Robot task-level emotion feature extraction method and system based on multi-evaluator distribution modeling

Abstract

The invention relates to a robot task level emotion feature extraction method and system based on multi-evaluator distribution modeling, and belongs to the technical field of artificial intelligence and robots. The method comprises the following steps: collecting emotion score data of multiple evaluators; calculating a plurality of emotion statistical characteristics; constructing an emotion prediction model and selecting an optimal feature subset based on the evaluation indexes; and outputting a structured emotion interface for a robot behavior learning system to call. According to the method, statistical modeling and feature selection mechanisms are introduced, so that the problem that an existing emotion recognition method lacks generalization and task-level coupling capability is solved, and the naturalness and decision-making flexibility of robot emotion interaction are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and robotics technology, specifically relating to an emotion computing method in robot task planning, and in particular to a robot task-level emotion feature extraction method and system based on multi-evaluator distribution modeling. Background Technology

[0002] As artificial intelligent agents, robots have significantly improved the logic and efficiency of their behavior during autonomous task planning and execution. However, traditional task-level control, typically based on logical programming or reinforcement learning models, focuses only on environmental states and physical constraints, completely ignoring the profound impact of human emotional states on behavior generation and the establishment of human-machine trust. Introducing emotional factors can significantly enhance the naturalness of human-machine interaction and the flexibility of task decision-making, transforming robots from efficient machines into trustworthy partners.

[0003] In recent years, with the development of multimodal emotion recognition technology, robots have begun to possess preliminary "emotion perception" capabilities. Existing technologies primarily use sensors such as cameras and microphones to collect signals from users' facial expressions and voice intonation, thereby identifying their emotion categories. Representative existing patents, such as CN114842504B and CN118238151A, are dedicated to improving the accuracy of emotion recognition and applying it to the robot's instantaneous responses.

[0004] However, these existing technologies have obvious limitations:

[0005] (1) Remaining at the perception level and lacking task-level modeling: Existing methods mostly treat emotion as an independent, instantaneous control signal, failing to parameterize emotional features and incorporate them as a continuous state variable into the task-level decision-making and planning process.

[0006] (2) Ignoring subjectivity and statistical characteristics: Human emotional responses to the same task have subjective differences. Existing technologies rely on a single modality or a single evaluator, failing to perform statistical modeling of multi-evaluator data and thus failing to characterize the group distribution characteristics and uncertainty of emotions.

[0007] (3) Lack of feature interface: The current system only outputs discrete emotion categories (such as happy and sad), lacking stable and structured emotion feature interfaces that can be directly called by the task planning model, resulting in a disconnect between the two layers of "emotion perception" and "task execution".

[0008] Therefore, constructing an algorithmic framework that can achieve emotion modeling at the task level and coordinate the optimization of emotion features with task scheduling has become an important direction for improving the cognitive intelligence level of robots. Summary of the Invention

[0009] To address the problems in the existing technologies, namely the lack of task-level statistical modeling capabilities and structured feature interfaces in current robot emotion computing methods, which leads to a disconnect between emotion perception and task planning, this invention aims to provide a method and system for extracting emotion features that can transform subjective emotion scores from multiple evaluators into quantifiable, generalizable, and directly applicable emotion features for task decision-making.

[0010] The technical solution of the present invention is as follows:

[0011] A method for extracting robot task-level emotional features based on multi-evaluator distribution modeling, characterized in that the method includes the following steps:

[0012] Step S10, Multi-evaluator sentiment data collection: In multiple task scenarios, invite multiple independent evaluators to give subjective sentiment scores to the robot's task performance and record the scene feature vector corresponding to each task scenario.

[0013] Step S20, Sentiment Statistical Feature Construction: Based on the multi-evaluator rating data collected in each task scenario, calculate a set of statistical features that can comprehensively describe the characteristics of sentiment distribution. The sentiment statistical features include, but are not limited to, the following ten: mean, standard deviation, skewness, kurtosis, entropy, positive sentiment percentage, mean absolute deviation, Gini coefficient, and bimodality.

[0014] Step S30, Sentiment Prediction Model Construction and Training: A lightweight neural network model, EMOPredictor, is constructed. It takes the feature vector of the task scene and the task encoding as input, and the sentiment statistical features calculated in step S20 as the prediction target, and is trained accordingly. The EMOPredictor model structure includes an input layer, at least one fully connected hidden layer (using the ReLU activation function), and an output layer. To satisfy the probability distribution characteristics, the output layer can be normalized using the Softmax function. The model is optimized by minimizing the mean squared error (MSE) between the predicted and true values.

[0015] Step S40, Optimal Sentiment Feature Selection: The trained sentiment prediction model is used to make predictions on the validation set, and multiple evaluation metrics are calculated between the predicted results and the true values, including mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²). 2 ) and Pearson correlation coefficient (r). Based on these indicators (e.g., R is preferred), 2 (Features with high r-values ​​and low MSE and MAE values) are selected from ten initial features to choose an optimal subset of sentiment features to eliminate redundancy and enhance the model's generalization ability.

[0016] Step S50, Structured Emotion Feature Output: The optimal subset of emotion features selected in step S40 is output as a standardized structured interface for direct use by the robot's behavior learning or task planning system. This module supports an online update mechanism, allowing for fine-tuning of the emotion prediction model when new tasks and scoring data are added, enabling continuous learning and adaptation.

[0017] Compared with the prior art, the technical solution provided by the present invention has the following significant advantages:

[0018] (1) It realizes the transformation from subjective emotions to objective features: by introducing multi-evaluator data and statistical distribution modeling, subjective and vague human emotions are transformed into objective and structured mathematical features, providing calculable emotional input for robots.

[0019] (2) Enhanced the generalization and robustness of emotional features: By calculating multiple statistical features and selecting features based on evaluation indicators, noise caused by individual evaluation differences is effectively eliminated, so that emotional features have good stability and generalization ability in different tasks and scenarios.

[0020] (3) It provides direct support for task-level decision-making: the generated structured emotion feature interface can be seamlessly integrated into the robot's reinforcement learning, decision tree and other task planning models, and truly realizes the closed loop of emotion perception and task execution.

[0021] (4) It combines interpretability and scalability: the statistical features used have clear mathematical meaning, and the model decision-making process is transparent. At the same time, the method framework is easy to extend and can be easily incorporated with features from more modalities or more complex prediction models. Attached Figure Description

[0022] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0023] Figure 1 This is an overall flowchart of the robot task-level emotion feature extraction method based on multi-evaluator distribution modeling of the present invention. Detailed Implementation

[0024] The following is combined Figure 1 The present invention will be further described in detail below with reference to the embodiments. These embodiments are implemented based on the technical solution of the present invention, providing detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0025] Example: Emotional Task Scheduling for Industrial Picking Robots

[0026] Application Scenario: In a smart warehouse, an industrial picking robot needs to collaborate with multiple human employees. To improve collaboration efficiency and human comfort, it is desirable for the robot to adjust its task execution strategy (such as movement speed and robotic arm range of motion) based on the collective emotional feedback of the employees.

[0027] Implementation steps:

[0028] Data Acquisition (S10): When the robot performs ten typical tasks, such as "picking fragile items," "high-speed sorting," and "close-range human-robot transfer," ten employees are invited as evaluators to rate their psychological feelings about the robot's behavior on a scale of 1 to 5 (1 point: very nervous and anxious, 5 points: very comfortable and reassured). Simultaneously, visual scene features (such as work area crowding) and semantic tags (such as task urgency) are recorded to form a scene feature vector.

[0029] Feature construction (S20): For the "high-speed sorting" task, assume that 10 ratings are collected: [2,3,2,4,2,3,1,2,3,2].

[0030] The calculated mean μ = 2.4 and the standard deviation σ ≈ 0.84.

[0031] Calculations of skewness, kurtosis, and entropy revealed a left-skewed (negatively skewed) distribution with low kurtosis, indicating that evaluators generally felt somewhat uneasy and had rather divergent opinions.

[0032] The positive sentiment percentage p_pos = 0.2 (only 2 ratings > 3) is calculated, indicating a high Gini coefficient and suggesting inequality in sentiment identification.

[0033] Model training and feature selection (S30, S40):

[0034] The datasets for all tasks are divided into training and validation sets.

[0035] The EMOPredictor model is constructed and trained with the goal of minimizing the MSE loss.

[0036] Evaluate the predictive performance of ten features on the validation set. Assume that R0 is found for four features: mean, standard deviation, positive sentiment percentage (p_pos), and entropy. 2 The correlation coefficient was >0.9 and the Pearson correlation coefficient was >0.85, while other features had poor predictive performance or were redundant.

[0037] Therefore, {mean, std, p_pos, entropy} is selected as the optimal subset of sentiment features.

[0038] Deployment and Application (S50):

[0039] The trained model and feature selector are deployed to the robotic system.

[0040] When the robot is running a new task, the emotion output module outputs a four-dimensional feature vector [mean, std, p_pos, entropy] in real time.

[0041] After receiving the vector, if the robot's task planner detects that mean is low, std is high, and p_pos is low, it will determine that the current task mode has caused widespread unease and significant disagreements within the team, and will automatically switch to a more conservative and safer task execution strategy.

[0042] Alternative solution:

[0043] The EMOPredictor network structure can be replaced with a Transformer encoder to better handle long-range dependencies in complex task scenarios.

[0044] The emotional features can be used to construct a reward function, which, combined with a reinforcement learning algorithm, can train a robot strategy that can proactively optimize human emotional experiences.

[0045] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for extracting robot task-level emotional features based on multi-evaluator distribution modeling, characterized in that, The method includes: • Step S10: Collect subjective emotional rating data from multiple evaluators in a multi-task scenario; • Step S20: Calculate various sentiment statistical features based on the scoring data, including mean, standard deviation, skewness, kurtosis, entropy, positive sentiment percentage, interquartile range, mean absolute deviation, Gini coefficient, and bimodality. • Step S30: Construct a sentiment prediction model, using task encoding and scene features as input, to predict the sentiment statistical features; • Step S40: Based on the evaluation metrics of the predicted results and the actual values, including mean squared error, mean absolute error, coefficient of determination and Pearson correlation coefficient, select the optimal subset of sentiment features; • Step S50: Output the selected emotional features as a structured interface for the robot behavior learning system to call.

2. The method according to claim 1, characterized in that, The calculation method for the sentiment statistical features in step S20 includes: • Mean: Standard deviation: • Skewness: • Kuroshi: Entropy: H = -∑ k p k log p k , where p k Let be the probability of the k-th class in the rating distribution; • Percentage of positive emotions: Where T is the emotion threshold; • Interquartile range: IQR = Q3 - Q1, where Q1 and Q3 are the first and third quartiles, respectively; • Mean absolute deviation: • Gini coefficient: • Bimodality: Where μ1, μ2, σ1, and σ2 are parameters of the bimodal distribution.

3. The method according to claim 1, characterized in that, In step S30, the sentiment prediction model is a lightweight neural network EMOPredictor, whose structure includes two fully connected layers and a ReLU activation function. The output layer uses Softmax normalization, and the loss function is the mean squared error. Where y i For the true value, These are predicted values.

4. The method according to claim 1, characterized in that, The calculation method for the evaluation index in step S40 includes: • Mean square error: • Mean absolute error: • Determination coefficient: • Pearson correlation coefficient:

5. The method according to claim 1, characterized in that, In step S50, the sentiment feature interface supports an online update mechanism. When new task data is added, the sentiment prediction model can be fine-tuned to quickly adapt to the new situation.

6. A robot task-level emotion feature extraction system based on multi-evaluator distribution modeling, characterized in that, The system includes: • Data acquisition module, used to collect subjective sentiment rating data from multiple evaluators; • Feature construction module, used to calculate various sentiment statistical features; • The sentiment prediction module is used to build and train sentiment prediction models; • Feature selection module: Selects the optimal subset of sentiment features based on evaluation metrics; • Output module, which outputs the selected sentiment features as a structured interface.

7. A storage device storing a plurality of programs, characterized in that, The program is adapted to be loaded and executed by a processor to implement the method of any one of claims 1-5.

8. A processing apparatus, comprising a processor and a storage device, characterized in that, The program is adapted to be loaded and executed by a processor to implement the method of any one of claims 1-5.

Images