Embodied intelligent operation robot operation level reasoning guide progressive diffusion strategy method

Abstract

A progressive diffusion strategy method for guiding hierarchical reasoning in embodied intelligent operation and maintenance robots is proposed. First, a demonstration trajectory dataset of operation and maintenance tasks of the robot is collected, and three-level operation and maintenance supervision labels are extracted, including global path, operation strategy, and fine-grained execution labels. Then, encoders for global path reasoning, operation strategy reasoning, and fine-grained execution reasoning are trained based on these labels. The demonstration trajectory is reasoned frame by frame, generating three-way reasoning features, which are then concatenated with the original observations to form an enhanced observation sequence. Next, using the enhanced observation sequence as conditional input, a progressive operation and maintenance strategy based on a diffusion model is trained in conjunction with a stage-aware gating network. Finally, during the robot deployment phase, the three-level reasoning features are calculated in real-time from the original environmental observations using an augmentation wrapper during reasoning and concatenated. The enhanced observations are then input into the trained progressive diffusion strategy to generate a sequence of operation and maintenance actions. This invention improves the success rate of fine-grained operation tasks of operation and maintenance robots.

Description

Technical Field

[0001] This invention belongs to the field of embodied intelligent robot operation strategy learning technology, specifically involving a Hierarchical Reasoning-guided Progressive Diffusion Policy (HRP-DP) method for guiding hierarchical reasoning in the operation of embodied intelligent operation and maintenance robots. Background Technology

[0002] With the continuous improvement of the intelligence level of industrial equipment, equipment condition monitoring and operation and maintenance (O&M) have created an urgent demand for automated robot operation. In high-risk or confined environments such as substations, aircraft engines, and petrochemical pipelines, O&M robots need to accurately complete complex tasks such as sensor placement (precisely installing vibration, temperature, and current sensors to designated locations on the equipment), equipment condition detection (manipulating detection probes to contact key measurement points on the equipment), and bolt tightening and loosening detection. These O&M operations are characterized by strong contact, high positioning accuracy requirements, and multi-stage structure, which greatly challenges the generalization and fine execution capabilities of traditional robot operation strategies.

[0003] Diffusion policy (DP) represents a significant advancement in robotics manipulation in recent years. Its core idea is to model policy learning as a conditional denoising diffusion process: given observations, an action sequence is generated from Gaussian noise through K-step iterative denoising. Diffusion policy demonstrates significant advantages in representing multimodal action distributions and handling high-dimensional continuous action spaces. However, standard diffusion strategies have the following fundamental limitations: First, there is the issue of condition consistency. In the K-step denoising process, each step uses the exact same observation conditions for guidance, lacking differentiated information for different denoising stages. Second, there is a lack of hierarchical cognitive structure. When performing tasks such as sensor mounting, maintenance robots naturally need to first determine the location of the target equipment (global path level), then determine the current maintenance stage and operation mode (strategy level), and finally precisely control the mounting contact force and position (fine execution level). Standard diffusion strategies mix all hierarchical information and cannot effectively utilize this hierarchical structure. Third, there is a mismatch between the needs of early and late denoising stages. The early stage of denoising (high noise stage) requires global path information to guide the macroscopic structure of the motion trajectory, while the late stage of denoising (low noise stage) requires fine execution information to optimize the details of contact force and position. Existing methods cannot provide targeted differentiated condition information.

[0004] To address the specific challenges of operation and maintenance tasks, existing methods suffer from the following shortcomings: While hierarchical reinforcement learning (RL) possesses a hierarchical structure, it is not suitable for diffusion denoising frameworks and struggles to handle fine-grained force control in continuous action spaces during contact-intensive operation and maintenance tasks; task decomposition methods only decompose at the task level and do not provide differentiated guidance for denoising steps within the diffusion process, failing to adaptively provide information of different granularities for high and low noise stages. The standard diffusion strategy proposed by Chi et al. (Chi, Cheng, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. "Diffusion policy: Visuomotor policy learning via action diffusion") The International Journal of Robotics Research 44, no. 10-11 (2025): 1684-1704.) provides an important foundation for the operation of maintenance robots, but all its denoising steps share the same observation conditions, making it difficult to distinguish the differentiated operational requirements of the "approaching equipment" and "precise mounting" stages in maintenance scenarios. Ma et al. explored a hierarchical diffusion strategy in multi-task robot operation (Ma, Xiao, Sumit Patidar, Iain Haughton, and Stephen James. "Hierarchical diffusion policy for kinematics-aware multi-task roboticmanipulation." In Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition (pp. 18081-18090. 2024.), but its hierarchical decomposition focuses on the separation of high-level task planning and low-level motion generation, without embedding hierarchical reasoning into different stages within the diffusion denoising process.

[0005] Therefore, there is an urgent need for a robot operation strategy learning method that can deeply integrate hierarchical operation and maintenance cognitive reasoning with diffusion denoising process, so as to achieve differentiated condition guidance for different denoising stages, thereby improving the success rate and robustness of operation and maintenance robots in delicate operation tasks such as sensor mounting and status detection. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, the purpose of this invention is to provide a progressive diffusion strategy method for guiding operational-level reasoning of embodied intelligent operation and maintenance robots. This method deeply integrates three-level operation and maintenance cognitive reasoning (global path reasoning, operation strategy reasoning, and fine execution reasoning) with the diffusion denoising process, thereby achieving differentiated condition guidance for different denoising stages and improving the success rate of fine operation tasks such as sensor mounting and status detection of operation and maintenance robots.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A progressive diffusion strategy method for guiding hierarchical reasoning in embodied intelligent operation and maintenance robots includes the following steps: Step 1: Collect a demonstration trajectory dataset of the operation and maintenance robot's operation tasks; Step 2: Automatically extract three-level operation and maintenance supervision labels from the demonstration trajectory dataset, including global path labels, operation strategy labels, and fine-grained execution labels; Step 3: Based on the three-level operation and maintenance supervision labels, train the global path reasoning encoder, the operation strategy reasoning encoder, and the fine-grained execution reasoning encoder respectively; Step 4: Using the three trained inference encoders, infer the demonstration trajectory dataset collected in Step 1 frame by frame, generate three-way inference features and splice them with the original observations to form an enhanced observation sequence; Step 5: Using the enhanced observation sequence as conditional input, and combining it with a learnable stage-aware gating network, train a progressive operation and maintenance strategy based on a diffusion model. Step 6: During the deployment phase of the operation and maintenance robot, the three-level inference features are calculated in real time from the original environmental observations by the augmentation wrapper during inference and then spliced ​​together. The augmented observations are input into the trained progressive diffusion strategy to generate a sequence of operation and maintenance actions.

[0008] In step 2, the global path label is composed of the unit direction vector from the end effector of the maintenance robot to the installation location of the target equipment and the Euclidean distance, with a dimension of 4, specifically represented as follows: in, Let be the unit direction vector from the end effector to the target equipment position. This is the Euclidean distance from the end effector to the target equipment position.

[0009] In step 2, the operation strategy label includes the one-hot encoding of the operation and maintenance phase and the offset vector from the end effector to the operation target point, with a dimension of 7. Specifically, it is represented as follows: in, One-hot encoding for the operation and maintenance phase includes four phases: location, proximity, operation, and verification. This is a three-dimensional offset vector from the end effector to the target point. During the operation and maintenance phase, the distance threshold and contact force threshold between the end effector and the target equipment surface are automatically determined without manual annotation.

[0010] In step 2, the fine-grained execution label is the velocity vector of the end effector at the current moment, which is automatically calculated through the action difference between adjacent frames. It has a 3-dimensional dimension and is specifically represented as follows: in, To ensure precise label execution, This is the velocity vector of the end effector, with a dimension of 3.

[0011] In step 3, the input observations of the three inference encoders are all standardized first: in, and The original observation vector and the standardized vector of the robot are given. and These are the mean and standard deviation of the training set observations, respectively; all three encoders adopt an MLP structure, mapping standardized observations to 32-dimensional inference features, and are trained using their respective supervised prediction heads; the parameters of the inference encoder are frozen during the diffusion strategy training phase.

[0012] In step 4, the enhanced observation sequence is constructed by concatenating the original observation sequence with the three-way inference features, specifically as follows: in, For the original observation dimension, , , These are global path, operation strategy, and fine-grained execution reasoning features, respectively.

[0013] In step 5, the stage-aware gating network takes the ratio of the current denoising step to the total number of steps as input and outputs the dynamic mixing weights of the three inference features through the softmax function: in, For the current denoising step, This represents the total number of denoising steps. It is a lightweight MLP gated network; this mechanism enables the diffusion strategy to be implemented early in the denoising process. Automatically emphasize global path reasoning, especially in the late stages of denoising. It automatically emphasizes fine-grained execution reasoning, thereby enabling progressive operation and maintenance guidance.

[0014] The training objective of the diffusion strategy in step 5 is to minimize the denoising score matching loss under enhanced observation conditions, specifically: in, To demonstrate the action, It is random noise. For the first The action after adding noise, To enhance observation A denoising network with conditions.

[0015] A progressive diffusion strategy system for guiding operational hierarchical reasoning in embodied intelligent operation and maintenance robots, comprising the following methods: Operation and maintenance trajectory acquisition module: used to collect demonstration data of the robot performing operation and maintenance tasks such as sensor mounting and status detection; Hierarchical tag extraction module: used to automatically extract global path tags, operation strategy tags, and fine-grained execution tags from the operation and maintenance demonstration trajectory; Global path reasoning encoder module: maps robot observations to 32-dimensional global path reasoning features; Operational strategy reasoning encoder module: maps robot observations to 32-dimensional operational strategy reasoning features; Fine-grained execution reasoning encoder module: maps robot observations to 32-dimensional fine-grained execution reasoning features; Stage-aware gating network module: dynamically adjusts the mixed weights of the three-layer inference features according to the denoising steps; Progressive diffusion strategy network module: Based on enhanced observation, it generates operation and maintenance action sequences through iterative denoising; The inference-time enhanced wrapper module calculates and assembles inference features in real time during the operation and maintenance deployment phase, while remaining compatible with the original operation and maintenance environment interface.

[0016] The aforementioned progressive diffusion strategy system for guiding operational hierarchy reasoning in embodied intelligent operation and maintenance robots enhances the observation dimension. satisfy: in, Indicates the original observation dimension. The output dimension of each inference encoder; the system is adaptable to any operation and maintenance strategy based on the diffusion model, without relying on a specific diffusion network backbone architecture; the inference-time enhancement wrapper supports inference-time ablation verification, which can evaluate the independent contribution of each module to the operation and maintenance performance by setting the corresponding inference components to zero without retraining.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention innovatively introduces the hierarchical cognitive structure of operation and maintenance into the diffusion strategy denoising process, filling the gap in the existing methods in guiding the differentiation stage; (2) This invention proposes a hierarchical operation and maintenance label automatic extraction method that does not require manual labeling, which automatically identifies the four operation stages of sensor mounting from the demonstration trajectory, significantly reducing implementation costs; (3) The inference-time enhanced wrapper design of this invention ensures full compatibility with any basic diffusion strategy and operation and maintenance environment interface; (4) Experimental verification shows that a 100% success rate was achieved in sensor mounting and maintenance tasks, which is 6% higher than the baseline. The ablation experiment fully verifies the necessity of each inference module. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall process of the HRP-DP operation and maintenance method according to an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram of the hierarchical reasoning encoder architecture and stage-aware gating mechanism in an embodiment of the present invention.

[0020] Figure 3 The following are training process curves for the three-level inference encoders in this embodiment of the invention: (a) is the direction cosine similarity of the global path inference encoder, (b) is the classification accuracy of the operation strategy inference encoder during the operation and maintenance phase, and (c) is the change of speed prediction MSE of the fine execution inference encoder with training rounds.

[0021] Figure 4 The chart shows a comparison of the success rate of operation and maintenance tasks in embodiments of the present invention, as well as the results of the ablation study of the inference module. Detailed Implementation

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

[0023] Reference Figure 1 A progressive diffusion strategy method for guiding operational-level reasoning in embodied intelligent operation and maintenance robots includes the following steps: Step 1: Collect a demonstration trajectory dataset of the operation and maintenance robot's operation tasks; This embodiment takes sensor mounting and maintenance task as a typical implementation scenario: In a simulated or real maintenance environment, a 7-DOF maintenance robot is controlled by remote operation or expert teaching to perform the operation task of accurately mounting vibration sensors to designated positions of rotating equipment (such as bearing housings, motor housings), and simultaneously collecting multimodal observation data and corresponding motion data of the robot to form a demonstration trajectory dataset. Each demonstration trajectory is stored as a time-step sequence, at time... Robot observations are composed of the following components: in, Image features (dimensions) extracted for visual encoder ), Joint angle vector ( (one joint) For the three-dimensional position of the end effector, For the end effector attitude quaternion, In gripper state ( (dimensional); action vector This involves splicing the incremental pose of the end effector with the control parameters of the gripper; a total of [number] data were collected. The demonstration trajectories constitute the dataset. .

[0024] Step 2: In this embodiment, three-level operation and maintenance supervision labels are automatically extracted from the demonstration trajectory dataset collected in Step 1, without the need for manual annotation; (1) Global path label: Reflects the global guidance information of the operation and maintenance robot to the installation location of the target equipment, with 4 dimensions: in, Let be the unit direction vector from the end effector to the target installation location. The three-dimensional coordinates of the target equipment installation point. The distance is Euclidean. (2) Operation strategy label: Encodes the relative offset between the current operation stage and the end effector to the operation target point, with 7 dimensions: The operation and maintenance phase is divided into four stages: the location phase (the distance from the end effector to the target is greater than a threshold). Approach phase (distance less than) And the contact force is zero), operation phase (contact force is greater than the force threshold) The process includes the placement action, the verification stage (returning to verification after placement), and the operation stage, which is automatically determined by distance threshold and force sensor signal, without the need for manual labeling. (3) Fine-grained execution label: captures the fine motion state of the end effector during mounting contact, with a dimension of 3: in, To ensure precise label execution, The velocity vector of the end effector is 3-dimensional; that is, the three-dimensional displacement vector of the end effector between adjacent frames, which is directly calculated from the trajectory data difference.

[0025] Step 3: Based on the three-level operation and maintenance supervision labels extracted in Step 2, train the global path reasoning encoder, the operation strategy reasoning encoder, and the fine-grained execution reasoning encoder respectively; Reference Figure 2 Dedicated MLP encoders were designed and trained for each of the three inference levels; all encoder inputs were first observation-normalized. in, and The original observation vector and the standardized vector of the robot are given. and These are the mean and standard deviation of the training set observations, respectively; all three encoders adopt an MLP structure, mapping standardized observations to 32-dimensional inference features, and are trained using their respective supervised prediction heads; the parameters of the inference encoder are frozen during the diffusion strategy training phase.

[0026] The network structure of all three encoders is: input layer ( (128-dimensional, ReLU) → Fully connected layer (128-dimensional, ReLU) → Fully connected layer (64-dimensional, ReLU) → Inference feature layer (32-dimensional). The global path reasoning encoder predicts a head dimension that changes from 32 to 4, and its training loss is the sum of the direction cosine similarity loss and the distance MSE loss. The operation policy inference encoder employs a dual-head design: a stage classification head (32→4, cross-entropy loss) and an operation point offset regression head (32→3, MSE loss). The fine-grained inference encoder, with a prediction head dimension of 32→3, employs the speed MSE loss: All three encoders use the Adam optimizer with an initial learning rate of A cosine annealing decay strategy was used, trained for 100 epochs, with a batch size of 512. (See reference...) Figure 3 The training results of the inference encoders are as follows: the global path inference encoder achieved a direction cosine similarity of 0.9931; the operation strategy inference encoder achieved a classification accuracy of 98.12% during the operation and maintenance phase; and the fine execution inference encoder achieved a speed prediction MSE as low as [missing value]. All three encoders reliably extracted operation and maintenance task-related information at each level from the raw observations.

[0027] Step 4: Using the three inference encoders trained in Step 3, infer the demonstration trajectory dataset collected in Step 1 frame by frame, generate three-way inference features and splice them with the original observations to form an enhanced observation sequence. in, For the original observation dimension, , , These are respectively global path, operation strategy, and fine-grained execution reasoning characteristics; The augmented observation dataset is stored in HDF5 format, and three inference feature keys, reasoning_global, reasoning_strategy, and reasoning_refine, are added to the original observation data for subsequent diffusion strategy training; the parameters of the inference encoder are frozen in this step and subsequent steps.

[0028] Step 5: Using the enhanced observation sequence generated in Step 4 as conditional input, and combining it with a learnable stage-aware gating network, train a progressive operation and maintenance strategy based on a diffusion model. This embodiment uses the enhanced observation sequence generated in step 4 as conditional input to train a progressive diffusion strategy network based on DDPM (U-Net architecture, 100-step inference). The core innovation lies in introducing a stage-aware gating network, which uses the ratio of the current denoising step to the total number of steps as input to achieve dynamic and differentiated utilization of different inference levels at different denoising stages.

[0029] The dynamic weights of the stage-aware gating network are calculated using the softmax function as follows: in, For the current denoising step, This represents the total number of denoising steps. It is a lightweight MLP gated network; Gated Networks It is a lightweight MLP. The inference features after dynamic weighting are: In the early stage of denoising (k / K→1, high noise stage), the gating network automatically emphasizes global path reasoning, providing guidance for the macroscopic structure of motion trajectory (the direction of robot movement towards the equipment); in the late stage of denoising (k / K→0, low noise stage), the gating network automatically emphasizes fine execution reasoning, optimizing the fine force and position control details in the sensor mounting contact stage.

[0030] The training objective of the diffusion strategy is to minimize the denoising score matching loss under augmented observation conditions. in, To demonstrate the action, It is random noise. For the first The action after adding noise, To enhance observation A conditional denoising network; Training configuration: Adam optimizer, initial learning rate Cosine annealing decay, training for 2000 epochs, batch size 256, EMA model parameter update, inference encoder parameters frozen and not participating in gradient update.

[0031] Step 6: During the deployment phase of the operation and maintenance robot, the three-level inference features are calculated in real time from the original environmental observations by the augmentation wrapper during inference and then spliced ​​together. The augmented observations are input into the trained progressive diffusion strategy to generate a sequence of operation and maintenance actions.

[0032] In this embodiment, during the deployment phase of the maintenance robot, the inference-time augmentation wrapper implements the following real-time augmentation process: Input: Raw observations of the operation and maintenance environment Normalization process: Three-way inference coding (parallel): , , Enhanced observation stitching: Progressive diffusion strategy inference (100-step iterative denoising): Output: Operation and maintenance sequence, directly controlling the robot to perform sensor mounting or status detection operations.

[0033] This design ensures full compatibility with the original operation and maintenance environment interface without requiring modification of the environment code; the inference incremental computation overhead is only three-way MLP forward inference (about 0.3ms), which has no significant impact on the real-time performance of operation and maintenance.

[0034] This embodiment uses the sensor mounting and maintenance task of industrial rotating equipment as a typical case to verify the effectiveness of the method of the present invention.

[0035] Experimental Platform: A 7-DOF Franka Emika Panda industrial robotic arm was used, with an eye-in-hand camera and an ATI Mini45 force / torque sensor mounted at the end effector. An operational scenario was built within the robosuite / robomimic simulation environment. Task Objective: To precisely mount a vibration acceleration sensor to a designated measurement point on a rotary bearing housing, encompassing four stages: positioning, approach, mounting operation, and retraction verification.

[0036] Observation space: Image features (64-dimensional features extracted by ResNet-18 encoder) + joint angles (7-dimensional) + end effector position (3-dimensional) + end effector quaternions (4-dimensional) + gripper state (2-dimensional) = 80 dimensions, i.e. Motion space: End effector incremental pose + gripper control = 7 dimensions.

[0037] Dataset: 200 expert teaching and maintenance demonstration trajectories were collected, totaling 18,640 time steps, of which 160 were used for training and 40 for validation. The hierarchical inference encoder was trained for 100 epochs (approximately 3 minutes / RTX 4090), and the diffusion strategy was trained for 2000 epochs (approximately 4.5 hours / RTX 4090). Hyperparameter settings: sequence length T=2, prediction window T_a=8, diffusion steps K=100, gating network is a two-layer MLP (1→16→3), inference feature dimension... .

[0038] Reference Figure 4 The experimental results are as follows: (1) Comparison of task success rates (50 independent test rounds): The baseline success rate of the standard diffusion strategy was 94.0% (3 failures, all due to force position control deviation in the fine mounting stage); the success rate of HRP-DP (this invention) was 100.0% (zero failures), an improvement of 6.0% compared to the baseline. (2) Ablation study of the inference module (zeroing during inference): After removing global path inference, the success rate dropped to 22.0% ( The success rate was 78.0%, indicating that global path guidance is crucial for robot target addressing; after removing operation strategy reasoning, the success rate dropped to 0.0%. The success rate was 100.0%, indicating that identification information is the most critical module during the operation and maintenance phase; without it, the robot completely loses its perception of the placement operation phase. After removing the fine-grained execution reasoning, the success rate dropped to 82.0%. (18.0%), indicating that precise execution information makes a significant contribution to the force-position control accuracy during the placement contact stage. Ablation experiments fully validated the necessity of each of the three inference modules.

Claims

1. A progressive diffusion strategy method for guiding hierarchical reasoning in embodied intelligent operation and maintenance robots, characterized in that, Includes the following steps: Step 1: Collect a demonstration trajectory dataset of the operation and maintenance robot's operation tasks; Step 2: Automatically extract three-level operation and maintenance supervision labels from the demonstration trajectory dataset, including global path labels, operation strategy labels, and fine-grained execution labels; Step 3: Based on the three-level operation and maintenance supervision labels, train the global path reasoning encoder, the operation strategy reasoning encoder, and the fine-grained execution reasoning encoder respectively; Step 4: Using the three trained inference encoders, infer the demonstration trajectory dataset collected in Step 1 frame by frame, generate three-way inference features and splice them with the original observations to form an enhanced observation sequence; Step 5: Using the enhanced observation sequence as conditional input, and combining it with a learnable stage-aware gating network, train a progressive operation and maintenance strategy based on a diffusion model. Step 6: During the deployment phase of the operation and maintenance robot, the three-level inference features are calculated in real time from the original environmental observations by the augmentation wrapper during inference and then spliced ​​together. The augmented observations are input into the trained progressive diffusion strategy to generate a sequence of operation and maintenance actions.

2. The method according to claim 1, characterized in that, In step 2, the global path label is composed of the unit direction vector from the end effector of the maintenance robot to the installation location of the target equipment and the Euclidean distance, with a dimension of 4, specifically represented as follows: in, Let be the unit direction vector from the end effector to the target equipment position. This is the Euclidean distance from the end effector to the target equipment position.

3. The method according to claim 1, characterized in that, In step 2, the operation strategy label includes the one-hot encoding of the operation and maintenance phase and the offset vector from the end effector to the operation target point, with a dimension of 7. Specifically, it is represented as follows: in, One-hot encoding for the operation and maintenance phase includes four phases: location, proximity, operation, and verification. This is a three-dimensional offset vector from the end effector to the target point. During the operation and maintenance phase, the distance threshold and contact force threshold between the end effector and the target equipment surface are automatically determined without manual annotation.

4. The method according to claim 1, characterized in that, In step 2, the fine-grained execution label is the velocity vector of the end effector at the current moment, which is automatically calculated through the action difference between adjacent frames. It has a 3-dimensional dimension and is specifically represented as follows: in, To ensure precise labeling, This is the velocity vector of the end effector, with a dimension of 3.

5. The method according to claim 1, characterized in that, In step 3, the input observations of the three inference encoders are all standardized first: in, and The original observation vector and the standardized vector of the robot are given. and These are the mean and standard deviation of the training set observations, respectively; all three encoders adopt an MLP structure, mapping standardized observations to 32-dimensional inference features, and are trained using their respective supervised prediction heads; the parameters of the inference encoder are frozen during the diffusion strategy training phase.

6. The method according to claim 1, characterized in that, In step 4, the enhanced observation sequence is constructed by concatenating the original observation sequence with the three-way inference features, specifically as follows: in, For the original observation dimension, , , These are global path, operation strategy, and fine-grained execution reasoning features, respectively.

7. The method according to claim 1, characterized in that, In step 5, the stage-aware gating network takes the ratio of the current denoising step to the total number of steps as input and outputs the dynamic mixing weights of the three inference features through the softmax function: in, For the current noise reduction step, This represents the total number of denoising steps. It is a lightweight MLP gated network; this mechanism enables the diffusion strategy to be implemented early in the denoising process. Automatically emphasize global path reasoning, especially in the late stages of denoising. It automatically emphasizes fine-grained execution reasoning, thereby enabling progressive operation and maintenance guidance.

8. The method according to claim 1, characterized in that, The training objective of the diffusion strategy in step 5 is to minimize the denoising score matching loss under enhanced observation conditions, specifically: in, To demonstrate the action, It is random noise. For the first The action after adding noise, To enhance observation A denoising network with conditions.

9. A progressive diffusion strategy system for guiding hierarchical reasoning in embodied intelligent operation and maintenance robots, implementing the progressive diffusion strategy method for guiding hierarchical reasoning in embodied intelligent operation and maintenance robots as described in any one of claims 1-8, characterized in that, include: Operation and maintenance trajectory acquisition module: used to collect demonstration data of the robot performing operation and maintenance tasks such as sensor mounting and status detection; Hierarchical tag extraction module: used to automatically extract global path tags, operation strategy tags, and fine-grained execution tags from the operation and maintenance demonstration trajectory; Global path reasoning encoder module: maps robot observations to 32-dimensional global path reasoning features; Operational strategy reasoning encoder module: maps robot observations to 32-dimensional operational strategy reasoning features; Fine-grained execution reasoning encoder module: maps robot observations to 32-dimensional fine-grained execution reasoning features; Stage-aware gating network module: dynamically adjusts the mixed weights of the three-layer inference features according to the denoising steps; Progressive diffusion strategy network module: Based on enhanced observation, it generates operation and maintenance action sequences through iterative denoising; The inference-time enhanced wrapper module calculates and assembles inference features in real time during the operation and maintenance deployment phase, while remaining compatible with the original operation and maintenance environment interface.

10. The system according to claim 9, characterized in that, Enhance observation dimensions satisfy: in, Indicates the original observation dimension. The output dimension of each inference encoder; the system is adaptable to any operation and maintenance strategy based on the diffusion model; the inference-time enhancement wrapper supports inference-time ablation verification, which can evaluate the independent contribution of each module to the operation and maintenance performance by setting the corresponding inference components to zero without retraining.

Images