Embodied intelligent operation robot multi-level safety constraint field guidance diffusion strategy method and system

Abstract

A diffusion strategy method and system for guiding multi-level safety constraint fields in embodied intelligent operation and maintenance robots is disclosed. First, a multi-level safety constraint field is constructed for equipment operation and maintenance scenarios, quantifying the safety risks of the safe operation zone, equipment sensitive access zone, personnel activity interference zone, and equipment protection prohibition zone in the form of a continuous differentiable scalar field. Then, a multi-modal state observation module for the operation and maintenance scenario is designed, integrating robot body state, target equipment geometry, and safety topology distance information. Next, the gradient of the multi-level constraint field is injected into each step of the diffusion strategy's denoising calculation using an embedded denoising guidance method, causing the denoising trajectory to deviate along the direction of decreasing constraint field value. Simultaneously, a learnable adaptive guidance intensity scheduling function is employed, outputting a larger guidance intensity in the macro-path planning stage and a minimal guidance intensity in the fine-grained control stage. Finally, an operation and maintenance action sequence that satisfies the multi-level safety constraints is output. This invention achieves an optimal balance between safety constraints and fine-grained control.

Description

Technical Field

[0001] This invention belongs to the field of embodied intelligence and robot operation and maintenance technology, specifically relating to a diffusion strategy method and system for guiding multi-level safety constraint fields for embodied intelligent operation and maintenance robots. Background Technology

[0002] Large industrial equipment requires continuous monitoring of various physical parameters such as vibration, temperature, and pressure during its service life. Traditional methods relying on manual sensor placement suffer from significant drawbacks, including high operational risks, poor consistency, and low efficiency. Meanwhile, maintenance sites commonly contain hazardous sources such as high-voltage electrical equipment, high-temperature pipelines, and rotating machinery, imposing stringent safety requirements on operators and automated equipment entering the work area. Embossed intelligent maintenance robots, as core equipment for achieving autonomous maintenance and replacing manual labor, are gradually being explored and applied in high-risk fields such as power, chemical, and nuclear power.

[0003] The current mainstream approach in robot policy learning—the diffusion strategy—models action generation as a conditional diffusion denoising process, demonstrating excellent generalization ability in complex manipulation tasks. However, the diffusion strategy has the following fundamental shortcomings in terms of operational safety constraints: First, it lacks specific safety awareness capabilities for operational scenarios, and the policy network fits dangerous and safe actions indiscriminately; second, existing safety constraints are mostly applied as post-processing after the diffusion process is completed, resulting in a disconnect between constraint correction and the diffusion denoising process, leading to abrupt changes in the action curve; third, safety metrics typically use binary judgment, failing to provide continuous gradient information to guide the denoising direction.

[0004] Fixed-strength safety constraints are insufficient to meet the phased requirements of operation and maintenance tasks: operation and maintenance tasks include two distinct phases: macro-path planning and fine-grained contact control. The macro-path planning phase requires strong spatial safety constraints, while the fine-grained control phase requires high-precision force control guidance. Fixed-strength safety constraints cannot meet the differentiated requirements of the two phases.

[0005] The diffusion strategy proposed by Chi et al., which characterizes robot visual motion strategies as a conditional denoising diffusion process, has achieved significantly better performance than existing methods in various manipulation tasks. However, it lacks explicit modeling of safety constraints (ChiC, Xu Z, Feng S, et al. Diffusion Policy: Visuomotor Policy Learning via Action Diffusion[J]. The International Journal of Robotics Research, 2024, 44(10-11), 1684-1704.). Siciliano et al., in their classic work on robot modeling and control, pointed out that the safety planning of collaborative maintenance robots in constrained operating spaces needs to simultaneously consider both contact force control and spatial collision avoidance (Siciliano B, Sciavicco L, Villani L, et al. Robotics: Modelling, Planning and Control[M]. Springer, 2009.). These limitations make it difficult for existing solutions to achieve practical engineering safety levels in high-value maintenance scenarios such as autonomous sensor mounting in industrial applications and hazardous area inspections.

[0006] Therefore, there is an urgent need for a multi-level safety constraint field-guided diffusion strategy method for embodied intelligent operation and maintenance robots, which can not only retain the high-quality action generation capability of the diffusion strategy, but also continuously apply continuous, differentiable, and stage-adaptive safety constraints within the denoising process, thereby breaking through the bottleneck of existing technologies in industrial operation and maintenance safety. Summary of the Invention

[0007] To overcome the shortcomings of the prior art, the present invention aims to provide a diffusion strategy method and system for guiding multi-level safety constraint fields for embodied intelligent operation and maintenance robots. The method models multi-level safety constraints in the form of a continuous scalar field and embeds them into the diffusion denoising process in a gradient-guided manner. This fundamentally solves the problems of existing diffusion strategies in industrial operation and maintenance scenarios, such as lack of safety perception, constraints acting outside the process, discretized safety measurements, and fixed constraint strength. It is applicable to intelligent operation and maintenance scenarios such as sensor mounting in industrial equipment, equipment status inspection, and operation parameter acquisition.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A diffusion strategy method for guiding multi-level safety constraint fields in embodied intelligent operation and maintenance robots includes the following steps: Step 1: Construct a multi-level safety constraint field for operation and maintenance scenarios: Collect successful operation trajectories of embodied operation and maintenance robots in operation and maintenance tasks, extract state-action pairs, train the constraint field network, take robot state and operation and maintenance actions as input, and output non-negative scalar constraint field values. Through multi-level annotation, the safe operation area, equipment sensitive approach area, personnel activity interference area, equipment protection prohibition area and spatial restricted area are continuously and uniformly encoded. Step 2, Construct a multimodal state observation for the operation and maintenance scenario: Integrate the robot end effector pose, joint state, local features of the target device's 3D point cloud, coordinates of the sensor installation target point, and the safety topology distance information of the current working area to form an operation and maintenance context observation vector. , serving as the unified condition input for the constraint field network and the diffusion strategy; Step 3, Constraint Field Gradient Embedded Denoising: In each step of the denoising calculation of the diffusion strategy, a constraint field gradient correction is applied after the standard denoising output, causing the denoising trajectory to shift along the direction of decreasing constraint field value. The correction formula is as follows: in The first after constraint field correction Noise reduction steps For the standard diffusion denoiser in the first The noise reduction action of the step output, Number the current denoising step. For adaptive guidance strength, To constrain the gradient of field operations and maintenance actions and direct it towards higher risk directions, subtracting this gradient automatically biases the denoising results towards safer operating areas. Step 4, Adaptive Guidance Strength Scheduling: A learnable guidance strength scheduling function is adopted. In the early stage of denoising, namely the macro-path planning stage, the maximum guidance strength value is output to strongly constrain the overall movement direction of the maintenance robot away from the danger zone. In the later stage of denoising, namely the fine control stage, the minimum guidance strength value is output to preserve the generation quality of the fine control action of the diffusion strategy. The minimum guidance strength value is shrunk to less than 10% of the maximum guidance strength value. Step 5, Output Safety Operation and Maintenance Action Sequence: After the complete denoising process guided by the constraint field, the operation and maintenance action sequence that meets the multi-level safety constraints is output. The robot then performs the operation and maintenance tasks of sensor mounting, equipment status acquisition, and inspection path following in sequence.

[0009] The specific method for constructing training data for the multi-level security constraint field in step 1 is as follows: Safety operation level sample label is State-action pairs are extracted directly from the demonstration trajectory of the embodied maintenance robot successfully completing maintenance tasks; Device proximity level sample label Add a standard deviation to the safe operating level actions. Gaussian perturbation; Personnel interference risk level sample label Add a standard deviation to the safe operating level actions. Gaussian perturbation; Equipment protection prohibition level sample label is Replace the actions of the safety operation level samples with actions of uniformly distributed random sampling; Space restricted area sample label is Extract the state-action pair when the end effector moves into a predefined prohibited operation area; in To constrain the field label value, indicating the safety risk level of the state-action pair, The standard deviation of the equipment's sensitive proximity level disturbance. The standard deviation of the disturbance level caused by human interference.

[0010] Step 1, the constraint field network, specifically includes: an input layer, which takes the operation and maintenance context state vector as input. Operation and maintenance action vectors The input is concatenated as a joint input; multiple fully connected hidden layers are used, each followed by a normalization layer and a SiLU activation function; the output layer uses the Softplus activation function to ensure the constraint field values. The nonnegativity of the constraint field network is observed; the constraint field network is trained under supervision using the Huber loss function. in The value of Huber's loss function. This is the Huber loss threshold. For the constrained field network prediction value, The value of the constraint field label is used for annotation.

[0011] Step 2: Multimodal state observation vectors in the operation and maintenance scenario The specific components are as follows: robot body state sub-vector, including the three-dimensional position of the end effector, quaternion pose, joint angle vector and mounting tool state; target device state sub-vector, including the three-dimensional coordinates of the target sensor mounting point, the device surface normal vector estimation and the device current operating status code; safety topology distance sub-vector, including the shortest distance estimation from the end effector to each predefined safety boundary; the sub-vectors are concatenated to form a unified operation and maintenance context observation vector.

[0012] In step 3, the constraint field gradient is calculated using an automatic differentiation mechanism: the operation and maintenance action tensor is... Set the state to require gradient calculation; forward calculation of constraint field values. ;right about Perform backpropagation to obtain gradients Gradients are clipped using the L2 norm before the correction is applied. in The cropping threshold, It is the L2 norm of the gradient vector to prevent excessive motion correction caused by sudden changes in constraint field values, which could lead to joint impact on the robot.

[0013] In step 4, the adaptive guidance strength scheduling function is implemented by the guidance strength scheduling network, which uses a normalized denoising step. As input, the shape factor is output through two fully connected layers and a sigmoid activation function, and then compared with the learnable maximum guided strength parameter. Multiplying them together yields the final guidance strength value: in For the first The guiding strength value for step-by-step noise reduction. The maximum learnable guidance strength parameter. For parameterized scheduling networks, For the trainable parameters of the scheduling network, It is the Sigmoid activation function. Number the current denoising step. This represents the total number of noise reduction steps; during the fine mounting stage. At that time, it automatically shrinks to within 10% of the maximum value; during the macro-path planning stage. At the same time, maintain maximum guidance intensity.

[0014] The diffusion strategy method also includes Candidate security screening inference mode: Observe and execute the current state of the maintenance robot. Sub-independent diffusion strategy sampling, to obtain Group maintenance action candidate sequences; use constraint field networks to analyze each group of candidates. Step motion calculation of average constraint field value: in For the first The average constraint field score of the candidate action sequence. For candidate sequence index, This represents the total number of candidate sequences sampled independently. The number of action steps contained in a single candidate sequence. For time step indexes within the action sequence, For the first The operational context state of the step. For the first The first group of candidate sequences Step movement, The constraint field label value is used; the candidate sequence with the lowest constraint field value is selected as the final action to be executed. The candidate security screening mode does not require modification of the basic diffusion strategy model structure and can be used as a training-free security enhancement plugin for already deployed and maintained diffusion strategies.

[0015] A diffusion strategy system for guiding multi-level safety constraint fields in embodied intelligent operation and maintenance robots, comprising the following diffusion strategy methods: The maintenance robot body module includes a robotic arm, an end effector, and a sensor mounting and actuator. The multimodal perception module includes a joint encoder, a vision sensor, a torque sensor, and a 3D point cloud acquisition unit; The security topology map module pre-stores 3D security partitioning information for the operation and maintenance area; The constraint field network module is used to learn and evaluate multi-level safety constraint field values ​​for state-action pairs in real time. The basic module for diffusion strategies includes a pre-trained neural network for diffusion strategies of operation and maintenance tasks and a noise scheduler; The guidance intensity scheduling module outputs adaptive guidance intensity based on the current denoising step and maintenance task stage. The constraint field-guided denoising module uses the constraint field gradient and adaptive guidance strength to safely correct the operation and maintenance action trajectory after each denoising calculation. The operation and maintenance task scheduling module is responsible for parsing operation and maintenance task instructions, planning inspection paths, and scheduling sensor mounting sequences. The modules collaborate through a unified operation and maintenance control interface.

[0016] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned diffusion strategy method for guiding multi-level safety constraint fields for embodied intelligent operation and maintenance robots.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention introduces the operation and maintenance safety constraints from the outside of the diffusion process to the inside, eliminates the sudden change in the action curve caused by post-processing projection, and ensures smooth contact force during the sensor mounting process; (2) The five-level continuous constraint field of this invention for operation and maintenance scenarios provides more refined risk quantification than binary safety boundaries; (3) The adaptive guidance intensity scheduling of the task phase of this invention achieves the optimal balance between macroscopic security and fine-grained control quality; (4) Experiments have shown that the safety violation rate of the present invention in the safety disturbance scenario is reduced from 35.9% to 0%, and the constraint field value is improved by 83.1%; (5) The N-candidate security screening mode of the present invention provides a training-free plug-and-play deployment option. Attached Figure Description

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

[0019] Figure 2 This is a schematic diagram of constrained field gradient embedded denoising and adaptive guidance intensity scheduling in an embodiment of the present invention.

[0020] Figure 3 This invention presents a comparison of the heatmap and adaptive guidance intensity scheduling curve of the multi-level safety constraint field in the operation and maintenance scenario of the embodiment of the invention with the constraint field value trajectory of the denoising process.

[0021] Figure 4 The results are experimental results of the safety performance of the sensor mounting task in an embodiment of the present invention. Detailed Implementation

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

[0023] Reference Figure 1 A diffusion strategy method for guiding multi-level safety constraint fields in embodied intelligent operation and maintenance robots includes the following steps: Step 1: Construct a multi-level safety constraint field for operation and maintenance scenarios: Collect successful operation trajectories of embodied operation and maintenance robots in operation and maintenance tasks, extract state-action pairs, train the constraint field network, take robot state and operation and maintenance actions as input, and output non-negative scalar constraint field values. Through multi-level annotation, the safe operation area, equipment sensitive approach area, personnel activity interference area, equipment protection prohibition area and spatial restricted area are continuously and uniformly encoded. This embodiment constructs a multi-level safety constraint field oriented towards equipment maintenance scenarios in the state-action space of the embodied maintenance robot. The constraint field network takes the concatenation of the robot's maintenance context state vector and maintenance action vector as input and outputs non-negative scalar constraint field values, which are mathematically defined as follows: in The scalar constraint field value output by the constraint field network. For the operation and maintenance context state space, For operational and maintenance action space, It is the set of nonnegative real numbers; The constraint field constructs a training set using multi-level labeled data. The labeling rules for the five levels are as follows: Safety Operational Grade Samples (Labels) ): Extract state-action pairs directly from the demonstration trajectory of the embodied maintenance robot successfully completing maintenance tasks; Device-sensitive proximity samples (labels) ): Add standard deviation to safe operating level actions. Gaussian perturbation ,in This is a random perturbation noise vector. It follows a Gaussian distribution. The standard deviation of the device's proximity-sensitive disturbance is used to simulate the slight positional jitter when the end effector approaches the device surface. Human interference risk level samples (labels) ): Add standard deviation to safe operating level actions. Gaussian perturbation ,in The standard deviation of the danger level disturbance caused by human interference is used to simulate the dangerous actions of the operation and maintenance robot's trajectory deviating into the human activity area; Equipment protection prohibited level samples (labels) ): Replace the action of the safety operation level sample with the action in The action of uniformly distributed random sampling simulates a dangerous operation that is completely out of control. Space exclusion zone samples (labels) ): Extract the state-action pairs when the end effector moves to a predefined prohibited operation area; the relevant restricted area boundaries are automatically generated from the 3D security topology map of the operation and maintenance scenario. The specific architecture of the constrained field network is as follows: the input layer concatenates the state vector and the action vector; multiple fully connected hidden layers, each followed by a normalization layer and a SiLU activation function; the output layer uses the Softplus activation function to ensure non-negativity of the output; training uses the Huber loss function. in The value of Huber's loss function. This is the Huber loss threshold. To predict the output of the constrained field network, the Adam optimizer is used in conjunction with a cosine annealing learning rate scheduling strategy for training. Step 2, Construct a multimodal state observation for the operation and maintenance scenario: Integrate the robot end effector pose, joint state, local features of the target device's 3D point cloud, coordinates of the sensor installation target point, and the safety topology distance information of the current working area to form an operation and maintenance context observation vector. , serving as the unified condition input for the constraint field network and the diffusion strategy; This embodiment performs structured fusion of the operation and maintenance robot state information involved in step 1 to construct an operation and maintenance context observation vector. It is composed of the following three sub-vectors: in For the operation and maintenance context observation vector, The robot body state subvector includes the three-dimensional position of the end effector, quaternion pose, joint angle vector, and the state of the mounting tool. The target device state subvector contains the three-dimensional coordinates of the target sensor mounting point, the estimated normal vector of the device surface, and the current operating status code of the device. The safety topology distance subvector contains the shortest distance estimates from the end effector to each predefined safety boundary; Step 3, Constraint Field Gradient Embedded Denoising: In each step of the denoising calculation of the diffusion strategy, the following operations are performed on the operational context observation vector obtained in Step 2 and the current denoising trajectory: First, standard denoising prediction is completed, then the constraint field gradient is calculated using the constraint field network trained in Step 1 and a safety correction is applied. The correction formula is: in The first after constraint field correction Noise reduction steps For the standard diffusion denoiser in the first The noise reduction action of the step output, Number the current denoising step. For the first Adaptive guiding strength for step-by-step noise reduction. To constrain the gradient of the field to the operation and maintenance actions, the gradient is pointed in the direction where the value of the constraint field increases, i.e., the more dangerous it is. Subtracting this gradient makes the denoising trajectory automatically biased towards a safer operation and maintenance area. The constraint field gradient is calculated using an automatic differentiation mechanism: the maintenance action tensor is... Set the state to require gradient calculation; forward calculation of constraint field values. ;right about Perform backpropagation to obtain gradients Gradients are clipped using the L2 norm before the correction is applied. in The cropping threshold, The L2 norm of the gradient vector is used to prevent excessively large action correction amplitudes caused by sudden changes in constraint field values. A schematic diagram of the embedded denoising and adaptive guidance intensity scheduling of the constraint field gradient in this embodiment is shown below. Figure 2As shown, in each denoising step, the diffusion denoiser first outputs a standard denoised predicted action based on the current noisy action and the operation and maintenance context observation vector. Then, the constraint field network calculates the constraint field value for the predicted action and obtains the gradient direction through automatic differentiation. This gradient is clipped by the L2 norm and multiplied by the adaptive guidance strength to obtain a safety correction amount, which is then applied to the denoised action. This causes the corrected action to shift to a safer operation and maintenance area along the direction of the decreasing constraint field value. At the same time, the adaptive guidance strength scheduling network takes the normalized denoising step as input. In the early stage of denoising (macro path planning stage), it outputs a larger guidance strength to ensure the safety of the overall motion trajectory. In the later stage of denoising (fine mounting stage), it automatically decays the guidance strength to retain the fine-grained accuracy of the action, thus achieving a dynamic balance between safety constraints and control accuracy. Step 4, Adaptive Guiding Strength Scheduling: Design a learnable guiding strength scheduling function to dynamically output guiding strength values ​​based on the current denoising step in Step 3; the guiding strength scheduling network normalizes the denoising steps. Input: Output guidance strength in The maximum learnable guidance strength parameter. It is a scheduling network consisting of two fully connected layers and a SiLU activation function. For the trainable parameters of the scheduling network, It is the Sigmoid activation function. The total number of denoising steps for the diffusion strategy; in the fine mounting stage ( ), Automatically shrinks to within 10% of its maximum value, protecting the fine-grained movements required for precise placement; during the macro-path planning stage ( ), Maintaining maximum guidance intensity ensures the overall movement trajectory of the maintenance robot stays away from personnel and hazardous equipment areas. This embodiment compares the heatmap of the multi-level safety constraint field in the end effector's motion space and the adaptive guidance intensity scheduling curve with the constraint field value trajectory during the denoising process. Figure 3As shown in the constraint field heatmap, the safe operation area corresponds to a lower constraint field value (cold color area), while the equipment sensitive approach area, personnel interference danger area, and equipment protection prohibition area correspond to progressively higher constraint field values ​​(warm color area). The spatial forbidden zone corresponds to the highest constraint field value, indicating that the constraint field network has encoded multi-level safety semantics in the end effector action space in the form of a continuous scalar field. In the adaptive guidance intensity scheduling curve, the guidance intensity gradually decreases from close to 1 (early stage of denoising) to close to 0 (late stage of denoising) with the normalized denoising step, showing a monotonically decreasing trend, which verifies the "safety first, accuracy later" guidance strategy learned by the scheduling network. In the comparison of constraint field value trajectories during the denoising process, the constraint field value of the diffusion strategy using constraint field guidance is always maintained within the safe operation level range throughout the denoising process, while the constraint field value of the standard diffusion strategy without constraint field guidance increases significantly when subjected to safety disturbances, further verifying the effectiveness of the method of the present invention in real-time safety correction during the denoising process. Step 5, Output Safety Operation and Maintenance Action Sequence: After the complete denoising process guided by the constraint field in Steps 3 and 4, the operation and maintenance action sequence that satisfies multi-level safety constraints is output. The robot then performs operation and maintenance tasks such as sensor mounting, equipment status acquisition, and inspection path following in sequence.

[0024] Furthermore, the present invention also includes Candidate security screening inference mode: Observe and execute the current state of the maintenance robot. The sub-independent diffusion strategy sampling utilizes the constraint field network trained in step 1 to sample each candidate group. Step motion calculation of average constraint field value: in For the first The average constraint field score of the candidate action sequence. For candidate sequence index, This represents the total number of candidate sequences sampled independently. For the first Step-by-step operation and maintenance context state, For the first The first group of candidate sequences Step movement, For time step indexes within the action sequence, The number of action steps contained in a single candidate sequence; the candidate sequence with the lowest constraint field value is selected as the final action to be executed. The candidate security screening mode does not require modification of the basic diffusion strategy model structure and can be used as a training-free security enhancement plugin for already deployed and maintained diffusion strategies.

[0025] This embodiment uses an embodied maintenance robot performing a vibration sensor mounting task in an industrial setting as an example. This task requires the robot's end effector to start from its initial docking position, navigate to the designated installation point on the target industrial equipment, and precisely attach the magnetic vibration sensor to the equipment's outer surface using appropriate contact force and posture, ensuring that the entire process does not intrude into the high-voltage area of ​​the equipment or interfere with on-site inspection personnel. The motion space is 7-dimensional continuous motion (6-DOF pose increment of the end effector and clamping force control of the mounting tool), and the observation space is 46-dimensional.

[0026] The constraint field training data was constructed from 300 successful vibration sensor installation trajectories demonstrated by experts, and labeled into five levels: 32,450 samples for safe operation, 32,450 samples for equipment sensitive proximity (Gaussian perturbation standard deviation 0.10), 32,450 samples for personnel interference hazard (Gaussian perturbation standard deviation 0.45), 32,450 samples for equipment protection prohibition (uniformly distributed random motion), and 92 samples for restricted space. The total number of training samples was 130,037, and the number of test samples was 14,420.

[0027] Constrained field network architecture: Input layer Linear(53, 256), Hidden layer 1 Linear(256, 256), Hidden layer 2 Linear(256, 128), Output layer Linear(128, 1) with Softplus activation. Approximately 128K parameters. Training configuration: Adam optimizer (learning rate...). Weight decay Huber's losses ( Cosine annealing learning rate scheduling, batch size 2048, training 200 rounds.

[0028] The basic diffusion strategy employs a conditional UNet architecture (approximately 66M parameters) and is pre-trained on 270 demonstration trajectories (2000 epochs). During inference, the DDPM noise scheduler is used, with a total denoising step count of [missing information]. The bootstrap strength scheduling network structure is: Linear(1,32) plus SiLU plus Linear(32,32) plus SiLU plus Linear(32,1) plus Sigmoid, and the output is multiplied by the maximum bootstrap strength parameter.

[0029] The experiment adopted an offline evaluation scheme and designed two evaluation scenarios: normal operation and maintenance scenario and security disturbance scenario (injecting standard deviation with a 40% probability). (dangerous disturbances). Experimental results show that: referring to Figure 4In normal operation and maintenance scenarios, the average constraint field value of the method of this invention is 0.0712, which is the same as the standard diffusion strategy of 0.0712, proving that the safety constraints do not affect the quality of normal operation. In safety disturbance scenarios, the average constraint field value of the method of this invention is 0.0695, which is significantly better than the standard diffusion strategy of 0.4203 (improvement of 83.1%), and the safety violation rate is reduced from 35.9% to 0%. All 19 dangerous disturbance injection events are 100% identified and corrected.

Claims

1. A diffusion strategy method for guiding multi-level safety constraint fields in embodied intelligent operation and maintenance robots, characterized in that, Includes the following steps: Step 1: Construct a multi-level safety constraint field for operation and maintenance scenarios: Collect successful operation trajectories of embodied operation and maintenance robots in operation and maintenance tasks, extract state-action pairs, train the constraint field network, take robot state and operation and maintenance actions as input, and output non-negative scalar constraint field values. Through multi-level annotation, the safe operation area, equipment sensitive approach area, personnel activity interference area, equipment protection prohibition area and spatial restricted area are continuously and uniformly encoded. Step 2, Construct a multimodal state observation for the operation and maintenance scenario: Integrate the robot end effector pose, joint state, local features of the target device's 3D point cloud, coordinates of the sensor installation target point, and the safety topology distance information of the current working area to form an operation and maintenance context observation vector. , serving as the unified condition input for the constraint field network and the diffusion strategy; Step 3, Constraint Field Gradient Embedded Denoising: In each step of the denoising calculation of the diffusion strategy, a constraint field gradient correction is applied after the standard denoising output, causing the denoising trajectory to shift along the direction of decreasing constraint field value. The correction formula is as follows: in The first after constraint field correction Noise reduction steps For the standard diffusion denoiser in the first The noise reduction action of the step output, Number the current denoising step. For adaptive guidance strength, To constrain the gradient of field operations and maintenance actions and direct it towards higher risk directions, subtracting this gradient automatically biases the denoising results towards safer operating areas. Step 4, Adaptive Guidance Strength Scheduling: A learnable guidance strength scheduling function is adopted. In the early stage of denoising, namely the macro-path planning stage, the maximum guidance strength value is output to strongly constrain the overall movement direction of the maintenance robot away from the danger zone. In the later stage of denoising, namely the fine control stage, the minimum guidance strength value is output to preserve the generation quality of the fine control action of the diffusion strategy. The minimum guidance strength value is shrunk to less than 10% of the maximum guidance strength value. Step 5, Output Safety Operation and Maintenance Action Sequence: After the complete denoising process guided by the constraint field, the operation and maintenance action sequence that meets the multi-level safety constraints is output. The robot then performs the operation and maintenance tasks of sensor mounting, equipment status acquisition, and inspection path following in sequence.

2. The diffusion strategy method according to claim 1, characterized in that, The specific method for constructing training data for the multi-level security constraint field in step 1 is as follows: Safety operation level sample label is State-action pairs are extracted directly from the demonstration trajectory of the embodied maintenance robot successfully completing maintenance tasks; Device proximity level sample label Add a standard deviation to the safe operating level actions. Gaussian perturbation; Personnel interference risk level sample label Add a standard deviation to the safe operating level actions. Gaussian perturbation; Equipment protection prohibition level sample label is Replace the actions of the safety operation level samples with actions of uniformly distributed random sampling; Space restricted area sample label is Extract the state-action pair when the end effector moves into a predefined prohibited operation area; in To constrain the field label value, indicating the safety risk level of the state-action pair, The standard deviation of the equipment's sensitive proximity level disturbance. The standard deviation of the disturbance level caused by human interference.

3. The diffusion strategy method according to claim 1, characterized in that, Step 1, the constraint field network, specifically includes: an input layer, which takes the operation and maintenance context state vector as input. Operation and maintenance action vectors The input is concatenated as a joint input; multiple fully connected hidden layers are used, each followed by a normalization layer and a SiLU activation function; the output layer uses the Softplus activation function to ensure the constraint field values. The nonnegativity of the constraint field network is observed; the constraint field network is trained under supervision using the Huber loss function. in The value of Huber's loss function. This is the Huber loss threshold. For the constrained field network prediction value, The value of the constraint field label is used for annotation.

4. The diffusion strategy method according to claim 1, characterized in that, Step 2: Multimodal state observation vectors in the operation and maintenance scenario The specific components are as follows: robot body state sub-vector, including the three-dimensional position of the end effector, quaternion pose, joint angle vector and mounting tool state; target device state sub-vector, including the three-dimensional coordinates of the target sensor mounting point, the device surface normal vector estimation and the device current operating status code; safety topology distance sub-vector, including the shortest distance estimation from the end effector to each predefined safety boundary; the sub-vectors are concatenated to form a unified operation and maintenance context observation vector.

5. The diffusion strategy method according to claim 1, characterized in that, In step 3, the constraint field gradient is calculated using an automatic differentiation mechanism: the operation and maintenance action tensor is... Set the state to require gradient calculation; forward calculation of constraint field values. ;right about Perform backpropagation to obtain gradients Gradients are clipped using the L2 norm before the correction is applied. in The cropping threshold, It is the L2 norm of the gradient vector to prevent excessive motion correction caused by sudden changes in constraint field values, which could lead to joint impact on the robot.

6. The diffusion strategy method according to claim 1, characterized in that, In step 4, the adaptive guidance strength scheduling function is implemented by the guidance strength scheduling network, which uses a normalized denoising step. As input, the shape factor is output through two fully connected layers and a sigmoid activation function, and then compared with the learnable maximum guided strength parameter. Multiplying them together yields the final guidance strength value: in For the first The guiding strength value for step-by-step noise reduction. The maximum learnable guidance strength parameter. For parameterized scheduling networks, For the trainable parameters of the scheduling network, It is the Sigmoid activation function. Number the current denoising step. This represents the total number of noise reduction steps; during the fine mounting stage. At that time, it automatically shrinks to within 10% of the maximum value; during the macro-path planning stage. At the same time, maintain maximum guidance intensity.

7. The diffusion strategy method according to claim 1, characterized in that, Also includes Candidate security screening inference mode: Observe and execute the current state of the maintenance robot. Sub-independent diffusion strategy sampling, to obtain Group maintenance action candidate sequences; use constraint field networks to analyze each group of candidates. Step motion calculation of average constraint field value: in For the first The average constraint field score of the candidate action sequence. For candidate sequence index, The total number of candidate sequences sampled independently. The number of action steps contained in a single candidate sequence. For time step indexes within the action sequence, For the first The operational context state of the step. For the first The first group of candidate sequences Step movement, The constraint field label value is used; the candidate sequence with the lowest constraint field value is selected as the final action to be executed. The candidate security screening mode does not require modification of the basic diffusion strategy model structure and can be used as a training-free security enhancement plugin for already deployed and maintained diffusion strategies.

8. A diffusion strategy system for guiding multi-level safety constraint fields of embodied intelligent operation and maintenance robots, implementing the diffusion strategy method according to any one of claims 1-7, characterized in that, include: The maintenance robot body module includes a robotic arm, an end effector, and a sensor mounting and actuator. The multimodal perception module includes a joint encoder, a vision sensor, a torque sensor, and a 3D point cloud acquisition unit; The security topology map module pre-stores 3D security partitioning information for the operation and maintenance area; The constraint field network module is used to learn and evaluate multi-level safety constraint field values ​​for state-action pairs in real time. The basic module for diffusion strategies includes a pre-trained neural network for diffusion strategies of operation and maintenance tasks and a noise scheduler; The guidance intensity scheduling module outputs adaptive guidance intensity based on the current denoising step and maintenance task stage. The constraint field-guided denoising module uses the constraint field gradient and adaptive guidance strength to safely correct the operation and maintenance action trajectory after each denoising calculation. The operation and maintenance task scheduling module is responsible for parsing operation and maintenance task instructions, planning inspection paths, and scheduling sensor mounting sequences. The modules collaborate through a unified operation and maintenance control interface.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the diffusion strategy method for guiding multi-level safety constraint fields for embodied intelligent operation and maintenance robots as described in any one of claims 1-7.

Images