A spatial intelligent three-dimensional pose estimation offset detection and directional correction method

Abstract

The application discloses a kind of space intelligent three-dimensional pose estimation offset detection and directional correction method, it is related to space intelligence field, solve the problem of insufficient accuracy of existing offset detection, comprising: step S1: obtaining candidate planning scheme, record the reward estimation value of each state, extract the pose parameter corresponding to state, generate space embedding vector, state is divided;Step S2: error analysis is carried out to state point, obtain real task completion degree, offset state is evaluated in combination with reward estimation value;Distribution amplification weight is generated from space embedding vector, state confidence is evaluated, in combination with offset state, calculate comprehensive offset risk score;Step S3: to be reestimated area is identified, and the path of to be reestimated area is shunted;Step S4: the processing result of to be reestimated area is recorded;Step S5: the processing process of candidate planning scheme is statistically archived;The application can effectively improve offset detection accuracy.

Description

Technical Field

[0001] This invention belongs to the field of spatial intelligence, and more specifically relates to digital twin task planning and reward modeling technology, specifically a spatial intelligent three-dimensional pose estimation offset detection and orientation correction method. Background Technology

[0002] Current research has made progress in learning-based reward modeling, model-based planning and decision-making, and learning reward functions from demonstration data. However, as these methods are applied to long-term, complex spatial task planning scenarios, existing methods still have significant shortcomings: 1. Existing methods treat the state input to the reward estimation model as an abstract feature vector, without considering the geometric proximity between the actual pose configuration of the task-related objects in the 3D workspace and the target space configuration. This results in the reward estimation offset being unable to be explained and measured from a spatial geometric perspective. If the planning algorithm continues to make decisions based on the offset reward estimate, it is easy to converge to a scheme where the actual pose achievement is not ideal but the estimated reward is inflated.

[0003] 2. Existing methods do not pay enough attention to the distribution characteristics of reward offset in the 3D pose distribution space. The reward estimation is relatively reliable for some types of pose configurations (common pose combinations with sufficient training data coverage), while the offset risk is significantly higher for other pose configurations (marginal pose combinations with sparse training data, such as unconventional placement angles and rare spatial arrangements). However, existing methods use a uniform reward estimation reliability standard for all pose regions and fail to adjust the degree of dependence of the planning algorithm on reward estimation according to the differences in spatial pose distribution characteristics.

[0004] 3. Existing methods lack a systematic coordination mechanism for the relationship between sparse but more reliable real task result feedback (such as whether the actual 3D pose of the object is close enough to the target space configuration, and whether the key assembly accuracy indicators meet the standards) and dense but risky learning-based reward estimation. The planning process mainly relies on dense reward estimation for guidance and fails to make full use of sparse real pose feedback to correct the systematic bias of the reward estimation model.

[0005] To address this, a spatial intelligent three-dimensional pose estimation offset detection and orientation correction method is proposed. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the purpose of this invention is to provide a spatial intelligent three-dimensional pose estimation offset detection and orientation correction method, aiming to improve the accuracy of offset detection.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a spatial intelligent three-dimensional pose estimation offset detection and orientation correction method, the specific working process of each step is as follows: Step S1: Obtain candidate planning schemes, collect the three-dimensional pose state sequence corresponding to the candidate planning schemes, traverse the three-dimensional pose state sequence, record the reward estimate of each state, extract the pose parameters corresponding to the state, generate spatial embedding vectors, divide the states based on the spatial embedding vectors, and label the division results. Step S2: Based on the 3D pose state sequence, perform error analysis on the state points to obtain the pose error, map the pose error to the reward estimate to obtain the actual task completion degree, and evaluate the offset state in combination with the reward estimate; generate distributed amplification weights from the spatial embedding vector, collect the number of samples of formal feedback, evaluate the state confidence, and calculate the comprehensive offset risk score by combining the offset state and the distributed amplification weights to perform regional localization. Step S3: Obtain the target offset tolerance, combine it with the comprehensive offset risk score, identify the area to be re-estimated, collect sparse real feedback of the area to be re-estimated, and perform path diversion in the area to be re-estimated. Step S4: Based on the path splitting of the region to be re-estimated, perform targeted local reward re-estimate execution or independent task result backtracking verification on the region to be re-estimated, and record the processing results of the region to be re-estimated; Step S5: Statistically archive the processing of candidate planning schemes to generate an experience rule base; retrieve the experience rule base through spatial embedding vectors, match historical offset status, reuse experience for subsequent tasks, and at the same time, use the processing results of each time to optimize the experience rule base.

[0008] Furthermore, the specific steps of step S1 are as follows: Step S11: Based on the candidate planning scheme, collect the three-dimensional position of the task object in space and denote the three-dimensional position as p_i; collect the pose of the task object and denote the pose as R_i; statistically analyze the three-dimensional position and pose at different times to obtain the object state s_t, s_t={(p_i,R_i)}; statistically analyze the object state to obtain the three-dimensional pose state sequence. Step S12: Call the learning-based reward estimation model to process each object state in the 3D pose state sequence and output the reward estimate R(s) corresponding to each state; collect the target space configuration of the corresponding state, extract the spatial embedding vector, evaluate the data distribution of each state by the spatial embedding vector, and divide and label the sparsity of the training data of the state.

[0009] Furthermore, the specific steps of step S12 are as follows: Step S121: Based on the target space configuration of the state, extract the spatial embedding vector, denote the spatial embedding vector as f(s), and calculate the average k-nearest neighbor distance d(s) of the state in the spatial embedding distribution of the training data from the spatial embedding vector. Step S122: Obtain a preset edge threshold, compare the preset edge threshold with the average distance of k nearest neighbors, if the average distance of k nearest neighbors is greater than the preset edge threshold, mark the state as an edge state with sparse coverage of training data; if the average distance of k nearest neighbors is not greater than the preset edge threshold, mark the state as a common state with sufficient coverage; record the average distance of k nearest neighbors as the coverage density.

[0010] Furthermore, the specific steps of step S2 are as follows: Step S21: Extract state points from the 3D pose state sequence, collect the target space configuration corresponding to the state points, calculate the pose error of the state points, and perform range consistency mapping between the pose error and the reward estimate through monotonically decreasing mapping to obtain the true task completion degree; evaluate the offset of the state points by the true task completion degree and the reward estimate, and perform corresponding state annotation on the state points. Step S22: Collect training data on the sparsity of state points and calculate the distribution amplification weights of state points; count the number of real feedback samples corresponding to state points, evaluate the reliability of state points, obtain state confidence, integrate the distribution amplification weights, state confidence, and offset to obtain the comprehensive offset risk score corresponding to the state points, and locate the three-dimensional pose region.

[0011] Furthermore, the specific steps of step S21 are as follows: Step S211: Collect the object state of the state point. Based on the object state s_t = {(p_i, R_i)}, record the object's 3D position p_i and object pose R_i; according to the target space configuration, record the target position p_i of the state point. goal With the target pose R_i goal Data is collected by combining the object's 3D position p_i and object's pose R_i with the target position p_i. goal With the target pose R_i goal The pose error e(s) is calculated. Step S212: Collect the upper limit of the actual output range of the learning reward estimation model to obtain the range upper limit R_max; map the pose error e(s) to the range upper limit, and perform range consistency mapping between the pose error and the reward estimate to obtain the true task completion degree R_true(s). The reward estimate R(s) of the state point is collected and calculated in combination with the actual task completion degree R_true(s) to obtain the offset Δ(s) of the state point. If Δ(s)≥0, it indicates that the reward estimate is optimistic relative to the actual pose achievement; if Δ(s)<0, it indicates that the reward estimate is pessimistic relative to the actual pose achievement.

[0012] Furthermore, the specific steps of step S22 are as follows: Step S221: Based on the sparsity of the training data of the state points, the coverage density d(s) is obtained, and the distribution amplification weight is calculated according to the coverage density to obtain the distribution amplification weight w(s). Collect available real feedback samples for each state point to obtain the number of real feedback samples; record the minimum number of samples required for confidence to obtain the number of sample requirements; calculate the state confidence level c(s) from the number of real feedback samples and the number of sample requirements. Step S222: Collect the distribution amplification weights, state confidence and offset, integrate the distribution amplification weights, state confidence and offset to obtain the comprehensive offset risk score CR(s); The specific 3D pose region with the highest comprehensive offset risk score in the 3D pose state sequence is located.

[0013] Furthermore, the specific steps of step S3 are as follows: Step S31: Compare the target offset tolerance with the comprehensive offset risk score. If the comprehensive offset risk score is greater than the target offset tolerance, record the spatial region corresponding to the state point to obtain the region to be re-estimated. If the comprehensive offset risk score is not greater than the target offset tolerance, adopt the reward estimate corresponding to the state point for subsequent planning decisions. Step S32: Statistically analyze the region to be re-estimated, collect the number of real feedback samples corresponding to the state points of the region to be re-estimated. If the number of feedback samples is greater than 0, it indicates that there is usable feedback. Perform targeted local reward re-estimate on the region to be re-estimated. For regions where there is no usable feedback, perform independent task result backtracking verification and divert the processing path.

[0014] Furthermore, the specific steps of step S4 are as follows: Step S41: Based on the path diversion of the region to be re-estimated, collect data on the region to be re-estimated for targeted local reward re-estimated, recalculate the reward estimate of the region to be re-estimated, collect data on the specific sub-targets in the region to be re-estimated, record the weight ratio of each sub-target, and adjust the weight ratio; perform statistics on the reward estimate and weight control results. Step S42: Collect data on the re-estimation region to be re-estimated for independent task result backtracking verification, access candidate solutions for execution, independently obtain the actual 3D pose configuration of the object achieved by the state points in the re-estimation region, obtain the target space configuration corresponding to the state points, and recalculate based on the actual 3D pose configuration of the object. If the calculation result meets the tolerance requirements, the verification of the re-estimation region is completed; if not, the re-estimation region is marked as needing further correction and transmitted manually.

[0015] Furthermore, the specific steps of step S41 are as follows: Step S411: Extract the reward estimate, state confidence, and offset of the state points in the region to be re-estimated, recalculate the reward estimate, and obtain the updated reward estimate; Step S412: Collect the specific sub-targets of the region to be re-estimated, and denote the sub-targets as k. Collect the weight ratio of each sub-target and denote it as μ_k. Obtain the offset state of the state points of the region to be re-estimated. If the region to be re-estimated is optimistically offset, adjust the weight ratio optimistically to obtain the adjustment weight. If the region to be re-estimated is pessimistically offset, adjust the weight ratio pessimistically to obtain the adjustment weight.

[0016] Furthermore, the specific steps of step S5 are as follows: Step S51: Statistically analyze the pose state, reward estimate, and comprehensive offset risk score of each state point in the candidate planning scheme processing process to obtain state processing data. Statistically analyze the calculation process of the comprehensive offset risk score, the diversion processing of the region to be re-estimated, and the verification and judgment process of the region to be re-estimated in the candidate planning scheme processing process to obtain processing record data. Construct an empirical rule base from the state processing data and the processing record data. Step S52: Collect spatial embedding features for subsequent tasks, retrieve the spatial embedding features of subsequent tasks from the empirical rule base, reuse the historical processing process and processing data for subsequent tasks, and simultaneously optimize the processing of subsequent tasks by feeding back the empirical rule base.

[0017] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention establishes reward offset evaluation based on the geometric error of the object's three-dimensional pose, and incorporates the degree of offset, distribution characteristics and offset direction into the quantitative evaluation, avoiding the risk of the overall acceptance of reward estimation masking the specific pose offset, and improving the accuracy of the judgment of offset status.

[0018] 2. This invention locates the specific 3D pose region with the largest comprehensive offset risk score in the 3D pose state sequence by using the comprehensive offset risk score, ensuring the accuracy of the location region. At the same time, it compares the target offset tolerance with the comprehensive offset risk score to accurately identify the region to be re-estimated, ensuring the accuracy of local re-estimated resources and reducing re-estimated costs.

[0019] 3. This invention provides a reliable verification method that does not rely on the reward estimation model itself by backtracking and verifying the results of independent tasks, using the actual obtained real 3D pose as the criterion, thus ensuring the accuracy of correction. By re-estimating the local results, the reward estimates of the remaining confirmed reliable pose regions are retained unchanged, thus avoiding the waste of resources caused by retraining the entire model. Attached Figure Description

[0020] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0021] Figure 1 This is a schematic diagram of the process of the present invention.

[0022] Figure 2 This is a schematic diagram illustrating the offset calculation of the present invention.

[0023] Figure 3 This is a schematic diagram of the path splitting process of the present invention. Detailed Implementation

[0024] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0025] This application provides a spatial intelligent 3D pose estimation offset detection and orientation correction method. The execution subject of this spatial intelligent 3D pose estimation offset detection and orientation correction method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application embodiment: a server, a terminal, etc. In other words, the spatial intelligent 3D pose estimation offset detection and orientation correction method can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster.

[0026] Reference Figure 1 The diagram shown is a flowchart illustrating a spatial intelligent 3D pose estimation offset detection and orientation correction method according to an embodiment of the present invention. In this embodiment, the spatial intelligent 3D pose estimation offset detection and orientation correction method includes: Step S1: Obtain candidate planning schemes, collect the three-dimensional pose state sequence corresponding to the candidate planning schemes, traverse the three-dimensional pose state sequence, record the reward estimate of each state, extract the pose parameters corresponding to the state, generate spatial embedding vectors, divide the states based on the spatial embedding vectors, and label the division results.

[0027] Step S11: Based on the candidate planning scheme, collect the three-dimensional position of the task object in space and denote the three-dimensional position as p_i; collect the pose of the task object and denote the pose as R_i; statistically analyze the three-dimensional position and pose at different times to obtain the object state s_t, s_t={(p_i,R_i)}; statistically analyze the object state to obtain the three-dimensional pose state sequence. Step S12: Call the learning-based reward estimation model to process each object state in the 3D pose state sequence and output the reward estimate R(s) corresponding to each state; collect the target space configuration of the corresponding state, extract the spatial embedding vector, evaluate the data distribution of each state by the spatial embedding vector, and divide and label the sparsity of the training data of the state. Step S121: Based on the target space configuration of the state, extract the spatial embedding vector, denote the spatial embedding vector as f(s), and calculate the average k-nearest neighbor distance d(s) of the state in the spatial embedding distribution of the training data from the spatial embedding vector. ; Where k is the number of neighborhood samples, j∈KNN(s) is the traversal of the neighborhood set, and f(s_j) is the spatial embedding vector of the j-th neighbor point; Step S122: Obtain a preset edge threshold, compare the preset edge threshold with the average distance of k nearest neighbors, if the average distance of k nearest neighbors is greater than the preset edge threshold, mark the state as an edge state with sparse coverage of training data; if the average distance of k nearest neighbors is not greater than the preset edge threshold, mark the state as a common state with sufficient coverage; record the average distance of k nearest neighbors as the coverage density.

[0028] Step S2: Based on the 3D pose state sequence, perform error analysis on the state points to obtain the pose error, map the pose error to the reward estimate to obtain the actual task completion degree, and evaluate the offset state in combination with the reward estimate; generate distributed amplification weights from the spatial embedding vector, collect the number of samples of formal feedback, evaluate the state confidence, and calculate the comprehensive offset risk score by combining the offset state and the distributed amplification weights to perform regional localization. Please see Figure 2Step S21: Extract state points from the 3D pose state sequence, collect the target space configuration corresponding to the state points, calculate the pose error of the state points, and perform range consistency mapping between the pose error and the reward estimate through monotonically decreasing mapping to obtain the true task completion degree; evaluate the offset of the state points by the true task completion degree and the reward estimate, and perform corresponding state annotation on the state points. Step S211: Collect the object state of the state point. Based on the object state s_t = {(p_i, R_i)}, record the object's 3D position p_i and object pose R_i; according to the target space configuration, record the target position p_i of the state point. goal With the target pose R_i goal Data is collected by combining the object's 3D position p_i and object's pose R_i with the target position p_i. goal With the target pose R_i goal The pose error e(s) is calculated. ; Where d_rot refers to the rotation geodesic distance, which is the angular distance between the two attitudes on SO(3), and α and β are the weighting coefficients of translation error and rotation error, respectively.

[0029] Step S212: Collect the upper limit of the actual output range of the learning reward estimation model to obtain the range upper limit R_max; map the pose error e(s) to the range upper limit, and perform range consistency mapping between the pose error and the reward estimate to obtain the true task completion degree R_true(s). ; Where σ is a calibrable scaling factor (which can be calibrated according to the allowable error reference level for this type of task, so that R_true(s) corresponding to the reference error level falls at the expected scale).

[0030] The reward estimate R(s) of the state point is collected and calculated in combination with the actual task completion degree R_true(s) to obtain the offset Δ(s) of the state point. ; If Δ(s)≥0, it indicates that the reward estimate is optimistic relative to the actual pose achievement; if Δ(s)<0, it indicates that the reward estimate is pessimistic relative to the actual pose achievement.

[0031] Step S22: Collect training data on the sparsity of state points and calculate the distribution amplification weights of state points; count the number of real feedback samples corresponding to state points, evaluate the reliability of state points, obtain state confidence, integrate the distribution amplification weights, state confidence, and offset to obtain the comprehensive offset risk score corresponding to the state points, and locate the three-dimensional pose region.

[0032] Step S221: Based on the sparsity of the training data of the state points, the coverage density d(s) is obtained, and the distribution amplification weight is calculated according to the coverage density to obtain the distribution amplification weight w(s). ; Where τ_edge is the density threshold, λ is the upper limit of magnification, κ is the smoothing scale parameter, and logistic is the calculation function, the internal calculation of which is as follows: ; Where x is the independent variable of the function and e is the natural base.

[0033] Collect available real feedback samples for each state point to obtain the number of real feedback samples; record the minimum number of samples required for confidence to obtain the number of sample requirements; calculate the state confidence level c(s) from the number of real feedback samples and the number of sample requirements. ; Where n_feedback(s) is the number of actual feedback samples, and n_min is the number of samples required.

[0034] Step S222: Collect the distribution amplification weights, state confidence and offset, integrate the distribution amplification weights, state confidence and offset to obtain the comprehensive offset risk score CR(s); ; The specific 3D pose region with the highest comprehensive offset risk score in the 3D pose state sequence is located.

[0035] Step S3: Obtain the target offset tolerance, combine it with the comprehensive offset risk score, identify the area to be re-estimated, collect sparse real feedback of the area to be re-estimated, and perform path diversion in the area to be re-estimated.

[0036] Please see Figure 3 Step S31: Compare the target offset tolerance with the comprehensive offset risk score. If the comprehensive offset risk score is greater than the target offset tolerance, record the spatial region corresponding to the state point to obtain the region to be re-estimated. If the comprehensive offset risk score is not greater than the target offset tolerance, adopt the reward estimate corresponding to the state point for subsequent planning decisions. Step S32: Statistically analyze the region to be re-estimated, collect the number of real feedback samples corresponding to the state points of the region to be re-estimated. If the number of feedback samples is greater than 0, it indicates that there is usable feedback. Perform targeted local reward re-estimate on the region to be re-estimated. For regions where there is no usable feedback, perform independent task result backtracking verification and divert the processing path.

[0037] Step S4: Based on the path splitting of the region to be re-estimated, perform targeted local reward re-estimate execution or independent task result backtracking verification on the region to be re-estimated, and record the processing results of the region to be re-estimated; Step S41: Based on the path diversion of the region to be re-estimated, collect data on the region to be re-estimated for targeted local reward re-estimated, recalculate the reward estimate of the region to be re-estimated, collect data on the specific sub-targets in the region to be re-estimated, record the weight ratio of each sub-target, and adjust the weight ratio; perform statistics on the reward estimate and weight control results. Step S411: Extract the reward estimate, state confidence, and offset of the state points in the region to be re-estimated, recalculate the reward estimate, and obtain the updated reward estimate; ; Step S412: Collect the specific sub-targets of the region to be re-estimated, and denote the sub-target as k. Collect the weight ratio of each sub-target and denote it as μ_k. Obtain the offset state of the state points of the region to be re-estimated. If the region to be re-estimated is an optimistic offset, perform optimistic adjustment on the weight ratio to obtain the adjustment weight μ_k'. ; Where ρ∈(0,1) is the adjustment coefficient, and CR_max is the normalized upper limit of the historical comprehensive offset risk of this type of sub-target; If the region to be re-estimated is pessimistically biased, the weight ratio is pessimistically adjusted to obtain the adjusted weight μ_k'; ; It should be noted that differentiated corrections are made based on two different types of bias risks: one overly optimistic and the other overly pessimistic, to ensure the stability and accuracy of the corrections.

[0038] Step S42: Collect data on the re-estimation region to be re-estimated for independent task result backtracking verification, access candidate solutions for execution, independently obtain the actual 3D pose configuration of the object achieved by the state points in the re-estimation region, obtain the target space configuration corresponding to the state points, and recalculate based on the actual 3D pose configuration of the object. If the calculation result meets the tolerance requirements, the verification of the re-estimation region is completed; if not, the re-estimation region is marked as needing further correction and transmitted manually.

[0039] Step S5: Statistically archive the processing of candidate planning schemes to generate an experience rule base; retrieve the experience rule base through spatial embedding vectors, match historical offset status, reuse experience for subsequent tasks, and at the same time, use the processing results of each time to optimize the experience rule base.

[0040] Step S51: Statistically analyze the pose state, reward estimate, and comprehensive offset risk score of each state point in the candidate planning scheme processing process to obtain state processing data. Statistically analyze the calculation process of the comprehensive offset risk score, the diversion processing of the region to be re-estimated, and the verification and judgment process of the region to be re-estimated in the candidate planning scheme processing process to obtain processing record data. Construct an empirical rule base from the state processing data and the processing record data. Step S52: Collect spatial embedding features for subsequent tasks, retrieve the spatial embedding features of subsequent tasks from the empirical rule base, reuse the historical processing process and processing data for subsequent tasks, and simultaneously optimize the processing of subsequent tasks by feeding back the empirical rule base.

[0041] It should be noted that the empirical rules of displacement accumulated through this process can be migrated and reused across multiple similar production lines within the same group, reducing the time and cost of accumulating real feedback samples that would otherwise be required for each production line to independently and repeatedly discover similar pose displacement patterns.

[0042] Compared to the problems described in the background art, this invention establishes reward offset evaluation based on the geometric error of the object's 3D pose, and incorporates the degree of offset, distribution characteristics, and offset direction into a unified quantitative evaluation. This avoids the overall acceptance of reward estimation masking the risk of specific pose offsets, thus improving the accuracy of offset judgment. Furthermore, this invention locates the specific 3D pose region with the largest comprehensive offset risk score in the 3D pose state sequence by using the comprehensive offset risk score, ensuring the accuracy of the location region. At the same time, it compares the target offset tolerance with the comprehensive offset risk score to accurately identify the region to be re-estimated, ensuring the accuracy of local re-estimated resources and reducing re-estimated costs. Finally, this invention provides a reliable verification method that does not rely on the reward estimation model itself, using the actual obtained 3D pose as the criterion, through independent task result backtracking verification, ensuring the accuracy of correction. By using directional local re-estimate, the reward estimate of other confirmed reliable pose regions remains unchanged, avoiding the waste of resources caused by retraining the entire model. Therefore, the spatial intelligent 3D pose estimation offset detection and directional correction method provided by this invention can improve the accuracy of offset detection and directional correction.

[0043] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0044] Finally, it should be noted that deleting any one of the above embodiments does not affect the technical solutions of other embodiments. The above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A spatial intelligent three-dimensional pose estimation offset detection and orientation correction method, characterized in that, include: Step S1: Obtain candidate planning schemes, collect the three-dimensional pose state sequence corresponding to the candidate planning schemes, traverse the three-dimensional pose state sequence, record the reward estimate of each state, extract the pose parameters corresponding to the state, generate spatial embedding vectors, divide the states based on the spatial embedding vectors, and label the division results. Step S2: Based on the three-dimensional pose state sequence, perform error analysis on the state points to obtain the pose error, map the pose error to the reward estimate to obtain the true task completion degree, and evaluate the offset state in combination with the reward estimate. The distributed amplification weights are generated from the spatial embedding vectors. The number of samples in the formal feedback is collected to evaluate the state confidence. The comprehensive offset risk score is calculated by combining the offset state and the distributed amplification weights to perform regional positioning. Step S3: Obtain the target offset tolerance, combine it with the comprehensive offset risk score, identify the area to be re-estimated, collect sparse real feedback of the area to be re-estimated, and perform path diversion in the area to be re-estimated. Step S4: Based on the path splitting of the region to be re-estimated, perform targeted local reward re-estimate execution or independent task result backtracking verification on the region to be re-estimated, and record the processing results of the region to be re-estimated; Step S5: Statistically archive the processing of candidate planning schemes to generate an experience rule base; By using spatial embedding vectors, the empirical rule base is retrieved, historical offset conditions are matched, and the experience is reused for subsequent tasks. At the same time, the results of each processing are fed back to the empirical rule base for optimization.

2. The spatial intelligent three-dimensional pose estimation offset detection and orientation correction method according to claim 1, characterized in that, The specific steps of step S1 are as follows: Step S11: Based on the candidate planning scheme, collect the three-dimensional position of the task object in space and denote the three-dimensional position as p_i; collect the pose of the task object and denote the pose as R_i; statistically analyze the three-dimensional position and pose at different times to obtain the object state s_t, s_t={(p_i,R_i)}; statistically analyze the object state to obtain the three-dimensional pose state sequence. Step S12: Call the learning-based reward estimation model to process each object state in the 3D pose state sequence and output the reward estimate R(s) corresponding to each state; collect the target space configuration of the corresponding state, extract the spatial embedding vector, evaluate the data distribution of each state by the spatial embedding vector, and classify and label the sparsity of the training data of the state.

3. The spatial intelligent three-dimensional pose estimation offset detection and orientation correction method according to claim 2, characterized in that, The specific steps of step S12 are as follows: Step S121: Based on the target space configuration of the state, extract the spatial embedding vector, denote the spatial embedding vector as f(s), and calculate the average k-nearest neighbor distance d(s) of the state in the spatial embedding distribution of the training data from the spatial embedding vector. Step S122: Obtain a preset edge threshold, compare the preset edge threshold with the average distance of k nearest neighbors, if the average distance of k nearest neighbors is greater than the preset edge threshold, mark the state as an edge state with sparse coverage of training data; if the average distance of k nearest neighbors is not greater than the preset edge threshold, mark the state as a common state with sufficient coverage; record the average distance of k nearest neighbors as the coverage density.

4. The spatial intelligent three-dimensional pose estimation offset detection and orientation correction method according to claim 1, characterized in that, The specific steps of step S2 are as follows: Step S21: Extract state points from the 3D pose state sequence, collect the target space configuration corresponding to the state points, calculate the pose error of the state points, and perform range consistency mapping between the pose error and the reward estimate through monotonically decreasing mapping to obtain the true task completion degree; evaluate the offset of the state points by the true task completion degree and the reward estimate, and perform corresponding state annotation on the state points. Step S22: Collect training data on the sparsity of state points and calculate the distribution amplification weights of state points; count the number of real feedback samples corresponding to state points, evaluate the reliability of state points, obtain state confidence, integrate the distribution amplification weights, state confidence, and offset to obtain the comprehensive offset risk score corresponding to the state points, and locate the three-dimensional pose region.

5. The spatial intelligent three-dimensional pose estimation offset detection and orientation correction method according to claim 4, characterized in that, The specific steps of step S21 are as follows: Step S211: Collect the object state of the state point. Based on the object state s_t = {(p_i, R_i)}, record the object's 3D position p_i and object pose R_i; according to the target space configuration, record the target position p_i of the state point. goal With the target pose R_i goal Data is collected by combining the object's 3D position p_i and object's pose R_i with the target position p_i. goal With the target pose R_i goal The pose error e(s) is calculated. Step S212: Collect the upper limit of the actual output range of the learning reward estimation model to obtain the range upper limit R_max; map the pose error e(s) to the range upper limit, and perform range consistency mapping between the pose error and the reward estimate to obtain the true task completion degree R_true(s). The reward estimate R(s) of the state point is collected and calculated in combination with the actual task completion degree R_true(s) to obtain the offset Δ(s) of the state point. If Δ(s)≥0, it indicates that the reward estimate has a positive bias relative to the actual pose achievement; if Δ(s)<0, it indicates that the reward estimate has a negative bias relative to the actual pose achievement.

6. The spatial intelligent three-dimensional pose estimation offset detection and orientation correction method according to claim 4, characterized in that, The specific steps of step S22 are as follows: Step S221: Based on the sparsity of the training data of the state points, the coverage density d(s) is obtained, and the distribution amplification weight is calculated according to the coverage density to obtain the distribution amplification weight w(s). Collect available real feedback samples for each state point to obtain the number of real feedback samples; record the minimum number of samples required for confidence to obtain the number of sample requirements; calculate the state confidence level c(s) from the number of real feedback samples and the number of sample requirements. Step S222: Collect the distribution amplification weights, state confidence and offset, integrate the distribution amplification weights, state confidence and offset to obtain the comprehensive offset risk score CR(s); The specific 3D pose region with the highest comprehensive offset risk score in the 3D pose state sequence is located.

7. The spatial intelligent three-dimensional pose estimation offset detection and orientation correction method according to claim 1, characterized in that, The specific steps of step S3 are as follows: Step S31: Compare the target offset tolerance with the comprehensive offset risk score. If the comprehensive offset risk score is greater than the target offset tolerance, record the spatial region corresponding to the state point to obtain the region to be re-estimated. If the overall offset risk score is not greater than the target offset tolerance, the reward estimate corresponding to the state point will be adopted for subsequent planning decisions. Step S32: Statistically analyze the region to be re-estimated, collect the number of real feedback samples corresponding to the state points of the region to be re-estimated. If the number of feedback samples is greater than 0, it indicates that there is usable feedback. Perform targeted local reward re-estimate on the region to be re-estimated. For regions without usable feedback, perform independent task result backtracking verification and divert the processing path.

8. The spatial intelligent three-dimensional pose estimation offset detection and orientation correction method according to claim 1, characterized in that, The specific steps of step S4 are as follows: Step S41: Based on the path splitting of the region to be re-estimated, collect data on the region to be re-estimated for targeted local reward re-estimated, recalculate the reward estimate of the region to be re-estimated, collect data on the specific sub-targets in the region to be re-estimated, record the weight ratio of each sub-target, and adjust the weight ratio; perform statistics on the reward estimate and weight control results. Step S42: Collect data on the re-estimation region to be re-estimated for independent task result backtracking verification, access candidate solutions for execution, independently obtain the actual 3D pose configuration of the object achieved by the state points in the re-estimation region, obtain the target space configuration corresponding to the state points, and recalculate based on the actual 3D pose configuration of the object. If the calculation result meets the tolerance requirements, the verification of the re-estimation region is completed; if not, the re-estimation region is marked as needing further correction and transmitted manually.

9. The spatial intelligent three-dimensional pose estimation offset detection and orientation correction method according to claim 8, characterized in that, The specific steps of step S41 are as follows: Step S411: Extract the reward estimate, state confidence, and offset of the state points in the region to be re-estimated, recalculate the reward estimate, and obtain the updated reward estimate; Step S412: Collect data on the specific sub-targets in the region to be re-estimated, denoted as k, and collect the weight percentage of each sub-target, denoted as μ_k; Obtain the offset state of the state points in the region to be re-estimated. If the region to be re-estimated is offset, adjust the weight ratio to obtain the adjustment weight.

10. The spatial intelligent three-dimensional pose estimation offset detection and orientation correction method according to claim 1, characterized in that, The specific steps of step S5 are as follows: Step S51: Statistically analyze the pose state, reward estimate, and comprehensive offset risk score of each state point in the candidate planning scheme processing process to obtain state processing data. Statistically analyze the calculation process of the comprehensive offset risk score, the diversion processing of the region to be re-estimated, and the verification and judgment process of the region to be re-estimated in the candidate planning scheme processing process to obtain processing record data. Construct an empirical rule base from the state processing data and the processing record data. Step S52: Collect spatial embedding features for subsequent tasks, retrieve the spatial embedding features of subsequent tasks from the empirical rule base, reuse the historical processing process and processing data for subsequent tasks, and simultaneously optimize the processing of subsequent tasks by feeding back the empirical rule base.

Images