Robot Active Data Acquisition Method Based on Uncertainty and Action Cost

Abstract

This application relates to the field of robot learning and control technology, and in particular to a method for active robot data acquisition based on uncertainty and action cost. The method includes: constructing a set of machine learning models to predict the robot's state and calculating an uncertainty metric; generating a multi-step sequence of actions to be executed based on a comprehensive trade-off between the uncertainty metric and the action execution cost; using a control barrier function to perform safety verification and correction on the action sequence; controlling the robot to execute the corrected action sequence and collecting sensor interaction data for automatic annotation; and updating the machine learning model and uncertainty estimate using the newly acquired dataset. This application can effectively solve the problems of low robot data acquisition efficiency, insufficient coverage of boundary conditions, and difficulty in ensuring the safety of the exploration process, achieving safe and efficient autonomous data acquisition and continuous learning.

Description

Technical Field

[0001] This application relates to the field of robot learning and control technology, and in particular to a method for active robot data acquisition based on uncertainty and action cost. Background Technology

[0002] In the fields of embodied intelligence and robot control, data-driven learning paradigms have become mainstream, with algorithms such as deep reinforcement learning and imitation learning widely applied to complex tasks such as grasping, assembly, and mobile navigation. The performance of these methods is highly dependent on the scale and quality of the training data. Existing data acquisition schemes mainly include manual teleoperation teaching, stochastic exploration strategies, and active learning methods based on static samples. Manual teaching allows operators to directly control the robot to complete specific trajectories and record state-action pairs, enabling the acquisition of high-quality successful demonstrations. Stochastic exploration strategies introduce noise into the action space or use probability distribution sampling, allowing the robot to autonomously try various action combinations in the environment to collect interaction data. Traditional active learning techniques are mainly applied to static data processing scenarios such as image classification, using the calculation of the uncertainty of unlabeled samples to filter data points that require manual labeling.

[0003] However, in existing technologies, when faced with robotic tasks involving high-dimensional continuous action spaces and complex physical constraints, there are often problems such as low efficiency in effective data acquisition and difficulty in covering critical boundary conditions. Due to the lack of effective quantification of blind spots in model cognition and comprehensive consideration of the cost of action execution, existing methods are prone to getting bogged down in a large number of meaningless repetitive attempts or failing to reach sparse reward regions. At the same time, it is difficult to ensure physical safety during autonomous exploration, leading to an increased risk of equipment damage or the need for expensive human monitoring throughout the process. Summary of the Invention

[0004] This application provides a robot active data acquisition method based on uncertainty and action cost to solve the above problems. The method includes: S1. Construct a group of machine learning models, input the robot's real-time running status, predict the subsequent running status or task completion results, and calculate the uncertainty measure based on the differences between multiple sets of prediction results. S2. Based on the uncertainty measure of the current state, combined with the action execution cost, simultaneously calculate the execution cost of various actions, comprehensively weigh the benefits of reducing the degree of uncertainty and the action consumption cost, and generate a multi-step continuous sequence of actions to be executed. S3. Retrieve the robot's preset physical safety boundary standard, and perform safety verification on the selected multi-step continuous action sequence to be executed. If the action may cause the robot to enter an unsafe state, then correct it through the control barrier function and output the action sequence that can be executed. S4. Control the robot to execute the sequence of possible actions, collect sensor interaction data during the execution process, and automatically label it to form a standardized sample dataset; S5. Using the newly collected standardized sample dataset, update the machine learning model and update the uncertainty estimate for subsequent action selection.

[0005] Optionally, the process of constructing a set of machine learning models and calculating an uncertainty metric includes: The machine learning model group consists of an ensemble model or a Bayesian neural network, and the machine learning model group contains several predictive sub-models with differential biases in their initialization parameters. The same frame of the robot's real-time running state is synchronously input into all the prediction sub-models, and each model outputs an independent subsequent running state and a prediction probability distribution of the task completion result. The discrete differences of multiple predicted probability distributions are quantified using prediction variance or prediction entropy. The parameter uncertainty and data random uncertainty are separated and weighted and fused together to obtain the uncertainty measure in a unified manner.

[0006] Optionally, the step of uniformly calculating the uncertainty measure based on the differences between multiple sets of prediction results includes: The prediction results output by each prediction sub-model are converted into numerical vectors of the same dimension, and the variance and mean Euclidean distance between the vectors are calculated as the basic discrete features. Based on the historical prediction errors of each prediction sub-model, weights are assigned to reduce the interference of sub-models with poor prediction accuracy on the calculation of the basic discrete features, thereby obtaining a corrected dispersion index. The superposition weights of the two types of uncertainty components are dynamically adjusted in conjunction with the current exploration phase of the robot task, and the standardized uncertainty metric is output by mapping it to a fixed numerical range.

[0007] Optionally, the comprehensive trade-off between the benefits of reduced uncertainty and the costs of the actions includes: A predefined complete robot motion library is provided, and the cost of motion execution includes at least one of time cost, energy cost, and collision risk cost. Corresponding loss parameters are pre-calibrated for each motion in the complete robot motion library. Estimate the benefit of reducing the degree of uncertainty after executing each candidate action, and configure an adaptive balance weight value for the benefit of reducing the degree of uncertainty and the cost of executing the action. The balance weight value is dynamically adjusted synchronously with the current uncertainty metric. By incorporating robot endurance constraints and maximum steps per planning iteration into the trade-off logic, candidate action combinations that exceed energy consumption limits or deviate from the main task objective are filtered out, and a preliminary action combination set is output.

[0008] Optionally, generating a multi-step sequence of actions to be executed includes: Multi-step look-ahead rolling planning is carried out using model predictive control or Monte Carlo tree search to deduce the expected decrease in uncertainty after different combinations of candidate actions are executed, and to quantify the cumulative information gain corresponding to each combination. Construct an objective function with the cumulative information gain minus the weighted cumulative action consumption cost as its core, and iteratively optimize the initial action combination set; The optimal combination of consecutive actions for the objective function is selected to generate the multi-step consecutive action sequence that satisfies the step count constraint.

[0009] Optionally, the process of security verification and correction of the multi-step continuous action sequence to be executed includes: A preset physical safety boundary standard constitutes a robot safety set, which is defined by at least one of the following: joint rotation limit, motion speed limit, end contact force limit, and minimum distance between the robot body and the obstacle. A control barrier function is built using the robot’s real-time pose and torque state as input, and the control barrier function is used as a unified safety metric to determine whether the physical safety boundary standard has been touched. The robot's state evolution trajectory corresponding to the multi-step continuous sequence of actions to be executed is analyzed segment by segment. The control barrier function is substituted to identify dangerous motion segments. The trajectory speed and path inflection points are adjusted while taking into account information gain and action consumption cost. The sequence of possible actions that falls within the robot's safety set at all times is output.

[0010] Optionally, the process of collecting and automatically labeling the sensor interaction data includes: During the execution of the action sequence, the robot synchronously collects multi-source raw sensor interaction data, which includes the robot's real-time status, issued action commands, and task interaction results. Simultaneously extract the corresponding auxiliary label parameters, which include sampling timestamp, real-time uncertainty metric, single-step action information gain, safety boundary trigger flag, and task success / failure label; The time-series matched sensor interaction data is bound one by one with all the auxiliary label parameters, the data storage format is unified, abnormal sensor data fragments are removed, and a complete standardized sample dataset is generated.

[0011] Optionally, the process of removing anomalous sensor data segments includes: For each type of sensor interaction data, an amplitude threshold and a sampling continuity threshold are set, and all time-series sensor interaction data are traversed in segments. Invalid data segments with excessive amplitude, interrupted sampling, or numerical abrupt changes exceeding the normal fluctuation range are directly removed; For valid data with only slight noise, a moving mean filter is used for smoothing, retaining valid samples corresponding to complete action interactions, thus completing the preprocessing of the standardized sample dataset.

[0012] Optionally, the process of updating the machine learning model includes: The standardized sample dataset is divided into multiple training mini-batch validation subsets, and mini-batch gradient descent is used to synchronously incrementally iterate all the prediction sub-models within the machine learning model group. The total number of newly added valid samples within a single collection cycle is counted in real time, an adaptive mapping relationship between the number of samples and the frequency of model iterations is established, and the update frequency of the machine learning model group is dynamically adjusted. The prediction accuracy of each prediction sub-model is evaluated and updated using the training mini-batch validation subset. The historical error weights corresponding to each prediction sub-model are updated synchronously to complete the overall parameter optimization of the machine learning model group.

[0013] Optionally, the process of updating the uncertainty estimate includes: Based on the ensemble model or Bayesian neural network after parameter iteration, re-solve the two types of components: parameter uncertainty and data random uncertainty. Adjust the fusion weights of the two types of uncertainty components based on the current task exploration depth, and recalculate the global total uncertainty metric. The updated uncertainty metric is cached in the action planning module as the core input parameter for generating the multi-step continuous action sequence to be executed in the next round.

[0014] This application provides a robot active data acquisition method based on uncertainty and action cost. This method constructs a set of machine learning models to predict the robot's real-time operating state using multiple models, and calculates a unified uncertainty metric based on the differences in the prediction results, thereby accurately quantifying the current knowledge blind spots. Based on this, the uncertainty metric is jointly optimized with the costs of action execution time, energy consumption, and risk. Multi-step continuous action sequences are generated through model predictive control or Monte Carlo tree search, enabling the robot to actively select action paths with the maximum information gain and reasonable execution cost, avoiding the resource waste caused by blind random exploration. Furthermore, a control barrier function is used to rigorously verify and correct the generated action sequences for safety, ensuring that all executed actions fall within a safe set defined by joint limits, speed limits, and contact force limits, effectively solving the problem of equipment damage or safety accidents that easily occur in traditional autonomous exploration. Subsequently, the robot executes the corrected safe action sequence and simultaneously collects multi-source sensor interaction data. A standardized sample dataset is formed by automatically labeling meta-information such as timestamps, uncertainty values, and task results. This dataset is then used to incrementally update the machine learning model and its uncertainty estimate online, achieving a closed-loop feedback between perception, decision-making, execution, and learning. Therefore, this solution not only improves the efficiency of high-value data collection and coverage of boundary conditions, but also ensures continuous and safe operation in real physical environments, providing data support for the training of embodied intelligence strategies. Attached Figure Description

[0015] 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.

[0016] Figure 1 A flowchart of a robot active data acquisition method based on uncertainty and action cost provided in an embodiment of this application; Figure 2 This is a schematic diagram of the uncertainty estimation of the integrated model in an embodiment of the present invention; Figure 3 This is a flowchart of the look-ahead planning action selection process based on MPC in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the safety correction of the control barrier function in an embodiment of the present invention; Figure 5 This is a timing diagram of online incremental data acquisition and model update in an embodiment of the present invention. Detailed Implementation

[0017] 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, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0018] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article, unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship.

[0019] The embodiments of this application will now be described in further detail with reference to the accompanying drawings.

[0020] The specific implementation method can be referred to in the following embodiments, wherein the data mentioned in the embodiments are only for reference and examples, so that relevant personnel can better understand them.

[0021] Figure 1 This is a flowchart illustrating a robot active data acquisition method based on uncertainty and action cost, provided in one embodiment of this application. The method of this embodiment can be applied to the server in the above scenario. (See also...) Figure 1 The process, specifically the implementation steps, are as follows: Example 1: In the field of robot learning and control technology, the performance of control strategies based on deep reinforcement learning and imitation learning is highly dependent on the scale and quality of training data. However, collecting data in real physical environments faces many challenges: traditional manual teaching methods require operators to collect successful trajectories through teleoperation or drag-and-drop teaching, which is extremely labor-intensive and inefficient; when using random exploration strategies, robots often perform a large number of meaningless actions, resulting in a very low percentage of effective data and making the equipment susceptible to damage due to collisions; furthermore, for sparse reward tasks, random exploration struggles to reach key success states, making it difficult for the strategy to learn core skills. Existing active learning methods mostly target label queries for static samples, failing to fully consider the physical constraints, execution costs, and real-time safety during the execution of robot action sequences. There is an urgent need for a method that can autonomously identify high-value states, proactively execute actions with the highest information content, and efficiently collect data within safe boundaries.

[0022] To address the aforementioned problems, this application provides a method for active robot data acquisition based on uncertainty and action cost, comprising the following steps: Step 1: Build a set of machine learning models, input the robot's real-time running status, predict the subsequent running status or task completion results, and calculate the uncertainty measure based on the differences between multiple sets of prediction results. The construction of a machine learning model group can refer to establishing a set of predictive sub-models with differentiated initialization parameters or training data order. This model group specifically includes ensemble models or Bayesian neural network structures. The robot's real-time operating state comes from real-time feedback from the robot's own sensors and external environment perception devices, including joint angles, end-effector pose, visual features, and torque information. Predicting subsequent operating states or task completion results can involve using the sub-models within the model group to process the same frame of input state in parallel, outputting independent predictive distributions of the state transition probability or task success rate at the next moment. Uncertainty measurement is obtained by quantifying the dispersion between multiple sets of predictive probability distributions. Its role is to characterize the current model's cognitive blind spots or areas of insufficient confidence for a specific state. Specifically, the system calculates the variance or entropy value between the output vectors of each sub-model, weightedly fusing parameter uncertainty and data random uncertainty to uniformly convert it into a scalar form of uncertainty measurement. For example, when the robot faces a smooth curved object it has never seen before, the predicted success rate of grasping the object by different sub-models in the ensemble model varies greatly, resulting in a high calculated prediction variance, indicating that the uncertainty measurement in this state is high. This multi-model differentiated prediction mechanism can accurately pinpoint knowledge gaps, providing a quantitative basis for subsequent proactive exploration. This step aims to estimate the uncertainty of the current state or model online, avoiding ineffective random exploration and providing core guiding indicators for action selection.

[0023] Step 2: Based on the uncertainty measurement of the current state, combined with the cost of action execution, simultaneously calculate the cost of various actions, comprehensively weigh the benefits of reducing the degree of uncertainty with the cost of action execution, and generate a multi-step continuous sequence of actions to be executed; The execution cost of an action includes at least one of the following: time cost, energy cost, and collision risk cost. This is a pre-calibrated loss parameter for each action within the robot's complete action library. The benefit of reduced uncertainty can refer to the estimated reduction in system uncertainty after executing candidate actions, i.e., information gain. The comprehensive trade-off can refer to introducing adaptive balancing weights, which are dynamically adjusted synchronously with the current uncertainty level to find the optimal solution between exploring high-value information and controlling execution costs. Specifically, the system employs model predictive control or Monte Carlo tree search to conduct multi-step look-ahead rolling planning, deduce the expected reduction in uncertainty after executing different combinations of candidate actions, construct an objective function with the cumulative information gain minus the weighted cumulative action consumption cost as its core, and iteratively optimize the initial set of action combinations. For example, when the remaining battery power is low, the system automatically increases the weight of energy cost, prioritizing action combinations that are short-path, low-power, but still significantly reduce uncertainty; while when the battery power is sufficient and the system is in a high-uncertainty region, it tends to select actions that maximize information gain, even if they take slightly longer. Through this dynamic trade-off mechanism, the generated multi-step sequence of actions to be executed satisfies both the step count constraint and ensures maximum data acquisition efficiency with limited resources. This step solves the technical challenge of selecting the action sequence that maximizes information gain while considering physical execution costs, thus realizing a shift from blind exploration to purposeful exploration.

[0024] Step 3: Retrieve the robot's preset physical safety boundary standards and perform safety checks on the selected multi-step continuous action sequence to be executed. If the action may cause the robot to enter an unsafe state, correct it by controlling the barrier function and output an action sequence that can be executed. The preset physical safety boundary standards constitute the robot's safety set, specifically defined by at least one of the following: joint rotation limits, motion speed limits, end-effector contact force limits, and minimum distance between the robot and obstacles. The control barrier function is a safety metric function built using the robot's real-time pose and torque state as input, used to determine whether a sequence of actions will violate the physical safety boundary. Specifically, the system analyzes the robot's state evolution trajectory corresponding to a multi-step sequence of actions to be executed, and substitutes it into the control barrier function for identification. If a trajectory segment is detected that would cause the robot's state to jump out of the safety set (e.g., the expected contact force exceeds the threshold or the joint angle exceeds the limit), then, without considering information gain and motion cost, a quadratic programming algorithm is used to adjust the velocity curve or path inflection point of the trajectory, forcibly pulling the state back into the safe area. For example, when the planned action sequence indicates that the robotic arm is rapidly approaching an obstacle, the control barrier function will intervene immediately, generating a deceleration and fine-tuning direction correction command to ensure that the distance between the end-effector and the obstacle is always greater than the minimum safe distance. Through this proactive safety correction mechanism, the output action sequence ensures that all possible actions fall within the robot's safety set at all times. This step ensures the safety of the active exploration process, prevents robots from performing destructive actions, and enables active data collection to be deployed on real robots without the need for constant human monitoring.

[0025] Step 4: Control the robot to execute a sequence of actionable actions, collect sensor interaction data during the execution process, and automatically label the data to form a standardized sample dataset; The sensor interaction data includes multi-source raw data such as the robot's real-time state, issued action commands, and task interaction results, which are collected synchronously during the robot's execution of a sequence of movable actions. Automatic annotation refers to the synchronous extraction of corresponding auxiliary label parameters and the binding of time-matched sensor interaction data with all auxiliary label parameters. The auxiliary label parameters specifically include sampling timestamps, real-time uncertainty metrics, single-step action information gain, safety boundary trigger flags, and task success / failure labels. Specifically, while recording the state-action-result triplet, the system automatically removes abnormal data segments with excessive amplitude or sampling interruptions, smooths valid data with slight noise using a moving average filter, and standardizes the data storage format. For example, in a grasping attempt, the system not only records joint current and camera images but also automatically labels data with high uncertainty, no safety boundary triggered, and grasping failure, forming a complete structured sample record. Through this synchronous acquisition and automatic annotation method, a high-quality standardized sample dataset is formed, which can be directly used for training downstream control strategies. This step automates and standardizes the data acquisition process, increases the proportion of valid samples, and avoids the high cost of manual annotation.

[0026] Step 5: Use the newly collected standardized sample dataset to update the machine learning model and the uncertainty estimate for subsequent action selection.

[0027] Updating the machine learning model can refer to using a mini-batch gradient descent algorithm to synchronously incrementally iterate all predictive sub-models within the model group. Updating the uncertainty estimate involves resolving the two components—parameter uncertainty and data random uncertainty—based on the model after parameter iteration, adjusting the fusion weights according to the current task exploration depth, and recalculating the global total uncertainty metric. Specifically, the system continuously counts the total number of newly added valid samples within a single collection cycle, establishes an adaptive mapping relationship between the number of samples and the model iteration frequency, and dynamically adjusts the update frequency. The prediction accuracy of each updated predictive sub-model is evaluated using a training mini-batch validation subset, and the historical error weights corresponding to each sub-model are updated synchronously to complete overall parameter optimization. For example, after collecting 50 new samples, the system automatically triggers a model fine-tuning, improving the model's prediction accuracy for previously high-uncertainty areas (such as object edge slippage), thereby reducing the uncertainty metric in that area and guiding the robot to move to other unknown areas in the next round of exploration. By caching the updated uncertainty metric to the action planning module as the core input parameter for generating a multi-step continuous action sequence in the next round, a closed-loop feedback of perception-decision-execution-learning is achieved. This step endows the system with continuous learning capabilities, enabling the exploration strategy to adaptively adjust as the model's accuracy improves, gradually transitioning from extensive exploration in the early stages to refined data collection in the later stages.

[0028] This application constructs an efficient closed-loop method for robot active data acquisition through the synergistic effect of the aforementioned technical features. By constructing a set of machine learning models and calculating uncertainty metrics, the robot can autonomously identify current knowledge blind spots. Based on this, a comprehensive trade-off and multi-step forward planning are performed, combined with the cost of action execution, to generate action sequences that combine high information gain and low execution cost. Furthermore, a control barrier function is introduced to perform real-time safety verification and correction of the action sequences, ensuring that the exploration process always stays within the physical safety boundary and avoiding the risk of equipment damage. Subsequently, multi-source data is collected synchronously during execution and automatically labeled to form a standardized sample dataset. Finally, the newly collected data is used to update the model and uncertainty estimation online, realizing dynamic optimization of the exploration strategy. This integrated mechanism not only improves data acquisition efficiency and avoids a large number of meaningless random exploration actions, but also effectively covers boundary cases in sparse reward scenarios. At the same time, it incorporates active safety protection, enabling the robot to autonomously complete data acquisition and continuous model evolution without human intervention.

[0029] Example 2: In another optional embodiment, this application provides a method for constructing a set of machine learning models and calculating an uncertainty metric, the method comprising the following steps: like Figure 2 As shown, this embodiment uses an ensemble model composed of multiple prediction sub-models to estimate the uncertainty of the robot's current operating state. The robot's real-time operating state ( The data is synchronously input into each prediction sub-model. Each sub-model, based on its differing initialization parameters or training process, outputs a predicted distribution of the next time step's running state or task completion result. Subsequently, the prediction variance, prediction entropy, and discrete differences among the prediction distributions are quantified, and corresponding weights are assigned based on the historical prediction errors of each sub-model to reduce the interference of sub-models with lower prediction accuracy on the estimation results. Furthermore, the parameter uncertainty reflecting model cognitive divergence and the data random uncertainty reflecting environmental noise and sensor perturbations are separated, and fusion weights are assigned according to the current task exploration stage to obtain an uncertainty metric mapped to a preset numerical range. This uncertainty metric is used to characterize the extent to which the machine learning model group has sufficient understanding of the current robot state, and serves as input for subsequent candidate action evaluation and prospective planning. After the model is updated using newly collected standardized sample data, the parameters and historical error weights of each prediction sub-model are updated synchronously, so that the subsequent uncertainty metric changes dynamically with the model learning process.

[0030] Step 1: The machine learning model group consists of an ensemble model or a Bayesian neural network. The machine learning model group contains several predictive sub-models with different initialization parameters. In this context, a machine learning model set refers to a collection of models used to quantify the uncertainty of key outputs in the robot's current state. Its specific architecture includes ensemble models or Bayesian neural networks. In an ensemble model approach, the model set consists of M predictive sub-models with identical structures but differing initialization parameters. For example, 5 to 10 neural networks are trained, with the initial weights of each network randomly set or fine-tuned using training data in different orders. This ensures that each sub-model produces differentiated predictions when faced with the same input. In a Bayesian neural network approach, a probability distribution is introduced for the network weights, and the effects of multiple sub-models are simulated using variational inference or Monte Carlo Dropout to approximate the posterior distribution. The role of the predictive sub-models is to independently predict subsequent operating states, task completion results, or physical quantities such as contact forces. For example, when the robot performs a grasping task, each sub-model in the model set receives the same joint angles and visual feature inputs. However, due to differences in internal parameters, sub-model A might predict a success probability of 0.9, while sub-model B might predict 0.6. This output discrepancy is the basis for quantifying uncertainty. By constructing a model group containing several predictive sub-models with differentiated biases in initialization parameters, we can capture cognitive blind spots caused by insufficient training data or model capacity limitations at the model structure level, providing necessary multi-source prediction benchmarks for distinguishing between parameter uncertainty and data random uncertainty.

[0031] Step 2: Synchronously input the same frame of the robot's real-time running status into all prediction sub-models, and output independent subsequent running status and prediction probability distribution of task completion results respectively; The robot's real-time operational state refers to its physical perception data at the current moment, specifically including its real-time body pose, joint angles, end effector speed, torque sensor readings, and environmental visual feature vectors. This step involves distributing the aforementioned state data from the same frame in parallel to all prediction sub-models within the model group. Each sub-model performs forward inference based on its unique parameter distribution, outputting an independent prediction probability distribution. The prediction probability distribution describes the possibility space of future state transitions or task outcomes (such as success / failure) given the current state. For example, in a scenario where a quadruped robot walks on unknown terrain, given the current body posture and foot contact force data, the five sub-models might each output a probability density function for the stability score at the next moment. Some models predict high stability (mean 0.8, small variance), while others predict low stability (mean 0.4, large variance). These independent outputs not only contain the expected value of the state but also retain complete distributional information, allowing the system to simultaneously observe the central tendency and dispersion of the prediction results, thus providing direct data support for quantifying the differences between multiple sets of prediction results.

[0032] Step 3: Quantify the discrete differences of multiple predicted probability distributions using prediction variance or prediction entropy, separate parameter uncertainty from data random uncertainty and weighted fusion, and uniformly convert to obtain the uncertainty measure.

[0033] Uncertainty metric is a standardized numerical measure of the degree of ambiguity in a robot's understanding of its current state. Specifically, prediction variance or prediction entropy is first used as a mathematical tool to quantify the discrete differences among multiple predicted probability distributions: prediction variance is characterized by the sum of squared Euclidean distances between the means of each sub-model's output, reflecting the degree of disagreement between different models; prediction entropy is characterized by the information entropy of the average predicted distribution, reflecting the degree of hesitation within a single model in determining the outcome. Based on this, total uncertainty is broken down into two components: parameter uncertainty and data random uncertainty. Parameter uncertainty mainly stems from the model's own lack of knowledge, manifested as significant differences in prediction results between sub-models (i.e., high variance), which can be reduced by increasing training data or optimizing the model structure. Data random uncertainty mainly originates from inherent environmental noise or unpredictability (such as slippery ground or sensor noise), manifested as distribution dispersion that cannot be eliminated even with a perfect model (i.e., high entropy). Subsequently, the weights of the two components are dynamically adjusted and weighted according to the current task exploration stage for fusion. For example, parameter uncertainty is given higher weight in the early stages of exploration to guide model learning, while data random uncertainty is emphasized in the later stages to identify environmental boundaries, ultimately mapping to a standardized uncertainty metric output within a fixed numerical range. By separating parameter uncertainty from data random uncertainty and then weighting and fusing them, this step can more accurately pinpoint the source of uncertainty, avoid misjudging environmental noise as model defects, or model ignorance as environmental complexity, thereby improving the interpretability and guiding value of uncertainty assessment.

[0034] This application constructs a machine learning model group consisting of several predictive sub-models with differentiated initialization parameters, and utilizes prediction variance or prediction entropy to quantify the discrete differences of multiple sets of predicted probability distributions, achieving effective decomposition and weighted fusion of parameter uncertainty and data random uncertainty. Specifically, the architecture design of the ensemble model or Bayesian neural network ensures that each sub-model can output an independent prediction distribution under the same frame of robot real-time operating state input. This structural difference provides a basis for capturing the model's cognitive blind spots. Furthermore, by mathematically quantifying the discreteness of multiple sets of prediction results, the system can distinguish between parameter uncertainty caused by insufficient training and data random uncertainty caused by environmental noise. Finally, the unified uncertainty measure obtained by weighted fusion of the two not only comprehensively reflects the degree of unknown in the current state, but also provides a decision-making basis for subsequently weighing the benefits of reducing the degree of unknown against the cost of action consumption. This allows the robot to prioritize exploration of high-value areas that can reduce parameter uncertainty, while avoiding ineffective actions caused by pure noise interference, thereby improving the efficiency and quality of data acquisition while ensuring safety.

[0035] Example 3: In another optional embodiment, this application provides a specific implementation method for uniformly calculating an uncertainty measure based on the differences between multiple sets of prediction results.

[0036] Step 1: Convert the prediction results output by each prediction sub-model into numerical vectors of the same dimension, and calculate the variance and mean Euclidean distance between the vectors as the basic discrete features. The prediction results can refer to the independent predicted values ​​output by each prediction sub-model in the aforementioned machine learning model group for the same input state. These predicted values ​​may be expressed as scalar probabilities, multidimensional state vectors, or task completion scores. To quantify the degree of divergence between different sub-models, it is first necessary to standardize the heterogeneous or high-dimensional outputs, mapping them to numerical vectors of a uniform dimension for mathematical statistical operations. Specifically, if the sub-model output is a probability distribution, its mean vector or key quantiles can be extracted to form feature vectors; if the output is the state coordinates at the next time step, its coordinate vector is directly taken. After obtaining a set of numerical vectors of the same dimension, the variance of all vectors in the set is calculated to measure the overall dispersion trend. At the same time, the Euclidean distance between any two vectors is calculated and the average value is obtained, which serves as the basic discrete feature. Variance reflects the fluctuation range of the prediction results around the central value, while the mean Euclidean distance intuitively represents the magnitude of the differences between each pair of sub-models. For example, in an integrated system containing five sub-models, if the capture success rates predicted by the five sub-models in a certain state are [0.8, 0.82, 0.79, 0.4, 0.81], after being converted into vectors, the significant difference between 0.4 and the other values ​​will lead to a substantial increase in the calculated variance and mean Euclidean distance, thus initially indicating a high degree of uncertainty in that state. By transforming multi-source predictions into statistical features in this way, abstract model divergences can be quantified into calculable physical indicators, providing a data foundation for subsequent uncertainty assessment.

[0037] Step 2: Based on the historical prediction errors of each prediction sub-model, assign weights to reduce the interference of sub-models with poor prediction accuracy on the calculation of basic discrete features, and obtain the corrected dispersion index. The weights are dynamically generated confidence coefficients based on the performance of each prediction sub-model over a past period. In actual operation, some sub-models may exhibit persistently low prediction accuracy due to excessive initialization bias, incomplete training data coverage, or overfitting to specific scenarios. Directly incorporating the outputs of these low-quality sub-models into discrete feature calculations would introduce noise and even mislead the direction of uncertainty estimation. Therefore, the system maintains a historical prediction error record table, continuously calculating the deviation (e.g., mean squared error or absolute error) between the predicted values ​​and actual observations of each sub-model over the most recent N time steps. Weights are then assigned in reverse based on this historical error, with sub-models exhibiting smaller historical errors receiving higher weights and those with larger historical errors receiving lower weights or even zero weight. When calculating the basic discrete features, a weighted variance or weighted distance formula is used, ensuring that the prediction results of high-precision sub-models dominate the discreteness calculation, while the influence of low-precision sub-models is significantly suppressed. For example, if the historical average error of sub-model A is 0.05 and the historical average error of sub-model B is 0.3, then when calculating the dispersion of the current state, the output weight of sub-model A might be set to 0.9, and the weight of sub-model B might be set to 0.1. The resulting corrected dispersion index can effectively eliminate spurious uncertainties caused by bad models, ensuring that the uncertainty measure truly reflects the randomness of the data itself or the blind spots in the model's cognition, rather than the failure of individual models.

[0038] Step 3: Dynamically adjust the superposition weights of the two types of uncertainty components in combination with the current robot task exploration phase, and map them to a fixed numerical range to output a standardized uncertainty metric.

[0039] The two types of uncertainty components refer to parameter uncertainty (originating from the uncertainty of the model's own parameters, which usually decreases as data increases) and data random uncertainty (originating from environmental noise or inherent randomness, which is usually impossible to eliminate). The robot's focus on these two types of uncertainty differs at different stages of task exploration: in the early stages, when model parameters have not yet converged and parameter uncertainty is dominant, the robot should be encouraged to explore areas unknown to the model for rapid learning; in the later stages, when the model is more mature and data random uncertainty becomes the main challenge, the focus should be on areas with complex and variable environments. Therefore, the system dynamically adjusts the weighting of these two components in the final uncertainty metric based on the current task exploration depth (e.g., the total number of samples collected, the degree of model convergence, or the time progress). Specifically, a parameter uncertainty weight coefficient that monotonically decreases with the exploration progress and a correspondingly increasing data random uncertainty weight coefficient can be set. The corrected dispersion index is then decomposed or associated with these two components and weighted and fused. Finally, to facilitate unified threshold judgment and cost calculation in subsequent action planning modules, the fused values ​​are mapped to a fixed numerical range (e.g., [0,1]) using a sigmoid function or linear normalization method, outputting a standardized uncertainty metric. For example, in the first 100 iterations of the task, the parameter uncertainty weight is set to 0.8, and the data random uncertainty weight is set to 0.2, allowing the robot to prioritize exploring the model's cognitive gaps. After more than 1000 iterations, the weights are automatically adjusted to 0.2 and 0.8, guiding the robot to finely characterize the random boundaries of the environment. Through this dynamic adjustment and standardized output mechanism, the uncertainty metric not only possesses robustness but also the ability to adaptively evolve with the learning process, thereby guiding the robot to maintain an efficient proactive data acquisition strategy throughout its entire lifecycle.

[0040] This application achieves highly robust and adaptive uncertainty quantification through the synergistic effect of the above steps. First, by vectorizing the prediction results and calculating the variance and Euclidean distance, a mathematical foundation for quantifying model divergence is established. Next, a weighting mechanism based on historical errors is introduced to effectively filter out interference from low-quality sub-models, ensuring the purity and reliability of the dispersion index. On this basis, by dynamically adjusting the fusion weights of parameter uncertainty and data random uncertainty during the task exploration phase, the uncertainty metric can sensitively perceive changes in the learning process, smoothly transitioning from early broad-based exploration to later intensive data collection. Finally, the uncertainty metric output after standardized mapping provides a unified, stable, and semantically meaningful input parameter for subsequent action cost trade-offs and safety verification, improving the overall efficiency and safety of the robot's active data collection.

[0041] Example 4: In one optional embodiment, this application provides a method for comprehensively balancing the benefits of reducing the degree of unknown and the cost of motion consumption. The method further includes steps such as predefining a complete robot motion library and calibrating loss parameters, estimating the benefits of reducing the degree of unknown and configuring adaptive balance weights, and introducing constraints to filter candidate motion combinations.

[0042] Step 1: Predefine a complete robot motion library. The cost of motion execution includes at least one of the following: time cost, energy cost, and collision risk cost. Pre-calibrate the corresponding loss parameters for each motion in the complete robot motion library. The complete robot motion library can refer to a collection of atomic actions or motion primitives that the robot controller can execute. Its source can be a discretized motion space generated in advance through motion planning algorithms, or typical motion patterns obtained by clustering teaching data. The cost of motion execution is used to quantify the physical resources and safety costs incurred in executing a single action. Specifically, this includes: time cost, i.e., the estimated time required for the action to move from the initial state to the final state; energy cost, usually estimated by integrating the product of joint torque and angular velocity over the time domain; and collision risk cost, calculated based on the reciprocal of the minimum distance between obstacles in the current environment map and the robot's expected trajectory. Pre-calibrating corresponding loss parameters for each action can refer to establishing an index for each action entry in the motion library during the offline or initialization phase through simulation or limited field testing, recording its baseline time, baseline energy consumption, and basic risk coefficient under standard operating conditions. For example, for a six-axis robotic arm to move forward 10cm, the pre-calibrated time cost is 0.5 seconds and the energy cost is 12 joules. If executed in a narrow passage, the collision risk cost will be dynamically amplified based on the real-time perceived distance to obstacles. Through this pre-calibration mechanism, the system can query the basic cost of any candidate action within milliseconds, providing data support for subsequent real-time optimization. This step aims to construct an action search space with cost attributes, ensuring that subsequent action selection not only considers information acquisition capabilities but also the economy and safety of execution.

[0043] Step 2: Estimate the benefit of reducing the degree of uncertainty after executing each candidate action, and configure adaptive balancing weight values ​​for the benefit of reducing the degree of uncertainty and the cost of executing the action. The balancing weight values ​​are dynamically adjusted in sync with the current uncertainty metric. The benefit of reduced uncertainty can refer to the reduction in the robot's uncertainty about its environment or itself after performing a candidate action. Its calculation depends on the above embodiments and the machine learning model set constructed therein. Specifically, the system utilizes the current uncertainty metric. With predicting the execution action Post-expectation uncertainty The difference is used as the profit value The adaptive balancing weight is a scalar coefficient. This weight is used to adjust the relative importance of the benefit and cost terms in the objective function. The weight is set using a dynamic mapping strategy: when the uncertainty of the current state is high (indicating that the model has a vague understanding of the current region and urgently needs exploration), it is automatically reduced. The value of the uncertainty metric is used to highlight the dominance of the benefit of reduced uncertainty in the objective function, encouraging the robot to perform actions with high information gain; conversely, when the uncertainty metric is small (indicating that the model is relatively accurate), it automatically increases the value of the uncertainty metric. The value makes the cost of performing actions a primary consideration, prompting robots to choose low-risk, low-energy, conservative actions. For example, setting... ,in As the benchmark weight, As a regulating factor, when When it decreases from 0.8 to 0.2, The corresponding increase in the value allows the system to smoothly switch from an aggressive exploration mode to a robust utilization mode. The benefits of reduced unknowns and the costs of action execution are combined through this adaptive weight to form a joint optimization objective function. This achieves an adaptive balance between exploration efficiency and operating costs at different cognitive stages.

[0044] Step 3: Introduce robot endurance constraints and maximum number of steps per planning session constraints into the trade-off logic, filter out candidate action combinations that exceed the energy consumption limit or deviate from the main task objective, and output a preliminary action combination set.

[0045] The robot's endurance constraint can be defined as a total energy consumption limit set based on the current remaining battery power or hydraulic system pressure, ensuring that the planned action sequence will not cause the robot to stop midway through the task due to energy depletion. The maximum number of steps per planning cycle constraint can be defined as the upper limit of the number of actions that can be planned continuously within a decision cycle, used to ensure real-time calculation and prevent error accumulation caused by excessively long planning horizons. In the trade-off logic, the system first sorts all candidate actions or short sequences based on the joint optimization objective function calculated above, and then applies the above hard constraints for screening: action combinations whose cumulative energy consumption exceeds the remaining battery power safety threshold (e.g., 20% of the remaining battery power) and long sequences whose length exceeds the preset maximum number of steps (e.g., 5 steps) are eliminated. In addition, candidate combinations that, although locally high, will cause the robot's overall pose to deviate significantly from the main task target area (e.g., the grasping work area) are also filtered out. For example, if a certain action combination can bring extremely high uncertainty reduction benefits, but its execution energy consumption will cause the remaining battery power to fall below the warning line, then the combination will be directly filtered out. After multi-level filtering, the remaining legal action combinations constitute the preliminary action combination set. The output of this step provides a candidate pool that has been verified for safety and feasibility for generating a multi-step sequence of actions to be executed in the subsequent embodiments described above, effectively avoiding the risk of energy crisis or task failure caused by blindly pursuing information gain.

[0046] This application achieves a deep synergy between the benefits of reducing the degree of uncertainty and the costs of actions by predefining an action library with loss parameters, constructing adaptive balancing weights that dynamically adjust with uncertainty, and introducing a multi-hard constraint filtering mechanism. Specifically, the predefined loss parameters provide a unified benchmark for cost quantification, allowing costs in different dimensions (time, energy consumption, and risk) to be uniformly incorporated into the optimization framework; the adaptive balancing weights act as an intelligent regulating valve, dynamically adjusting the aggressiveness of the exploration strategy based on the model's current cognitive confidence, solving the problem that fixed weights cannot adapt to changing environments; and the endurance and step constraints serve as a bottom-level safety net, ensuring the physical feasibility of the optimization results. The combined use of these three elements enables the robot to boldly explore in high-uncertainty areas to quickly acquire key data, while automatically switching to a conservative strategy in low-uncertainty areas or under resource-constrained conditions. This ensures data acquisition quality while extending the robot's continuous operating time and reducing the risk of equipment damage.

[0047] Example 5: In one optional embodiment, this application provides a method for generating a multi-step continuous sequence of actions to be executed, the method comprising the following steps: like Figure 3 As shown, when using model predictive control for look-ahead planning, the robot's current state is first obtained ( ) and the uncertainty measure corresponding to the current state ( ), and generate a set of candidate actions based on a predefined complete robot action library. For each candidate action sequence, a set of machine learning models is used to extrapolate the future within a preset rolling time domain. The robot's state trajectory is calculated for each step, and the uncertainty changes corresponding to each step state and the cost of action execution are predicted. The cost of action execution includes at least one of time cost, energy cost, and collision risk cost. Under the conditions of satisfying the robot's endurance limit, the planned number of steps limit, and the main task objective constraints, the cumulative information gain of each candidate action sequence is calculated. ) and cumulative action consumption cost ( ), and in accordance with ( Construct a comprehensive evaluation objective for candidate action sequences, where, ( The weights are dynamically adjusted based on the current uncertainty measure and the robot's remaining resources. By comparing the objective function values ​​of each candidate action sequence, the action sequence with the optimal objective function value is selected, and a multi-step action sequence to be executed, arranged consecutively from the current moment, is output. After the robot performs at least one of the actions, the next round of rolling prediction and action sequence optimization is carried out based on the latest acquired robot state to reduce the impact of long-term prediction errors on the action selection results.

[0048] Step 1: Employ model predictive control or Monte Carlo tree search to conduct multi-step look-ahead rolling planning, deduce the expected decrease in uncertainty after the execution of different candidate action combinations, and quantify the cumulative information gain corresponding to each combination; Multi-step forward rolling planning refers to the process of simulating and extrapolating the evolution of the system state within a finite future time horizon using a pre-built set of machine learning models as internal dynamic models. Specifically, Model Predictive Control (MPC) solves the open-loop optimal control problem within a finite time domain and re-optimizes based on the latest state feedback at each sampling time, achieving closed-loop rolling optimization; Monte Carlo Tree Search (MCTS) constructs a search tree through random sampling and uses statistical simulation to evaluate the value of different action paths. During the extrapolation process, the system takes the preliminary action combination set selected in the above embodiments as input, substitutes it into the machine learning model set in sequence, and predicts the uncertainty measure value at each future time after executing the combination. The expected uncertainty reduction is the difference between the uncertainty measure at the current time and the predicted uncertainty measure at the future time. The cumulative information gain is a scalar value obtained by summing the expected uncertainty reduction of each step within the planning horizon, used to characterize the total contribution of the action sequence to the improvement of the model's cognitive ability over a long period. For example, when the planning steps are set to 3 steps, the system calculates the uncertainty reduction after the execution of the first, second, and third steps respectively. Then the cumulative information gain This multi-step deduction mechanism overcomes the shortcomings of single-step greedy strategies that focus only on immediate gains while ignoring long-term effects, ensuring the information exploration value of the generated action sequences on a global scale.

[0049] Step 2: Construct an objective function with the cumulative information gain minus the weighted cumulative action cost as its core, and iteratively optimize the initial action combination set; The objective function is a mathematical expression used to evaluate the overall quality of candidate action sequences. Its core logic lies in balancing exploration gains and execution costs. Cumulative information gain represents the knowledge increment brought by the action sequence, while the weighted cumulative action cost is derived from the linear or nonlinear weighted sum of the time cost, energy cost, and collision risk cost predefined in the above embodiment. The weight coefficients can be dynamically adjusted according to the current task stage; for example, information gain is emphasized when the battery is sufficient, while energy cost is emphasized when the battery is low. Iterative optimization can refer to traversing or heuristically searching each candidate sequence in the initial action combination set, substituting its corresponding cumulative information gain and weighted cumulative action cost into the objective function to calculate the score, and recording the current highest score and its corresponding action sequence. Specifically, if a gradient optimization algorithm is used, the gradient of the objective function with respect to the action parameters is calculated and the action sequence is updated along the ascending direction; if a discrete search algorithm is used, obviously inferior branches are removed through pruning strategies. For example, assuming the information gain of candidate sequence A is 0.8, the cost is 0.2, and the weight coefficient is 0.5, then the score is... Candidate sequence B has an information gain of 0.6, a cost of 0.1, and a score of [missing information]. During the iteration process, sequence A is retained as the current optimal solution. This step achieves rapid sorting and optimization of action sequences under complex constraint spaces by uniformly mapping multi-dimensional evaluation indicators to a single scalar value.

[0050] Step 3: Select the optimal combination of continuous actions for the objective function and generate a multi-step continuous sequence of actions to be executed that satisfies the step count constraint.

[0051] In this context, selection refers to choosing the continuous action combination with the largest objective function value from all candidate solutions generated during the iterative optimization process as the final output. Meeting the step constraint means that the length of the generated action sequence must be strictly equal to the preset planning horizon (e.g., fixed at N steps), or fall within the allowed minimum to maximum step range, to ensure the continuity and integrity of robot control commands. The generated multi-step continuous action sequence to be executed is an ordered list of action vectors containing specific control commands (such as joint angle increments, end-effector velocity vectors, etc.) for the next N time steps. For example, after thousands of iterations, the system identifies a three-step action combination consisting of moving forward 10cm, closing the gripper, and lifting upward 5cm. This combination has a significantly higher objective function score than other randomly or heuristically generated combinations, and the total number of steps meets the preset 3-step limit; therefore, it is determined as the final executable sequence. This step, as the endpoint of the planning phase, directly produces the input for subsequent safety verification and control execution, ensuring that the robot can execute globally optimized high-value exploratory actions.

[0052] This application achieves a leap from local optima to global optima by employing model predictive control or Monte Carlo tree search for multi-step look-ahead rolling planning, and combining the joint optimization objective function of cumulative information gain and action cost. Specifically, the aforementioned deduction of the reduction in future uncertainty provides accurate profit prediction data for the construction of the objective function, avoiding ineffective exploration caused by short-sighted decisions; the introduced weighted cumulative action cost effectively suppresses the blind selection of high-energy-consuming or high-risk actions, achieving a balance between economy and safety in the generated action sequence; furthermore, the optimal continuous action combination selected based on the iterative optimization results not only meets strict step constraints but also ensures the continuous acquisition of high-information sensor interaction data during execution. This multi-step look-ahead and cost-sensitive collaborative mechanism improves the robot's active data acquisition efficiency in complex physical environments, especially in task scenarios requiring continuous action coordination to trigger critical state changes, generating a coordinated and highly information-dense action flow, thereby accelerating the convergence speed of machine learning models.

[0053] Example 6: In one optional implementation, this application provides a method for security verification and correction of a multi-step continuous action sequence to be executed. The method further includes strict security boundary constraints and security correction processing on the aforementioned generated multi-step continuous action sequence to be executed.

[0054] Step 1: Pre-set physical safety boundary standards to form a robot safety set. The robot safety set is defined by at least one of the following: joint rotation limit, motion speed limit, end contact force limit, and minimum distance between the robot body and obstacles. The robot safety set is a legal subspace of the robot's operational states, defined based on the robot's physical hardware limits and the safety constraints of the task environment. Specifically, joint rotation limits refer to the permissible mechanical travel range of each degree of freedom joint angle, preventing damage to the mechanical structure due to overshoot; motion speed limits are the maximum linear and angular velocity thresholds set based on the motor's rated speed and dynamic stability; the end-effector contact force limit is the maximum permissible contact force value set for interactive tasks, used to avoid destructive impacts on the manipulated object or the robot itself; and the minimum distance between the robot body and obstacles is a collision avoidance buffer distance set based on the sensor's sensing range. These indicators together constitute a multi-dimensional safety set. ,in Represents the robot's state vector. This is a scalar function characterizing the level of safety. For example, for a six-axis collaborative robotic arm, its safety set can be specifically set as follows: joint angles are limited to ±175 degrees, the end effector linear velocity does not exceed 0.5 m / s, the normal contact force during grasping is less than 20 N, and the Euclidean distance between any part of the robotic arm and surrounding obstacles is maintained at more than 5 cm. By uniformly mapping these discrete physical constraints to a mathematical safety set, a clear criterion is provided for subsequent formal safety verification.

[0055] Step 2: Construct a control barrier function using the robot's real-time pose and torque status as input, and use the control barrier function as a unified safety metric to determine whether the physical safety boundary has been touched. The Control Barrier Function (CBF) is a mathematical tool based on Lyapunov theory, used to transform the static safety set defined above into real-time safety constraints in a dynamic process. The input to this function comes directly from the robot's real-time sensor data, including the current joint pose read by the encoder, joint torques obtained by the torque sensor, and the relative obstacle positions calculated by vision or LiDAR. Its purpose is to quantify the safety margin of the current state from the unsafe region, and to measure its time derivative. With decay rate Relationship (i.e.) This constrains the system's evolution direction, ensuring that the state trajectory always tends towards or remains within a safe set. Specifically, when the robot approaches joint limits or obstacles, The value of the threshold gradually decreases, and the control barrier function generates an exclusive constraint term accordingly, forcing the controller to adjust its output. For example, when the distance between the end effector and the obstacle is detected to rapidly decrease from 10cm to 6cm, the control barrier function calculates a negative safety margin trend based on the current approach speed and immediately generates a corresponding constraint inequality, requiring the control input at the next moment to reduce the rate of change of distance, thereby preventing a collision. This mechanism transforms the originally passive threshold alarm into an active dynamic constraint, enabling early intervention against potential risks.

[0056] Step 3: Analyze the robot's state evolution trajectory corresponding to the multi-step continuous sequence of actions to be executed segment by segment, substitute it into the control barrier function to identify dangerous motion segments, and adjust the trajectory speed and path inflection points while taking into account information gain and action consumption cost, and output the sequence of feasible actions that fall within the robot's safety set at all times.

[0057] like Figure 4 As shown, for a multi-step sequence of actions to be executed obtained from look-ahead planning, a machine learning model is first used to deduce the robot's state evolution trajectory corresponding to the action sequence, and each predicted state in the trajectory is sequentially substituted into the control barrier function for verification. The robot safety set is defined by at least one of the following: joint rotation limit, motion speed limit, end-effector contact force limit, and minimum distance between the robot body and the obstacle; when the control barrier function value corresponding to the predicted state satisfies ( When the control barrier function value corresponding to a trajectory segment is less than zero, or its trend indicates that the robot is about to cross the safety boundary, the trajectory segment is identified as a dangerous motion segment. For the dangerous motion segment, the original action command is modified for safety, using the control barrier function constraint as a hard constraint. This modification includes reducing the execution speed of the corresponding trajectory segment, adjusting the direction of movement, resetting path inflection points, or limiting the end-effector contact force. During the modification process, the optimization goal is to reduce the deviation between the modified action and the original planned action, ensuring that the modified action sequence satisfies safety constraints while preserving as much information gain as possible from the original action sequence and controlling the cost of additional actions. Thus, the trajectory evolution of a state that might have entered the obstacle safety boundary or exceeded other physical limits is modified so that all predicted moments satisfy (…). The sequence of possible actions.

[0058] This step involves the specific application of the constructed control barrier function in the temporal dimension. The system first discretizes the generated preliminary multi-step continuous sequence of actions to be executed, and uses a machine learning model to deduce the robot's expected state evolution trajectory within each short time interval. Then, each deduced intermediate state point is substituted into the control barrier function for verification. If a certain trajectory segment leads to... or violate If a condition is met, the segment is determined to be a dangerous motion segment. For the identified dangerous segments, the system uses quadratic programming or other optimization algorithms to correct the original action sequence. The correction process is not simply truncating the action, but rather constructing a new objective function while satisfying the hard safety constraints. This function aims to minimize the deviation between the corrected action sequence and the original sequence in terms of information gain and action cost. Specific adjustment strategies include reducing the execution speed of the trajectory segment to extend the reaction time, fine-tuning the path inflection points to avoid high-risk areas, or changing the direction of force application to reduce the peak contact force. For example, if the planned look-ahead sequence includes a high-speed movement through a narrow passage, and the control barrier function checks for a risk of lateral collision, the system will automatically adjust the velocity profile of that path segment from a trapezoid to a triangle (reducing the maximum speed) and slightly shift the path centerline to increase the distance to the wall, while ensuring that the corrected action still covers high uncertainty areas to maintain the effectiveness of data acquisition. The final output action sequence guarantees that the robot state strictly falls within the aforementioned safety set at all predicted moments, achieving a dynamic balance between safety and exploration efficiency.

[0059] This application establishes a robust safety protection loop by pre-setting physical safety boundary standards, constructing control barrier functions, and analyzing and correcting motion sequences segment by segment. The pre-set physical safety boundary standards define inviolable physical red lines for the robot, clearly defining specific limitations on joints, speed, force, and distance. Based on this, the control barrier functions transform these static standards into dynamic mathematical constraints, enabling real-time perception of state change trends and prediction of potential risks. Furthermore, by analyzing the state evolution trajectory segment by segment and substituting it into the control barrier functions, the system can accurately locate dangerous motion segments and intelligently adjust trajectory speed and path inflection points while balancing information gain and motion consumption costs. This collaborative mechanism not only ensures that the output action sequence always falls within the safety set, effectively avoiding equipment damage and personal injury, but also overcomes the shortcomings of traditional passive limiting methods, such as poor flexibility and susceptibility to task interruption. This allows the robot to autonomously and safely conduct efficient data acquisition and exploration in real and complex physical environments.

[0060] Example 7: In one embodiment, this application provides a method for collecting and automatically labeling sensor interaction data, the method comprising the following steps: Step 1: During the execution of the robot's action sequence, multi-source raw sensor interaction data is collected synchronously. The multi-source raw sensor interaction data includes the robot's real-time status, issued action commands, and task interaction results. The multi-source raw sensor interaction data refers to the raw signal stream captured in real time by the onboard sensor array when the robot performs actions in the physical environment. The real-time state of the robot specifically includes the angle and angular velocity fed back by the joint encoders, the acceleration and attitude quaternions output by the inertial measurement unit (IMU), and environmental contact information acquired by visual or force sensors. The issued action commands can refer to the motor torque commands or position target values ​​generated by the controller in the previous planning cycle. The task interaction results can refer to the physical feedback at the end of the action execution, such as whether the object was successfully grasped or whether a collision occurred with the end effector—binary or continuous variables. This data is obtained by directly reading from the underlying hardware interface and caching it in shared memory through high-frequency sampling (e.g., 500Hz to 1kHz). Its purpose is to completely reproduce the dynamic behavior and environmental response during the robot's action execution, providing a realistic physical trajectory basis for subsequent model training. For example, in a robotic arm grasping scenario, the real-time state of the machine can be specifically represented by the angle sequence of the six joints and the opening and closing degree of the end effector gripper. The action commands issued are the expected torques of the motors of each joint, and the task interaction results are the peak contact force detected by the force sensor and the object displacement determined by the vision system. Through this multi-dimensional synchronous acquisition, it is possible to ensure strict alignment of the raw data on the time axis and avoid state-action mismatch problems caused by sensor delays.

[0061] Step 2: Synchronously extract the corresponding auxiliary label parameters, which include sampling timestamp, real-time uncertainty metric, single-step action information gain, safety boundary trigger flag, and task success / failure label; Among them, the supporting auxiliary label parameters refer to metadata used to characterize the semantic value and safety attributes of the data, which are generated in parallel with the original sensor data. The sampling timestamp is the system's absolute time at the moment of data acquisition, used to establish a time-series index for multi-source data; the real-time uncertainty metric is derived from the variance or entropy of the machine learning model group's prediction of the state at the current moment, reflecting the ambiguity of the model's perception of the environment under that state; the single-step action information gain is calculated based on the difference in uncertainty metrics before and after executing the action, quantifying the action's contribution to reducing model uncertainty; the safety boundary trigger flag is a Boolean variable indicating whether the current action has triggered the correction logic of the control barrier function; the task success / failure label is the result of automatic determination based on preset task completion criteria (such as stable object grasping for more than 2 seconds). These parameters are extracted in real time by calling the intermediate variables of the uncertainty estimation module and the safety verification module in the data acquisition loop. Their function is to give the original data intelligent labels, enabling the subsequent training process to distinguish between high-value samples and ordinary samples, and to identify potential dangerous working conditions. For example, when a robot attempts to grasp a smooth cylinder, if the model's predicted success probability distribution has a large dispersion, the real-time uncertainty metric is high (e.g., 0.8). If the attempt ultimately causes the object to slip, the task success / failure label is marked as failure. Simultaneously, if a velocity correction is triggered due to approaching joint limits, the safety boundary trigger flag is set to True. By introducing these auxiliary labels, the data not only records what happened but also how the model viewed the event at that time and its value.

[0062] Step 3: Bind the time-series matched sensor interaction data to all auxiliary label parameters one by one, unify the data storage format, remove abnormal sensor data fragments, and generate a complete standardized sample dataset. The process involves several key steps: temporal matching (using sampling timestamps as primary keys to align and merge the collected multi-source raw data with the extracted auxiliary label parameters at the microsecond level); unifying the data storage format (encapsulating heterogeneous data such as image tensors, scalar values, and Boolean flags into structured data objects such as TFRecord or HDF5 format to ensure consistency in field definitions); and removing anomalous sensor data fragments (filtering out invalid data segments caused by sensor failure, communication packet loss, or extreme noise using preset amplitude thresholds and continuity check rules). This process is automatically executed through a data preprocessing pipeline. First, time-aligned data frames are constructed. Then, a sliding window detection algorithm is applied to identify and remove fragments with amplitudes exceeding physical limits (such as joint angle abrupt changes exceeding 90 degrees / millisecond) or interrupted timestamps. Finally, a cleaned and standardized sample dataset is output. For example, when processing a grasping trajectory containing 1000 time steps, if the system detects non-physical spike noise (exceeding the upper limit of the range) in the force sensor readings from step 500 to 505, the preprocessing module will automatically truncate and remove this segment, retaining only the normal and valid data before and after it, and repackage and store the remaining part with the corresponding uncertainty label. This improves the purity and usability of the dataset, ensuring that every sample input into the machine learning model has accurate spatiotemporal context and reliable semantic annotation, thus effectively avoiding interference from noisy data on model convergence.

[0063] This application achieves intelligent and standardized data acquisition by coordinating multi-source raw sensor interaction data and supporting auxiliary label parameters in the aforementioned steps. Specifically, the real-time state of the robot, issued action commands, and task interaction results provide a true mapping of the physical world, while auxiliary labels such as real-time uncertainty metrics and single-step action information gain endow the data with semantic information at the cognitive level. By strictly binding these two types of data in time sequence, not only are the problems of labeling lag and information loss in traditional data acquisition solved, but the generated standardized sample dataset also naturally possesses uncertainty perception attributes. On this basis, the mechanism for removing abnormal sensor data fragments further ensures data quality and prevents erroneous signals from misleading model updates. This end-to-end data organization method allows subsequent incremental model updates using this dataset to prioritize learning features in high information gain regions while avoiding noise interference in unsafe conditions, thereby improving the efficiency of active data acquisition and the convergence speed of model training while ensuring the safety of robot exploration.

[0064] Example 8: In one optional embodiment, this application provides a specific implementation method for removing anomalous sensor data fragments. Based on the acquisition of multi-source raw sensor interaction data and extraction of auxiliary label parameters in the above embodiments, this method further performs hierarchical cleaning and smoothing processing on the time-series data to generate a high-quality standardized sample dataset.

[0065] Step 1: Set amplitude thresholds and sampling continuity thresholds for each type of sensor interaction data, and iterate through all time-series sensor interaction data in segments; The amplitude threshold is a pre-calibrated numerical limit based on the physical range of the robot sensor and the normal fluctuation range under the task scenario, used to define the effective range of sensor readings. The sampling continuity threshold is the upper limit of the time interval for determining whether data stream loss or interruption has occurred, usually set as an integer multiple of the sensor's rated sampling period. Specifically, the system first loads the corresponding threshold parameters from the preset configuration file according to the classification characteristics of data types (such as joint angles, end torques, visual depth maps, etc.). For example, for a six-dimensional force sensor with a range of ±100N, its amplitude threshold can be set to ±120N to accommodate instantaneous impacts, while the sampling continuity threshold is set to 5ms. If the timestamp difference between two adjacent data packets exceeds this value, it is determined to be an interruption. Subsequently, the processor divides the acquired long-term sensor interaction data into several fixed-length time window segments and sequentially traverses and scans them to provide basic data units for subsequent anomaly identification. By setting thresholds by type and traversing segments in this way, the system can adapt to the characteristic differences of multi-source heterogeneous sensors, ensuring the pertinence and accuracy of subsequent judgment logic.

[0066] Step 2: Directly remove invalid data segments with amplitude exceeding the limit, sampling interruption, or numerical abrupt changes exceeding the normal fluctuation range; Invalid data segments refer to data sequences that have completely lost their physical meaning due to sensor malfunction, electromagnetic interference, or communication packet loss. The specific judgment logic includes: marking a segment as having an amplitude exceeding the threshold when the absolute value of any data point within the segment exceeds the amplitude threshold; marking a segment as having a sampling interruption when there are discontinuous timestamps within the segment and the interval is greater than the sampling continuity threshold; and marking an abnormal mutation when the rate of change (i.e., the slope of numerical abrupt change) of adjacent data points within the segment exceeds the upper limit of the normal fluctuation range calculated based on historical data statistics. For example, when a robot performs a grasping task, if the force sensor abruptly changes from 5N to 500N within 1ms and then recovers, and this rate of change far exceeds the rate of change of force corresponding to the maximum allowable acceleration of the robotic arm's dynamics, then the segment is judged as invalid data caused by electrical noise. Once a segment is marked as invalid data of any of the above types, the system will directly delete it from the cache queue and not proceed to subsequent processing. This mechanism effectively prevents severely distorted data from polluting model training and prevents machine learning models from learning incorrect state-action mapping relationships.

[0067] Step 3: Apply moving mean filtering to smooth the valid data with only slight noise, retain the valid samples corresponding to complete action interactions, and complete the preprocessing of the standardized sample dataset.

[0068] Slight noise can refer to data whose overall trend conforms to physical laws, but is superimposed with small-amplitude deviations due to high-frequency random jitter. Moving average filtering is a linear smoothing algorithm that suppresses random noise by replacing the value of the center point with the arithmetic mean of multiple consecutive data points within a window. Specifically, for data segments not eliminated above, the system further calculates their local variance or signal-to-noise ratio. If the indicator falls within a preset slight noise range, the filtering procedure is initiated. For example, setting the sliding window length to 5 sampling points for joint angle data sequences... Processing and updating the angle value. After this processing step, the force perception curve, which was originally jagged, becomes smoother, more realistically reflecting the changing trend of contact force, while retaining the characteristics of key transient events such as collisions and slippage. Finally, all the effective samples that have been filtered and smoothed are repackaged and bound to auxiliary label parameters such as timestamps and uncertainty measures extracted in the above embodiments to form a complete standardized sample dataset for subsequent model updates.

[0069] This application constructs a hierarchical data cleaning mechanism through the synergistic effect of the above steps. By setting amplitude thresholds and sampling continuity thresholds for different types of sensor interaction data, it achieves fine-grained control over multi-source heterogeneous data. Combining the strategy of directly removing severely anomalous segments with smoothing the slightly noisy data using moving average filtering, it decisively discards invalid data that could cause model divergence while retaining as many valid samples containing real physical interaction information as possible. This processing method not only improves the purity and consistency of the standardized sample dataset, avoiding model mislearning problems caused by dirty data, but also ensures the complete recording of boundary conditions (such as instantaneous collision force and rapid slip) during the data acquisition process, laying a data foundation for improving the accuracy of subsequent machine learning model groups and the reliability of uncertainty estimation.

[0070] Example 9: In one possible implementation, embodiments of this application provide a method for updating a machine learning model, the method comprising the following steps: Step 1: Divide the standardized sample dataset into multiple training mini-batch validation subsets, and use mini-batch gradient descent to synchronously and incrementally iterate all prediction sub-models within the machine learning model group; The standardized sample dataset can refer to the high-quality interactive data set formed after automatic annotation and removal of outlier segments. Dividing the dataset into multiple training mini-batches and validation subsets can mean randomly shuffling the sample order according to a preset ratio (e.g., 8:2 or 9:1), extracting consecutive data blocks as a mini-batch, and using the remaining portion as a validation subset. This partitioning method aims to balance the memory usage of a single computation with the stability of gradient estimation. Employing mini-batch gradient descent for synchronous incremental iteration of all predictive sub-models within the machine learning model group can mean using the current mini-batch data to calculate the gradient of the loss function with respect to the model parameters, and updating the weight parameters of all sub-models within the ensemble model or Bayesian neural network in the above embodiment along the reverse direction of the gradient. For example, if the ensemble model contains 5 neural network sub-models with identical structures, when 60 new standardized samples are collected, they are divided into 6 mini-batches, each containing 10 samples. These samples are then sequentially input into these 5 sub-models for forward and backward propagation, simultaneously completing a round of parameter fine-tuning, rather than waiting for the full dataset to accumulate before offline training. This small-batch synchronous incremental iteration mechanism can reduce the instantaneous pressure on the memory resources of the robot's edge computing device while ensuring that the model can respond quickly to new distributed data, thus achieving low-latency online learning.

[0071] Step 2: Real-time statistics of the total number of newly added valid samples within a single collection cycle, establishing an adaptive mapping relationship between the number of samples and the frequency of model iterations, and dynamically adjusting the update frequency of the machine learning model group; In this context, a single data collection cycle can refer to the time window within which the robot completes a multi-step sequence of actions and finishes data annotation. Real-time statistics of the total number of newly added valid samples can refer to the counter accumulating the number of samples that pass quality verification after each action sequence is executed. Establishing an adaptive mapping relationship between the number of samples and the frequency of model iterations can refer to pre-setting a dynamic trigger threshold function, which defines a non-linear correspondence between the number of newly accumulated samples and the triggering of model update operations. For example, setting an initial threshold of 50 samples means that a full mini-batch iteration is triggered every 50 newly added valid samples; when the cumulative number of valid samples exceeds 500, the trigger threshold is dynamically increased to 100 samples, or the update interval is adjusted from updating immediately after each collection to updating once every two collections. Dynamically adjusting the update frequency of the machine learning model group can refer to automatically scaling the training rhythm according to the exploration density of the current task stage: high-frequency updates in the early stages of exploration when data is scarce to quickly build basic cognition, and reducing the update frequency in the data-rich period to avoid overfitting and wasting computational resources. This step, in conjunction with the aforementioned small-batch processing, ensures that the model update rhythm matches the data inflow speed through adaptive mapping, preventing both computational resource congestion caused by frequent updates and policy deviations caused by update lags.

[0072] Step 3: Use the training mini-batch validation subset to evaluate the prediction accuracy of each prediction sub-model after the update, and synchronously update the historical error weights corresponding to each prediction sub-model to complete the overall parameter optimization of the machine learning model group.

[0073] The process of evaluating and updating the prediction accuracy of each prediction sub-model using a small-batch validation subset can refer to immediately testing each sub-model with a reserved validation subset after completing the parameter iterations described above, calculating its mean squared error (MSE) or cross-entropy loss value under the current data distribution, thereby quantifying the real-time prediction capability of each sub-model. Synchronously updating the historical error weights corresponding to each prediction sub-model can refer to recalculating the contribution weight of each sub-model in the ensemble voting based on the latest accuracy results obtained from the evaluation, combined with an exponential moving average algorithm. For example, if a sub-model performs well in this validation, its historical error weight coefficient will increase from 0.15 to 0.20, while the weights of sub-models that performed poorly in the validation will be reduced accordingly, thereby reducing the interference of low-precision sub-models in the calculation of variance or entropy values ​​in the subsequent uncertainty metric calculations described in the above embodiments. Optimizing the overall parameters of the machine learning model group can refer to achieving optimal configuration in terms of both internal parameters and structural combination through the closed-loop process of iterative update-validation evaluation-weight reallocation described above.

[0074] This application constructs an efficient online continuous learning mechanism for robots through the synergistic effect of the aforementioned technical features. By dividing the standardized sample dataset into mini-batch validation subsets and combining them with the mini-batch gradient descent algorithm, the model achieves rapid absorption of new knowledge and low-resource-consumption updates. Furthermore, an adaptive mapping relationship between sample size and iteration frequency is introduced, enabling the system to intelligently adjust the computational load based on the density of data accumulation, avoiding resource waste or learning lag caused by fixed-frequency updates. Further, the validation subset is used to evaluate and dynamically adjust the historical error weights of each prediction sub-model in real time, ensuring that the ensemble model maintains high-confidence uncertainty estimation even when facing distributed offset data. This dual mechanism of parameter optimization and weight adjustment allows the machine learning model group to continuously evolve as data collection deepens, not only improving the prediction accuracy of individual models but also enhancing the robustness of the entire model group in complex dynamic environments, providing more accurate uncertainty guidance for the generation of subsequent action sequences.

[0075] Example 10: In one optional embodiment, this application provides a method for updating uncertainty estimates, the method comprising the following steps: Step 1: Based on the ensemble model or Bayesian neural network after parameter iteration, re-solve for the two types of components: parameter uncertainty and data random uncertainty. Parameter uncertainty refers to the degree of uncertainty in the distribution of model weights due to limited training data or initialization differences. It originates from the dispersion of prediction results of each sub-model in the ensemble model, or the variance of the posterior distribution of the Bayesian neural network. Data random uncertainty refers to the inherent fluctuations in data caused by irreducible factors such as sensor noise and random environmental disturbances, which affect the mapping process from model input to output. Specifically, after the incremental iteration of model parameters in the above embodiment is completed, the system immediately freezes the current model weights. For the ensemble model scheme, the same current state sample is input to all sub-models again, and the variance of the predicted output values ​​of each sub-model is calculated as a quantitative indicator of parameter uncertainty. At the same time, the entropy value of the residual distribution is statistically analyzed as a quantitative indicator of data random uncertainty. For the Bayesian neural network scheme, through multiple forward propagations using Monte Carlo Dropout, the total prediction variance is decomposed into a weight variance term (parameter uncertainty) and a noise variance term (data random uncertainty). For example, when an ensemble model contains five sub-networks, if the predictions of the success rate of crawling by the five networks in a certain state are [0.2, 0.8, 0.3, 0.7, 0.4], then its high variance indicates high parameter uncertainty. Conversely, if all network predictions fluctuate around 0.5 and have large historical residuals, it indicates high data stochastic uncertainty. Through this decomposition mechanism, it is possible to accurately identify whether the current cognitive blind spot stems from insufficient model fitting or inherent environmental noise, providing a basis for subsequent differentiated weighting.

[0076] Step 2: Adjust the fusion weights of the two types of uncertainty components based on the current task exploration depth, and recalculate the global total uncertainty metric; The task exploration depth refers to a comprehensive representation of the number of data collection cycles completed by the robot, the cumulative effective sample size, or the average uncertainty level across the entire state space. It measures the stage of the learning process. The fusion weights are coefficients assigned to parameter uncertainty and data random uncertainty, and their setting is dynamically adjusted as the exploration depth increases based on a preset decay function or segmentation strategy. Specifically, in the early stages of task exploration, samples are scarce and the model has not yet converged. At this time, a higher weight is assigned to parameter uncertainty (e.g., 0.8), and a lower weight is assigned to data random uncertainty (e.g., 0.2) to encourage the robot to explore areas where the model's understanding is ambiguous. As the collected data accumulates and the exploration depth increases, the model gradually matures, parameter uncertainty decreases, and the system automatically lowers its weight and correspondingly increases the weight of data random uncertainty, thereby guiding the robot to focus on complex or noisy boundary conditions in the environment. The global total uncertainty is measured using a weighted summation formula. The standardized values ​​obtained from the calculation, where For dynamically adjusted weights, For example, when the cumulative number of collected samples reaches 1,000, the system determines that it has entered the intermediate exploration phase and adjusts the weights to 0.5:0.5; when the number of samples exceeds 5,000 and the average prediction error is below the threshold, the weights are adjusted to 0.2:0.8. Through this dynamic adjustment mechanism, a smooth transition from reducing model ignorance to strategies for dealing with complex environments is achieved, avoiding the failure of a single indicator in long-term learning.

[0077] Step 3: Cache the updated uncertainty metric to the action planning module as the core input parameter for generating a multi-step continuous sequence of actions to be executed in the next round.

[0078] like Figure 5 As shown, the robot's active data acquisition and model updates are performed cyclically in continuous time steps. In time step ( The robot executes a sequence of actions that has been safety-corrected. ), synchronously collecting real-time sensor data such as the body's status, issuing action commands, and task interaction results. The system automatically labels samples based on sampling timestamps, real-time uncertainty metrics, single-step action information gain, safety boundary trigger flags, and task success or failure results, generating corresponding standardized samples. Subsequently, at time step ( ), ( )to( The process of repeating actions, data collection, and automatic annotation is performed, and the standardized sample increments formed at each time step are written to the data buffer pool. When the number of newly added valid samples in the data buffer pool reaches the preset update condition, or reaches the model update cycle adapted to the current task exploration stage, the training mini-batch and validation subset are extracted from the data buffer pool, and incremental training is performed on each prediction sub-model in the machine learning model group to obtain the updated model parameters. After the model parameters are updated, the parameter uncertainty and data random uncertainty are recalculated. The fusion weights of these two types of uncertainty components are then updated based on the current task's exploration depth to obtain the updated uncertainty metric. The updated uncertainty metric is written into the action planning module for candidate action evaluation and multi-step look-ahead planning in the next time step, thereby forming an online closed-loop time sequence of "action execution - data collection - automatic annotation - sample accumulation - model update - uncertainty update - replanning".

[0079] The action planning module refers to the computational unit that executes the S2 step and the multi-step look-ahead planning logic described above. It is responsible for generating the optimal action sequence based on the uncertainty metric. The core input parameters are the key variables driving the objective function calculation of the planning algorithm, directly determining the estimation benchmark for information gain. Specifically, the recalculated global total uncertainty metric is written to a specified cache area in shared memory or a message queue. Before starting the next round of rolling time-domain planning (such as MPC or MCTS), the action planning module prioritizes reading this latest cached value, replacing the old uncertainty estimate from the previous round, to extrapolate the magnitude of uncertainty reduction in future states. For example, when generating the action sequence for the next three steps, the planner uses the latest uncertainty metric to calculate the expected information gain for each step. If the updated values ​​are not used, invalid actions may be selected based on outdated high-uncertainty assessments. This step ensures data synchronization between the sensing end (model update) and the decision-making end (action planning), forming a closed-loop control link of acquisition-update-planning-execution, effectively preventing deviations in the exploration direction caused by changes in model state.

[0080] This application achieves real-time synchronization and adaptive evolution of uncertainty estimation through the aforementioned steps. By resolving the two components of parameter uncertainty and data stochastic uncertainty, and dynamically adjusting the fusion weights in conjunction with the task exploration depth, it not only maintains the consistency between the uncertainty measure and the model's current cognitive level, but also endows the system with the ability to differentiate its exploration focus at different learning stages. Based on this, the updated uncertainty measure is cached in real-time and used as the core input for action planning, enabling the generation of subsequent multi-step action sequences to be optimized based on the latest knowledge state. This improves the targeting and efficiency of active data collection, ensuring that the robot always focuses on the most valuable state region throughout the entire learning cycle.

[0081] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A robot active data acquisition method based on uncertainty and action cost, characterized in that, include: S1. Construct a group of machine learning models, input the robot's real-time running status, predict the subsequent running status or task completion results, and calculate the uncertainty measure based on the differences between multiple sets of prediction results. S2. Based on the uncertainty measure of the current state, combined with the action execution cost, simultaneously calculate the execution cost of various actions, comprehensively weigh the benefits of reducing the degree of uncertainty and the action consumption cost, and generate a multi-step continuous sequence of actions to be executed. S3. Retrieve the robot's preset physical safety boundary standard, and perform safety verification on the selected multi-step continuous action sequence to be executed. If the action may cause the robot to enter an unsafe state, then correct it through the control barrier function and output the action sequence that can be executed. S4. Control the robot to execute the sequence of possible actions, collect sensor interaction data during the execution process, and automatically label it to form a standardized sample dataset; S5. Using the newly collected standardized sample dataset, update the machine learning model and update the uncertainty estimate for subsequent action selection.

2. The method according to claim 1, characterized in that, The process of constructing a set of machine learning models and calculating uncertainty metrics includes: The machine learning model group consists of an ensemble model or a Bayesian neural network, and the machine learning model group contains several predictive sub-models with differential biases in their initialization parameters. The same frame of the robot's real-time running state is synchronously input into all the prediction sub-models, and each model outputs an independent subsequent running state and a prediction probability distribution of the task completion result. The discrete differences of multiple predicted probability distributions are quantified using prediction variance or prediction entropy. The parameter uncertainty and data random uncertainty are separated and weighted and fused together to obtain the uncertainty measure in a unified manner.

3. The method according to claim 2, characterized in that, The uncertainty measure, calculated uniformly based on the differences between multiple sets of prediction results, includes: The prediction results output by each prediction sub-model are converted into numerical vectors of the same dimension, and the variance and mean Euclidean distance between the vectors are calculated as the basic discrete features. Based on the historical prediction errors of each prediction sub-model, weights are assigned to reduce the interference of sub-models with poor prediction accuracy on the calculation of the basic discrete features, thereby obtaining a corrected dispersion index. The superposition weights of the two types of uncertainty components are dynamically adjusted in conjunction with the current exploration phase of the robot task, and the standardized uncertainty metric is output by mapping it to a fixed numerical range.

4. The method according to claim 1, characterized in that, The comprehensive trade-off between the benefits of reducing the degree of uncertainty and the costs of the actions includes: A predefined complete robot motion library is provided, and the cost of motion execution includes at least one of time cost, energy cost, and collision risk cost. Corresponding loss parameters are pre-calibrated for each motion in the complete robot motion library. Estimate the benefit of reducing the degree of uncertainty after executing each candidate action, and configure an adaptive balance weight value for the benefit of reducing the degree of uncertainty and the cost of executing the action. The balance weight value is dynamically adjusted synchronously with the current uncertainty metric. By incorporating robot endurance constraints and maximum steps per planning iteration into the trade-off logic, candidate action combinations that exceed energy consumption limits or deviate from the main task objective are filtered out, and a preliminary action combination set is output.

5. The method according to claim 4, characterized in that, The generation of a multi-step continuous sequence of actions to be executed includes: Multi-step look-ahead rolling planning is carried out using model predictive control or Monte Carlo tree search to deduce the expected decrease in uncertainty after different combinations of candidate actions are executed, and to quantify the cumulative information gain corresponding to each combination. Construct an objective function with the cumulative information gain minus the weighted cumulative action consumption cost as its core, and iteratively optimize the initial action combination set; The optimal combination of consecutive actions for the objective function is selected to generate the multi-step consecutive action sequence that satisfies the step count constraint.

6. The method according to claim 1, characterized in that, The process of security verification and correction of the multi-step continuous action sequence to be executed includes: A preset physical safety boundary standard constitutes a robot safety set, which is defined by at least one of the following: joint rotation limit, motion speed limit, end contact force limit, and minimum distance between the robot body and the obstacle. A control barrier function is built using the robot’s real-time pose and torque state as input, and the control barrier function is used as a unified safety metric to determine whether the physical safety boundary standard has been touched. The robot's state evolution trajectory corresponding to the multi-step continuous sequence of actions to be executed is analyzed segment by segment. The control barrier function is substituted to identify dangerous motion segments. The trajectory speed and path inflection points are adjusted while taking into account information gain and action consumption cost. The sequence of possible actions that falls within the robot's safety set at all times is output.

7. The method according to claim 1, characterized in that, The process of collecting and automatically labeling the sensor interaction data includes: During the execution of the action sequence, the robot synchronously collects multi-source raw sensor interaction data, which includes the robot's real-time status, issued action commands, and task interaction results. Simultaneously extract the corresponding auxiliary label parameters, which include sampling timestamp, real-time uncertainty metric, single-step action information gain, safety boundary trigger flag, and task success / failure label; The time-series matched sensor interaction data is bound one by one with all the auxiliary label parameters, the data storage format is unified, abnormal sensor data fragments are removed, and a complete standardized sample dataset is generated.

8. The method according to claim 7, characterized in that, The process of removing anomalous sensor data segments includes: For each type of sensor interaction data, an amplitude threshold and a sampling continuity threshold are set, and all time-series sensor interaction data are traversed in segments. Invalid data segments with excessive amplitude, interrupted sampling, or numerical abrupt changes exceeding the normal fluctuation range are directly removed; For valid data with only slight noise, a moving mean filter is used for smoothing, retaining valid samples corresponding to complete action interactions, thus completing the preprocessing of the standardized sample dataset.

9. The method according to claim 2, characterized in that, The process of updating the machine learning model includes: The standardized sample dataset is divided into multiple training mini-batch validation subsets, and mini-batch gradient descent is used to synchronously incrementally iterate all the prediction sub-models within the machine learning model group. The total number of newly added valid samples within a single collection cycle is counted in real time, an adaptive mapping relationship between the number of samples and the frequency of model iterations is established, and the update frequency of the machine learning model group is dynamically adjusted. The prediction accuracy of each prediction sub-model is evaluated and updated using the training mini-batch validation subset. The historical error weights corresponding to each prediction sub-model are updated synchronously to complete the overall parameter optimization of the machine learning model group.

10. The method according to claim 9, characterized in that, The process of updating the uncertainty estimate includes: Based on the ensemble model or Bayesian neural network after parameter iteration, re-solve the two types of components: parameter uncertainty and data random uncertainty. Adjust the fusion weights of the two types of uncertainty components based on the current task exploration depth, and recalculate the global total uncertainty metric. The updated uncertainty metric is cached in the action planning module as the core input parameter for generating the multi-step continuous action sequence to be executed in the next round.

Images