of power equipment, because it can quickly adapt to new tasks and environments with minimal data - a feature that is very important when robots are expected to operate reliably and safely in complex and dynamically changing environments. It integrates successful technologies such as visual guidance, trajectory generation and compliance control to comprehensively solve the precision automation tasks used by industrial operators and achieve higher levels of efficiency and safety.

[0427] C. Robustness and Failure Modes

[0428] Although DMIL performs well in terms of force feedback, trajectory deviation, and success rate, it may face limitations when applied to more complex tasks or complicated environments.

[0429] Robustness: DIL is robust in maintaining high success rates and low force feedback in different types of button locking tasks. It changes its strategy according to the complexity of the task and maintains relatively stable performance in Class A, Class B, and Class C tasks. This stability depends largely on the efficiency of modeling and the adaptability of DMIL to the task environment. When DMIL faces unfamiliar environments or extreme conditions, the performance of the algorithm will deteriorate.

[0430] Failure Modes: Slightly more complex mechanical interaction tasks may be challenging for DIL, for example, tasks that require coordination of multiple tools or highly precise mechanical control. In such cases, larger trajectory deviations or unnecessarily high force feedback may experience DIL. If there are sudden or unpredictable circumstances where the motion pattern of button locking changes or external forces are applied, DIL may be slow to adapt to such changes. It can result in trajectory control errors, longer completion times, or even failure to complete the task. In addition, when insufficient or low-quality training data is available, DMIL will degrade significantly, with increased deviations and reduced success rates.

[0431] Although DMIL is very robust, its performance can suffer in very complex mechanical tasks, dynamic environments, or if the algorithm is not trained with enough data. In other words, this means that the algorithm needs further validation and improvement in real-world applications to increase the range of uncertain tasks it can effectively handle.

[0432] D. Limitations

[0433] Despite their great potential, vision-guided techniques such as DMIL and YOLOv8 also have significant limitations. YOLOv8 works very fast in recognizing objects on images, although it overestimates - or can show very large drops in accuracy and stability when complex scenes involve cluttered backgrounds, high occlusions, or poor lighting conditions. In addition, YOLOv8 may not recognize new objects or objects it has never been well trained on, and thus may not generalize to the dynamic real world. In addition, the DMIL algorithm allows for some generalization between different tasks; however, adaptability and flexibility to unexpected changes in the environment or changes in the type of tasks may become scarce or insufficient. This is mainly because DMIL relies heavily on high-quality training data to develop an effective learning model. In practice, this can be difficult and expensive to obtain in frequently changing environments or highly specialized industrial settings. Future work must be directed towards making the algorithm robust, reducing its dependence on high-quality training data, and enhancing its adaptability to new environments.

[0434] E. Ethical and social impact

[0435] In this study, the practical application prospects of DIL and vision-guided technology are huge, and real deployment requires comprehensive consideration of safety, ethics, and social impact. It becomes crucial to improve safety standards, avoid model bias, solve employment challenges, and protect data privacy leakage.

[0436] Considering the ethical and societal impacts, especially regarding the implementation of robotics and automated systems, DMIL excels in task adaptability and automation; however, it must carefully consider various ethical issues.

[0437] 1. Specific application fields or related products of the present invention.

[0438] The technical solution of the present invention has a wide range of application value and is mainly applicable to multiple fields and related products. First, in the field of industrial automation, the mobile robot arm control technology of the present invention can be applied to automated production lines, especially tasks such as precision assembly and material handling. By dynamically adjusting impedance parameters and real-time learning functions, the operating efficiency and accuracy are significantly improved, and the need for manual intervention is reduced. In addition, in terms of power substation maintenance, the technology of the present invention provides a solution for automated maintenance operations, especially in the button lock cabinet door operation scenario, the robot arm completes precise unlocking and maintenance tasks through intelligent control, improving safety and work efficiency. In the medical field, the mobile robot arm technology of the present invention can automatically adjust the operating force according to the mechanical requirements of different surgical tasks, and is applied to surgical robots and rehabilitation equipment to ensure the accuracy and safety of the operation. Finally, logistics and warehousing management can also benefit from the present invention, through the automated handling and sorting system to achieve fast and efficient operation and reduce human operation errors.

[0439] 2. Relevant evidence of the technical effects obtained by the embodiments of the present invention.

[0440] The present invention has demonstrated significant technical effects through experimental data and practical application verification. Experiments were conducted in the CoppeliaSim simulation environment and on a real mobile robotic arm platform. The results show that the present invention can significantly improve the task adaptability and execution efficiency of the robotic arm. Compared with traditional methods, the present invention has improved the operation accuracy and execution speed in complex tasks. By integrating variable impedance control (VIC), the robotic arm can dynamically adjust the stiffness and damping parameters according to task requirements, thereby reducing misoperation. The process of the robotic arm opening a door is as follows: Fig.17 shown.

[0441] In practical applications, the robot arm of the present invention has demonstrated excellent operational capabilities in multiple button lock cabinet door tasks, and the task success rate is much higher than that of traditional methods. Through real-time force feedback and stiffness adjustment mechanisms, the robot arm can improve operational accuracy and shorten response time in different physical environments. In addition, with the help of latent embedded optimization (LEO) technology, the present invention has demonstrated the ability to quickly adapt to new tasks in multi-task learning, shortened learning time, and significantly improved the deployment efficiency of new tasks.

[0442] Table 6. Performance evaluation of the algorithm in different button locking tasks

[0443]

[0444] In summary, through experimental verification, the present invention has achieved technical breakthroughs in optimizing intelligent control strategies, improving operational efficiency, and enhancing task adaptability, demonstrating high practical application value.

[0445] It should be noted that the embodiments of the present invention can be implemented by a combination of software and hardware. The hardware portion can be implemented using dedicated logic; the software portion can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. A person of ordinary skill in the art will appreciate that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0446] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A mobile robot arm control method based on deep meta-imitation learning, characterized in that: The following steps are involved: Step 1: Introduce the deep meta-imitation learning DIL algorithm to combine deep meta-learning with imitation learning IL technology; Step 2: Use a deep learning 6D posture estimation method to determine the position and orientation of the button lock and intuitively identify key operating reference points; Step 3, in the whole IL stage, the adversarial inverse reinforcement learning AIRL and variable impedance control VIC technology are integrated; Step 4, introduce latent embedding optimization (LEO) into the deep meta-learning architecture.

2. The mobile robot arm control method based on deep meta-imitation learning as claimed in claim 1, characterized in that: The 6D posture estimation method using deep learning: (1) Visually guided trajectory based on deep learning; (2) IL combined with VIC; (3)DMIL.

3. The mobile robot arm control method based on deep meta-imitation learning as claimed in claim 2, characterized in that: The deep learning-based visual guidance trajectory: a. Mobile robot end position and posture analysis The electric cabinet door includes a door frame and a door leaf, and the button door lock includes a button and a handle; the coordinate system is defined as follows: {W} is the world coordinate system, fixed at the lower corner of the door frame; {O1} represents the lock head coordinate system, and {O2} represents the handle coordinate system, both of which are associated with the button door lock; all coordinate systems follow the right-hand rule; {O_1} and {O_2} are attached to the button door lock, and their relative position relationship |O_1O_2|=O_12 remains unchanged; the homogeneous transformation matrix from {O_1} to {O_2} is expressed as: {W} is fixed on the door frame, and {O_1} is fixed on the cabinet door; their relative posture changes with the cabinet door rotation angle θ_o; assuming that the origin O_1 of the {O_1} coordinate system to the door hinge is w_o and the height is z_o, then the coordinates ^WO_1 (x_o, y_o, z_o) are: The homogeneous transformation matrix of the button lock's pressed position relative to {O_1} relative to {W} is: The unlocking trajectory includes pressing the button to lock and rotating the handle; after pressing the button to a preset distance along the x-axis, release the handle; the transformation matrix of {O2} in {W} is: In order to model the motion of the handle, a feature point P_s at the end of the handle is selected, whose motion follows a cylindrical arc with O_2 as the vertex and the handle length l as the generatrix; to depict this arc, assume that the bounce angle of the handle is The coordinates of P_s in {O_2} are: The unlocking trajectory "T"_1 rotating around the x-axis is represented by the arc in {O_2}}: To open the door, the cabinet door needs to be rotated around the door hinge and unlocked (Figure 1); it is assumed that {W} is located at the bottom of the door hinge and is axially aligned with {O1} and {O2}; when the handle is pulled, the cabinet door rotates clockwise along the {W}z axis; the opening motion is represented by the trajectory of the handle end point P_s; b. Visually guided trajectory generation The visual guidance system uses the YOLOv8 instance segmentation algorithm to accurately identify target objects and perform real-time image segmentation in the complex environment of the substation; the system is compatible with the capture of high-precision depth and image data. Realsense TM The D435 depth camera is used in conjunction with the robot to accurately determine the position and posture of key components that the robot arm needs to operate.

4. The mobile robot arm control method based on deep meta-imitation learning as claimed in claim 2, characterized in that: The IL United VIC: a. Contact force model between the mobile robot and the door lock environment The impedance model formula of the mobile robot in the task space is: As shown in Equation 8, x_e=x-x_d represents the position error between the actual position x and the desired position x_d; Mr is the inertia matrix, Dr is the damping matrix, and Kr is the stiffness matrix; for an n-degree-of-freedom robot, these are n×n positive definite symmetric matrices, which are usually assumed to be positive definite diagonal matrices; F_e is the contact force of the door lock on the end effector of the robot; the one-dimensional impedance model formula in the joint space is expressed as: Where q, q˙ and q¨ are the vectors of generalized joint coordinates, angular velocity and angular acceleration respectively, and τ_e represents the external force / torque in the joint space. In addition, using the Jacobian matrix J(q) of the robot, the correlation between τ_e and F_e is as follows: τe=J(q)TFe (10) The dynamic equations of the mobile manipulator describe in detail the dynamic relationship between torque and state variables in different spaces, providing a basis for the force control of the manipulator. In the presence of contact impedance, the dynamics of the joint space is usually expressed by the Lagrange equation: Where C represents the coupling matrix of Coriolis and centripetal forces; F is the friction force; G is the gravity load; where is the joint driving torque; The door lock mechanism is modeled as a first-order spring system, the lock stiffness is represented by k_i, and the contact pressure of the lock is: f e =k i (x o -x c ) (12) Where x_o is the natural position of the door lock contact surface when unloaded, f_e can be measured using a six-axis force sensor in the actual system; the expected tracking force is represented by f_d, and the difference between the expected tracking force and the actual contact force is expressed as: Δf=f e -f d (13) b. Mobile Robotic Arm VIC VIC parameters K and D are further introduced; VIC adjusts stiffness K and damping D to find the optimal values ​​that ensure the compliance and stability of the arm in different tasks; For the target impedance, the input is the force deviation Δf, and the output is the trajectory deviation e(t); the second-order transfer function of the position-based impedance control model is: The model describes the dynamic trajectory deviation e(t) and force deviation Δf as a "mass-damper-spring" second-order differential system: M_r, B_r and K_r represent the mass, damping and stiffness coefficient matrices of the target impedance model, respectively; c. AIRL combined with VIC.

5. The mobile robot arm control method based on deep meta-imitation learning as claimed in claim 4, characterized in that: The AIRL and VIC combine to: AIRL consists of a generator G that tries to learn from expert behavior and a discriminator D that tries to distinguish between trajectories obtained from generated policies and actual expert trajectories. AIRL learns from the reward functions of expert behavior and then uses these rewards to guide the optimal control strategy of the robot arm to achieve precise door opening action. In AIRL, the reward function r θ (s,a,K,D) is summarized as follows r θ (s,a,K,D)=f(s,a)-λ1(K-K t ) 2 -λ2(D-D t ) 2 (16) f(s, a)F(s, a) represents the basic reward function based on state and action to ensure that the position and posture of the robot arm meet the expected trajectory of opening the door button lock; K_t and D_t are the target values ​​of the damping coefficient and stiffness coefficient, which are determined by the system design requirements; λ_1 and λ_2 are regularization weights used to balance different aspects of the reward; The discriminator D_φ and the generator π_θ have objective functions that are at the heart of adversarial training and aim to ensure that the generated policy closely replicates the expert behavior; the goal of the discriminator is to distinguish between trajectories produced by the generator π_θ and true trajectories from the expert π_E; Its optimization objective is defined as: It consists of the log-likelihood of the generated trajectory and the log-likelihood of the expert trajectory, which improves the discriminator's ability to correctly distinguish between the generated and expert trajectories; the generator aims to generate trajectories similar to the expert's behavior; the optimization objective of the generator is: This means that the generator π_θ tries to minimize the discriminator The log-likelihood of is used to simulate expert behavior, effectively "fooling" the discriminator; VIC allows the robot arm to dynamically change its stiffness and damping according to different interactions and task requirements, which is critical when performing high-precision operations (such as opening a push-button lock); Impedance control force formula: Where K and D represent the stiffness matrix and damping matrix respectively, p_"D" and p˙D represent the target position and velocity, and p and p˙ represent the current position and velocity; Incorporating the VIC parameters K and D into the AIRL framework, the optimized reward function is: λ_p and λ_d are weight factors for position and velocity changes, which aim to balance the trajectory learning objective and dynamics requirements.

6. The mobile robot arm control method based on deep meta-imitation learning as claimed in claim 2, characterized in that: The DMIL: DMIL: Integrates two meta-learning techniques: AIRL and LEO; The essence of trajectory generalization technology is to integrate the optimization goal of AIRL with the dynamic adjustment of LEO, and obtain the following optimization formula: Where γ is the discount factor, r_φ represents the reward function derived from the discriminator, and λ∥θ∥^2 acts as a regularization term to reduce model complexity and promote generalization; In addition, the dynamic adjustment of LEO is given by the regularized formula: Here, λ_K and λ_D are regularization parameters that regulate the adjustment of stiffness K and damping D, ensuring that the performance in new environments is not affected by previous learning; Training loss calculation: For each task T_i, the loss function on the training set D^tr is defined as: where r_(θ, φ) represents the reward function derived by the discriminator, Q^π(s, a) represents the expected action value according to the current policy π; after calculating the training loss, the encoding vector z^' is updated by gradient descent: The modified z^' is then decoded using the decoder g_(φ_d)(z^') to extract the new parameters θ_i^': After completing a training cycle, calculate the loss on the validation set D^val: In successive rounds of training, the validation loss of each task is used to update the global parameters 7. A mobile manipulator control research system based on deep meta-imitation learning that implements the mobile manipulator control method based on deep meta-imitation learning as described in any one of claims 1 to 6, characterized in that: The mobile manipulator control research system based on deep meta-imitation learning includes: The combination module is used to introduce the deep meta-imitation learning DIL algorithm and combine deep meta-learning with imitation learning IL technology; A determination module is used to determine the position and orientation of the button lock using a deep learning 6D posture estimation method to intuitively identify key operating reference points; Integration module for integration of adversarial inverse reinforcement learning (AIRL) and variable impedance control (VIC) techniques throughout the IL phase; Introducing a module for introducing latent embedding optimization (LEO) in deep meta-learning architectures.

8. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the mobile robot arm control method based on deep meta-imitation learning as described in any one of claims 1-6.

9. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the mobile robot arm control method based on deep meta-imitation learning as described in any one of claims 1 to 6.

10. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the mobile robotic arm control research system based on deep meta-imitation learning as described in claim 7.

Images