Visual navigation method and device based on self-imitating diffusion

Abstract

The embodiment of the application provides a visual navigation method and device based on self-imitation diffusion, effective perception of an environment is realized through feature extraction and state coding. An evaluation mechanism is constructed, a reliable sample selection strategy is established by combining reward calculation and weight screening. Strategy optimization is introduced, and continuous improvement of navigation is ensured through self-imitation learning and parameter updating. The method effectively solves the deficiencies of traditional technologies in trajectory generation, quality evaluation and strategy optimization, and provides technical support for visual navigation.

Description

Technical Field

[0001] This application relates to the field of computer vision, specifically to a visual navigation method and apparatus based on self-imitating diffusion. Background Technology

[0002] Existing visual navigation methods have significant shortcomings. Traditional systems perform poorly in trajectory generation and feature extraction, failing to effectively achieve environmental perception and path planning, thus affecting navigation performance.

[0003] Furthermore, existing technologies face bottlenecks in trajectory evaluation and sample selection. Most systems lack robust quality assessment mechanisms and weight allocation strategies, resulting in suboptimal sample quality.

[0004] Existing systems have technical shortcomings in policy optimization. The lack of in-depth analysis of training samples makes it difficult to achieve efficient policy updates through self-imitation mechanisms, thus affecting learning performance. Solving these problems is crucial for improving visual navigation capabilities. Summary of the Invention

[0005] To address the problems in the prior art, this application provides a visual navigation method and apparatus based on self-imitation diffusion, which can effectively solve the shortcomings of traditional technologies in trajectory generation, quality assessment and strategy optimization, and provide technical support for visual navigation.

[0006] To solve at least one of the above problems, this application provides the following technical solution: In a first aspect, this application provides a visual navigation method based on self-imitating diffusion, comprising: Based on the navigation strategy, a candidate trajectory set is generated in the environment. The environmental visual information is extracted according to the preset perception rules to generate a perception data table. The perception data table is state encoded to generate a state feature set. The state feature set is sampled according to the preset diffusion rules to generate a candidate trajectory set. The candidate trajectory set is used for subsequent evaluation. The candidate trajectory set is quality evaluated to generate a high-quality trajectory set. The candidate trajectory set is then used to calculate rewards according to preset evaluation rules to generate a reward value table. The reward value table is then filtered by threshold to generate a trajectory weight table. Samples are selected from the candidate trajectory set according to the trajectory weight table to generate a high-quality trajectory set. The high-quality trajectory set is then used for strategy updates. Based on the high-quality trajectory set, the navigation strategy is updated by self-imitation to generate an optimized strategy. The high-quality trajectory set is used to construct learning samples according to preset training rules to generate a training dataset. The training dataset is denoised and trained to generate a diffusion parameter table. The navigation strategy is updated according to the diffusion parameter table to generate an optimized strategy. The optimized strategy is then applied to the visual navigation task.

[0007] Furthermore, it also includes: generating an original dataset based on visual information collected from the navigation environment; acquiring environmental images according to preset sampling rules to generate an image data table; performing data preprocessing on the image data table to generate a feature map; performing multi-scale analysis on the feature map according to preset feature extraction rules to generate a perceptual feature set; and using the perceptual feature set for subsequent encoding. The sensory feature set is used to generate a candidate trajectory set. The sensory feature set is then mapped to a state vector table according to a preset encoding rule. A trajectory generator is constructed according to a preset diffusion sampling rule to generate a sampling parameter table. Based on the sampling parameter table, trajectory sampling is performed on the state vector table to generate a candidate trajectory set. The candidate trajectory set is then used for subsequent evaluation.

[0008] Furthermore, it also includes: generating a feature vector set based on the state representation of the perception data table, preprocessing the perception data table according to a preset encoding rule to generate an encoded data table, performing feature transformation on the encoded data table according to a preset mapping rule to generate a transformation matrix, performing state mapping on the encoded data table according to the transformation matrix to generate a feature vector set, and using the feature vector set for subsequent sampling; Trajectory sampling is performed on the feature vector set to generate a candidate trajectory set. Noise is injected into the feature vector set according to a preset diffusion rule to generate an initial state table. The initial state table is denoised iteratively according to the diffusion time sequence template to generate a trajectory sequence table. A trajectory generator is constructed based on the trajectory sequence table to generate a candidate trajectory set. The candidate trajectory set is used for quality evaluation.

[0009] Furthermore, it also includes: constructing an evaluator based on the candidate trajectory set to generate a scoring data table; performing feature decomposition on the candidate trajectory set according to preset evaluation rules to generate a trajectory feature table; performing multi-dimensional analysis on the trajectory feature table according to a preset indicator system to generate an evaluation indicator set; performing scoring calculation on the trajectory feature table according to the evaluation indicator set to generate a scoring data table; and using the scoring data table for reward calculation. The scoring data table is used to calculate reward values ​​to generate a reward value table. The scoring data table is then normalized according to a preset reward rule to generate a normalized score table. The normalized score table is then weighted and combined according to a preset weight template to generate a combined score table. A reward calculator is then constructed based on the combined score table to generate a reward value table. The reward value table is then used for trajectory filtering.

[0010] Furthermore, it also includes: constructing a filter based on the reward value table to generate a trajectory weight table, performing distribution analysis on the reward value table according to a preset threshold rule to generate a distribution feature table, dividing the distribution feature table into intervals according to a preset filtering template to generate an interval mapping table, calculating the importance of the reward value table based on the interval mapping table to generate a trajectory weight table, and using the trajectory weight table for sample selection; A high-quality trajectory set is generated by selecting samples from the trajectory weight table. The trajectory weight table is then sorted by priority according to a preset selection rule to generate a ranking index table. The candidate trajectory set is sampled based on the ranking index table to generate a sampling sequence table. A selector is constructed based on the sampling sequence table to generate a high-quality trajectory set. The high-quality trajectory set is then used for self-imitation update.

[0011] Furthermore, it also includes: building a trainer based on a high-quality trajectory set to generate a training dataset; performing data augmentation on the high-quality trajectory set according to a preset sampling rule to generate an augmented data table; performing feature annotation on the augmented data table according to a preset annotation rule to generate an annotated data table; organizing the annotated data table according to a preset training template to generate a training dataset; and using the training dataset for parameter updates. The training dataset is used to learn parameters to generate a model parameter table. The training dataset is divided into batches according to a preset learning rule to generate a training batch table. Gradients are calculated on the training batch table according to a preset loss function to generate a gradient vector set. A parameter optimizer is constructed based on the gradient vector set to generate a model parameter table. The model parameter table is used for policy updates.

[0012] Furthermore, it also includes: constructing a denoiser to generate a diffusion parameter table based on the training dataset; injecting noise into the training dataset according to preset noise rules to generate a noise data table; configuring the sampling step size of the noise data table according to preset time rules to generate a time series parameter table; performing denoising iteration on the noise data table according to the time series parameter table to generate an iteration result table; constructing a parameter extractor based on the iteration result table to generate a diffusion parameter table; and using the diffusion parameter table for policy updates. The diffusion parameter table is optimized to generate an optimization strategy. The diffusion parameter table is mapped according to a preset update rule to generate a strategy parameter table. The navigation strategy is restructured according to the strategy parameter table to generate an optimization model. A navigation controller is built based on the optimization model to generate an optimization strategy. The optimization strategy is then deployed to the visual navigation system.

[0013] Secondly, this application provides a visual navigation device based on self-imitating diffusion, comprising: The trajectory sampling module is used to generate a candidate trajectory set in the environment based on the navigation strategy, extract features from the environmental visual information according to the preset perception rules to generate a perception data table, perform state encoding on the perception data table to generate a state feature set, perform trajectory sampling on the state feature set according to the preset diffusion rules to generate a candidate trajectory set, and use the candidate trajectory set for subsequent evaluation. The trajectory filtering module is used to perform quality evaluation on the candidate trajectory set to generate a high-quality trajectory set, calculate the reward of the candidate trajectory set according to the preset evaluation rules to generate a reward value table, perform threshold filtering on the reward value table to generate a trajectory weight table, select samples from the candidate trajectory set according to the trajectory weight table to generate a high-quality trajectory set, and use the high-quality trajectory set for strategy update. The visual navigation module is used to perform self-imitation update of the navigation strategy based on the high-quality trajectory set to generate an optimized strategy. The high-quality trajectory set is used to construct learning samples according to preset training rules to generate a training dataset. The training dataset is denoised and trained to generate a diffusion parameter table. The navigation strategy is updated according to the diffusion parameter table to generate an optimized strategy. The optimized strategy is then applied to the visual navigation task.

[0014] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the self-imitating diffusion-based visual navigation method.

[0015] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the self-imitating diffusion-based visual navigation method described above.

[0016] Fifthly, this application provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the self-imitating diffusion-based visual navigation method.

[0017] As can be seen from the above technical solution, this application provides a visual navigation method and apparatus based on self-imitation diffusion, which achieves effective environmental perception through feature extraction and state encoding. An evaluation mechanism is constructed, combining reward calculation and weight selection to establish a reliable sample selection strategy. Strategy optimization is introduced, ensuring continuous improvement of navigation through self-imitation learning and parameter updates. This method effectively solves the shortcomings of traditional technologies in trajectory generation, quality assessment, and strategy optimization, providing technical support for visual navigation. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the visual navigation method based on self-imitating diffusion in an embodiment of this application. Figure 2 This is a structural diagram of the visual navigation device based on self-imitating diffusion in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of the electronic device in the embodiments of this application.

[0020] Figure label: Electronic device 9600, central processing unit 9100, memory 9140, communication module 9110, input unit 9120, audio processor 9130, display 9160, power supply 9170, buffer memory 9141, application / function storage unit 9142, data storage unit 9143, driver storage unit 9144, antenna 9111, speaker 9131, microphone 9132. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] The acquisition, storage, use, and processing of data in this application comply with relevant laws and regulations.

[0023] To address the problems existing in current technologies, this application provides a visual navigation method and apparatus based on self-imitation diffusion. Through feature extraction and state encoding, it achieves effective environmental perception. An evaluation mechanism is constructed, combining reward calculation and weight selection to establish a reliable sample selection strategy. Policy optimization is introduced, ensuring continuous improvement of navigation through self-imitation learning and parameter updates. This method effectively solves the shortcomings of traditional technologies in trajectory generation, quality assessment, and policy optimization, providing technical support for visual navigation.

[0024] To effectively address the shortcomings of traditional technologies in trajectory generation, quality assessment, and strategy optimization, and to provide technical support for visual navigation, this application provides an embodiment of a visual navigation method based on self-imitating diffusion. See [link to embodiment]. Figure 1 The visual navigation method based on self-imitation diffusion specifically includes the following: Step S101: Generate a candidate trajectory set in the environment based on the navigation strategy, extract features from the environmental visual information according to the preset perception rules to generate a perception data table, encode the perception data table to generate a state feature set, sample the state feature set according to the preset diffusion rules to generate a candidate trajectory set, and use the candidate trajectory set for subsequent evaluation. First, the raw visual information from the navigation environment is input, and then temporal and camera pose alignment is performed to generate an image sequence for perception. Based on preset perception rules, distortion correction, brightness normalization, and scale resampling are performed on each frame, and optical flow consistency filtering is conducted within the time window to obtain a set of frames with lower noise. Subsequently, a multi-scale convolutional perceptron is invoked to extract three types of features from this set: edge texture, semantic regions, and passable space. These features are then summarized into a perception data table using the frame number as the key, serving as the first intermediate value in this step.

[0025] Based on the aforementioned perception data table, state representation is performed, mapping the three types of features of each frame to unified state coordinates. Specifically, the semantic region is first sparsified by instance segmentation masking, then the passable space is converted into an occupied raster representation, and edge textures are encoded as directional histogram entries. After dimensional alignment, these entries are written into the state encoder, outputting a state feature set. The state feature set corresponds one-to-one with the original frame timestamp, facilitating subsequent sampling and reading in chronological order.

[0026] A diffusion sampling timing template is configured based on the temporal order of the state feature set to determine the mapping relationship between noise injection intensity and denoising steps. To this end, a trajectory diffusion sampler is constructed, whose input is a sliding window of the state feature set, and whose output is candidate trajectory segments. Before sampling, environmental constraint markers, including intersection topology and restricted areas, are added to each window as sampling feasibility masking signals to avoid generating unreachable action sequences.

[0027] Based on the aforementioned time-series template, noise injection and denoising iterations are initiated, generating several trajectory segments window by window. Overlapping and splicing, as well as velocity consistency checks, are performed at the window boundaries. For segments that do not meet consistency requirements, one-order denoising is performed at the splicing point, and the direction is re-estimated until the velocity and acceleration changes fall within a set range. This process yields a preliminary candidate trajectory set in segment form, retaining the window start and end indices within each entry to ensure that subsequent evaluations can trace back to the corresponding state features.

[0028] Trajectory-level resampling is performed on the candidate segments to improve global coherence. Specifically, rigid body registration is performed on the end pose and start pose of adjacent segments, and key points within the trigger segment of the trajectory with accumulated deviation exceeding a threshold are backfilled. During backfilling, adjustments are only allowed within the traversable space. After processing, a complete temporal domain coverage set of candidate trajectories is output, with each trajectory including three statistics: sampling steps, number of environmental constraint hits, and traversable percentage.

[0029] To clarify the causal relationship between sampling and environmental constraints, a control variable can be optionally recorded as a prior score for subsequent quality assessment during the generation process of a single trajectory. This control variable is calculated using the following formula: Qu = a1·Ru + a2·Cu a3·Vu.

[0030] In the formula, Qu is the recorded score, Ru is the passable percentage, Cu is the inverse measure of the number of times the environmental constraint is hit, and Vu is the speed fluctuation; a1, a2, and a3 are non-negative weights, set according to the environmental complexity and fixed in the current batch. Qu does not directly determine the quality, but it is read as a sorting reference during the evaluation stage.

[0031] After the candidate trajectory set is generated, the corresponding state feature index and timestamp are written into the trajectory metadata to establish a mapping from trajectory to state feature set. This mapping is used for re-retrieval of keyframes for cost calculation during subsequent quality assessment, and also for state restoration when replaying high-quality samples during the policy update stage, avoiding the disconnect between state and action.

[0032] Based on the aforementioned mapping, anchor points are further extracted from each candidate trajectory at fixed time intervals to label the semantic region sequence traversed by the route and the grid crossing rate, which are used for feature decomposition in the subsequent evaluation module. Anchor point extraction does not change the trajectory ontology; it only serves as a simplified representation of the evaluation input and points back to the original trajectory index, ensuring that the scoring result can locate the specific trajectory entry.

[0033] Then, a consistency check is performed on the generation process to check whether the start and end poses of all entries fall within the time range of the perception data table, and whether the environmental constraint markers are consistent with the passable space at that time. Inconsistent entries are removed from this batch, and the reason for removal and the corresponding window are recorded to avoid affecting the subsequent threshold analysis.

[0034] Finally, the candidate trajectory set, along with trajectory metadata, anchor point sequences, and control variables Qu, are packaged into an evaluation input, and a query interface is exposed for subsequent steps to perform quality evaluation and reward calculation on the candidate trajectory set. This interface also provides filtering capabilities based on time and environmental constraints, enabling subsequent steps to reproduce the experiment under the same conditions and ensuring accurate replay of the corresponding state sequence during the self-imitation update phase.

[0035] Step S102: Perform quality assessment on the candidate trajectory set to generate a high-quality trajectory set; calculate the reward of the candidate trajectory set according to a preset assessment rule to generate a reward value table; perform threshold screening on the reward value table to generate a trajectory weight table; select samples from the candidate trajectory set according to the trajectory weight table to generate a high-quality trajectory set; and use the high-quality trajectory set for strategy update. First, the candidate trajectory set, its trajectory metadata, and anchor point sequence output in step S101 are read and aligned according to timestamps and environmental constraint markers. For each trajectory, the corresponding state feature sliding window is retrieved to restore the passable space and semantic region of the keyframe, which are used as the context for cost calculation. To avoid data drift, missing anchor points are first backfilled with adjacent time-series occupied raster segments, and the backfilling marker is written to the trajectory-side record. Subsequent calculations apply a reduced weight to this marker.

[0036] Based on the aligned data, feature decomposition is performed, breaking down the trajectory into three evaluation dimensions: path reachability, geometric smoothness, and target relevance. Specifically, path reachability is used to count the proportion of passable grid crossings and the number of environmental constraint violations; geometric smoothness measures the continuity of velocity and curvature; and target relevance measures the spatial proximity of the trajectory's termination state to the task objective. These three indicators form an evaluation feature set, which is mapped one-to-one with the trajectory index and serves as direct input for scoring calculation.

[0037] A scorer is constructed based on the aforementioned set of evaluation features to calculate multi-dimensional scores for each item. The scorer employs a weighted summation and penalty-based approach, using accessibility and smoothness as primary items and violations and significant jitter as deduction items. For paragraphs containing backfill markers, a weighting coefficient is introduced to reduce their contribution. The scoring results, along with dimensional details, are written back to the scoring records to form the basic data for reward calculation. To ensure comparability, different scenarios are aligned within batches to avoid any particular scenario dominating the score.

[0038] Reward calculations are performed on the scoring records, outputting a reward value table. The reward calculation first normalizes the scores of each dimension by batch percentile, then weights them according to a preset weight template, while adding a sparsity penalty to prevent a single dimension from being too high and masking other defects. Entries with severe conflicts with environmental constraints are truncated with a lower limit to ensure their rewards do not exceed a safe threshold. Finally, the reward value for each trajectory is obtained, and the weight template version used in the calculation is retained for subsequent traceability.

[0039] To balance stability and diversity during sample selection, reward smoothing and suppression of excessive bias can be optionally controlled. In this embodiment, this control employs the following formula: Rk = b1·Ak + b2·Gk b3·Pk, Where Rk is the smoothing reward for the k-th trajectory, Ak is the reachability sub-score, Gk is the geometric smoothing sub-score, and Pk is the violation penalty sub-score; b1, b2, and b3 are non-negative weights, fixed in the current batch. Rk is written into an auxiliary column of the reward value table for sorting during subsequent threshold analysis, but does not replace the main reward value.

[0040] After the reward value table is generated, the threshold screening process begins. First, the reward values ​​are analyzed for distribution. Candidate threshold intervals are determined based on the density valleys within each batch and historical threshold patterns. Then, an auxiliary column Rk is used to eliminate extreme values ​​within these intervals, resulting in a trajectory weight table. The trajectory weight table assigns an importance weight and a label indicating whether each trajectory should be included in the sample pool. It also retains the scene number and constraint hit count for bias control during subsequent sampling.

[0041] The sample selection process is triggered based on the trajectory weight table. The selector prioritizes samples according to their importance weight and performs stratified extraction based on scene number to avoid an excessive number of samples in a single scene. A similarity suppression strategy is used for similar trajectories under the same task objective, retaining only representative entries and marking the rest as redundant for later supplementation. After extraction, a high-quality trajectory set is formed and referenced to the original trajectory index to ensure that it can be replayed to the corresponding state and perception data.

[0042] After the high-quality trajectory set is determined, a training list for policy updates is generated, which includes trajectory sequences, anchor times, and environmental constraint labels. The training list serves as the input for the next step, used for constructing learning samples and for denoising training in self-imitation learning. Simultaneously, the reward value table and trajectory weight table are retained as update logs, providing a basis for threshold adaptation and weight template revision in subsequent batches.

[0043] Step S103: Based on the high-quality trajectory set, the navigation strategy is updated by self-imitation to generate an optimized strategy. The high-quality trajectory set is used to construct learning samples according to preset training rules to generate a training dataset. The training dataset is denoised and trained to generate a diffusion parameter table. The navigation strategy is updated according to the diffusion parameter table to generate an optimized strategy. The optimized strategy is applied to the visual navigation task.

[0044] First, the high-quality trajectory set and training list output in step S102 are read and aligned according to the trajectory sequence and anchor point time to recover the corresponding perceptual features and environmental constraint labels. Each trajectory is divided into fixed-length segments, and the start and end poses and passable space sections are added to both ends of the segments as context for subsequent learning samples. For segments with missing frames, only time backfilling is performed and a backfilling mark is recorded. Subsequent training applies a weight reduction to this mark to avoid bias updates.

[0045] Based on the high-quality trajectory set, training samples are constructed. Trajectory segments and their contexts are packaged as input items, and action sequences and pose changes within segments are recorded as target items. Sample organization follows pre-defined training rules: first, samples within the same scene are scattered according to time; second, cross-scene sampling is stratified according to environmental complexity; and third, the sampling probability of samples containing backfilled tags is reduced. This organization yields the training dataset, and segment indices are retained for playback and error localization.

[0046] The denoising training process is configured based on the training dataset, establishing a diffusion denoising network as a "trajectory denoiser." The input consists of the state context and the action sequence perturbed by noise, and the output is the denoised action residual. During training, time-related noise intensity is injected into each sample, employing multi-step denoising iterations, with fragment-level loss as the optimization objective. To avoid overfitting to a single scene, balanced sampling is used for batches across scenes to ensure a stable gradient source distribution.

[0047] In the denoising training, a reward-guided weighting mechanism is introduced, mapping the reward scores of high-quality trajectories to sample weights to adjust the loss contribution. Simultaneously, a reduction factor is applied to samples containing backfilled labels. Optionally, the following formula is introduced to describe the weighting structure of the single-batch loss: Ld = c1·Eres + c2·Ereg c3·Erob.

[0048] In the formula, Ld is the batch denoising loss, Eres is the expected error of the action residual term, Ereg is the trajectory smoothing regularization term, and Erob is the penalty term for violating environmental constraints; c1, c2, and c3 are non-negative weights that remain unchanged in the current training round. Ld is used to update the denoising network parameters and record the weight values ​​and batch index in the training log.

[0049] After the denoising training converges, the network parameters and temporal noise table are extracted and summarized into a diffusion parameter table. The diffusion parameter table includes step configuration, noise intensity mapping, and denoiser weight version, and exposes a read interface. Subsequently, the diffusion parameter table is written into the navigation strategy to complete parameter-level replacement and structural verification; incompatible old parameters are rolled back to ensure stability during the inference phase.

[0050] After updating the strategy according to the diffusion parameter table, an optimized strategy is generated, and then closed-loop playback is performed in the verification environment. Playback uses the same perceptual data and environmental constraints as in step S101 to generate a navigation trajectory at a fixed number of sampling steps, checking whether the trajectory continuity and reachability meet preset boundaries. For any discovered boundary conflicts, a problem segment index is recorded, and incremental corrections are performed on the training dataset.

[0051] Finally, the optimized strategy is deployed to the visual navigation task. The deployment process retains the strategy version number, diffusion parameter table number, and training batch record for easy retesting and backtracking in a real-world environment. Simultaneously, the organization of the high-quality trajectory set and training dataset is saved as a template for direct reuse in the next round of self-imitation updates, forming a closed-loop chain from sample construction and denoising training to parameter updates.

[0052] As described above, the self-imitating diffusion-based visual navigation method provided in this application can achieve effective environmental perception through feature extraction and state encoding. An evaluation mechanism is constructed, combining reward calculation and weight selection to establish a reliable sample selection strategy. Policy optimization is introduced, ensuring continuous improvement of navigation through self-imitating learning and parameter updates. This method effectively addresses the shortcomings of traditional technologies in trajectory generation, quality assessment, and policy optimization, providing technical support for visual navigation.

[0053] In one embodiment of the visual navigation method based on self-imitation diffusion in this application, the following specific content may also be included: Step S201: Generate an original dataset by collecting visual information based on the navigation environment, acquire environmental images according to preset sampling rules to generate an image data table, perform data preprocessing on the image data table to generate a feature map, perform multi-scale analysis on the feature map according to preset feature extraction rules to generate a perceptual feature set, and use the perceptual feature set for subsequent encoding. Step S202: Generate a candidate trajectory set from the perceived feature set, map the perceived feature set to a state vector table according to a preset encoding rule, construct a trajectory generator to generate a sampling parameter table according to a preset diffusion sampling rule, perform trajectory sampling on the state vector table based on the sampling parameter table to generate a candidate trajectory set, and use the candidate trajectory set for subsequent evaluation.

[0054] First, after accessing the multi-source image streams from the navigation environment, timestamps and device numbers are aligned to obtain a raw dataset arranged chronologically. Based on preset sampling rules, keyframes are extracted from each camera within a given time window, and the approximate pose, exposure range, and field of view coverage are recorded to form an image data table. For segments with temporal jitter, interpolation is performed using adjacent keyframes, and interpolation markers are added. Subsequent processing reduces the weight of these markers to avoid interpolation errors affecting feature statistics.

[0055] Based on the image data table, data preprocessing is performed sequentially, including distortion correction, brightness normalization, and color correction, and the projection with consistent scale is restored using camera intrinsic parameters. Then, edge gradients and optical flow stability are calculated for each frame, strongly reflective frames are filtered out, and their indices are retained for traceability. This process outputs a feature map, with entries including texture intensity, initial semantic segmentation, and walkable probability raster, which serves as input for subsequent multi-scale analysis.

[0056] Multi-scale analysis is performed on the feature map, using a pyramid window to extract descriptors from three layers: local texture, regional semantics, and global geometry. The local layer focuses on the stability of corner points and line segments, the regional layer focuses on the boundary consistency of roads, obstacles, and landmarks, and the global layer infers reachable corridors based on the horizon and vanishing perspective points. The results from these three layers are fused into a perceptual feature set, and a one-to-one mapping is established with the original timestamps and camera numbers to ensure that subsequent encoding can be read in chronological order and from the perspective.

[0057] Based on the perceived feature set, a state mapping process is constructed to align local, regional, and global descriptors to unified state coordinates. Specifically, the passability probability grid is compressed into block occupancy rate, the relative orientation of landmarks is encoded into polar coordinate pairs, and texture stability is summarized into direction vectors. After dimensional alignment, the above encodings output a state vector table, retaining the source mapping, which is used as an index for sampler reading.

[0058] A trajectory generator is constructed on the state vector table. The trajectory generator follows a preset diffusion sampling rule, first injecting time-related noise into each fixed-length time window, and then performing denoising iterations according to the step template to gradually recover the action increment. During the iteration process, reachability masking is enabled, allowing only actions that fall within the passable occupied block to be retained; out-of-bounds steps are backtracked and the direction is re-estimated to ensure the physical feasibility of the generated sequence.

[0059] To reduce cross-scene offset, the trajectory generator maintains noise intensity curves and step configurations for different camera viewpoints, and performs pose alignment and velocity continuity checks when stitching cross-viewpoint windows. This process outputs several trajectory segments, and records the number of out-of-bounds backtracking and occupancy statistics at the end of each segment as a priori reference for subsequent evaluation.

[0060] Based on the trajectory segments, segment-level stitching and conflict resolution are performed. Rigid body registration is performed on the end and start poses of adjacent segments. If the cumulative error exceeds the limit, the side with the higher occupancy rate is maintained and the other side is locally resampled. After stitching, a candidate trajectory set is obtained, along with segment source, sliding window index, and reachability statistics, to ensure that the original state and perceived features can be located during the evaluation stage.

[0061] Finally, the candidate trajectory set, along with the reference relationships in the state vector table, is packaged into an evaluation input, and time and viewpoint retrieval interfaces are exposed. This input is read in subsequent quality assessment and reward calculation to generate reward value tables and trajectory weight tables, further supporting the construction of training data for sample selection and self-imitating updates.

[0062] In one embodiment of the visual navigation method based on self-imitation diffusion in this application, the following specific content may also be included: Step S301: Based on the perception data table, perform state representation to generate a feature vector set. Preprocess the perception data table according to the preset encoding rules to generate an encoded data table. Perform feature transformation on the encoded data table according to the preset mapping rules to generate a transformation matrix. Perform state mapping on the encoded data table according to the transformation matrix to generate a feature vector set. Use the feature vector set for subsequent sampling. Step S302: Perform trajectory sampling on the feature vector set to generate a candidate trajectory set; inject noise into the feature vector set according to a preset diffusion rule to generate an initial state table; perform denoising iteration on the initial state table according to the diffusion time sequence template to generate a trajectory sequence table; construct a trajectory generator based on the trajectory sequence table to generate a candidate trajectory set; and use the candidate trajectory set for quality evaluation.

[0063] First, after accessing the perception data table established in step S201, a one-time alignment is performed according to frame time and viewpoint number, filtering out entries with missing timestamps and abnormal camera intrinsic parameters. Based on preset encoding rules, the semantic mask, passability probability grid, and edge gradient of each frame are standardized, and the camera pose is converted into a pose vector under a unified reference coordinate system, forming an encoded data table. For frames with interpolation markers, only the marker bits are recorded without changing the pixel distribution; subsequent state mapping will reduce the weight of this marker to avoid the distribution deviation introduced by interpolation affecting feature stability.

[0064] Feature transformation is performed based on the encoded data table, and mapping sub-channels are constructed for the three types of fields—raster, semantic, and geometric—according to preset mapping rules. Specifically, the occupancy vector of the passability probability raster is obtained by dividing it into blocks, the semantic mask is encoded into a region descriptor according to the category boundary length and region connectivity, and the edge gradients are aggregated to obtain directional histogram entries. After scale normalization and dimension alignment, the three types of entries are assembled into a transformation matrix, and the source field and frame index are recorded in the matrix metadata to ensure that the specific sensing source can be traced back in the future.

[0065] State mapping is performed based on the transformation matrix, and the state characterizer is invoked to push entries from different sources into a unified state coordinate system. Two constraints are introduced during the mapping process: first, an reachability constraint, which prohibits mapping blocks with low occupancy rates to high-confidence states; and second, a continuity constraint, which requires that the state differences between adjacent frames be continuous within the velocity and acceleration ranges. Mapping results satisfying these constraints are output as a feature vector set, corresponding one-to-one with frame timestamps, and used as direct input for diffusion sampling.

[0066] Once the feature vector set is available, the diffusion sampling process is initialized. First, time-related noise is injected into the feature vectors according to a preset diffusion rule to generate an initial state table. The correspondence between noise intensity and time steps is given by a time-series template, which is fixed within the current batch. The injection amplitude is reduced for frames containing interpolation markers to minimize drift amplified by uncertainty. The initial state table retains the noise identifier and constraint mask bit for each time step, providing a basis for subsequent denoising and reachability verification.

[0067] The denoising iteration is initiated based on the initial state table, gradually restoring state increments and accumulating them into trajectory points. Before each denoising step, reachability mask bits are read, and backtracking and direction reestimation are performed on candidate increments falling on impassable blocks. After each denoising step, a coherence check is performed, and local reestimation is triggered for entries with abrupt changes in velocity and curvature. The trajectory points restored at each time step are concatenated in their original time order to obtain a trajectory sequence table, and the number of out-of-bounds backtracking and mask hit rate are recorded in each entry as prior indicators for subsequent quality assessment.

[0068] To clarify the relationship between iterative convergence and constraint satisfaction, an iterative quality metric can optionally be calculated for each trajectory segment, serving as a ranking reference during the evaluation phase. This metric is expressed by a unique formula: Zp = d1·Up + d2·Sp d3·Hp.

[0069] In the formula, Zp is the iterative quality score of the p-th segment; Up is the denoising convergence stability, reflecting the decay trend of the residual in consecutive steps; Sp is the coherence score, derived from the smoothing statistics of velocity and curvature; Hp is the constraint conflict degree, derived from out-of-bounds backoff and masking hit situations; d1, d2, and d3 are non-negative weights, fixed in the current batch. Zp is written into an auxiliary column of the trajectory sequence table for quality assessment reading and does not directly participate in sampling decisions.

[0070] After the trajectory sequence list is generated, a trajectory generator is constructed to perform rigid body registration and stitching on adjacent segments, correcting pose differences caused by cross-viewpoints. If cumulative drift exceeds the limit during stitching, segments with higher iteration quality metrics are retained first, and the other segment is locally resampled within its start and end neighborhoods, then converged again through denoising iterations. Upon completion, a candidate trajectory set is output, with each trajectory carrying three pieces of information: source window, registration residual, and constraint statistics, ensuring that the source of the problem can be located in the subsequent evaluation stage.

[0071] After the candidate trajectory set is determined, it is registered together with the reference relationship between the feature vector set and the trajectory sequence table to form a replayable data bundle. This data bundle is read in the quality evaluation stage. First, constraint statistics and iterative quality metrics are used as priors, and then the relationship between the termination state and the task objective is comprehensively scored to finally generate a reward value table and a trajectory weight table. After completing the above process, the candidate trajectory set enters the evaluation and screening stage in step S102. High-quality trajectories will be replayed according to the training list for self-imitation update.

[0072] In one embodiment of the visual navigation method based on self-imitation diffusion in this application, the following specific content may also be included: Step S401: Based on the candidate trajectory set, construct an evaluator to generate a scoring data table. Decompose the candidate trajectory set according to preset evaluation rules to generate a trajectory feature table. Perform multi-dimensional analysis on the trajectory feature table according to a preset index system to generate an evaluation index set. Calculate the score of the trajectory feature table according to the evaluation index set to generate a scoring data table. Use the scoring data table for reward calculation. Step S402: Calculate reward values ​​for the scoring data table to generate a reward value table; normalize the scoring data table according to preset reward rules to generate a normalized score table; weight the normalized score table according to a preset weight template to generate a combined score table; construct a reward calculator based on the combined score table to generate a reward value table; and use the reward value table for trajectory filtering.

[0073] First, the candidate trajectory set, its source window, registration residual, and constraint statistics output in step S302 are read. Trajectories within the same batch are aligned by time and scene number and used as evaluation input. Based on this input, an evaluator is instantiated. The evaluator decomposes each trajectory into three basic elements at the trajectory level: reachability, geometric coherence, and target fit. Simultaneously, the number of out-of-bounds backtracking and the masking hit rate in the trajectory sequence table are read as prior signals to participate in subsequent calculations, preventing the score from deviating from the generation constraints.

[0074] After the evaluator is initialized, feature decomposition is performed, outputting a trajectory feature table. Specifically, the following steps are taken: the passability percentage is calculated using a passability probability grid, and the number of constraint conflicts is counted; geometric coherence is characterized by velocity, acceleration, and curvature sequences; and target fit is described by the spatial distance from the final pose to the target and the orientation difference. These fields correspond one-to-one with the trajectory index, and anchor point time is retained for cross-frame retrieval of perception entries for verification. The trajectory feature table then proceeds to index construction.

[0075] Based on the trajectory feature table, a multi-dimensional analysis is performed according to a preset indicator system to generate an evaluation indicator set. This analysis first performs interval processing on reachability-related fields to eliminate isolated boundary crossings caused by short-term occlusion; then, window smoothing is introduced for geometric coherence to reduce the impact of single-point anomalies on the overall result; finally, scene difficulty conversion is introduced for target fit to relax the proximity requirements of complex scenes to a reasonable range. The evaluation indicator set retains detailed values ​​and confidence markers for each dimension, providing a traceable basis for scoring calculations.

[0076] Based on the evaluation index set, the evaluator calculates a score for each trajectory and generates a score data table. The scoring follows a structure of parallel weighting and deduction of primary items: accessibility and geometric coherence are accumulated as primary items, while constraint conflicts and large curvature abrupt changes are deducted; when prior signals indicate frequent regressions during the generation stage, the weight of that item is further reduced. The scoring results, along with dimensional contribution and prior reading records, are written back to the score data table as direct input for subsequent reward value calculation.

[0077] Once the scoring data table is available, the reward value calculation process is initiated. First, the scores for each dimension are normalized batch by batch to eliminate scale differences between scenarios, resulting in a normalized score table. Then, weighted combinations are performed according to a preset weight template to form a combined score table. The weight template is fixed within the current batch and distinguishes the relative importance of primary and secondary items, while allowing for downweighting of dimensions with insufficient confidence markers, making the reward more closely reflect the quality of available information.

[0078] To prevent high scores in one dimension from masking other deficiencies, an equilibrium term described by the formula can be optionally introduced during the combination stage: Wm = e1·Hc + e2·Gs e3·Pc.

[0079] In the formula, Wm is the intermediate balance score; Hc is the reachability summary score, derived from the proportion and conflict-free rate; Gs is the geometric coherence score, derived from velocity and curvature smoothing statistics; Pc is the penalty score, derived from constraint conflict and high-frequency backoff; e1, e2, and e3 are non-negative weights, fixed in this batch. Wm is written into an auxiliary column of the combined score table, used in the reward calculator to jointly determine the reward order with the main combined score.

[0080] After the combined score table is generated, a reward calculator is constructed. It comprehensively reads the main combined score and the Wm, performs robustness processing on extreme entries, and outputs a reward value table. The robustness processing includes density valley truncation and anomaly deduction recovery to prevent the threshold from being influenced by tail samples. The reward value table provides a reward value, the number of participating dimensions, and key deduction sources for each trajectory, facilitating the interpretation of boundary samples in subsequent screening stages.

[0081] Based on the reward value table, a threshold analysis is performed before trajectory screening. The analysis process involves peak and valley segmentation of the reward distribution to determine candidate threshold intervals. Combining the dimensional confidence markers of the entries with prior conflict statistics, a suggestion is given regarding whether to include the entry into the sample pool. This suggestion, along with the reward value, is written into the trajectory screening input, providing a decision-making basis for the threshold screening and importance sampling in step S102.

[0082] Finally, the scoring data table, normalized score table, and reward value table are linked by trajectory index and delivered to the subsequent screening module. Based on this, the screening module generates trajectory weights and a high-quality trajectory set. The relevant results are read during self-imitation updates to construct training data and set sample weights, achieving a seamless transition from evaluation quantification to policy updates.

[0083] In one embodiment of the visual navigation method based on self-imitation diffusion in this application, the following specific content may also be included: Step S501: Based on the reward value table, construct a filter to generate a trajectory weight table, perform distribution analysis on the reward value table according to a preset threshold rule to generate a distribution feature table, divide the distribution feature table into intervals according to a preset filtering template to generate an interval mapping table, calculate the importance of the reward value table according to the interval mapping table to generate a trajectory weight table, and use the trajectory weight table for sample selection. Step S502: Select samples from the trajectory weight table to generate a high-quality trajectory set; sort the trajectory weight table according to preset selection rules to generate a ranking index table; sample the candidate trajectory set according to the ranking index table to generate a sampling sequence table; construct a selector based on the sampling sequence table to generate a high-quality trajectory set; and use the high-quality trajectory set for self-imitation update.

[0084] First, the reward value table output in step S402 is read and aligned with the trajectory index and scene number, serving as input to the filter. Based on this input, the morphological characteristics of the reward distribution are statistically analyzed according to preset threshold rules, specifically including three quantitative indicators: peak position, tail decay slope, and density valley position. To reduce the perturbation of small-sample scenes, the scene number is used as a stratification condition to calculate the above indicators separately, and then weighted and merged at the batch level to obtain a distribution feature table. This table also retains the sample size and confidence marker for each stratum, providing constraints for subsequent interval division.

[0085] Based on the distribution feature table, intervals are divided according to a preset screening template, and an interval mapping table is output. Interval division starts from the density valley value, dividing the reward axis into three segments: a high region, a critical region, and a low region. Simultaneously, scene scale constraints are introduced to prevent extreme threshold shifts due to insufficient samples in a certain stratum. A buffer zone is set for the critical region; trajectories within the buffer zone will have their decisions delayed, determined by importance calculation combined with stability priors. The interval mapping table uses the trajectory index as the key, registering the interval label and stratification source corresponding to each trajectory.

[0086] Based on the interval mapping table, importance calculation is performed on the reward value table to generate a trajectory weight table. Importance consists of three parts: first, interval base weights, assigning higher base weights to higher intervals and lower base weights to lower intervals; second, stability correction, reading the number of out-of-bounds backtrackings and the masking hit rate recorded during the candidate trajectory generation stage, and reducing the weight of samples with high backtracking rates; and third, diversity adjustment, statistically analyzing the similarity of trajectories near the same scene and the same target, performing intra-group suppression on highly similar entries, retaining only representative trajectories with full weights, and marking the remaining entries as spares with small weights. The processed trajectory weight table is output, retaining detailed weight composition for easy verification.

[0087] Once the trajectory weight table is available, the sorting stage of the sample selection process begins. First, trajectories are globally sorted according to their importance weights, while stratified quotas are implemented at the scene level to ensure that trajectories are selected for each scene. Then, a sorting index table is constructed, which, in addition to recording the sorting position, also registers the trajectory's interval label, stability correction amount, and diversity group number, used to control the intra-group interval during sampling. The sorting index table serves as the direct input for importance sampling.

[0088] Importance sampling is performed based on the ranking index table, and a sampling sequence list is output. Sampling employs a segmented annealing strategy, prioritizing sampling from high-performing regions at group intervals, interspersed with a small number of critical region samples to maintain boundary information, and then extracting a very small number of redundant samples from low-performing regions as negative examples, depending on the sample size. Intra-group similarity is checked during each sampling to prevent the continuous selection of similar trajectories; entries with excessively large stability corrections are only used as fillers when samples are insufficient. After sampling, a sequence with a relatively balanced coverage of scenes, targets, and action patterns is obtained.

[0089] A selector is instantiated based on the sampling sequence list. The candidate trajectory set and its metadata are called, and the sequence is replayed in order to verify whether the reachability statistics are consistent with those during evaluation. For entries found to be inconsistent, resampling is triggered, and the reason and location are recorded in the sampling log to avoid repeating errors. Entries that pass verification are written into a high-quality trajectory set, along with their anchor time, environmental constraint markers, and source index, ensuring that subsequent training can be traced back to these points.

[0090] To improve the usability of subsequent self-imitating updates, a training-side summary can be optionally generated for the high-quality trajectory set, annotating each trajectory with usage suggestions, such as whether it is suitable for long-term window training or whether smoothing regularization needs to be emphasized in the denoising loss. This summary does not change the sample set, but only serves as a reference for training scheduling, and establishes a reference relationship with the trajectory weight table.

[0091] Finally, the high-quality trajectory set, ranking index table, and sampling logs are output to the self-imitation update entry. Subsequent steps read this set to construct learning samples and denoising training batches, read the ranking index table to determine sample weights and appearance frequencies; at the same time, the trajectory weight table and interval mapping table are retained as the basis for threshold adaptation, used for fine-tuning and backtracking verification of the selection rules for the next batch.

[0092] In one embodiment of the visual navigation method based on self-imitation diffusion in this application, the following specific content may also be included: Step S601: Based on the high-quality trajectory set, a trainer is constructed to generate a training dataset. The high-quality trajectory set is augmented according to a preset sampling rule to generate an augmented data table. The augmented data table is labeled according to a preset labeling rule to generate a labeled data table. The labeled data table is organized according to a preset training template to generate a training dataset. The training dataset is used for parameter updates. Step S602: Perform parameter learning on the training dataset to generate a model parameter table, divide the training dataset into batches according to a preset learning rule to generate a training batch table, perform gradient calculation on the training batch table according to a preset loss function to generate a gradient vector set, construct a parameter optimizer based on the gradient vector set to generate a model parameter table, and use the model parameter table for policy update.

[0093] First, the high-quality trajectory set output in step S502, along with its anchor point time, environmental constraint markers, and source index, is read, and a training list is created according to scene number and time order. Based on preset sampling rules, each trajectory is windowed and temporally shuffled to generate a segment set. For segments containing interpolation or out-of-bounds backoff markers, their sampling frequency is reduced, and a weighting factor is recorded as a reference for subsequent loss weighting. The above processing yields the original segment queue for enhancement.

[0094] Data augmentation is performed on the original fragment queue. Specifically, amplitude perturbation and temporal warp are applied to the action increment, and a small affine perturbation is performed on the pose trajectory, ensuring that the perturbed position remains within the passable grid. Color jitter and slight blurring are applied to the context-aware features, while the semantic mask remains unchanged to ensure consistent target boundaries. The augmented samples, along with the original fragment indexes, are written back to form an augmented data table, recording the perturbation type and amplitude range for subsequent annotation and training phases.

[0095] Feature annotation is performed on the enhanced data table, generating three types of labels for each segment according to preset annotation rules: first, a trajectory smoothness label, derived from velocity and curvature statistics; second, a reachability label, derived from traversal occupancy rate and constraint conflict count; and third, a target fit label, derived from the spatial relationship between the termination pose and the target. These labels are aligned with the action sequence at the segment level, with confidence markers added to outlier fields. After annotation, an annotation data table is output, maintaining a one-to-one mapping between segments and labels.

[0096] Data is organized according to the labeled data table, and the fragment context, action sequence, and three types of labels are packaged into training samples. Cross-scene balanced sampling and difficulty stratification are performed according to the training template. The samples are arranged into fixed-length temporal blocks, and the scene coverage and label confidence statistics within each batch are recorded to form the training dataset. The training dataset provides a fast retrieval interface by fragment index, supporting playback and error localization.

[0097] Once the training dataset is available, batches are generated according to preset learning rules to create a training batch table. During the batching process, the overlap ratio between adjacent batches is limited to avoid overlearning of a single scene in a short period. Samples with insufficient confidence labels are distributed to later batches to reduce early gradient oscillations. Subsequently, the parameter learning phase begins, where a parameter optimizer is established to calculate the loss for each batch and output a gradient vector set.

[0098] To standardize the loss calculation, this embodiment employs a unique weighting structure on the training side, incorporating both fragment-level error and constraint penalties in parallel. The optional loss expression is as follows: Jt = f1·Er + f2·Es f3·Ec.

[0099] In the formula, Jt is the total loss of the current batch; Er is the action reconstruction error term, which statistically represents the action residual after denoising; Es is the trajectory smoothing regularization term, which constrains the continuity of velocity and curvature; Ec is the reachability penalty term, which addresses constraint conflicts and out-of-bounds backoff; f1, f2, and f3 are non-negative weights, which are fixed in the current training round. Jt is used for backpropagation to generate a gradient vector set, and the weight values ​​and corresponding batch numbers are recorded in the training log.

[0100] After the gradient vector set is generated, the parameter optimizer performs multi-step updates and records the learning rate plan and momentum state, outputting a model parameter table. The model parameter table includes the weights of the denoising network, temporal noise mapping, and regularization coefficient version numbers. It is compared with the interface of previous strategies through a consistency check to ensure structural compatibility. If an incompatible layer is found, a local reset is triggered and rollback information is registered to prevent unpredictable behavior during the inference phase.

[0101] After completing the parameter replacement on the policy side according to the model parameter table, fast replay verification is initiated. Verification uses independent hold-out segments from a high-quality trajectory set, performs forward propagation with the same temporal length as training, and checks whether the reconstruction error and reachability metrics are within the set boundaries. If verification passes, the model parameter table is written into the policy update pipeline, deployable policy parameters are generated, and the source training dataset and weight version are bound in the metadata.

[0102] Finally, the updated strategy is applied to the visual navigation task. Version number, batch number, and scene coverage statistics are recorded during deployment to facilitate incremental learning and backtracking in new environments. Simultaneously, the organizational structure of the training dataset and augmentation data table is retained as a training template and directly reused in the next round of self-imitation updates to maintain closed-loop consistency from sample collection and label generation to parameter learning.

[0103] In one embodiment of the visual navigation method based on self-imitation diffusion in this application, the following specific content may also be included: Step S701: Construct a denoiser based on the training dataset to generate a diffusion parameter table. Inject noise into the training dataset according to preset noise rules to generate a noise data table. Configure the sampling step size of the noise data table according to preset time rules to generate a time series parameter table. Perform denoising iteration on the noise data table according to the time series parameter table to generate an iteration result table. Construct a parameter extractor based on the iteration result table to generate a diffusion parameter table. Use the diffusion parameter table for policy update. Step S702: Optimize the diffusion parameter table to generate an optimization strategy. Map the parameters of the diffusion parameter table according to a preset update rule to generate a strategy parameter table. Reconstruct the navigation strategy according to the strategy parameter table to generate an optimization model. Construct a navigation controller based on the optimization model to generate an optimization strategy. Deploy the optimization strategy to the visual navigation system.

[0104] First, the training dataset and batch partitioning information output from step S602 are read. The context and action sequence are restored according to the segment index, serving as the input source for the denoiser. Based on this, the denoiser is instantiated, taking the state context and the disturbed action as input and the action residual as output. To initiate the diffusion process, temporal-related noise is injected into the training dataset according to preset noise rules. Short-window and long-window intensity curves are distinguished, and the injection amplitude is reduced for samples marked with insufficient confidence. A noise data table is generated, retaining the mapping between noise labels and sample indices for subsequent traceability.

[0105] Based on the noise data table, temporal parameters are configured to map the denoising steps and step size of each sample to a fixed template according to preset temporal rules, outputting a temporal parameter table. During mapping, the scene complexity label is read, and a finer later step size is allocated for complex scenes, while a uniform step size is maintained for simple scenes, ensuring that the denoising iteration can handle detail jitter at the end of convergence. The temporal parameter table and the noise data table correspond one-to-one, serving as the scheduling basis for the iteration process.

[0106] Denoising iterations are performed based on the stated time-series parameter table. For each step, the reachability mask bit is read first to prevent the action increment from being decoded into an impassable grid. Then, residual regression is performed and the result is accumulated into the trajectory increment. A coherence check is added between adjacent steps, and local reestimation is triggered for entries with abrupt changes in velocity and curvature. After the iteration is completed, an iteration result table is output, recording the residual trajectory, convergence curve, and constraint hit statistics for each sample, serving as the original basis for parameter extraction.

[0107] A parameter extractor is built on the iteration result table to summarize the denoiser weight snapshot, noise intensity mapping, and step configuration to form a diffusion parameter table. During extraction, the gradient contributions for different scenarios are normalized to avoid a single scenario dominating parameter updates; samples with abnormal convergence curves are removed from their corresponding statistical entries to ensure parameter stability. The diffusion parameter table exposes its version number and the dependent time-series template number for policy-side reading.

[0108] The strategy side is updated according to the diffusion parameter table, the parameters are loaded into the trajectory generation branch, and consistency verification is performed. The verification includes structural compatibility, accessibility masking effectiveness, and basic coherence boundary checks. After the verification passes, the data is quickly replayed on the validation set to confirm that the trajectory coherence and constraint compliance meet the preset boundaries under a fixed number of sampling steps. Then, the diffusion parameter table is submitted to the strategy optimization stage.

[0109] In the strategy optimization phase, the diffusion parameter table is mapped to generate a strategy parameter table. During mapping, the stage weights of the denoiser are mapped to the trajectory decoding layer and step size scheduler in the strategy, and the noise intensity curve is transcribed into the sampling temperature configuration on the inference side. After mapping, the navigation strategy is restructured accordingly, including updating the decoding layer channel layout and introducing constraint-aware reachability gating, forming an optimized model. The optimized model undergoes a closed-loop test in a sandbox environment to verify its stability in conjunction with the timing template.

[0110] A navigation controller is built based on the optimized model. This controller is responsible for linking perceived features and policy parameters into actual control commands. Specifically, it reads state features using a fixed-length sliding window, generates action sequences through the optimized model, and then applies reachability gating and a speed smoother at the control layer to output navigation commands for execution. The controller retains a playback interface to facilitate pinpointing online issues to specific timing windows and parameter versions.

[0111] Finally, the optimization strategy is deployed to the visual navigation system. Before deployment, the strategy version, diffusion parameter table number, and timing template number are recorded, and the effective time and rollback strategy are registered on the platform side. After deployment, runtime logs are continuously collected, including statistics on the coherence of generated trajectories and constraint hit statistics, which are used for threshold revision and template fine-tuning during the next round of self-imitation updates. The above logs and diffusion parameter tables form a corresponding relationship, supporting reproducible experiments and incremental iterations in different environments.

[0112] To effectively address the shortcomings of traditional technologies in trajectory generation, quality assessment, and strategy optimization, and to provide technical support for visual navigation, this application provides an embodiment of a self-imitating diffusion-based visual navigation device for implementing all or part of the aforementioned self-imitating diffusion-based visual navigation method. See [link to embodiment]. Figure 2 The visual navigation device based on self-imitation diffusion specifically includes the following: The trajectory sampling module 10 is used to generate a candidate trajectory set in the environment based on the navigation strategy, extract features from the environmental visual information according to the preset perception rules to generate a perception data table, perform state encoding on the perception data table to generate a state feature set, perform trajectory sampling on the state feature set according to the preset diffusion rules to generate a candidate trajectory set, and use the candidate trajectory set for subsequent evaluation. The trajectory filtering module 20 is used to perform quality evaluation on the candidate trajectory set to generate a high-quality trajectory set, calculate the reward of the candidate trajectory set according to the preset evaluation rules to generate a reward value table, perform threshold filtering on the reward value table to generate a trajectory weight table, select samples from the candidate trajectory set according to the trajectory weight table to generate a high-quality trajectory set, and use the high-quality trajectory set for strategy update. The visual navigation module 30 is used to perform self-imitation update of the navigation strategy based on the high-quality trajectory set to generate an optimized strategy, construct learning samples from the high-quality trajectory set according to preset training rules to generate a training dataset, perform denoising training on the training dataset to generate a diffusion parameter table, update the navigation strategy parameters according to the diffusion parameter table to generate an optimized strategy, and apply the optimized strategy to the visual navigation task.

[0113] As described above, the self-imitating diffusion-based visual navigation device provided in this application can achieve effective environmental perception through feature extraction and state encoding. An evaluation mechanism is constructed, combining reward calculation and weight selection to establish a reliable sample selection strategy. Strategy optimization is introduced, ensuring continuous improvement of navigation through self-imitating learning and parameter updates. This method effectively addresses the shortcomings of traditional technologies in trajectory generation, quality assessment, and strategy optimization, providing technical support for visual navigation.

[0114] To further illustrate this solution, this application also provides a specific application example of implementing self-imitating diffusion-based visual navigation using the above-mentioned self-imitating diffusion-based visual navigation, which specifically includes the following: The method includes: 1. Generate multiple candidate trajectories in the environment using the current navigation strategy; 2. Evaluate the quality of the candidate trajectories based on environmental feedback to obtain corresponding trajectory rewards; 3. Based on the trajectory rewards, select or weight the candidate trajectories to obtain a high-quality trajectory set; 4. Use the high-quality trajectory as a self-imitation learning sample to update the navigation strategy; •in: 1. The navigation strategy is a diffusion strategy; 2. The update process uses reward-guided self-imitation learning to make the trajectory distribution of the policy output concentrated towards a high-quality trajectory distribution.

[0115] Optionally, in this embodiment, the self-imitation learning samples are generated autonomously by the current navigation strategy and do not rely on pre-collected expert example trajectory data.

[0116] The trajectory reward is used to sort candidate trajectories; only trajectories with reward values ​​higher than a preset threshold are selected for policy updates.

[0117] The candidate trajectories are assigned importance weights based on the trajectory rewards; the policy training loss is then weighted using these importance weights.

[0118] The strategy update uses the complete trajectory as the learning object; it does not perform progressive reinforcement updates on single-step actions.

[0119] During the training of the diffusion strategy: the denoising loss function is weighted based on trajectory rewards, so that the diffusion model focuses on fitting high-quality trajectory samples.

[0120] By exponentializing or normalizing the trajectory rewards, the policy update magnitude is limited to prevent excessive deviation in trajectory distribution.

[0121] Specifically, in at least part of the training phase: target location information is not provided to the navigation strategy; target-independent candidate trajectories are generated; the target-independent trajectories are used to enhance the diversity of self-imitating samples.

[0122] Both target-independent trajectories and target-oriented trajectories are used for policy updates to improve the robustness of the policy in visual navigation tasks.

[0123] The difficulty of the training scenario is assessed based on the reward statistics of the candidate trajectories; self-imitation updates are only performed on training scenarios that meet the preset reward conditions.

[0124] Specifically, in the inference phase: no additional trajectory scoring network or rule filtering module is used; the navigation trajectory is directly output by the diffusion strategy.

[0125] Self-imitation learning causes the trajectory distribution output by the navigation strategy to concentrate on high-quality trajectories; the inference phase reduces the number of diffusion samplings to generate navigation trajectories, thus improving inference speed.

[0126] In addition, a visual navigation system may be included, comprising: a perception module for acquiring visual information; a trajectory generation module for generating candidate trajectories based on a diffusion strategy; a reward evaluation module for calculating rewards for candidate trajectories; and a self-imitation learning module for updating the navigation strategy based on high-quality trajectories.

[0127] From a hardware perspective, in order to effectively address the shortcomings of traditional technologies in trajectory generation, quality assessment, and strategy optimization, and to provide technical support for visual navigation, this application provides an embodiment of an electronic device for implementing all or part of the self-imitating diffusion-based visual navigation method. The electronic device specifically includes the following components: The system comprises a processor, memory, a communications interface, and a bus; wherein the processor, memory, and communications interface communicate with each other via the bus; the communications interface is used to realize information transmission between the self-imitating diffusion-based visual navigation device and core business systems, user terminals, and related databases and other related devices; the logic controller can be a desktop computer, tablet computer, or mobile terminal, etc., and this embodiment is not limited to these. In this embodiment, the logic controller can be implemented with reference to the embodiments of the self-imitating diffusion-based visual navigation method and the embodiments of the self-imitating diffusion-based visual navigation device, the contents of which are incorporated herein, and repeated details will not be described again.

[0128] It is understood that the user terminal may include smartphones, tablet computers, network set-top boxes, portable computers, desktop computers, personal digital assistants (PDAs), in-vehicle devices, smart wearable devices, etc. Among these, the smart wearable devices may include smart glasses, smartwatches, smart bracelets, etc.

[0129] In practical applications, parts of the self-imitating diffusion-based visual navigation method can be executed on the electronic device side as described above, or all operations can be completed in the client device. The choice can be made based on the processing power of the client device and the limitations of the user's usage scenario. This application does not impose any limitations on this. If all operations are completed in the client device, the client device may further include a processor.

[0130] The aforementioned client device may have a communication module (i.e., a communication unit) that can communicate with a remote server to achieve data transmission with the server. The server may include a server on the task scheduling center side; in other implementation scenarios, it may also include a server on an intermediate platform, such as a server on a third-party server platform that has a communication link with the task scheduling center server. The server may include a single computer device, a server cluster consisting of multiple servers, or a distributed server structure.

[0131] Figure 3 This is a schematic block diagram illustrating the system configuration of the electronic device 9600 according to an embodiment of this application. Figure 3 As shown, the electronic device 9600 may include a central processing unit 9100 and a memory 9140; the memory 9140 is coupled to the central processing unit 9100. It is worth noting that... Figure 3 This is an example; other types of structures can also be used to supplement or replace this structure to achieve telecommunications functions or other functions.

[0132] In one embodiment, the visual navigation method based on self-imitating diffusion can be integrated into the central processing unit 9100. The central processing unit 9100 can be configured to perform the following control: Step S101: Generate a candidate trajectory set in the environment based on the navigation strategy, extract features from the environmental visual information according to the preset perception rules to generate a perception data table, encode the perception data table to generate a state feature set, sample the state feature set according to the preset diffusion rules to generate a candidate trajectory set, and use the candidate trajectory set for subsequent evaluation. Step S102: Perform quality assessment on the candidate trajectory set to generate a high-quality trajectory set; calculate the reward of the candidate trajectory set according to a preset assessment rule to generate a reward value table; perform threshold screening on the reward value table to generate a trajectory weight table; select samples from the candidate trajectory set according to the trajectory weight table to generate a high-quality trajectory set; and use the high-quality trajectory set for strategy update. Step S103: Based on the high-quality trajectory set, the navigation strategy is updated by self-imitation to generate an optimized strategy. The high-quality trajectory set is used to construct learning samples according to preset training rules to generate a training dataset. The training dataset is denoised and trained to generate a diffusion parameter table. The navigation strategy is updated according to the diffusion parameter table to generate an optimized strategy. The optimized strategy is applied to the visual navigation task.

[0133] As described above, the electronic device provided in this application embodiment achieves effective environmental perception through feature extraction and state encoding. An evaluation mechanism is constructed, combining reward calculation and weighted selection to establish a reliable sample selection strategy. Strategy optimization is introduced, ensuring continuous improvement of navigation through self-imitation learning and parameter updates. This method effectively addresses the shortcomings of traditional technologies in trajectory generation, quality assessment, and strategy optimization, providing technical support for visual navigation.

[0134] In another embodiment, the visual navigation device based on self-imitating diffusion can be configured separately from the central processing unit 9100. For example, the visual navigation device based on self-imitating diffusion can be configured as a chip connected to the central processing unit 9100, and the functions of the visual navigation method based on self-imitating diffusion can be realized through the control of the central processing unit.

[0135] like Figure 3 As shown, the electronic device 9600 may further include: a communication module 9110, an input unit 9120, an audio processor 9130, a display 9160, and a power supply 9170. It is worth noting that the electronic device 9600 does not necessarily need to include these components. Figure 3 All components shown; in addition, the electronic device 9600 may also include Figure 3 For components not shown, please refer to existing technologies.

[0136] like Figure 3 As shown, the central processing unit 9100, sometimes also referred to as a controller or operating control, may include a microprocessor or other processor device and / or logic device, which receives inputs and controls the operation of various components of the electronic device 9600.

[0137] The memory 9140 may be, for example, one or more of a cache, flash memory, hard drive, removable media, volatile memory, non-volatile memory, or other suitable devices. It may store the aforementioned failure-related information, and also store a program for executing that information. The central processing unit 9100 may execute the program stored in the memory 9140 to perform information storage or processing, etc.

[0138] Input unit 9120 provides input to central processing unit 9100. Input unit 9120 may be, for example, a keypad or touch input device. Power supply 9170 provides power to electronic device 9600. Display 9160 displays images and text. Display may be, for example, an LCD display, but is not limited thereto.

[0139] The memory 9140 can be a solid-state memory, such as a read-only memory (ROM), random access memory (RAM), a SIM card, etc. It can also be a memory that retains information even when power is off, can be selectively erased, and contains more data; examples of this type of memory are sometimes referred to as EPROMs. The memory 9140 can also be some other type of device. The memory 9140 includes a buffer memory 9141 (sometimes referred to as a buffer). The memory 9140 may include an application / function storage unit 9142 for storing application programs and function programs or processes for executing the operation of the electronic device 9600 via the central processing unit 9100.

[0140] The memory 9140 may also include a data storage unit 9143 for storing data, such as contacts, digital data, pictures, sounds, and / or any other data used by the electronic device. The driver storage unit 9144 of the memory 9140 may include various drivers for the electronic device for communication functions and / or for performing other functions of the electronic device (such as messaging applications, address book applications, etc.).

[0141] The communication module 9110 is a transmitter / receiver that sends and receives signals via the antenna 9111. The communication module 9110 (transmitter / receiver) is coupled to the central processing unit 9100 to provide input signals and receive output signals, which is the same as in a conventional mobile communication terminal.

[0142] Based on different communication technologies, multiple communication modules 9110 can be configured in the same electronic device, such as cellular network modules, Bluetooth modules, and / or wireless LAN modules. The communication module 9110 (transmitter / receiver) is also coupled to a speaker 9131 and a microphone 9132 via an audio processor 9130 to provide audio output via the speaker 9131 and receive audio input from the microphone 9132, thereby realizing typical telecommunications functions. The audio processor 9130 may include any suitable buffer, decoder, amplifier, etc. Additionally, the audio processor 9130 is coupled to a central processing unit 9100, enabling on-device recording via the microphone 9132 and on-device playback of stored audio via the speaker 9131.

[0143] Embodiments of this application also provide a computer-readable storage medium capable of implementing all steps of the self-imitating diffusion-based visual navigation method with a server or client as the execution subject in the above embodiments. The computer-readable storage medium stores a computer program that, when executed by a processor, implements all steps of the self-imitating diffusion-based visual navigation method with a server or client as the execution subject in the above embodiments. For example, when the processor executes the computer program, it implements the following steps: Step S101: Generate a candidate trajectory set in the environment based on the navigation strategy, extract features from the environmental visual information according to the preset perception rules to generate a perception data table, encode the perception data table to generate a state feature set, sample the state feature set according to the preset diffusion rules to generate a candidate trajectory set, and use the candidate trajectory set for subsequent evaluation. Step S102: Perform quality assessment on the candidate trajectory set to generate a high-quality trajectory set; calculate the reward of the candidate trajectory set according to a preset assessment rule to generate a reward value table; perform threshold screening on the reward value table to generate a trajectory weight table; select samples from the candidate trajectory set according to the trajectory weight table to generate a high-quality trajectory set; and use the high-quality trajectory set for strategy update. Step S103: Based on the high-quality trajectory set, the navigation strategy is updated by self-imitation to generate an optimized strategy. The high-quality trajectory set is used to construct learning samples according to preset training rules to generate a training dataset. The training dataset is denoised and trained to generate a diffusion parameter table. The navigation strategy is updated according to the diffusion parameter table to generate an optimized strategy. The optimized strategy is applied to the visual navigation task.

[0144] As described above, the computer-readable storage medium provided in this application embodiment achieves effective environmental perception through feature extraction and state encoding. An evaluation mechanism is constructed, combining reward calculation and weighted selection to establish a reliable sample selection strategy. Strategy optimization is introduced, ensuring continuous improvement of navigation through self-imitation learning and parameter updates. This method effectively addresses the shortcomings of traditional technologies in trajectory generation, quality assessment, and strategy optimization, providing technical support for visual navigation.

[0145] Embodiments of this application also provide a computer program product capable of implementing all steps of the self-imitating diffusion-based visual navigation method in the above embodiments, where the execution subject is a server or a client. When executed by a processor, this computer program / instruction implements the steps of the self-imitating diffusion-based visual navigation method. For example, the computer program / instruction implements the following steps: Step S101: Generate a candidate trajectory set in the environment based on the navigation strategy, extract features from the environmental visual information according to the preset perception rules to generate a perception data table, encode the perception data table to generate a state feature set, sample the state feature set according to the preset diffusion rules to generate a candidate trajectory set, and use the candidate trajectory set for subsequent evaluation. Step S102: Perform quality assessment on the candidate trajectory set to generate a high-quality trajectory set; calculate the reward of the candidate trajectory set according to a preset assessment rule to generate a reward value table; perform threshold screening on the reward value table to generate a trajectory weight table; select samples from the candidate trajectory set according to the trajectory weight table to generate a high-quality trajectory set; and use the high-quality trajectory set for strategy update. Step S103: Based on the high-quality trajectory set, the navigation strategy is updated by self-imitation to generate an optimized strategy. The high-quality trajectory set is used to construct learning samples according to preset training rules to generate a training dataset. The training dataset is denoised and trained to generate a diffusion parameter table. The navigation strategy is updated according to the diffusion parameter table to generate an optimized strategy. The optimized strategy is applied to the visual navigation task.

[0146] As described above, the computer program product provided in this application achieves effective environmental perception through feature extraction and state encoding. An evaluation mechanism is constructed, combining reward calculation and weighted selection to establish a reliable sample selection strategy. Strategy optimization is introduced, ensuring continuous improvement of navigation through self-imitation learning and parameter updates. This method effectively addresses the shortcomings of traditional technologies in trajectory generation, quality assessment, and strategy optimization, providing technical support for visual navigation.

[0147] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0148] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (devices), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0149] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0150] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0151] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A visual navigation method based on self-imitation diffusion, characterized in that, The method includes: Based on the navigation strategy, a candidate trajectory set is generated in the environment. The environmental visual information is extracted according to the preset perception rules to generate a perception data table. The perception data table is state encoded to generate a state feature set. The state feature set is sampled according to the preset diffusion rules to generate a candidate trajectory set. The candidate trajectory set is used for subsequent evaluation. The candidate trajectory set is quality evaluated to generate a high-quality trajectory set. The candidate trajectory set is then used to calculate rewards according to preset evaluation rules to generate a reward value table. The reward value table is then filtered by threshold to generate a trajectory weight table. Samples are selected from the candidate trajectory set according to the trajectory weight table to generate a high-quality trajectory set. The high-quality trajectory set is then used for strategy updates. Based on the high-quality trajectory set, the navigation strategy is updated by self-imitation to generate an optimized strategy. The high-quality trajectory set is used to construct learning samples according to preset training rules to generate a training dataset. The training dataset is denoised and trained to generate a diffusion parameter table. The navigation strategy is updated according to the diffusion parameter table to generate an optimized strategy. The optimized strategy is then applied to the visual navigation task.

2. The visual navigation method based on self-imitation diffusion according to claim 1, characterized in that, The process of generating a candidate trajectory set in the environment based on a navigation strategy, and extracting features from environmental visual information according to preset perception rules to generate a perception data table, includes: Based on the visual information collected from the navigation environment, an original dataset is generated. The environmental images are acquired according to a preset sampling rule to generate an image data table. The image data table is preprocessed to generate a feature map. The feature map is analyzed at multiple scales according to a preset feature extraction rule to generate a perceptual feature set. The perceptual feature set is used for subsequent encoding. The sensory feature set is used to generate a candidate trajectory set. The sensory feature set is then mapped to a state vector table according to a preset encoding rule. A trajectory generator is constructed according to a preset diffusion sampling rule to generate a sampling parameter table. Based on the sampling parameter table, trajectory sampling is performed on the state vector table to generate a candidate trajectory set. The candidate trajectory set is then used for subsequent evaluation.

3. The visual navigation method based on self-imitation diffusion according to claim 1, characterized in that, The process of generating a state feature set by state encoding the perception data table, generating a candidate trajectory set by trajectory sampling of the state feature set according to a preset diffusion rule, and using the candidate trajectory set for subsequent evaluation includes: A feature vector set is generated based on the state representation of the perception data table. The perception data table is preprocessed according to the preset encoding rules to generate an encoded data table. The encoded data table is transformed according to the preset mapping rules to generate a transformation matrix. The encoded data table is then mapped according to the transformation matrix to generate a feature vector set. The feature vector set is used for subsequent sampling. Trajectory sampling is performed on the feature vector set to generate a candidate trajectory set. Noise is injected into the feature vector set according to a preset diffusion rule to generate an initial state table. The initial state table is denoised iteratively according to the diffusion time sequence template to generate a trajectory sequence table. A trajectory generator is constructed based on the trajectory sequence table to generate a candidate trajectory set. The candidate trajectory set is used for quality evaluation.

4. The visual navigation method based on self-imitation diffusion according to claim 1, characterized in that, The step of performing quality evaluation on the candidate trajectory set to generate a high-quality trajectory set, and calculating rewards for the candidate trajectory set according to preset evaluation rules to generate a reward value table, includes: An evaluator is constructed based on a candidate trajectory set to generate a scoring data table. The candidate trajectory set is decomposed into a trajectory feature table according to a preset evaluation rule. The trajectory feature table is analyzed in multiple dimensions according to a preset index system to generate an evaluation index set. The trajectory feature table is scored and calculated based on the evaluation index set to generate a scoring data table. The scoring data table is used for reward calculation. The scoring data table is used to calculate reward values ​​to generate a reward value table. The scoring data table is then normalized according to a preset reward rule to generate a normalized score table. The normalized score table is then weighted and combined according to a preset weight template to generate a combined score table. A reward calculator is then constructed based on the combined score table to generate a reward value table. The reward value table is then used for trajectory filtering.

5. The visual navigation method based on self-imitation diffusion according to claim 1, characterized in that, The step of performing threshold filtering on the reward value table to generate a trajectory weight table, selecting samples from the candidate trajectory set according to the trajectory weight table to generate a high-quality trajectory set, and using the high-quality trajectory set for policy update includes: A filter is constructed based on the reward value table to generate a trajectory weight table. The reward value table is then analyzed according to a preset threshold rule to generate a distribution feature table. The distribution feature table is then divided into intervals according to a preset filtering template to generate an interval mapping table. The importance of the reward value table is calculated based on the interval mapping table to generate a trajectory weight table. The trajectory weight table is then used for sample selection. A high-quality trajectory set is generated by selecting samples from the trajectory weight table. The trajectory weight table is then sorted by priority according to a preset selection rule to generate a ranking index table. The candidate trajectory set is sampled based on the ranking index table to generate a sampling sequence table. A selector is constructed based on the sampling sequence table to generate a high-quality trajectory set. The high-quality trajectory set is then used for self-imitation update.

6. The visual navigation method based on self-imitation diffusion according to claim 1, characterized in that, The step of generating an optimized navigation strategy by self-imitation based on the high-quality trajectory set, and constructing a training dataset from the high-quality trajectory set according to preset training rules, includes: A training dataset is generated by constructing a trainer based on a high-quality trajectory set. The high-quality trajectory set is augmented according to a preset sampling rule to generate an augmented data table. The augmented data table is labeled with features according to a preset labeling rule to generate a labeled data table. The labeled data table is organized according to a preset training template to generate a training dataset. The training dataset is used for parameter updates. The training dataset is used to learn parameters to generate a model parameter table. The training dataset is divided into batches according to a preset learning rule to generate a training batch table. Gradients are calculated on the training batch table according to a preset loss function to generate a gradient vector set. A parameter optimizer is constructed based on the gradient vector set to generate a model parameter table. The model parameter table is used for policy updates.

7. The visual navigation method based on self-imitation diffusion according to claim 1, characterized in that, The process of denoising and training the training dataset to generate a diffusion parameter table, updating the navigation strategy parameters based on the diffusion parameter table to generate an optimized strategy, and applying the optimized strategy to the visual navigation task includes: A denoiser is constructed based on the training dataset to generate a diffusion parameter table. The training dataset is injected with noise according to preset noise rules to generate a noise data table. The noise data table is configured with a sampling step size according to preset time rules to generate a time series parameter table. The noise data table is denoised iteratively according to the time series parameter table to generate an iteration result table. A parameter extractor is constructed based on the iteration result table to generate a diffusion parameter table. The diffusion parameter table is used for policy updates. The diffusion parameter table is optimized to generate an optimization strategy. The diffusion parameter table is mapped according to a preset update rule to generate a strategy parameter table. The navigation strategy is restructured according to the strategy parameter table to generate an optimization model. A navigation controller is built based on the optimization model to generate an optimization strategy. The optimization strategy is then deployed to the visual navigation system.

8. A visual navigation device based on self-imitation diffusion, characterized in that, The device includes: The trajectory sampling module is used to generate a candidate trajectory set in the environment based on the navigation strategy, extract features from the environmental visual information according to the preset perception rules to generate a perception data table, perform state encoding on the perception data table to generate a state feature set, perform trajectory sampling on the state feature set according to the preset diffusion rules to generate a candidate trajectory set, and use the candidate trajectory set for subsequent evaluation. The trajectory filtering module is used to perform quality evaluation on the candidate trajectory set to generate a high-quality trajectory set, calculate the reward of the candidate trajectory set according to the preset evaluation rules to generate a reward value table, perform threshold filtering on the reward value table to generate a trajectory weight table, select samples from the candidate trajectory set according to the trajectory weight table to generate a high-quality trajectory set, and use the high-quality trajectory set for strategy update. The visual navigation module is used to perform self-imitation update of the navigation strategy based on the high-quality trajectory set to generate an optimized strategy. The high-quality trajectory set is used to construct learning samples according to preset training rules to generate a training dataset. The training dataset is denoised and trained to generate a diffusion parameter table. The navigation strategy is updated according to the diffusion parameter table to generate an optimized strategy. The optimized strategy is then applied to the visual navigation task.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the visual navigation method based on self-imitating diffusion as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the visual navigation method based on self-imitating diffusion as described in any one of claims 1 to 7.

Images