A spatial intelligent long-term rolling prediction error accumulation and re-anchoring method
By constructing a prediction error accumulation field and a path diversion mechanism, the problems of unquantified error accumulation rate and insufficient re-anchoring mechanism in long-term planning are solved, achieving accurate error processing and efficient model improvement, reducing re-anchoring costs and improving prediction accuracy.
Patent Information
- Application Number
- CN202611592570.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-15
- Publication Date
- 2026-11-17
AI Technical Summary
Existing methods fail to effectively quantify the error accumulation rate in long-term planning, lack a mechanism for re-anchoring the trajectory, and have unclear attribution of error sources, resulting in high re-anchoring costs and low accuracy.
By collecting real-time observation data and actions, predictive trajectories are constructed, deviation and cumulative error are analyzed, trajectories to be re-anchored are marked, rolling predictions are re-executed, and feedback is fed back to model training, thereby achieving accurate error handling and model improvement.
It reduces the resource investment required for re-anchoring, improves the reliability of error judgment and prediction accuracy, reduces the cost of re-prediction, and provides targeted improvement directions.
Smart Images

Figure CN123115699A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of spatial intelligence, and more specifically relates to the dynamic prediction technology of digital twin environment, specifically a spatial intelligence long-term rolling prediction error accumulation and re-anchoring method. Background Technology
[0002] Current research has made progress in areas such as world model learning, model-based reinforcement learning for long-term planning, and dynamic prediction of digital twin environments. However, as these methods are applied to real-world embodied agent tasks requiring long-term planning, existing methods still have significant shortcomings.
[0003] 1. Existing methods typically directly adopt the entire rolling prediction trajectory without gradually quantifying and evaluating the specific degree of accumulated error. They do not pay enough attention to the differences in error accumulation rates under different environmental dynamic complexities. In scenarios with complex environmental interactions and high nonlinearity, the growth rate of accumulated error is significantly faster than in simple scenarios. However, existing methods use a uniform upper limit for the number of rolling prediction steps for all scenarios and fail to dynamically adjust it according to the actual accumulation rate.
[0004] 2. When existing methods discover that the imagined trajectory of a rolling prediction deviates significantly from subsequent actual observations, they can usually only choose to completely abandon the trajectory and re-predict or continue to use the biased trajectory. They lack a processing mechanism to use actual observations to orient and re-anchor the trajectory, correct only the deviation part and retain the verified part.
[0005] 3. Existing methods lack a detailed attribution mechanism for the sources of prediction errors in world models. They fail to distinguish whether the errors are mainly due to the quantification loss of the state representation itself or the inaccurate modeling of the dynamic transfer function. General error assessments are difficult to guide targeted model improvement directions.
[0006] To address this, a spatial intelligent long-term rolling prediction error accumulation and re-anchoring method is proposed. Summary of the Invention
[0007] To address the shortcomings of existing technologies, the purpose of this invention is to provide a spatial intelligent long-term rolling prediction error accumulation re-anchoring method. This invention aims to reduce re-anchoring costs and ensure adjustment accuracy.
[0008] To achieve the above objectives, the present invention adopts the following technical solution: a spatial intelligent long-term rolling prediction error accumulation and re-anchoring method, the specific working process of each step is as follows: Step S1: Collect the real-time observation status and execution tasks of the world model, decompose the execution actions based on the execution tasks, construct the execution action sequence, combine the real-time observation status, predict the subsequent state, and generate a complete prediction trajectory. Step S2: Collect the real-time observation status corresponding to the complete predicted trajectory, analyze the deviation between the predicted status and the real-time observation status, obtain the single-step deviation degree, statistically analyze the changes in the single-step deviation degree, evaluate the cumulative error, and construct the prediction error cumulative field; at the same time, extrapolate the prediction based on the changes in the single-step deviation degree, and record the extrapolation prediction deviation. Step S3: Obtain the task deviation standard constraint, mark the trajectory to be re-anchored in combination with the prediction error cumulative field, and perform path diversion on the trajectory to be re-anchored based on the real-time observation status corresponding to the complete prediction trajectory. Step S4: Based on the path diversion of the trajectory to be re-anchored, re-execute rolling prediction for the subsequent predicted trajectory to generate re-anchoring results; for the trajectory to be re-anchored with missing real-time observation status, re-perform deviation analysis and determine the anchoring results. Step S5: Archive the rolling prediction process, re-anchoring records, and verification conclusions. At the same time, feed back the deviation pattern between the actual observation and the imagined trajectory to the world model training process for subsequent model iteration and improvement.
[0009] Furthermore, the specific steps of step S1 are as follows: Step S11: Record the world model currently observed by the robot to obtain the real-time observation state, obtain the robot's execution plan, and statistically analyze the robot's execution actions based on the robot's execution plan to construct a sequence of execution actions; Step S12: Based on the sequence of executed actions, the displacement status of the robot is collected to obtain spatial change information. The spatial change information is used as a change parameter to map the world model. Combined with the real-time observation status, the world model observed after the robot performs actions is predicted to obtain the predicted state. The predicted state after completing different actions is recorded by the sequence of executed actions to generate a complete predicted trajectory.
[0010] Furthermore, the specific steps of step S2 are as follows: Step S21: Based on the complete predicted trajectory, extract the predicted state after the action is performed, collect the real-time observation state after the robot performs the corresponding action, feed back the real-time observation state to the predicted state, evaluate the deviation between the predicted state and the real-time observation state, and obtain the single-step deviation degree; statistically analyze the single-step deviation degree after performing different actions and construct a list of deviation changes. Step S22: Record the changes in single-step deviation from the deviation change list, evaluate the cumulative error generated by the action execution, and construct the prediction error cumulative field; analyze the error accumulation rate based on the prediction error cumulative field, and extrapolate the single-step deviation from the deviation change list to generate the extrapolated prediction deviation, and mark the extrapolated prediction deviation on the deviation change list.
[0011] Furthermore, the specific steps of step S21 are as follows: Step S211: Collect data on static fixed objects on the world model and establish spatial anchor points; extract the predicted state after the action is performed from the complete predicted trajectory, map the spatial anchor points to the predicted state, record the anchor point coordinates in the predicted state, and obtain the predicted coordinates; collect the real-time observation state corresponding to the predicted state from the robot's action, record the anchor point coordinates in the real-time observation state, and obtain the observation coordinates; calculate the coordinate deviation between the predicted coordinates and the observation coordinates to obtain the single-step deviation degree. Step S212: Traverse the predicted state based on the complete predicted trajectory, collect real-time observations of the corresponding actions after execution based on the predicted state, calculate the single-step deviation of different actions, and statistically analyze the single-step deviation based on the action execution order of the complete predicted trajectory to construct a list of deviation changes.
[0012] Furthermore, the specific steps of step S22 are as follows: Step S221: Extract the single-step deviation from the deviation change list, record the difference between adjacent single-step deviations to obtain the cumulative error value, statistically analyze the cumulative error value, and construct the prediction error cumulative field; from the prediction error cumulative field, iterate through the cumulative error value, extract the execution action corresponding to the cumulative error value, record the order of the execution actions, associate the order of the execution actions with the cumulative error, perform linear regression on the cumulative error, and record the error accumulation rate; Step S222: Predict the error change of the next action by the error accumulation rate to obtain the extrapolated prediction accumulation; extract the known single-step deviation of the last execution from the deviation change list, combine it with the extrapolated prediction accumulation to obtain the extrapolated prediction deviation, add the extrapolated prediction accumulation as an extension item to the deviation change list, and mark the extrapolated prediction deviation.
[0013] Furthermore, the specific steps of step S3 are as follows: Step S31: Based on the task deviation standard constraint, obtain the robot's error tolerance, judge the overall trajectory based on the deviation change list and the error tolerance, and combine the prediction error accumulation field to specifically locate and mark the anchor trajectory to be re-anchored. Step S32: Collect the predicted state corresponding to the action execution of the trajectory to be re-anchored, collect newly added external observation data, identify the newly added external observation data and the predicted state corresponding to the action execution, if there is a real-time observation state corresponding to the predicted state in the newly added external observation data, set the trajectory to be re-anchored to the orientation state re-anchoring; if there is no real-time observation state corresponding to the predicted state in the newly added external observation data, set the trajectory to be re-anchored to the historical trajectory re-anchoring, and mark the trajectory to be re-anchored by the orientation state re-anchoring and the historical trajectory re-anchoring.
[0014] Furthermore, the specific steps of step S31 are as follows: Step S311: Iterate through the elements in the deviation change list, extract the robot's deviation status, compare the robot's deviation status with the error tolerance. If the deviation is higher than the error tolerance, it is determined that the trajectory is abnormal and the trajectory is recorded; if the deviation is not higher than the error tolerance, the trajectory is determined to be normal. Step S312: Extract abnormal trajectories, collect the cumulative prediction error field of abnormal trajectories, extract the maximum value of the cumulative error in the cumulative prediction error field, locate the execution action corresponding to the cumulative error value, extract the corresponding trajectory of the execution action in the complete prediction trajectory, and obtain the trajectory to be re-anchored.
[0015] Furthermore, the specific steps of step S4 are as follows: Step S41: Extract the trajectory to be re-anchored for the directional state re-anchoring marked by the labeling result, record the execution action corresponding to the trajectory to be re-anchored, collect the real-time observation state corresponding to the execution action, take the real-time observation state as the trajectory starting point, extract the subsequent predicted trajectory of the trajectory starting point in the complete predicted trajectory, record the execution action of the subsequent predicted trajectory, map the trajectory starting point, re-execute rolling prediction on the subsequent predicted trajectory, and generate the re-anchoring result. Step S42: Extract the trajectory to be re-anchored from the historical trajectory re-anchored, obtain the subsequent predicted trajectory of the trajectory to be re-anchored in the complete predicted trajectory, collect the real-time observation status of the corresponding action of the subsequent predicted trajectory, independently verify the real-time observation status and the prediction status corresponding to the subsequent predicted trajectory, record the deviation between the real-time observation status and the prediction status, if the deviation is greater than the deviation of the action corresponding to the complete predicted trajectory, mark it as manual verification.
[0016] Furthermore, the specific steps of step S42 are as follows: Step S421: Obtain the cumulative error value of the re-anchored trajectory in the prediction error accumulation field, perform reverse processing on the re-anchored trajectory based on the cumulative error value, eliminate the error corresponding to the trajectory, map the elimination result to the real-time observation state of the subsequent predicted trajectory corresponding to the execution action, and optimize the real-time observation state. Step S422: Re-evaluate the error between the optimized real-time observation status and the subsequent predicted trajectory status, record the optimization deviation, compare the optimized deviation with the original deviation, if the deviation decreases, adopt the optimization result; if the deviation increases, mark the re-anchor point trajectory as a manually checked trajectory.
[0017] Furthermore, the specific steps of step S5 are as follows: Step S51: Archive the rolling prediction process, re-anchoring records and verification conclusions, record the trajectory for re-anchoring in the rolling prediction process, extract the re-anchoring judgment process corresponding to the trajectory, build a re-anchoring experience base, and provide feedback on the robot's processing process.
[0018] Step S52: Connect the re-anchoring experience base to the robot. When the robot needs to re-anchor its trajectory, search the re-anchoring experience base to find the corresponding re-anchoring process and implement it on the robot. Verify the implementation result with the corresponding re-anchoring result in the re-anchoring experience base, perform re-anchoring control on the robot, and train and optimize the world model based on the re-anchoring control of the robot.
[0019] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention converts single-step deviation, cumulative drift rate and re-anchoring interval into a unified quantitative index, avoiding the use of the entire trajectory to cover up the specific source of deviation, accurately extracting the main stages of error generation, achieving precise error processing, and making the investment of re-anchoring resources more precise.
[0020] 2. This invention verifies the prediction process by independently replaying the trajectory, thereby improving the reliability of error judgment. Based on the extracted error trajectory, the verified part of the trajectory is retained unchanged, avoiding the computational waste caused by completely re-predicting from the starting point of the task and reducing the cost of re-prediction.
[0021] 3. This invention distinguishes between two components: characterization loss and dynamic modeling error, providing a clear direction for targeted model improvement. It performs path diversion on the re-anchored trajectory based on specific attribution, balancing the accuracy of trajectory correction with computational cost. Attached Figure Description
[0022] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.
[0023] Figure 1 This is a schematic diagram of the process of the present invention.
[0024] Figure 2 This is a schematic diagram of the cumulative field processing of prediction error in this invention.
[0025] Figure 3This is a schematic diagram of the trajectory processing of the anchor point to be re-anchored according to the present invention. Detailed Implementation
[0026] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
[0027] This application provides a spatial intelligent long-term rolling prediction error accumulation and re-anchoring method. The executing entity of this spatial intelligent long-term rolling prediction error accumulation and re-anchoring 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: a server, a terminal, etc. In other words, the spatial intelligent long-term rolling prediction error accumulation and re-anchoring 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.
[0028] Reference Figure 1 The diagram shown is a flowchart illustrating a spatial intelligent long-term rolling prediction error accumulation and re-anchoring method according to an embodiment of the present invention. In this embodiment, the spatial intelligent long-term rolling prediction error accumulation and re-anchoring method includes: Step S1: Collect the real-time observation status and execution tasks of the world model, decompose the execution actions based on the execution tasks, construct the execution action sequence, combine the real-time observation status, predict the subsequent state, and generate a complete prediction trajectory. Step S11: Record the world model currently observed by the robot to obtain the real-time observation state, obtain the robot's execution plan, and statistically analyze the robot's execution actions based on the robot's execution plan to construct a sequence of execution actions; Step S12: Based on the sequence of executed actions, the displacement status of the robot is collected to obtain spatial change information. The spatial change information is used as a change parameter to map the world model. Combined with the real-time observation status, the world model observed after the robot performs actions is predicted to obtain the predicted state. The predicted state after completing different actions is recorded by the sequence of executed actions to generate a complete predicted trajectory.
[0029] It should be noted that: predicting the subsequent model state by performing actions and the resulting spatial changes provides data support for the prediction results under theoretical conditions, ensuring the accuracy of the prediction. Based on the spatial prediction, a complete trajectory is provided for a series of robot actions, providing a reference standard for subsequent error calculation.
[0030] Step S2: Collect the real-time observation status corresponding to the complete predicted trajectory, analyze the deviation between the predicted status and the real-time observation status, obtain the single-step deviation degree, statistically analyze the changes in the single-step deviation degree, evaluate the cumulative error, and construct the prediction error cumulative field; at the same time, extrapolate the prediction based on the changes in the single-step deviation degree, and record the extrapolation prediction deviation. Please see Figure 2 Step S21: Based on the complete predicted trajectory, extract the predicted state after the action is performed, collect the real-time observation state after the robot performs the corresponding action, feed back the real-time observation state to the predicted state, evaluate the deviation between the predicted state and the real-time observation state, and obtain the single-step deviation degree; statistically analyze the single-step deviation degree after performing different actions and construct a list of deviation changes. Step S211: Collect data on static fixed objects on the world model and establish spatial anchor points; extract the predicted state after the action is performed from the complete predicted trajectory, map the spatial anchor points to the predicted state, record the anchor point coordinates in the predicted state, and obtain the predicted coordinates; collect the real-time observation state corresponding to the predicted state from the robot's action, record the anchor point coordinates in the real-time observation state, and obtain the observation coordinates; calculate the coordinate deviation between the predicted coordinates and the observation coordinates to obtain the single-step deviation degree. ; Where dpd is the single-step deviation, yzb(a) is the predicted coordinate of the a-th spatial anchor point, gzb is the observed coordinate of the a-th spatial anchor point, and as is the number of spatial anchor points.
[0031] Step S212: Traverse the predicted state based on the complete predicted trajectory, collect real-time observations of the corresponding actions after execution based on the predicted state, calculate the single-step deviation of different actions, and statistically analyze the single-step deviation based on the action execution order of the complete predicted trajectory to construct a list of deviation changes.
[0032] Step S22: Record the changes in single-step deviation from the deviation change list, evaluate the cumulative error generated by the action execution, and construct the prediction error cumulative field; analyze the error accumulation rate based on the prediction error cumulative field, and extrapolate the single-step deviation from the deviation change list to generate the extrapolated prediction deviation, and mark the extrapolated prediction deviation on the deviation change list.
[0033] Step S221: Extract the single-step deviation from the deviation change list, record the difference between adjacent single-step deviations to obtain the cumulative error value, statistically analyze the cumulative error value, and construct the prediction error cumulative field; from the prediction error cumulative field, iterate through the cumulative error value, extract the execution action corresponding to the cumulative error value, record the order of the execution actions, associate the order of the execution actions with the cumulative error, perform linear regression on the cumulative error, and record the error accumulation rate; Step S222: Predict the error change of the next action by the error accumulation rate to obtain the extrapolated prediction accumulation; extract the known single-step deviation of the last execution from the deviation change list, combine it with the extrapolated prediction accumulation to obtain the extrapolated prediction deviation, add the extrapolated prediction accumulation as an extension item to the deviation change list, and mark the extrapolated prediction deviation.
[0034] It should be noted that: based on the existing error changes and the single-step deviation, the subsequent deviation is predicted. By extrapolating the predicted deviation, abnormal predicted trajectories are detected in a timely manner, and early warning processing is carried out for deviations to ensure the safe and stable operation of the robot.
[0035] Step S3: Obtain the task deviation standard constraint, mark the trajectory to be re-anchored in combination with the prediction error cumulative field, and perform path diversion on the trajectory to be re-anchored based on the real-time observation status corresponding to the complete prediction trajectory. Please see Figure 3 Step S31: Based on the task deviation standard constraint, obtain the robot's error tolerance, judge the overall trajectory based on the deviation change list and the error tolerance, and combine the prediction error accumulation field to specifically locate and mark the anchor trajectory to be re-anchored. Step S311: Iterate through the elements in the deviation change list, extract the robot's deviation status, compare the robot's deviation status with the error tolerance. If the deviation is higher than the error tolerance, it is determined that the trajectory is abnormal and the trajectory is recorded; if the deviation is not higher than the error tolerance, the trajectory is determined to be normal. Step S312: Extract abnormal trajectories, collect the cumulative prediction error field of abnormal trajectories, extract the maximum value of the cumulative error in the cumulative prediction error field, locate the execution action corresponding to the cumulative error value, extract the corresponding trajectory of the execution action in the complete prediction trajectory, and obtain the trajectory to be re-anchored.
[0036] Step S32: Collect the predicted state corresponding to the action execution of the trajectory to be re-anchored, collect newly added external observation data, identify the newly added external observation data and the predicted state corresponding to the action execution, if there is a real-time observation state corresponding to the predicted state in the newly added external observation data, set the trajectory to be re-anchored to the orientation state re-anchoring; if there is no real-time observation state corresponding to the predicted state in the newly added external observation data, set the trajectory to be re-anchored to the historical trajectory re-anchoring, and mark the trajectory to be re-anchored by the orientation state re-anchoring and the historical trajectory re-anchoring.
[0037] Step S4: Based on the path diversion of the trajectory to be re-anchored, re-execute rolling prediction for the subsequent predicted trajectory to generate re-anchoring results; for the trajectory to be re-anchored with missing real-time observation status, re-perform deviation analysis and determine the anchoring results. Step S41: Extract the trajectory to be re-anchored for the directional state re-anchoring marked by the labeling result, record the execution action corresponding to the trajectory to be re-anchored, collect the real-time observation state corresponding to the execution action, take the real-time observation state as the trajectory starting point, extract the subsequent predicted trajectory of the trajectory starting point in the complete predicted trajectory, record the execution action of the subsequent predicted trajectory, map the trajectory starting point, re-execute rolling prediction on the subsequent predicted trajectory, and generate the re-anchoring result. Step S42: Extract the trajectory to be re-anchored from the historical trajectory re-anchored, obtain the subsequent predicted trajectory of the trajectory to be re-anchored in the complete predicted trajectory, collect the real-time observation status of the corresponding action of the subsequent predicted trajectory, independently verify the real-time observation status and the prediction status corresponding to the subsequent predicted trajectory, record the deviation between the real-time observation status and the prediction status, and determine the anchor point result.
[0038] Step S421: Obtain the cumulative error value of the re-anchored trajectory in the prediction error accumulation field, perform reverse processing on the re-anchored trajectory based on the cumulative error value, eliminate the error corresponding to the trajectory, map the elimination result to the real-time observation state of the subsequent predicted trajectory corresponding to the execution action, and optimize the real-time observation state. Step S422: Re-evaluate the error between the optimized real-time observation status and the subsequent predicted trajectory status, record the optimization deviation, compare the optimized deviation with the original deviation, if the deviation decreases, adopt the optimization result; if the deviation increases, mark the re-anchor point trajectory as a manually checked trajectory.
[0039] Step S5: Archive the rolling prediction process, re-anchoring records, and verification conclusions. At the same time, feed back the deviation pattern between the actual observation and the imagined trajectory to the world model training process for subsequent model iteration and improvement.
[0040] Step S51: Archive the rolling prediction process, re-anchoring records and verification conclusions, record the trajectory for re-anchoring in the rolling prediction process, extract the re-anchoring judgment process corresponding to the trajectory, build a re-anchoring experience base, and provide feedback on the robot's processing process.
[0041] Step S52: Connect the re-anchoring experience base to the robot. When the robot needs to re-anchor its trajectory, search the re-anchoring experience base to find the corresponding re-anchoring process and implement it on the robot. Verify the implementation result with the corresponding re-anchoring result in the re-anchoring experience base, perform re-anchoring control on the robot, and train and optimize the world model based on the re-anchoring control of the robot.
[0042] It should be noted that by building a re-anchoring experience base, data references are provided for the robot's trajectory re-anchoring process, reducing computing costs. At the same time, this method can accurately identify abnormal error trajectories, ensuring the robot's operational stability and safety.
[0043] Compared to the problems described in the background technology, this invention unifies single-step deviation, cumulative drift rate, and re-anchoring interval into quantitative indicators, avoiding the use of the entire trajectory to mask the specific source of deviation, accurately extracting the main stages of error generation, and achieving precise error processing, making the re-anchoring resource allocation more accurate. Furthermore, this invention verifies the prediction process through independent trajectory replay, improving the reliability of error judgment. Based on the extracted error trajectory, the verified part of the trajectory is retained unchanged, avoiding the computational waste caused by completely re-predicting from the task starting point and reducing the cost of re-prediction. Finally, this invention distinguishes between two components: characterization loss and dynamic modeling error, providing a clear direction for targeted model improvement. Based on specific attribution, the path of the re-anchoring trajectory is divided, balancing the accuracy of trajectory correction and computational cost. Therefore, the spatial intelligent long-term domain rolling prediction error accumulation re-anchoring method provided by this invention can reduce re-anchoring costs and ensure adjustment accuracy.
[0044] 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.
[0045] 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 spatially intelligent long time domain rolling prediction error accumulation re-anchoring method, characterized in that, include: Step S1: Collect the real-time observation status and execution tasks of the world model, decompose the execution actions based on the execution tasks, construct the execution action sequence, combine the real-time observation status, predict the subsequent state, and generate a complete prediction trajectory. Step S2: Collect the real-time observation status corresponding to the complete predicted trajectory, analyze the deviation between the predicted status and the real-time observation status, obtain the single-step deviation degree, statistically analyze the changes in the single-step deviation degree, evaluate the cumulative error, and construct the prediction error cumulative field; at the same time, extrapolate the prediction based on the changes in the single-step deviation degree, and record the extrapolation prediction deviation. Step S3: Obtain the task deviation standard constraint, mark the trajectory to be re-anchored in combination with the prediction error cumulative field, and perform path diversion on the trajectory to be re-anchored based on the real-time observation status corresponding to the complete prediction trajectory. Step S4: Based on the path splitting of the trajectory to be re-anchored, re-execute rolling prediction on the subsequent predicted trajectories to generate re-anchoring results; For trajectories that are missing real-time observation data and need to be re-anchored, a new deviation analysis is performed to determine the anchoring results; Step S5: Archive the rolling prediction process, re-anchoring records, and verification conclusions, and feed back the deviation patterns between actual observations and imagined trajectories to the world model training process.
2. The spatially intelligent long time domain rolling prediction error accumulation re-anchoring method according to claim 1, characterized in that, The specific steps of step S1 are as follows: Step S11: Record the world model currently observed by the robot to obtain the real-time observation state, obtain the robot's execution plan, and statistically analyze the robot's execution actions based on the robot's execution plan to construct a sequence of execution actions; Step S12: Based on the sequence of executed actions, the displacement status of the robot is collected to obtain spatial change information. The spatial change information is used as a change parameter to map the world model. Combined with the real-time observation status, the world model observed after the robot performs actions is predicted to obtain the predicted state. The predicted state after completing different actions is recorded by the sequence of executed actions to generate a complete predicted trajectory.
3. The spatially intelligent long time domain rolling prediction error accumulation re-anchoring method according to claim 1, wherein, The specific steps of step S2 are as follows: Step S21: Based on the complete predicted trajectory, extract the predicted state after the action is performed, collect the real-time observation state after the robot performs the corresponding action, feed back the real-time observation state to the predicted state, evaluate the deviation between the predicted state and the real-time observation state, and obtain the single-step deviation degree; statistically analyze the single-step deviation degree after performing different actions and construct a list of deviation changes. Step S22: Record the changes in single-step deviation from the deviation change list, evaluate the cumulative error generated by the action execution, and construct the prediction error cumulative field; analyze the error accumulation rate based on the prediction error cumulative field, and extrapolate the single-step deviation from the deviation change list to generate the extrapolated prediction deviation, and mark the extrapolated prediction deviation on the deviation change list.
4. The spatially intelligent long time domain rolling prediction error accumulation re-anchoring method according to claim 3, wherein, The specific steps of step S21 are as follows: Step S211: Collect data on static fixed objects on the world model and establish spatial anchor points; extract the predicted state after the action is performed from the complete predicted trajectory, map the spatial anchor points to the predicted state, record the anchor point coordinates in the predicted state, and obtain the predicted coordinates; collect the real-time observation state corresponding to the predicted state from the robot's action, record the anchor point coordinates in the real-time observation state, and obtain the observation coordinates; calculate the coordinate deviation between the predicted coordinates and the observation coordinates to obtain the single-step deviation degree. Step S212: Traverse the predicted state based on the complete predicted trajectory, collect real-time observations of the corresponding actions after execution based on the predicted state, calculate the single-step deviation of different actions, and statistically analyze the single-step deviation based on the action execution order of the complete predicted trajectory to construct a list of deviation changes.
5. The spatially intelligent long time domain rolling prediction error accumulation re-anchoring method according to claim 3, wherein, The specific steps of step S22 are as follows: Step S221: Extract the single-step deviation from the deviation change list, record the difference between adjacent single-step deviations, obtain the cumulative error value, perform statistics on the cumulative error value, and construct the prediction error cumulative field; The cumulative error field is used to iterate through the cumulative error values, extract the corresponding actions, record the order of the actions, associate the order of actions with the cumulative error, perform linear regression on the cumulative error, and record the rate of error accumulation. Step S222: Predict the error change of the next action by the error accumulation rate to obtain the extrapolated prediction accumulation; extract the known single-step deviation of the last execution from the deviation change list, combine it with the extrapolated prediction accumulation to obtain the extrapolated prediction deviation, add the extrapolated prediction accumulation as an extension item to the deviation change list, and mark the extrapolated prediction deviation.
6. The spatially intelligent long time domain rolling prediction error accumulation re-anchoring method according to claim 1, wherein, The specific steps of step S3 are as follows: Step S31: Based on the task deviation standard constraint, obtain the robot's error tolerance, judge the overall trajectory based on the deviation change list and the error tolerance, and combine the prediction error accumulation field to specifically locate and mark the anchor trajectory to be re-anchored. Step S32: Collect the predicted state corresponding to the action execution of the trajectory to be re-anchored, collect newly added external observation data, identify the newly added external observation data and the predicted state corresponding to the action execution, if there is a real-time observation state corresponding to the predicted state in the newly added external observation data, set the trajectory to be re-anchored to the orientation state re-anchoring; if there is no real-time observation state corresponding to the predicted state in the newly added external observation data, set the trajectory to be re-anchored to the historical trajectory re-anchoring, and mark the trajectory to be re-anchored by the orientation state re-anchoring and the historical trajectory re-anchoring.
7. The spatially intelligent long time domain rolling prediction error accumulation re-anchoring method according to claim 6, wherein, The specific steps of step S31 are as follows: Step S311: Iterate through the elements in the deviation change list, extract the robot's deviation status, compare the robot's deviation status with the error tolerance, and if the deviation is higher than the error tolerance, determine that the trajectory is abnormal and record the trajectory. If the deviation is not higher than the error tolerance, the trajectory is considered normal. Step S312: Extract abnormal trajectories, collect the cumulative prediction error field of abnormal trajectories, extract the maximum value of the cumulative error in the cumulative prediction error field, locate the execution action corresponding to the cumulative error value, extract the corresponding trajectory of the execution action in the complete prediction trajectory, and obtain the trajectory to be re-anchored.
8. The spatially intelligent long time domain rolling prediction error accumulation re-anchoring method of claim 1, wherein, The specific steps of step S4 are as follows: Step S41: Extract the trajectory to be re-anchored for the directional state re-anchoring marked by the labeling result, record the execution action corresponding to the trajectory to be re-anchored, collect the real-time observation state corresponding to the execution action, take the real-time observation state as the trajectory starting point, extract the subsequent predicted trajectory of the trajectory starting point in the complete predicted trajectory, record the execution action of the subsequent predicted trajectory, map the trajectory starting point, re-execute rolling prediction on the subsequent predicted trajectory, and generate the re-anchoring result. Step S42: Extract the trajectory to be re-anchored from the historical trajectory re-anchored, obtain the subsequent predicted trajectory of the trajectory to be re-anchored in the complete predicted trajectory, collect the real-time observation status of the corresponding action of the subsequent predicted trajectory, independently verify the real-time observation status and the prediction status corresponding to the subsequent predicted trajectory, record the deviation between the real-time observation status and the prediction status, if the deviation is greater than the deviation of the action corresponding to the complete predicted trajectory, mark it as manual verification.
9. The spatially intelligent long time domain rolling prediction error accumulation re-anchoring method of claim 8, wherein, The specific steps of step S42 are as follows: Step S421: Obtain the cumulative error value of the re-anchored trajectory in the prediction error accumulation field, perform reverse processing on the re-anchored trajectory based on the cumulative error value, eliminate the error corresponding to the trajectory, map the elimination result to the real-time observation state of the subsequent predicted trajectory corresponding to the execution action, and optimize the real-time observation state. Step S422: Re-evaluate the error between the optimized real-time observation status and the subsequent predicted trajectory status, record the optimization deviation, compare the optimized deviation with the original deviation, if the deviation decreases, adopt the optimization result; if the deviation increases, mark the re-anchor point trajectory as a manually checked trajectory.
10. The spatially intelligent long time domain rolling prediction error accumulation re-anchoring method of claim 1, wherein, The specific steps of step S5 are as follows: Step S51: Archive the rolling prediction process, re-anchoring records and verification conclusions, record the trajectory for re-anchoring in the rolling prediction process, extract the re-anchoring judgment process corresponding to the trajectory, build a re-anchoring experience base, and provide feedback on the robot's processing process. Step S52: Connect the re-anchoring experience base to the robot. When the robot needs to re-anchor its trajectory, search the re-anchoring experience base to find the corresponding re-anchoring process and implement it on the robot. Verify the implementation result with the corresponding re-anchoring result in the re-anchoring experience base, perform re-anchoring control on the robot, and train and optimize the world model based on the re-anchoring control of the robot.
