Embodied intelligent operation and maintenance oriented world model based mechanical arm operation diffusion strategy domain adaptation method

Abstract

The application discloses a body-aware intelligent operation and maintenance robot arm operation diffusion strategy domain adaptation method based on a world model, and belongs to the technical field of body-aware intelligence and industrial robots. First, expert demonstration data of robot arm operation and maintenance operation is collected in a source domain controlled laboratory environment, and a diffusion strategy based on a denoising diffusion probability model and a one-dimensional convolutional neural network is trained. Then, a state world model is constructed, the environment state transition law of the robot arm operation and maintenance operation is learned, and the world model is fine-tuned by using a small amount of real operation and maintenance field data. Subsequently, a large amount of synthetic training data of the target domain is expanded by using the fine-tuned world model through an incremental disturbance pseudo-trajectory generation mechanism. The pseudo-trajectory and a small amount of real operation and maintenance data are mixed, and the diffusion strategy is fine-tuned to adapt to the target domain. Finally, the strategy is deployed in the operation and maintenance field through an observation offset estimation and domain compensation mechanism. The application significantly improves the success rate of the robot arm in performing fine operation tasks in the real operation and maintenance field.

Description

Technical Field

[0001] This invention belongs to the field of embodied intelligence and industrial robot technology, specifically relating to a domain adaptation method for robotic arm operation diffusion strategy based on a world model for embodied intelligent operation and maintenance. Background Technology

[0002] High-end equipment (including rotating machinery, pressure vessels, precision manufacturing equipment, aero-engines, and construction machinery) is the core support of a nation's manufacturing and defense industries. Real-time monitoring and preventative maintenance of its operational status are of significant strategic importance for ensuring production safety, extending equipment lifespan, and reducing unplanned downtime losses. Traditional maintenance methods heavily rely on manual labor: manually installing vibration acceleration sensors on rotating main bearings and impeller hubs, periodically checking and tightening pipe flange bolts, regularly replacing vulnerable seals in sealing cavities, and performing ultrasonic flaw detection on aero-engine compressor blades. These operations are not only labor-intensive and skill-intensive, but also often require personnel to enter dangerous areas with high temperatures, high pressures, and high speeds, posing significant safety risks.

[0003] Embossed intelligent robots, with robotic arms as their core execution unit, possess the potential to autonomously complete the aforementioned complex maintenance operations. However, deploying robot strategies from laboratory environments to real-world high-end equipment maintenance sites faces significant domain shift challenges: sensor installation tolerances cause observational shifts in the initial pose of the end effector relative to the workpiece; variations in lighting and vibration noise in industrial environments result in different statistical distributions of sensor readings compared to laboratory settings; and equipment manufacturing tolerances and assembly errors introduce non-negligible workpiece position deviations. The combined effect of these systemic differences significantly reduces the success rate of directly transferring robot strategies trained in laboratory environments to real-world maintenance sites.

[0004] Diffusion policy introduces diffusion generation models into robot policy learning, representing the policy as an iterative denoising process from noise to action through a denoised diffusion probability model (DDPM). It possesses advantages such as modeling multi-modal action distributions and strong expressive power for fine-grained tasks (Chi C., Feng S., Du Y., et al. DiffusionPolicy: Visuomotor Policy Learning via Action Diffusion[C]. Robotics: Science and Systems, 2023.). However, diffusion policy requires a large amount of high-quality demonstration data for training in a controlled laboratory environment, while re-collecting large amounts of data in real-world operations is extremely costly—the operation window is limited, professional operators are scarce, and each data collection results in equipment downtime. Existing domain adaptation methods mainly suffer from the following shortcomings: direct transfer methods ignore inter-domain differences, leading to a significant decrease in operational success rate (only 16% in typical scenarios); full-data fine-tuning requires the collection of a large amount of field data, resulting in long downtime for operations; and domain randomization methods struggle to accurately align systematic shift features.

[0005] Therefore, there is an urgent need for a method to efficiently complete the diffusion strategy domain adaptation under the condition of limited real operation and maintenance field data. A typical application is to use robotic arms to perform fine operation and maintenance operations such as sensor mounting, bolt tightening, and component replacement for high-end equipment, which can significantly reduce the field deployment threshold and cost of intelligent operation and maintenance robots. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, the purpose of this invention is to provide a world model-based method for adapting the diffusion strategy of robotic arms in intelligent operation and maintenance. In high-end equipment intelligent operation and maintenance scenarios, only a small amount of demonstration data needs to be collected at the actual operation and maintenance site. By using the world model to assist in the enhancement of pseudo-trajectory data, the diffusion strategy can achieve efficient target domain adaptation, which significantly improves the success rate of the robotic arm in performing fine operation tasks such as sensor installation, bolt tightening, and seal replacement at the actual operation and maintenance site.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A world-model-based method for adapting the diffusion strategy of a robotic arm for embodied intelligent operation and maintenance includes the following steps: Step 1: In a controlled laboratory source domain environment, collect expert demonstration data of the robotic arm performing high-end equipment operation and maintenance tasks to construct a source domain dataset; Step 2: Train a source domain diffusion strategy based on the source domain dataset, using a denoised diffusion probability model and a one-dimensional convolutional neural network to construct the diffusion strategy network; Step 3: Train a state world model based on the source domain dataset, using a multilayer perceptron residual prediction structure to obtain a pre-trained world model, with parameters denoted as... Step 4: Collect a small amount of target domain demonstration data at a real operation and maintenance site. Fine-tune the pre-trained world model obtained in Step 3 on the target domain transfer samples with a small learning rate to obtain the fine-tuned world model of the target domain. The parameters are denoted as follows. Step 5: Fine-tune the world model using the target domain Step 5: Apply incremental perturbations to the initial state of the target domain, generate multiple pseudo-trajectories through forward expansion of the world model, and construct a pseudo-trajectory dataset; Step 6: Merge the pseudo-trajectory dataset generated in Step 5 with the real target domain demonstration data collected in Step 4 to construct a hybrid dataset; Step 7: Initialize the diffusion strategy network with the source domain diffusion strategy parameters, and fine-tune and train the diffusion strategy network on the hybrid dataset to obtain a diffusion strategy network adapted to the target operation and maintenance site; Step 8: Deploy the fine-tuned diffusion strategy network at the target operation and maintenance site, and estimate the observation offset between the target domain and the source domain by statistically analyzing the mean difference. After applying domain compensation to the real-time observations, the data is input into the policy network, which outputs a sequence of robotic arm operation actions.

[0008] Step 1 specifically involves: in a controlled laboratory source domain environment, guiding a robotic arm to perform maintenance tasks via a teleoperation system, collecting expert demonstration trajectories, and forming a source domain dataset. = {τ i}, i = 1,…, Each demonstration trajectory = {( , It consists of a time series and an observation vector. It is composed of the position and attitude state of the equipment target object, the position of the robotic arm end effector, the quaternion of the end effector attitude, and the opening and closing state of the gripper, and has dimensions. = 19; Action Vector For joint position control commands, dimensions = 7.

[0009] Step 2 specifically involves: using a diffusion strategy to observe historical sequences. As a condition, = 2 steps, 38 dimensions in total, training a one-dimensional convolutional denoising network The training loss function is: in, For policy network parameters, Standard Gaussian noise, For noise prediction networks, For the first The action after the step spread, Uniform sampling from {1,…, }, = 100, The observed sequence is used; the policy predicts a 16-step action sequence with n_action_steps = 8 steps per execution.

[0010] Step 3 specifically involves: the state-world model. Using a residual prediction structure, the prediction formula is: in, for Time-state vector For action vectors, For world model parameters; It consists of a 4-layer fully connected network with 256 hidden dimensions, using SiLU activation function, and outputting an increment vector with the same dimension as the state; it is trained in a normalized space with mean squared error as the loss function. in, The training sample size is given by norm(·), and the normalization function is given by norm(·). Extract all adjacent frame transfer samples ( , , The AdamW optimizer was used to train the world model for 200 epochs to obtain the pre-trained world model parameters. .

[0011] Step 4 specifically involves: collecting a small amount of target domain demonstration data at a real high-end equipment operation and maintenance site. = { }, j = 1,…, There is a systematic observational bias between the actual operation and maintenance site and the laboratory environment. ;from Extract transition samples from adjacent frames, using a small learning rate lr = 5×10. -5 For pre-trained world models Fine-tuning for 200 epochs yields the fine-tuned world model for the target domain. .

[0012] Step 5 specifically involves: [The text abruptly ends here, likely due to an incomplete sentence or a missing section.] The initial state of each demonstration trajectory Apply Random perturbations of different scales, the first The initial state after the perturbation is: The perturbation scale Determined using the following incremental formula: in, = 0.01 is the basic perturbation scale. = 0.5 is the increment factor. = 8 represents the augmentation amount per demonstration; perturbations are applied only to physically variable state components, namely the target object position and the end effector position, without perturbing the attitude quaternion and gripper state; the original motion sequence of the demonstration is then preserved. ,use Stepwise forward unfolding generates pseudo-trajectories {( , , )},t = 0,…, -1, This represents the time step of the trajectory.

[0013] When constructing the hybrid dataset in step 6, the pseudo-trajectory dataset will be... Demonstration dataset with real target domain Merge into a hybrid dataset = ∪ The number of pseudo-trajectories in the mixed dataset is 6 to 10 times that of the real data. This mixing ratio effectively suppresses overfitting while significantly expanding the training data.

[0014] Step 7 specifically involves: using source domain diffusion strategy parameters The model parameters, including the exponential moving average (EMA), are initialized based on a mixed dataset. The normalized statistics of observations and actions are recalculated and the normalization layer of the policy network is updated; the AdamW optimizer is used with a learning rate of 1×10⁻⁶. -4 Weight decay 1×10 -6 Cosine annealing learning rate scheduling is used to train the model for 200 epochs while maintaining the EMA model with an EMA decay coefficient of 0.75, resulting in an adaptive diffusion strategy for the target domain. .

[0015] Observation offset in step 8 The estimation formula is: in, The first observation vector One portion, (·) and (·) Take the mean of all samples in the target domain and source domain datasets respectively; the domain compensation formula is: = o real - The compensated observations Input fine-tuning diffusion strategy, and generate robotic arm operation sequence after DDIM acceleration and denoising; the strategy is executed in a rolling manner: maintain 2-step observation history, and execute 8 steps each time.

[0016] The high-end equipment operation and maintenance tasks include: ① attaching vibration acceleration sensors or temperature sensors to key monitoring points of rotating machinery (wind turbine main bearings, steam turbine rotors, etc.); ② performing bolt tightening and torque detection operations at pressure vessels and pipeline connections; ③ replacing vulnerable seals or picking up and installing precision parts on precision manufacturing equipment (CNC machine tool spindles, precision grinding machines); ④ positioning and attaching non-destructive testing probes to the surface of aero-engine fan blades or compressor blades. The source domain is the laboratory simulation platform, and the target domain is the actual on-site operation and maintenance environment of the aforementioned high-end equipment. The sources of systematic observational deviation between the two include differences in sensor installation, industrial lighting and vibration interference, and equipment attitude tolerance.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention efficiently utilizes a small amount of field data to complete domain adaptation: only requires = 20 real-world operation and maintenance demonstration data points, enhanced with world model-assisted pseudo-trajectory ( × = 160 synthetic trajectories), which increases the success rate of robotic arm operation and maintenance from 16% of direct migration to 46% (a relative increase of 187.5%), and significantly reduces the on-site deployment cost of intelligent operation and maintenance robots for high-end equipment; (2) The two-stage training mechanism of the world model of this invention (source domain pre-training + target domain fine-tuning) makes full use of the large-scale laboratory data to learn the environmental dynamics prior of equipment operation and maintenance, and then uses very little field data to accurately adapt the state transition distribution of the real operation and maintenance scenario. (3) The incremental perturbation pseudo-trajectory generation mechanism of the present invention ( = 8, = 0.01, = 0.5) Covers different degrees of initial state offset, generating synthetic operation and maintenance trajectories with rich diversity; (4) Domain compensation deployment mechanism of this invention ( The estimation + DDIM 20-step inference method eliminates systematic observation bias in a lightweight way without modifying the policy network structure, and the inference speed is 5 times faster than during training. (5) The method of the present invention is universal and can be widely applied to various operation and maintenance scenarios such as sensor mounting for rotating machinery condition monitoring, bolt tightening for precision equipment, and probe positioning for non-destructive testing of aerospace equipment. Attached Figure Description

[0018] Figure 1This is a flowchart of a method according to an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram of the two-stage training (source domain pre-training and target domain fine-tuning) architecture of the state-world model in an embodiment of the present invention.

[0020] Figure 3 This is a schematic diagram of the incremental perturbation pseudo-trajectory generation strategy of the world model in an embodiment of the present invention.

[0021] Figure 4 This chart compares the success rates of three domain adaptation methods in target domain operation and maintenance tasks. Detailed Implementation

[0022] The present invention will now be described in detail with reference to the embodiments and accompanying drawings.

[0023] Reference Figure 1 This embodiment takes the installation of vibration acceleration sensors on key parts of high-end rotating machinery as a typical application scenario. A method for adapting the domain of robotic arm operation diffusion strategy based on a world model for embodied intelligent operation and maintenance includes the following steps: Step 1: In a controlled laboratory source domain environment, collect expert demonstration data of robotic arm performing high-end equipment operation and maintenance tasks to construct a source domain dataset. This embodiment simulates a high-end equipment operation and maintenance scenario in a controlled laboratory source domain environment. A robotic arm is guided by a remote control handle to perform sensor placement tasks, and expert demonstration trajectories are collected to form a source domain dataset. = { }, i = 1,…, , = 200; each demonstration track = {( , )},t = 0,…, , Trajectory time step; observation vector It is composed of four types of components: ① Position and attitude of the equipment target mounting point (10-dimensional); ② Three-dimensional position of the end effector (3-dimensional); ③ Attitude quaternion of the end effector (4-dimensional); ④ Gripper opening and closing state (2-dimensional); motion vector These are joint position control commands.

[0024] Step 2: Train a source domain diffusion strategy based on the source domain dataset, and construct the diffusion strategy network using a denoised diffusion probability model and a one-dimensional convolutional neural network; this embodiment is based on The training diffusion strategy models action generation as a denoising diffusion process: the forward process iterates through the real action sequence. Gaussian noise is added gradually, the first... The action after the first diffusion step is: in For the first Step-by-step cumulative noise figure Uniform sampling from {1,…, }, = 100; Reverse process to train the denoising network (1D U-Net), the training loss function is: in, For policy network parameters, Standard Gaussian noise, For noise prediction networks, The observed sequence is used; the policy predicts a 16-step action sequence (n_action_steps = 8 steps) each time. The policy is conditioned on 2 historical observations (n_obs_steps = 38 dimensions in total), and training uses the AdamW optimizer with a learning rate of 1×10^6. -4 The batch size is 256, the training is conducted for 300 epochs, and an EMA model (with a decay factor of 0.75) is maintained as the inference model.

[0025] Step 3: Train the state-world model based on the source domain dataset, and use a multilayer perceptron residual prediction structure to obtain a pre-trained world model. The parameters are denoted as follows. ;reference Figure 2 Constructing a state-world model Using a residual prediction structure, the prediction formula is as follows: in, for Time-state vector For action vectors, For world model parameters; It consists of 4 fully connected layers, with 256 hidden dimensions, and uses the SiLU activation function. The input is a concatenated state-action vector. The output is an increment vector with the same dimension as the state. .

[0026] Training is performed in the normalized space using mean squared error as the loss function: in, The training sample size is given by norm(·), and the normalization function is given by norm(·). Extract all adjacent frame transfer samples ( , , The AdamW optimizer was used to train the machine for 200 epochs with a learning rate of 1×10⁻⁶. -3 The training loss decreased from 0.023 to 0.003, and the parameters of the pre-trained world model were obtained. .

[0027] Step 4: Collect a small amount of target domain demonstration data in a real operation and maintenance site. Fine-tune the pre-trained world model obtained in Step 3 on the target domain transfer samples with a small learning rate to obtain the fine-tuned world model of the target domain. The parameters are denoted as follows. This embodiment collects a small amount of target domain demonstration data at a real high-end equipment operation and maintenance site. , = 20 entries. There is a systematic observational bias between the actual operational site and the laboratory environment. The typical offset of the target installation point is approximately [0.03, 0.02, 0.01] meters (caused by sensor installation tolerances and field vibration), and the initial position offset of the end effector is approximately [0.02, 0.01, 0.02] meters (caused by positioning error of the robotic arm base). Extract all adjacent frame transition samples, with a learning rate lr = 5×10 -5 For pre-trained world models Fine-tuning for 200 epochs yields the fine-tuned world model for the target domain. After fine-tuning, the training loss on the target domain data decreased from 0.006 to 0.001.

[0028] Step 5: Fine-tune the world model using the target domain An incremental perturbation is applied to the initial state of the target domain, and multiple pseudo-trajectories are generated through forward expansion of the world model to construct a pseudo-trajectory dataset; (Refer to...) Figure 3 This embodiment utilizes the target domain to fine-tune the world model. Generate pseudo-trajectories, for The initial state of each demonstration trajectory Apply = 8 increasing perturbations, the first The initial state after the perturbation is: The perturbation scale Determined using the following incremental formula: in, = 0.01 is the basic perturbation scale. = 0.5 is the increment factor. = 8 represents the augmentation amount per demo; perturbations are applied only to the target mount point location component (dimension index 0:3) and the end effector location component (dimension index 10:13). The original demo motion sequence is preserved. ,use Stepwise forward unfolding generates pseudo-trajectories: in This represents the time step of the trajectory. For all... = 20 demos generated = 8 pseudo-trajectories, generating a total of 160 pseudo-trajectories, constituting a pseudo-trajectory dataset. .

[0029] Step 6: Merge the pseudo-trajectory dataset generated in Step 5 with the real target domain demonstration data collected in Step 4 to construct a hybrid dataset; in this embodiment, and Merge into a hybrid dataset = ∪ It contains a total of 180 trajectories (160 pseudo-trajectories + 20 real-world demonstrations); the mixed ratio design principle is that the number of pseudo-trajectories is 6 to 10 times the amount of real data, which significantly expands the training data while retaining a certain proportion of real target domain trajectories as distribution anchors.

[0030] Step 7: Using the source domain diffusion strategy parameters as initialization, fine-tune the diffusion strategy network on the mixed dataset to obtain a diffusion strategy network adapted to the target operation and maintenance site; in this embodiment, the source domain diffusion strategy parameters are used as initialization parameters. (Including EMA model parameters) is used for initialization, based on Recalculate the normalized statistics of observations and actions and update the normalization layer of the policy network; employ the AdamW optimizer (learning rate 1×10⁻⁶). -4 Weight decay 1×10 -6 Cosine annealing learning rate scheduling was used to train the model for 200 epochs; the EMA model was maintained throughout the process (EMA decay coefficient 0.75), and the EMA model weights were used as the final deployment strategy. .

[0031] Step 8: Deploy the fine-tuned diffusion strategy network at the target operation and maintenance site, and estimate the observation offset between the target domain and the source domain by statistically analyzing the mean difference. After applying domain compensation to the real-time observations, the data is input into the policy network, which outputs a sequence of robotic arm operation actions.

[0032] This embodiment is deployed at the target maintenance site. The observed offset was calibrated offline through a small number of comparative experiments. ; Observation offset The estimation formula is: in, The first observation vector One portion, (·) and (·) Take the mean of all samples in the target domain and source domain datasets respectively; the domain compensation formula is: = o real - The compensated observations Input fine-tuning diffusion strategy, enter DDIM acceleration denoising (20 inference steps, compared to training time) = 100 steps, inference speed improved by 5 times) Generates a sequence of robotic arm maintenance operations; the strategy is executed in a rolling manner: maintains 2-step observation history, and executes 8 steps each time. Drives the robotic arm to complete fine maintenance operations such as sensor mounting and bolt tightening.

[0033] This embodiment takes the task of mounting a rotary mechanical vibration sensor by a robotic arm as a specific application scenario. It was verified on the RoboMimic robot operation benchmark (MuJoCo physical simulation environment, 7-DOF Panda robotic arm). The experimental server was configured with NVIDIA A40 GPU, Ubuntu 20.04, Python 3.9, and PyTorch 2.0.

[0034] The source domain dataset was collected in a simulation laboratory environment, totaling [number missing]. = Demonstration of sensor mounting operations with 200 lines. Target domain dataset was collected by introducing systematic observation offsets (dimensions). = 19, offset vector A virtual operations and maintenance environment (e.g., [0.03, 0.02, 0.01, 0,…]) is used only. = 20 real target domain demonstrations. ConditionalUnet1D channel configuration is [256, 512, 1024].

[0035] Experimental hyperparameters: Source domain diffusion strategy training for 300 epochs; world model source domain pre-training for 200 epochs (batch size 1024, learning rate 1×10⁻⁶). -3 ); fine-tuning the world model target domain for 200 epochs (batch size 20, learning rate 5×10⁻⁶). -5 ); Pseudo-trajectory generation parameters = 8, = 0.01, = 0.5; Policy target domain fine-tuning for 200 epochs (batch size 256, learning rate 1×10⁻⁶) -4 ); Observation offset Offline calibration was performed by comparing the difference in observation statistics between 20 demonstrations from the source domain and 20 from the target domain.

[0036] Experimental results (refer to) Figure 4 The success rate of performing 100 operations in the target domain was evaluated using rollout: the direct migration (without adaptation) baseline was 16%; the target domain full data fine-tuning baseline was 38%; and the WM-DA method of this invention was 46%, which is a relative improvement of 187.5% compared with the direct migration baseline and exceeds the full data fine-tuning baseline by 8 percentage points, fully verifying the effectiveness of this method in the deployment scenario of high-end equipment maintenance robots.

Claims

1. A domain adaptation method for robotic arm operation diffusion strategy based on a world model for embodied intelligent operation and maintenance, characterized in that, Includes the following steps: Step 1: In a controlled laboratory source domain environment, collect expert demonstration data of robotic arms performing high-end equipment maintenance tasks to construct a source domain dataset; Step 2: Train a source domain diffusion strategy based on the source domain dataset, using a denoised diffusion probability model and a one-dimensional convolutional neural network to construct the diffusion strategy network; Step 3: Train a state-world model based on the source domain dataset, using a multilayer perceptron residual prediction structure to obtain a pre-trained world model, with parameters denoted as... ; Step 4: Collect a small amount of target domain demonstration data in a real operation and maintenance site. Fine-tune the pre-trained world model obtained in Step 3 on the target domain transfer samples with a small learning rate to obtain the fine-tuned world model of the target domain. The parameters are denoted as follows. Step 5: Fine-tune the world model using the target domain Step 5: Apply incremental perturbations to the initial state of the target domain, generate multiple pseudo-trajectories through forward expansion of the world model, and construct a pseudo-trajectory dataset; Step 6: Merge the pseudo-trajectory dataset generated in Step 5 with the real target domain demonstration data collected in Step 4 to construct a hybrid dataset; Step 7: Initialize the diffusion strategy network with the source domain diffusion strategy parameters, and fine-tune and train the diffusion strategy network on the hybrid dataset to obtain a diffusion strategy network adapted to the target operation and maintenance site; Step 8: Deploy the fine-tuned diffusion strategy network at the target operation and maintenance site, and estimate the observation offset between the target domain and the source domain by statistically analyzing the mean difference. After applying domain compensation to the real-time observations, the data is input into the policy network, which outputs a sequence of robotic arm operation and maintenance actions.

2. The robotic arm operation diffusion strategy domain adaptation method according to claim 1, characterized in that, Step 1 specifically involves: in a controlled laboratory source domain environment, guiding a robotic arm to perform maintenance tasks via a teleoperation system, collecting expert demonstration trajectories, and forming a source domain dataset. = {τ i }, i = 1,…, Each demonstration trajectory = {( , It consists of a time series and an observation vector. It is composed of the position and attitude state of the equipment target object, the position of the robotic arm end effector, the quaternion of the end effector attitude, and the opening and closing state of the gripper, and has dimensions. = 19; Action Vector For joint position control commands, dimensions = 7.

3. The robotic arm operation diffusion strategy domain adaptation method according to claim 2, characterized in that, Step 2 specifically involves: using a diffusion strategy to observe historical sequences. As a condition, = 2 steps, 38 dimensions in total, training a one-dimensional convolutional denoising network The training loss function is: in, For policy network parameters, Standard Gaussian noise, For noise prediction networks, For the first The action after the step spread, Uniform sampling from {1,…, }, = 100, The observed sequence is used; the policy predicts a 16-step action sequence with n_action_steps = 8 steps per execution.

4. The robotic arm operation diffusion strategy domain adaptation method according to claim 1, characterized in that, Step 3 specifically involves: State World Model Using a residual prediction structure, the prediction formula is as follows: in, for Time-state vector For action vectors, For world model parameters; It consists of a 4-layer fully connected network with 256 hidden dimensions, using SiLU activation function, and outputting an increment vector with the same dimension as the state; it is trained in a normalized space with mean squared error as the loss function. in, The training sample size is given by norm(·), and the normalization function is given by norm(·). Extract all adjacent frame transfer samples ( , , The AdamW optimizer was used to train the world model for 200 epochs to obtain the pre-trained world model parameters. .

5. The robotic arm operation diffusion strategy domain adaptation method according to claim 4, characterized in that, Step 4 specifically involves: collecting a small amount of target domain demonstration data at a real high-end equipment operation and maintenance site. = { }, j = 1,…, ; There is a systematic observational bias between real-world maintenance sites and laboratory environments. ;from Extract transition samples from adjacent frames, using a small learning rate lr = 5×10. -5 For pre-trained world models Fine-tuning for 200 epochs yields the fine-tuned world model for the target domain. .

6. The robotic arm operation diffusion strategy domain adaptation method according to claim 4, characterized in that, Step 5 specifically involves: [The text abruptly ends here, likely due to an incomplete sentence or a missing section.] The initial state of each demonstration trajectory Apply Random perturbations of different scales, the first The initial state after the perturbation is: Among the perturbation scale Determined using the following incremental formula: in, = 0.01 is the basic perturbation scale. = 0.5 is the increment factor. = 8 represents the augmentation amount per demonstration; perturbations are applied only to physically variable state components, namely the target object position and the end effector position, without perturbing the attitude quaternion and gripper state; the original motion sequence of the demonstration is then preserved. ,use Stepwise forward unfolding to generate pseudo-trajectories {( , , )},t = 0,…, -1, This represents the time step of the trajectory.

7. The robotic arm operation diffusion strategy domain adaptation method according to claim 1, characterized in that, When constructing the hybrid dataset in step 6, the pseudo-trajectory dataset will be... Demonstration dataset with real target domain Merge into a hybrid dataset = ∪ The number of pseudo-trajectories in the mixed dataset is 6 to 10 times that of the real data. This mixing ratio effectively suppresses overfitting while significantly expanding the training data.

8. The robotic arm operation diffusion strategy domain adaptation method according to claim 1, characterized in that, Step 7 specifically involves: using source domain diffusion strategy parameters The model parameters, including the exponential moving average (EMA), are initialized based on a mixed dataset. The normalized statistics of observations and actions are recalculated and the normalization layer of the policy network is updated; the AdamW optimizer is used with a learning rate of 1×10⁻⁶. -4 Weight decay 1×10 -6 Cosine annealing learning rate scheduling is used to train the model for 200 epochs while maintaining the EMA model with an EMA decay coefficient of 0.75, resulting in an adaptive diffusion strategy for the target domain. .

9. The robotic arm operation diffusion strategy domain adaptation method according to claim 1, characterized in that, Observation offset in step 8 The estimation formula is: in, The first observation vector One portion, (·) and (·) Take the mean of all samples in the target domain and source domain datasets respectively; the domain compensation formula is: = o real - The compensated observations Input fine-tuning diffusion strategy, and generate robotic arm operation sequence after DDIM acceleration and denoising; the strategy is executed in a rolling manner: maintain 2-step observation history, and execute 8 steps each time.

10. The robotic arm operation diffusion strategy domain adaptation method according to claim 1, characterized in that, The high-end equipment operation and maintenance tasks include: ① installing vibration acceleration sensors or temperature sensors at key monitoring points of rotating machinery; ② performing bolt tightening and torque testing operations at pressure vessels and pipeline connections; ③ replacing vulnerable seals or picking up and installing precision parts on precision manufacturing equipment; ④ positioning and attaching non-destructive testing probes on the surface of aero-engine fan blades or compressor blades. The source domain is a laboratory simulation platform, and the target domain is the actual on-site operation and maintenance environment of the aforementioned high-end equipment. The sources of systematic observational deviation between the two include differences in sensor installation, industrial lighting and vibration interference, and equipment attitude tolerances.

Images