Robot exploration method and system based on diffusion-enhanced sequential topology planning

Abstract

The application discloses a robot detection method and system based on diffusion-enhanced serialized topological planning, and the method comprises the following steps: acquiring real-time point cloud observation data of an environment, performing incremental topological abstraction processing and hybrid graph coding reasoning, and generating a space embedding vector; performing history-enhanced serialized autoregressive path generation based on the space embedding vector to obtain serialized topological path data; sequentially performing reverse diffusion denoising, uncertainty quantization sampling and reliability gating mapping on the serialized topological path data to obtain trust score data; and constructing a composite objective function based on the trust score data, performing hybrid-driven optimization on the serialized topological path data, and realizing robot detection path planning. The application can capture the value distribution of a complex environment and meet the detection needs of robots in unknown dynamic environments. The application can be widely applied to the technical field of robot detection trajectory planning as a robot detection method and system based on diffusion-enhanced serialized topological planning.

Description

Technical Field

[0001] This invention relates to the field of robot detection trajectory planning technology, and in particular to a robot detection method and system based on diffusion-enhanced sequential topology planning. Background Technology

[0002] With the rapid development of mobile robot technology, the autonomous exploration capability of robots in complex, unknown, and unstructured 3D environments has become fundamental to key applications such as search and rescue, and industrial inspection. In related technologies, traditional autonomous exploration schemes mainly rely on classical planners, which are typically based on occupancy nets or frontier detection mechanisms. While such methods demonstrate some effectiveness in structured or constrained scenarios, when faced with large-scale unstructured environments, due to a lack of long-term predictability, these heuristic-based methods often struggle to balance detection coverage efficiency and computational scalability, easily getting trapped in local optima and exhibiting frequent backtracking movements due to decision redundancy.

[0003] In recent years, deep reinforcement learning (RL), as a data-driven paradigm, has been attempted to learn complex detection strategies from experience. However, existing online reinforcement learning methods face severe challenges in 3D environment detection: on the one hand, high-fidelity sensor simulation and large-scale 3D scenes result in extremely high training data acquisition costs, low sample efficiency, and highly unstable training processes. On the other hand, most existing RL methods simplify the detection task to a single-step viewpoint selection, ignoring the contextual relationships of historical decisions, leading to inconsistent detection paths and a large number of duplicate detections.

[0004] To improve data efficiency and trajectory quality, recent research has attempted to introduce diffusion models to learn from expert demonstrations and generate multi-step trajectories through action division. However, such methods typically treat discrete viewpoint selection as continuous motion generation, increasing task complexity. Furthermore, the iterative denoising process of diffusion models introduces significant computational latency, and the generated trajectories often lack safety guarantees. More importantly, the performance ceiling of these imitation-based methods is strictly limited to the referenced heuristic expert strategy, making it difficult to surpass it. In addition, regarding map representation, existing explicit geometric maps (such as occupancy grids or octree maps) suffer from excessive memory consumption or loss of geometric details in large-scale scenes. While topological representations offer good scalability, related techniques typically rely on expensive intermediate voxelization processes to extract graph structure, resulting in low computational efficiency.

[0005] In summary, when dealing with complex 3D environment detection tasks, the relevant technologies generally suffer from drawbacks such as high computational complexity, lack of temporal consistency in decision-making, strong dependence on online interaction, and performance limitations imposed by expert experience. Summary of the Invention

[0006] To address the aforementioned technical problems, the present invention aims to provide a robot detection method and system based on diffusion-enhanced sequential topology planning, which can capture the value distribution in complex environments and meet the real-time detection needs of robots in unknown dynamic environments.

[0007] The first technical solution adopted in this invention is: a robot detection method based on diffusion-enhanced sequential topology planning, comprising the following steps:

[0008] Acquire real-time point cloud observation data of the environment, perform incremental topological abstraction and hybrid graph encoding reasoning, and generate spatial embedding vectors;

[0009] Based on the spatial embedding vector, a history-enhanced serialized autoregressive path is generated to obtain serialized topological path data;

[0010] The serialized topology path data is sequentially subjected to inverse diffusion denoising, uncertainty quantization sampling, and reliability gating mapping to obtain trust score data;

[0011] Based on the trust score data, a composite objective function is constructed to perform hybrid-driven optimization on the serialized topological path data, thereby realizing robot exploration path planning.

[0012] Furthermore, the step of acquiring real-time point cloud observation data of the environment, performing incremental topological abstraction processing and hybrid graph encoding inference, and generating spatial embedding vectors specifically includes:

[0013] Real-time point cloud observation data of the environment is acquired, and the skeleton line of the geometric center of the environment is extracted through an incremental topology mapping mechanism. Nodes are defined, and a target node dataset is constructed. The target node dataset includes the robot's current position, the frontal viewpoint of the detection, and the connection points of the environmental skeleton.

[0014] The feature matrix is ​​obtained by performing feature matrix encoding calculation on the target node dataset;

[0015] Obtain the adjacency matrix of the target node dataset, which is used to reflect the physical connectivity between target nodes;

[0016] Hybrid graph encoding inference is performed on the feature matrix and the adjacency matrix to generate a spatial embedding vector.

[0017] Furthermore, the step of performing feature matrix encoding calculation on the target node dataset to obtain the feature matrix specifically includes:

[0018] The absolute positions of the target node dataset are centered and normalized to obtain relative 3D coordinates;

[0019] The unknown voxel volume of the target node dataset is quantized using a visibility algorithm to obtain an information gain scalar.

[0020] Based on the type of the target node dataset, binary mapping is performed to generate one-hot encodings of semantic types;

[0021] The relative 3D coordinates, the information gain scalar, and the one-hot encoding of the semantic type are concatenated to obtain the feature matrix.

[0022] Furthermore, the step of generating a serialized autoregressive path based on the spatial embedding vector to obtain serialized topological path data specifically includes:

[0023] Based on the multi-head attention mechanism, the robot's current position is used as the query vector, and the spatial embedding vector is used as the key vector and value vector. Attention weights are calculated and features are weighted and aggregated to obtain global geometric background data.

[0024] The trajectory prefix is ​​constructed by explicitly recording the planned node index sequence and inputting it into the converter encoder for temporal feature extraction and pooling processing to obtain the current trajectory prefix feature data.

[0025] The global geometric background data and the current trajectory prefix feature data row vector are concatenated and processed, and then linearly transformed through a mapping layer to obtain belief state data that characterizes the current decision moment.

[0026] Based on the belief state data, a masked policy distribution calculation is performed to obtain the selection probability distribution data of each candidate node;

[0027] Based on the selection probability distribution data of each candidate node, node sampling is performed, and the selected node is updated in real time to the trajectory prefix as the input for the next decision. This process is repeated until a path chain of a preset length is generated, resulting in serialized topological path data.

[0028] Furthermore, the step of calculating the masked policy distribution based on the belief state data to obtain the selection probability distribution data of each candidate node specifically includes:

[0029] Construct a dynamic validity mask based on the adjacency matrix;

[0030] The belief state data and the robot's current position are vector-concatenated and linearly projected to generate a query vector;

[0031] Calculate the matrix product of the query vector and the spatial embedding vector, and then divide by the square root of the feature dimension for scaling to obtain the spatial node feature attention retrieval calculation result.

[0032] The spatial node feature attention retrieval calculation result is added to the dynamic validity mask, and then normalized using the Softmax function to obtain the selection probability distribution data of each candidate node.

[0033] Furthermore, the specific calculation expression for the masked policy distribution is as follows:

[0034]

[0035] In the above formula, Data representing belief states, Indicates the robot's current position. Represents the query vector. Represents a spatial embedding vector. Indicates a dynamic validity mask. Represents the square root of the feature dimension. This represents the Softmax function. Indicates the first The selected candidate nodes in the step, This represents the probability distribution of a candidate node being selected, given the belief state and the robot's current position.

[0036] Furthermore, the step of sequentially performing reverse diffusion denoising, uncertainty quantization sampling, and reliability gating mapping on the serialized topology path data to obtain trust score data specifically includes:

[0037] Based on a pre-trained denoising network, the robot's current environmental observation state is obtained. With random noise as the initial state and the serialized topological path data as input, multi-step iterative reverse diffusion denoising prediction is performed to obtain action value prediction data that fits the multimodal value distribution of the action.

[0038] Uncertainty quantification sampling processing is performed on the action value prediction data, and its standard deviation is calculated to obtain value deviation assessment data that reflects the uncertainty of current value perception.

[0039] Based on the value deviation assessment data, combined with a preset penalty coefficient, the trust score data is obtained by mapping calculation through a negative exponential transformation function.

[0040] Furthermore, the specific expression for the multi-step iterative reverse diffusion denoising prediction is as follows:

[0041]

[0042] In the above formula, This represents the learned denoising network. Indicates the reverse diffusion denoising method. Step motion value prediction data, This represents the preset noise scheduling parameters. Indicates the first Step motion value prediction data, This represents the cumulative product term of the noise scheduling parameters. This indicates the robot's current environmental observation status. This represents serialized topology path data (i.e., actions). This indicates the current number of back diffusion steps. This represents the posterior variance of the reverse process. This represents standard Gaussian noise.

[0043] Furthermore, the step of constructing a composite objective function based on the trust score data and performing hybrid-driven optimization on the serialized topological path data to achieve robot exploration path planning specifically includes:

[0044] The trust score data is calculated by weighting the trust value prediction data with the action value prediction data to obtain the trust weighted target value data.

[0045] Construct diffusion reconstruction loss data and value regression loss data, update the weight parameters of the pre-trained denoising network to obtain an optimized denoising network. The diffusion reconstruction loss data represents the mean square error between the noise predicted by the pre-trained denoising network and the actual injected noise, and the value regression loss data represents the mean square error between the action value prediction data and the trust-weighted target value data.

[0046] A sequence-level trust reinforcement loss data is constructed, and the sequenced topology path data is backpropagated using the optimized denoising network to achieve robot exploration path planning.

[0047] The second technical solution adopted in this invention is: a robot detection system based on diffusion-enhanced sequential topology planning, comprising:

[0048] The first module is used to acquire real-time point cloud observation data of the environment, perform incremental topological abstraction processing and hybrid graph encoding reasoning, and generate spatial embedding vectors.

[0049] The second module is used to generate a historically enhanced serialized autoregressive path based on the spatial embedding vector, thereby obtaining serialized topological path data;

[0050] The third module is used to sequentially perform reverse diffusion denoising, uncertainty quantization sampling, and reliability gating mapping on the serialized topology path data to obtain trust score data;

[0051] The fourth module is used to construct a composite objective function based on the trust score data, perform hybrid-driven optimization on the serialized topological path data, and realize robot detection path planning.

[0052] The beneficial effects of the method and system of this invention are as follows: This invention acquires real-time point cloud observation data of the environment, performs incremental topological abstraction processing and hybrid graph encoding inference to generate spatial embedding vectors; further, based on the spatial embedding vectors, it performs historical enhancement-enhanced serialized autoregressive path generation to obtain serialized topological path data, and uses an autoregressive approach to predict long-range node sequences. This serialization generation method fully considers the historical decision context, eliminating frequent robot turns and path jitter. The generated trajectory maintains high logical consistency in the time dimension, which not only improves motion efficiency but also reduces the mechanical wear of the actuator; furthermore, the serialized topological path data is sequentially subjected to inverse diffusion denoising, uncertainty quantization sampling, and reliability gating mapping to obtain trust score data. An evaluator based on a diffusion probability model is introduced, which can accurately model the multimodal value distribution of detection actions and output a "trust score" representing the reliability of the decision. This mechanism effectively solves the problem of inaccurate value estimation in out-of-distribution areas or high-risk areas. The system uses "trust score" as a reliability gate to effectively filter out erroneous predictions with high uncertainty, significantly enhancing the robustness of the detection strategy in complex and unpredictable scenarios. Finally, based on the trust score data, a composite objective function is constructed to perform hybrid-driven optimization on the serialized topological path data, realizing robot detection path planning and meeting the robot's real-time detection needs in unknown dynamic environments. Attached Figure Description

[0053] Figure 1 This is a flowchart of the steps of the robot detection method based on diffusion-enhanced serialized topology planning of the present invention;

[0054] Figure 2 This is a structural block diagram of the robot detection system based on diffusion-enhanced serialized topology planning of the present invention;

[0055] Figure 3 This is a schematic diagram of diffusion-enhanced serialization topology planning provided in a specific embodiment of the present invention;

[0056] Figure 4 This is a schematic diagram of the 3D simulation environment provided in a specific embodiment of the present invention;

[0057] Figure 5 This is a schematic diagram of a complex 3D geometric scene provided in a specific embodiment of the present invention. Detailed Implementation

[0058] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.

[0059] First, the technical terms used in the embodiments of this invention will be explained:

[0060] 1) 3D Skeleton Graph: A high-level topological abstraction representation method for large-scale unstructured spaces. By extracting the skeleton lines at the geometric center of the environment, complex raw point clouds or voxel maps are transformed into lightweight graph structures composed of discrete nodes (representing key spatial locations) and edges (representing connectivity between locations). This representation can exponentially reduce the computational dimensionality during robot planning while preserving the basic topological features of the environment and obstacle constraints, thus supporting real-time computation in large-scale scenes.

[0061] 2) Diffusion Evaluator: An evaluation architecture that utilizes the inverse denoising process of a generative diffusion model to fit the action value distribution. Instead of directly outputting a single score, this evaluator accurately captures the potential multimodal value function in complex exploration scenarios by simulating the regression process from Gaussian noise to the target action distribution. This mechanism allows the evaluator to objectively assess the exploration value of each potential path when faced with multiple equally attractive unknown regions, thereby improving the accuracy and robustness of value estimation.

[0062] 3) Serialized Actor: A decision-making model with temporal foresight. Based on the current environmental state characteristics, this model uses a neural network to simultaneously predict a set of future positions or control parameter sequences arranged in chronological order, rather than just a single next action. This serialized generation method can directly integrate long-range planning logic into a single decision, ensuring that the robot's generated motion trajectory is spatially continuous and logically consistent, thereby eliminating decision jitter and improving the smoothness of motion.

[0063] 4) Advantage-Weighted Offline Reinforcement Learning: A learning paradigm that trains policies using pre-stored static trajectory datasets without real-time sampling and interaction with the physical environment. Its core lies in introducing an advantage function to measure the superiority of a specific action relative to the average level in the dataset, and using this advantage as weights to weight the loss function, causing the model to prioritize simulating those superior exploration behaviors. This method solves the problem of unstable training in online reinforcement learning and can effectively extract complex exploration strategies from historical data.

[0064] 5) Multimodal value distribution: Multiple non-overlapping high-probability (high-value) peak states exist in the robot's exploration decision space. In complex exploration tasks, when a robot is at the intersection of multiple unknown areas (such as a crossroads), different directions of motion often correspond to their own independent exploration potential. This distribution describes the value characteristic of this "multi-pronged approach" in the action space, reflecting the uncertainty of the decision-making environment and the objective physical property of multiple potential optimal solutions coexisting.

[0065] 6) Topology Planner: An autonomous navigation and exploration decision-making framework operating on a topological graph spatial scale. It decouples and integrates high-level global topological reasoning (determining "which area to go to") with low-level local geometric planning (determining "how to get there"). By searching for path schemes that maximize spatial coverage in an abstract network of nodes, this planner enables robots to avoid interference from local obstacles and achieve long-term task coordination in complex environments with a wide range and multiple levels.

[0066] The following problems still exist in the relevant technologies:

[0067] 1) Lack of global reasoning ability, prone to getting trapped in local optima: Traditional heuristic planners based on frontier-detection or occupancy grids mainly rely on local geometric information for decision-making. When faced with large and complex scenes, these methods lack long-horizontal predictive capabilities, making it difficult for robots to balance detection efficiency and computational scalability. They are prone to frequent backtracking in dead ends or complex connected regions, getting trapped in local optima.

[0068] 2) Low sample efficiency and unstable training in online reinforcement learning: Existing online reinforcement learning (Online RL)-based probing methods require massive real-time interactions with the environment, which leads to extremely high computational resource overhead and long training times in 3D high-fidelity simulation environments. Furthermore, since the reward signal in probing tasks is typically extremely sparse, online RL often faces problems such as training non-convergence or drastic performance fluctuations.

[0069] 3) Inaccurate value assessment in multimodal decision-making scenarios: In actual exploration, robots often face multiple exploration directions with potential value (such as multi-branch corridors). Related technologies mostly use unimodal policy networks, which are difficult to accurately model the value distribution in such multimodal scenarios, causing decisions to oscillate between multiple optimal candidate points and reducing the robustness of the exploration.

[0070] 4) Lack of temporal continuity in decision-making, resulting in uneven motion trajectories: Most existing solutions simplify the exploration task to single-step viewpoint selection. This discrete decision-making mechanism ignores the temporal correlation between historical decisions and future plans, leading to severe decision jitter and frequent turning during the execution of the robot-generated path, increasing mechanical wear and reducing motion efficiency.

[0071] 5) Performance is limited by the level of expert demonstrations: Some methods based on imitation learning or diffusion strategies have their performance strictly limited to the level of expert datasets. These methods can usually only passively imitate trajectories generated by heuristic algorithms and cannot surpass expert strategies through data-driven approaches.

[0072] Based on this, the present invention proposes a robot detection method based on diffusion-enhanced sequential topology planning, which achieves efficient and smooth detection in large-scale 3D environments by abstracting the environment into a topology graph and introducing a diffusion probability model.

[0073] Reference Figure 1 This invention provides a robot detection method based on diffusion-enhanced sequential topology planning, which includes the following steps:

[0074] S100: Acquire real-time point cloud observation data of the environment, perform incremental topological abstraction processing and hybrid graph encoding reasoning, and generate spatial embedding vectors;

[0075] It should be noted that in some embodiments, step S100 may include steps S110 to S140.

[0076] S110. Acquire real-time point cloud observation data of the environment, and extract the skeleton line of the geometric center of the environment through an incremental topology mapping mechanism, define the nodes, and construct the target node dataset. The target node dataset includes the robot's current position, the front-end viewpoint of the detection, and the connection points of the environmental skeleton.

[0077] In this embodiment, the system first uses airborne sensors to acquire real-time point cloud observations of the environment. The system extracts the skeleton lines of the geometric center of the environment through an incremental topology mapping mechanism, thereby defining nodes. By classifying and sampling the free space, obstacle boundaries, and skeleton structure in the environment, it identifies and extracts the robot's current position, the detection front viewpoint, and the connection points of the environmental skeleton, thus constructing a node set representing key spatial distribution features. Node set These represent the robot's current position, the frontal viewpoint of the detection, and the connection point of the environmental skeleton, respectively.

[0078] In some specific embodiments, an incremental topology mapping mechanism is used to map real-time observations. Abstracted as a dynamic topology graph .

[0079] S120. Perform feature matrix encoding calculation on the target node dataset to obtain the feature matrix;

[0080] Specifically, the absolute position of the target node dataset is centered and normalized to obtain relative 3D coordinates; the unknown voxel volume of the target node dataset is quantized using a visibility algorithm to obtain an information gain scalar; based on the type of the target node dataset, binary mapping is performed to generate a one-hot encoding of the semantic type; the relative 3D coordinates, information gain scalar, and one-hot encoding of the semantic type are concatenated to obtain a feature matrix.

[0081] In this embodiment, attribute extraction and vectorization are performed on each node in the node set, followed by feature matrix encoding: First, the absolute position of the node is centered and normalized to obtain relative 3D coordinates; second, the volume of the unknown voxel from the node's perspective is quantized using a visibility algorithm to obtain an information gain scalar; finally, binary mapping is performed according to the node type to generate one-hot encoding of the semantic type; by concatenating the above parameters, a feature matrix that can comprehensively characterize the environmental structure and task attributes is finally constructed. Each node feature It includes one-hot encoding of relative 3D coordinates, information gain scalars, and semantic types.

[0082] S130. Obtain the adjacency matrix of the target node dataset. The adjacency matrix is ​​used to reflect the physical connectivity between target nodes.

[0083] S140. Perform hybrid graph encoding reasoning on the feature matrix and the adjacency matrix to generate a spatial embedding vector.

[0084] In this embodiment, the feature matrix is ​​combined with the adjacency matrix reflecting the physical connectivity between nodes. The input is taken together and subjected to hybrid graph encoding inference: using a shared encoder, the long-range topological dependencies of the environment are captured sequentially across geographical distances through a linear global attention mechanism, and then integrated with the neighborhood geometric structure features by combining local graph convolution operations, thereby enhancing the perception of environmental structure and finally generating spatial embedding vectors. Its computation process combines linear global attention with local graph convolution:

[0085]

[0086] This embedding vector provides a unified global spatial context for subsequent actor and evaluator modules.

[0087] In summary, in some specific embodiments, environmental point cloud data is first acquired in real time through sensors, and the skeleton line of the geometric center of the environment is extracted using an incremental topology mapping mechanism. Nodes are then defined based on this. By identifying the robot's current position, the detection front viewpoint, and the connection points of the environmental skeleton, a raw node dataset representing key spatial locations is constructed. Next, attributes are extracted from each node in this dataset. By calculating the relative 3D coordinates, detection information gain, and one-hot encoding of semantic types for each node, these attributes are vectorized and concatenated to construct a feature matrix that characterizes the topological attributes of the environment. Finally, this feature matrix, along with an adjacency matrix reflecting node connectivity, is input into a shared encoder. Long-range dependencies are captured sequentially through a linear global attention mechanism, and neighborhood geometric features are integrated using local graph convolution operations. Finally, a unified spatial embedding vector is generated through feature fusion and pooling, providing a global spatial context for subsequent decision-making and evaluation modules.

[0088] S200. Based on spatial embedding vectors, a history-enhanced serialized autoregressive path is generated to obtain serialized topological path data;

[0089] It should be noted that in some embodiments, step S200 may include steps S210 to S250.

[0090] S210. Based on the multi-head attention mechanism, the robot's current position is used as the query vector, and the spatial embedding vector is used as the key vector and value vector. Attention weights are calculated and features are weighted and aggregated to obtain global geometric background data.

[0091] In this embodiment, during spatiotemporal feature fusion, a multi-head attention mechanism (MHA) is used, with the node embedding vector representing the robot's current position serving as the query vector. The spatial embedding vector output by the shared encoder is used as the key vector. ) and value vector ( Attention weights are calculated and features are aggregated using weighted aggregation. This process extracts global geometric background data reflecting the geometric relationship between the current location and its surrounding environment from the global space. :

[0092]

[0093] In the above formula, Represents global geometric background data. This indicates a multi-head attention mechanism. Represents the query vector. Represents the key vector. Represents a value vector.

[0094] S220. The planned node index sequence is explicitly recorded to form the trajectory prefix, and then input into the converter encoder for temporal feature extraction and pooling to obtain the current trajectory prefix feature data.

[0095] In this embodiment, a trajectory prefix is ​​constructed by explicitly recording the planned node index sequence, and its corresponding feature vector is input into the Transformer encoder for temporal feature extraction and pooling processing to obtain the current trajectory prefix feature data. (i.e., the historical decision-making context).

[0096] S230. The global geometric background data and the current trajectory prefix feature data row vector are concatenated and processed, and then linearly transformed through the mapping layer to obtain the belief state data used to characterize the current decision moment.

[0097] In this embodiment, the global geometric background data obtained through the above processing is... With current trajectory prefix feature data Perform vector concatenation and then pass the concatenated vector through a mapping layer. Perform a linear transformation to obtain the representation of the current i-th... Belief state data at each step of the decision-making process Its expression is:

[0098]

[0099] In the above formula, Data representing belief states, Represents the mapping layer. Represents global geometric background data. This represents the prefix feature data of the current trajectory.

[0100] S240. Calculate the masked policy distribution based on the belief state data to obtain the selection probability distribution data of each candidate node.

[0101] Specifically, a dynamic validity mask is constructed based on the adjacency matrix; the belief state data and the robot's current position are concatenated and linearly projected to generate a query vector; the matrix product of the query vector and the spatial embedding vector is calculated and scaled by dividing by the square root of the feature dimension to obtain the spatial node feature attention retrieval calculation result; the spatial node feature attention retrieval calculation result is added to the dynamic validity mask and normalized using the Softmax function to obtain the selection probability distribution data of each candidate node.

[0102] In this embodiment, to ensure the physical connectivity of the trajectory, a dynamic validity mask is first constructed based on the topological adjacency matrix. The position weights of non-adjacent or visited nodes are reset to negative infinity for masking filtering; simultaneously, the current belief state data is... After being embedded and stitched together with the robot's current position, the vector is processed through a linear projection layer to generate a query vector. Subsequently, the query vector is calculated. With spatial embedding vector The matrix product, divided by the square root of the feature dimension. Scaling is performed, and the calculation results are compared with a dynamic validity mask. The results are summed and then normalized using the Softmax function. The candidate node selection probability distribution data output by the policy head is defined as follows:

[0103]

[0104] In the above formula, Data representing belief states, Indicates the robot's current position. Represents the query vector. Represents a spatial embedding vector. Indicates a dynamic validity mask. Represents the square root of the feature dimension. This represents the Softmax function. Indicates the first The selected candidate nodes in the step, This represents the probability distribution of a candidate node being selected, given the belief state and the robot's current position.

[0105] S250. Based on the selection probability distribution data of each candidate node, node sampling is performed and the selected node is updated in real time to the trajectory prefix as the input for the next decision. The process is iterated until a path chain of a preset length is generated, and serialized topological path data is obtained.

[0106] In this embodiment, the next node is sampled based on the probability distribution data and updated in the historical trajectory prefix in a loop. This mechanism ensures that the generated action sequence forms a continuous and passable path chain.

[0107] In summary, in some specific embodiments, firstly, a multi-head attention mechanism is used to weight the spatial embedding vector output by the shared encoder. Using the robot's current node embedding as the query vector, global geometric background data reflecting the geometric relationship between the current position and the environment is extracted from global spatial features. Next, the selected historical node index sequence is explicitly extracted to form a trajectory prefix. Feature extraction is performed through the transformer encoder and its pooling operation to obtain compact sequence embedding data. Subsequently, the obtained global geometric background data and sequence embedding data are vector-concatenated and linearly transformed to synthesize the comprehensive belief state data at the current decision moment. Based on this, a dynamic validity mask is introduced to filter non-adjacent or visited nodes, and attention retrieval of spatial node features is performed using the query vector generated based on the comprehensive belief state to calculate the selection probability distribution data of each candidate node. Finally, node sampling is performed according to this probability distribution, and the selected nodes are updated in real time to the trajectory prefix as input for the next decision. Through iterative processing, a complete path chain of a preset length is generated, thereby obtaining serialized topological path data with long-range logical consistency.

[0108] S300. The serialized topology path data is sequentially subjected to reverse diffusion denoising, uncertainty quantization sampling, and reliability gating mapping to obtain trust score data.

[0109] It should be noted that in some embodiments, step S300 may include steps S310 to S330.

[0110] S310. Based on a pre-trained denoising network, the current environmental observation state of the robot is obtained. Random noise is used as the initial state, and serialized topological path data is used as input. Multi-step iterative reverse diffusion denoising prediction is performed to obtain action value prediction data that fits the multimodal value distribution of the action.

[0111] In this embodiment, reverse diffusion denoising is performed. The evaluator module takes random noise as the initial state and combines the robot's current environmental observation state and the action sequence generated by the actor as the conditional input. It performs multi-step iterative denoising prediction processing through a pre-trained denoising network to accurately capture the highly nonlinear and multi-peak distribution characteristics of the value function in the detection scene, thereby obtaining action value prediction data that fits the multimodal value distribution of the action.

[0112] In this process, diffusion inverse process modeling is used, and the evaluator predicts action value through iterative noise reduction. Its single-step update formula is:

[0113]

[0114] in It is a learned denoising network used to capture the highly nonlinear and multimodal distribution characteristics of the value function.

[0115] S320. Perform uncertainty quantification sampling on the action value prediction data and calculate its standard deviation to obtain value deviation assessment data that reflects the uncertainty of current value perception.

[0116] In this embodiment, uncertainty quantification sampling processing is performed: for the same input state and action, the system introduces independent random noise and performs multiple parallel Monte Carlo sampling processes through the above-mentioned diffusion inverse process. The value estimation results of multiple samplings are statistically analyzed and their standard deviation is calculated, thereby obtaining value deviation assessment data that reflects the uncertainty of current value perception.

[0117] S330. Based on the value deviation assessment data and combined with the preset penalty coefficient, the trust score data is obtained by mapping calculation through the negative exponential transformation function.

[0118] In this embodiment, a reliability-gated mapping process is performed: the obtained value deviation assessment data is combined with a preset penalty coefficient, and a mapping calculation is performed using a negative exponential transformation function to finally derive trust score data for measuring decision certainty. This trust score data is then used as an evaluation index to effectively identify and penalize the robot's strategy output when it is in an unknown or high-risk area, ensuring the robustness of its exploration behavior.

[0119] Among them, the reliability gating (Trust Score) is achieved through Monte Carlo sampling. The standard deviation of the value estimate obtained the next time This leads to the trust score. :

[0120]

[0121] This score is used to measure the certainty of decision-making, effectively identifying and penalizing strategy outputs in unknown or high-risk areas.

[0122] S400: Construct a composite objective function based on trust score data, perform hybrid-driven optimization on serialized topological path data, and realize robot exploration path planning.

[0123] It should be noted that in some embodiments, step S400 may include steps S410 to S430.

[0124] S410. Calculate the weighted average of the trust score data using the action value prediction data to obtain the trust weighted target value data.

[0125] In this embodiment, a trust-weighted target value construction process is performed. The system extracts the instant reward data and the state data of the next moment from the offline dataset, combines the conservative value prediction of each candidate action in the next moment output by the evaluator, and uses the trust score data obtained in the previous step to calculate the weighted average of these candidate values. This filters out unreliable estimates with high uncertainty in the calculation, and finally generates trust-weighted target value data to guide the stable updating of the network.

[0126] Among them, for trust-weighted value propagation, in calculating the target value At that time, a weighted average is used based on trust scores to filter out unreliable estimates:

[0127]

[0128] In the above formula, This represents the immediate reward value. Indicates the discount factor. Indicates the first Trust score for each candidate action. Indicates the state at the next moment. Indicates the sampling at the next moment. One candidate action.

[0129] S420. Construct diffusion reconstruction loss data and value regression loss data, update the weight parameters of the pre-trained denoising network to obtain the optimized denoising network. The diffusion reconstruction loss data represents the mean square error between the noise predicted by the pre-trained denoising network and the actual injected noise, and the value regression loss data represents the mean square error between the action value prediction data and the trust-weighted target value data.

[0130] In this embodiment, the system performs evaluation and diffusion network loss calculations. On the one hand, it extracts diffusion reconstruction loss data by calculating the mean square error between the noise predicted by the denoising network and the actual injected noise. On the other hand, it extracts value regression loss data by calculating the mean square error between the conservative predicted value output by the dual-Q network in the current state and the aforementioned trust-weighted target value data. The system combines these two loss data for gradient backpropagation processing to update the network weights, thereby obtaining the optimized evaluation network parameter data.

[0131] S430. Construct sequence-level trust reinforcement loss data, and combine it with the optimized denoising network to perform backpropagation processing on the serialized topology path data to realize robot exploration path planning.

[0132] In this embodiment, a sequence-level composite strategy optimization process is performed. The system extracts expert trajectory data and calculates behavioral cloning loss data through a behavioral cloning mechanism to absorb the structured priors of experts. At the same time, it extracts the long-range path sequences generated by the actors themselves, and uses the corresponding trust score data to perform a weighted summation of the prediction values ​​of each node in the self-generated sequence to construct sequence-level trust reinforcement loss data. The behavioral cloning loss data and the sequence-level trust reinforcement loss data are combined for backpropagation processing, which forces the system to maximize the reliable expected returns along the entire generated long-range path, thereby avoiding local short-sighted decision-making and finally obtaining convergent and highly robust sequenced detection strategy model data.

[0133] For sequence-level policy optimization, the actor surpasses the expert policy by optimizing the following objective function:

[0134]

[0135] This optimization mechanism forces the system to maximize the reliable expected reward along the entire generated long-distance path, thereby ensuring the efficiency and robustness of the detection behavior.

[0136] Furthermore, this embodiment will be described in conjunction with the accompanying drawings:

[0137] The model training in this embodiment was completed on a computing platform equipped with an Intel i7-14700KF CPU and dual RTX 2080 Ti GPUs.

[0138] The first step was dataset construction, where 500 expert trajectories were collected in typical 3D scenes such as "Block" and "Forest" using an expert planner. To train the sequential actors, the solution sequences of all Traveling Salesman Problem (TSP) problems were retained as supervision signals.

[0139] Further parameter configuration and reward settings are required. The path cost weight is 100. The learning rate is set to 1. The network architecture uses a shared heterogeneous graph encoder, employs the Adam optimizer, and sets the learning rate to 1. The batch size is 256.

[0140] Through 10,000 iterative optimization steps, the model converged on the offline dataset. This process combined behavioral cloning with trust-weighted reinforcement learning, ensuring that the model could both incorporate expert priors and identify and avoid out-of-distribution risks through a diffusion evaluator.

[0141] like Figure 4 As shown, the environment types include randomly generated partitioned scenes and forest scenes (scale: [size missing]). ), and large-scale realistic scene hotels (multi-layered structure, large scale) ) and highly unstructured caves (narrow grids, scale) ).

[0142] Experimental results show that in all scenarios, the present invention significantly outperforms the benchmark in both total detection time and travel distance. In complex cave scenarios, the path generated by this scheme is smoother and has global consistency, effectively avoiding the frequent backtracking and path oscillations common in traditional methods. This demonstrates the advantages of sequential actors in long-term planning and the robustness of the diffusion evaluator in handling complex topological connections.

[0143] To verify the robustness and versatility of this method, this embodiment underwent real-world testing on two different hardware platforms:

[0144] The experiment used a coaxial octocopter unmanned aerial vehicle (UAV) and a two-wheeled differential mobile robot (UGV). Both were equipped with Intel NUC computing units and Livox Mid-360 LiDAR.

[0145] like Figure 5 As shown, in the "garden" and "school gate" scenes with complex 3D geometry and significant height variations, the drone successfully navigated between narrow arches and multiple layers of vegetation using a topological path generated by the 3D skeleton map.

[0146] This embodiment deploys a strategy trained entirely in a 3D environment directly on a UGV platform with only 2D planar topological constraints. Test results show that in a "circular plaza" scene with long corridors and narrow passages, UGV can achieve efficient detection without any fine-tuning and avoids getting trapped in local optima.

[0147] Real-world experiments have demonstrated that the framework proposed in this invention can extract the essential connectivity of the environmental space independent of the robot's dynamic characteristics, and has extremely strong cross-platform adaptability and transferability from simulation to reality.

[0148] In summary, the embodiments of the present invention have the following advantages compared with the prior art:

[0149] 1) Significantly improved detection efficiency and optimized path cost: This embodiment uses SequenceActor for long-term reasoning, which can effectively identify the global connectivity of the environment and reduce the robot's redundant backtracking and hesitation during the detection process.

[0150] 2) Enhanced robustness under multimodal decision-making: An evaluator based on a diffusion probability model is introduced, which can accurately model the multimodal value distribution of detection actions and output a "trust score" representing the reliability of the decision. This mechanism effectively solves the problem of inaccurate value estimation in out-of-distribution areas or high-risk areas. By using the "trust score" as a reliability gating, the system effectively filters out erroneous predictions with high uncertainty, significantly enhancing the robustness of the detection strategy in complex and unpredictable scenarios.

[0151] 3) Extremely high motion smoothness and temporal consistency: This embodiment changes the traditional "single-step viewpoint selection" mode and adopts an autoregressive method to predict long-range node sequences. This serialization generation method fully considers the historical decision context, eliminating frequent robot turns and path jitter. The generated trajectory maintains high logical consistency in the time dimension, which not only improves motion efficiency but also reduces the mechanical wear of the actuators.

[0152] 4) Excellent zero-shot transfer learning and cross-platform adaptability: Thanks to the offline reinforcement learning framework and lightweight topological graph representation, this embodiment extracts the inherent spatial connectivity of the environment, rather than merely mimicking a specific dynamic model. The strategy implemented in this invention is trained entirely in a simulation environment and can be directly deployed on real-world unmanned aerial vehicle (UAV) and unmanned vehicle (UGV) platforms without any fine-tuning. Experiments demonstrate that this strategy maintains extremely high detection efficiency across different motion constraints (from omnidirectional 3D flight to planar 2D driving) and different geographical topologies, exhibiting strong versatility and practical application value.

[0153] 5) Balancing computational efficiency and real-time performance: By decoupling generative modeling (training phase) and autoregressive inference (execution phase), this invention avoids the iterative denoising delay common in diffusion models in real-time tasks. In practical deployment, this algorithm can achieve high-frequency replanning on lightweight onboard computing units (such as Intel NUC), meeting the real-time detection needs of robots in unknown dynamic environments.

[0154] Reference Figure 2 A robotic detection system based on diffusion-enhanced sequential topology planning includes:

[0155] The first module 201 is used to acquire real-time point cloud observation data of the environment, perform incremental topological abstraction processing and hybrid graph encoding reasoning, and generate spatial embedding vectors.

[0156] The second module 202 is used to generate a historically enhanced serialized autoregressive path based on the spatial embedding vector to obtain serialized topological path data;

[0157] The third module 203 is used to sequentially perform reverse diffusion denoising, uncertainty quantization sampling and reliability gating mapping on the serialized topology path data to obtain trust score data;

[0158] The fourth module 204 is used to construct a composite objective function based on the trust score data, perform hybrid-driven optimization on the serialized topological path data, and realize robot detection path planning.

[0159] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0160] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A robot detection method based on diffusion-enhanced sequential topology planning, characterized in that, Includes the following steps: Acquire real-time point cloud observation data of the environment, perform incremental topological abstraction and hybrid graph encoding reasoning, and generate spatial embedding vectors; Based on the spatial embedding vector, a history-enhanced serialized autoregressive path is generated to obtain serialized topological path data; The serialized topology path data is sequentially subjected to inverse diffusion denoising, uncertainty quantization sampling, and reliability gating mapping to obtain trust score data; Based on the trust score data, a composite objective function is constructed to perform hybrid-driven optimization on the serialized topological path data, thereby realizing robot exploration path planning.

2. The robot detection method based on diffusion-enhanced sequential topology planning according to claim 1, characterized in that, The step of acquiring real-time point cloud observation data of the environment, performing incremental topological abstraction processing and hybrid graph encoding inference, and generating spatial embedding vectors specifically includes: Real-time point cloud observation data of the environment is acquired, and the skeleton line of the geometric center of the environment is extracted through an incremental topology mapping mechanism. Nodes are defined, and a target node dataset is constructed. The target node dataset includes the robot's current position, the frontal viewpoint of the detection, and the connection points of the environmental skeleton. The feature matrix is ​​obtained by performing feature matrix encoding calculation on the target node dataset; Obtain the adjacency matrix of the target node dataset, which is used to reflect the physical connectivity between target nodes; Hybrid graph encoding inference is performed on the feature matrix and the adjacency matrix to generate a spatial embedding vector.

3. The robot detection method based on diffusion-enhanced sequential topology planning according to claim 2, characterized in that, The step of performing feature matrix encoding calculation on the target node dataset to obtain the feature matrix specifically includes: The absolute positions of the target node dataset are centered and normalized to obtain relative 3D coordinates; The unknown voxel volume of the target node dataset is quantized using a visibility algorithm to obtain an information gain scalar. Based on the type of the target node dataset, binary mapping is performed to generate one-hot encodings of semantic types; The relative 3D coordinates, the information gain scalar, and the one-hot encoding of the semantic type are concatenated to obtain the feature matrix.

4. The robot detection method based on diffusion-enhanced sequential topology planning according to claim 1, characterized in that, The step of generating serialized autoregressive paths based on the spatial embedding vector to obtain serialized topological path data specifically includes: Based on the multi-head attention mechanism, the robot's current position is used as the query vector, and the spatial embedding vector is used as the key vector and value vector. Attention weights are calculated and features are weighted and aggregated to obtain global geometric background data. The trajectory prefix is ​​constructed by explicitly recording the planned node index sequence and inputting it into the converter encoder for temporal feature extraction and pooling processing to obtain the current trajectory prefix feature data. The global geometric background data and the current trajectory prefix feature data row vector are concatenated and processed, and then linearly transformed through a mapping layer to obtain belief state data that characterizes the current decision moment. Based on the belief state data, a masked policy distribution calculation is performed to obtain the selection probability distribution data of each candidate node; Based on the selection probability distribution data of each candidate node, node sampling is performed, and the selected node is updated in real time to the trajectory prefix as the input for the next decision. This process is repeated until a path chain of a preset length is generated, resulting in serialized topological path data.

5. The robot detection method based on diffusion-enhanced sequential topology planning according to claim 4, characterized in that, The step of calculating the masked policy distribution based on the belief state data to obtain the selection probability distribution data of each candidate node specifically includes: Construct a dynamic validity mask based on the adjacency matrix; The belief state data and the robot's current position are vector-concatenated and linearly projected to generate a query vector; Calculate the matrix product of the query vector and the spatial embedding vector, and then divide by the square root of the feature dimension for scaling to obtain the spatial node feature attention retrieval calculation result. The spatial node feature attention retrieval calculation result is added to the dynamic validity mask, and then normalized using the Softmax function to obtain the selection probability distribution data of each candidate node.

6. The robot detection method based on diffusion-enhanced sequential topology planning according to claim 5, characterized in that, The specific calculation expression for the masked policy distribution is as follows: In the above formula, Data representing belief states, Indicates the robot's current position. Represents the query vector. Represents a spatial embedding vector. Indicates a dynamic validity mask. Represents the square root of the feature dimension. This represents the Softmax function. Indicates the first The selected candidate nodes in the step, This represents the probability distribution of a candidate node being selected, given the belief state and the robot's current position.

7. The robot detection method based on diffusion-enhanced sequential topology planning according to claim 1, characterized in that, The step of sequentially performing reverse diffusion denoising, uncertainty quantization sampling, and reliability gating mapping on the serialized topology path data to obtain trust score data specifically includes: Based on a pre-trained denoising network, the robot's current environmental observation state is obtained. With random noise as the initial state and the serialized topological path data as input, multi-step iterative reverse diffusion denoising prediction is performed to obtain action value prediction data that fits the multimodal value distribution of the action. Uncertainty quantification sampling processing is performed on the action value prediction data, and its standard deviation is calculated to obtain value deviation assessment data that reflects the uncertainty of current value perception. Based on the value deviation assessment data, combined with a preset penalty coefficient, the trust score data is obtained by mapping calculation through a negative exponential transformation function.

8. The robot detection method based on diffusion-enhanced sequential topology planning according to claim 7, characterized in that, The specific expression for the multi-step iterative reverse diffusion denoising prediction is as follows: In the above formula, This represents the learned denoising network. Indicates the reverse diffusion denoising method. Step motion value prediction data, This represents the preset noise scheduling parameters. Indicates the first Step motion value prediction data, This represents the cumulative product term of the noise scheduling parameters. This indicates the robot's current environmental observation status. This represents serialized topology path data (i.e., actions). This indicates the current number of back diffusion steps. This represents the posterior variance of the reverse process. This represents standard Gaussian noise.

9. The robot detection method based on diffusion-enhanced sequential topology planning according to claim 7, characterized in that, The step of constructing a composite objective function based on the trust score data and performing hybrid-driven optimization on the serialized topological path data to achieve robot exploration path planning specifically includes: The trust score data is calculated by weighting the trust value prediction data with the action value prediction data to obtain the trust weighted target value data. Construct diffusion reconstruction loss data and value regression loss data, update the weight parameters of the pre-trained denoising network to obtain an optimized denoising network. The diffusion reconstruction loss data represents the mean square error between the noise predicted by the pre-trained denoising network and the actual injected noise, and the value regression loss data represents the mean square error between the action value prediction data and the trust-weighted target value data. A sequence-level trust reinforcement loss data is constructed, and the sequenced topology path data is backpropagated using the optimized denoising network to achieve robot exploration path planning.

10. A robot detection system based on diffusion-enhanced sequential topology planning, characterized in that, Includes the following modules: The first module is used to acquire real-time point cloud observation data of the environment, perform incremental topological abstraction processing and hybrid graph encoding reasoning, and generate spatial embedding vectors. The second module is used to generate a historically enhanced serialized autoregressive path based on the spatial embedding vector, thereby obtaining serialized topological path data. The third module is used to sequentially perform reverse diffusion denoising, uncertainty quantization sampling, and reliability gating mapping on the serialized topology path data to obtain trust score data; The fourth module is used to construct a composite objective function based on the trust score data, perform hybrid-driven optimization on the serialized topological path data, and realize robot detection path planning.

Images