Humanoid robot action generation method based on cognitive feedback mechanism

Abstract

The invention discloses a humanoid robot action generation method based on a cognitive feedback mechanism. The humanoid robot action generation method comprises the following steps: S1, acquiring and preprocessing multi-modal sensing data; s2, performing intention recognition and cognitive feedback based on the standardized perception feature vector; s3, retrieving candidate actions according to the intention label and constructing an action parameter space; s4, globally optimizing the action parameters through a beetle antennae search algorithm to obtain an optimal action sequence; s5, the optimal action sequence is issued and executed, and real-time feedback information is collected; and S6, according to the feedback adjustment intention confidence and the cognitive reasoning result, judging whether to re-execute or output a final result. According to the method, intelligent perception, self-adaptive action generation and dynamic optimization of the humanoid robot on a complex environment and a user intention are realized, and the intelligence and safety of man-machine interaction are improved.

Description

Technical Field

[0001] The present invention relates to the field of robotics technology, and in particular to a method for generating humanoid robot motions based on a cognitive feedback mechanism. Background Art

[0002] With the rapid development of artificial intelligence and robotics, humanoid robots, as representatives of service robots and intelligent agents, have been widely used in numerous fields, including industrial manufacturing, home services, medical assistance, and educational companionship. Traditional humanoid robot motion generation methods are primarily based on preset motion templates, finite state machines, or trajectory planning algorithms. These methods rely on static rules to select and execute motions based on user commands or environmental signals. While these methods can achieve basic motion planning, they often lack a deep understanding of dynamic environments, complex tasks, and individual user intentions, as well as the ability to adapt in real time. In existing technologies, some research has introduced deep learning-based motion recognition and generation networks, improving the naturalness and diversity of robots' movements in specific scenarios. However, most approaches remain stuck in a one-way information flow, a linear process of perception-decision-execution, lacking high-level cognitive feedback regulation and adaptive closed-loop mechanisms.

[0003] When it comes to intent recognition, existing humanoid robot systems typically rely solely on a single modality for simple command parsing, neglecting the integration of multimodal information and in-depth modeling of user behavior history. This results in limited responsiveness to complex and changing user needs. While existing motion optimization algorithms, such as genetic algorithms and particle swarm algorithms, possess certain global optimization capabilities, they lack real-time performance, convergence speed, and adaptability to high-dimensional motion parameter spaces. Furthermore, they rarely incorporate cognitive-level intention reasoning and feedback mechanisms into the optimization process. Traditional approaches to the actual execution of robotic motions and environmental adaptability rely primarily on passive adjustments, failing to achieve feedback-based dynamic optimization and self-correction.

[0004] Current technology still has several key flaws: First, humanoid robots lack a complete closed-loop mechanism that integrates multimodal perception, intention recognition, cognitive feedback and action generation, making it difficult to adapt to dynamic interactions in complex environments; second, the optimization of action parameters is disconnected from high-level cognitive information, and there is a lack of global optimization and adaptive adjustment based on cognitive feedback, resulting in insufficient intelligence, personalization and flexibility of robot actions; third, the response to environmental and user changes during real-time action execution is not agile enough, making it difficult to ensure the naturalness and robustness of human-computer interaction.

[0005] Therefore, how to provide a humanoid robot motion generation method based on a cognitive feedback mechanism is an urgent problem that those skilled in the art need to solve. Summary of the Invention

[0006] One purpose of the present invention is to propose a method for generating humanoid robot motions based on a cognitive feedback mechanism. This method fully integrates advanced technologies such as multimodal perception, intent recognition, cognitive feedback, and beetle whisker search optimization. It describes in detail the complete process for achieving a humanoid robot's accurate understanding of and dynamic response to user intent in complex environments through multi-level intent modeling, adaptive cognitive feedback, and global optimization of motion parameters. This method boasts high interactive intelligence, natural and smooth motion, strong environmental adaptability, and outstanding self-optimization capabilities, effectively enhancing the efficiency of human-machine collaboration and user experience in diverse scenarios.

[0007] A method for generating humanoid robot motions based on a cognitive feedback mechanism according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect multimodal perception data of the humanoid robot's environment, preprocess the multimodal perception data, and generate a standardized perception feature vector;

[0009] S2. Based on the standardized perceptual feature vector, using the intent recognition and cognitive feedback method, through multimodal deep neural network fusion analysis, identify the user's current intent and intent confidence, generate an intent label, and output a cognitive feedback signal;

[0010] S3. According to the intent label, the candidate action sequence corresponding to the identified intent is retrieved from the action database, and the action parameter space is constructed based on the current environment information;

[0011] S4. Use the beetle whisker search algorithm to globally optimize the action parameter space, take the current candidate action plan as the search individual, and optimize the action parameters based on the beetle whisker search mechanism to obtain the optimal action sequence;

[0012] S5. Send the optimal action sequence to the humanoid robot for execution, collect real-time feedback information during the action execution process, and use it as feedback input to the intention recognition and cognitive feedback link;

[0013] S6. Adjust the intention confidence and cognitive reasoning results based on real-time feedback information. If it is detected that the action does not fully meet the recognition intention or the environment has changed significantly, return to step S2 and re-execute the intention recognition and action optimization process until the action completion meets the set requirements. If the action completion meets the set requirements, output the final action execution result.

[0014] Optionally, the multimodal perception data specifically includes image information, voice signals, human posture data, spatial position information and environmental physical parameters of the robot's environment.

[0015] Optionally, the preprocessing of the multimodal perception data specifically includes denoising and feature extraction of the image information of the multimodal perception data, framing and feature encoding of the voice signal of the multimodal perception data, normalizing the posture data of the multimodal perception data, and standardizing the spatial position information and environmental physical parameters of the multimodal perception data.

[0016] Optionally, the S2 specifically includes:

[0017] S21. Input the obtained standardized perception feature vectors into the short-term intent recognition network, the long-term intent recognition network, and the contextual intent reasoning network respectively. Each network adopts a multimodal deep neural network structure to extract feature information related to the corresponding intent type.

[0018] S22, the short-term intention recognition network outputs the short-term intention vector I s , the long-term intention recognition network outputs the long-term intention vector I l , the contextual intention inference network outputs the contextual intention vector I c ;

[0019] S23, I s , I l , I c Input adaptive fusion process, calculate the weight coefficient α based on the current environment complexity, interaction history length and user personality characteristics s , α l , α c , perform weighted fusion on the three types of intention vectors to obtain a multi-level comprehensive intention vector I fusion ;

[0020] S24, multi-level integrated intention vector I fusion Input intent classification layer, calculate the probability distribution P(I k |I fusion ) and obtain the probability values ​​corresponding to all categories;

[0021] S25, based on the probability distribution P(I k |I fusion ), a confidence gating mechanism and historical intent transfer probability are introduced to perform confidence-weighted scoring on the current candidate intent category:

[0022] S k =λ1P(I k |I fusion )+λ2Q(I k |I hist );

[0023] Among them, S kis the comprehensive score of the k-th type of intention, Q(I k |I hist ) is the historical intention transfer probability, λ1 and λ2 are weight coefficients, and the intention category with the highest score is finally selected as the current optimal intention label I * , and the corresponding score value is used as the confidence C;

[0024] S26, introduce the knowledge graph into the intention recognition process, and convert the current multi-level comprehensive intention vector I fusion , historical scene information and environmental data are used to perform semantic reasoning with the knowledge graph, and the semantically corrected intent label I is output through knowledge graph completion and verification. ** ;

[0025] S27, the obtained intention label I ** The confidence level C is input into the cognitive feedback generation process, and the real-time feedback information of the robot's current action execution, the rate of environmental change, and the user's emotional feedback signal are simultaneously collected;

[0026] S28, the cognitive feedback generation process dynamically adjusts the feedback intensity and feedback content of the cognitive feedback signal based on the collected real-time feedback data, combined with the adaptive gain adjustment mechanism and historical interaction information, to form the cognitive feedback signal R adpt ;

[0027] S29, the cognitive feedback signal R adpt It is transmitted back to the short-term intent recognition network, long-term intent recognition network and contextual intent reasoning network to correct the intent recognition results and various intent weights in real time, realize closed-loop adaptive adjustment of intent recognition and cognitive feedback, and output the corrected final intent label, confidence and cognitive feedback signal.

[0028] Optionally, the S3 specifically includes:

[0029] S31, according to the output final intention label I ** , confidence C and cognitive feedback signal R adpt , combined with the current environment information E to build the retrieval conditions and determine the candidate action set corresponding to the intention label

[0030] S32. For each candidate action A i Extracting action description vectors The action description vector includes action type, movement amplitude, speed, energy consumption, gait, and spatial position;

[0031] S33, according to the cognitive feedback signal R adpt And environmental information E, calculate the fitness score of each candidate action with the current intention and environment, using the weighted fitness function:

[0032]

[0033] Among them, i is the fitness of the i-th candidate action, Indicates the matching degree between intention and action, Q(A i ,E) represents the adaptability of the action to the environment, β1, β2, β3, β4 are weighting coefficients;

[0034] S34, normalizing the fitness scores of all candidate actions to form an action optimization vector Ψ;

[0035] S35. According to the action optimization vector Ψ, select the top K candidate actions ranked by fitness to form an action candidate sequence

[0036] S36, combining the current environment information E and the historical action execution data, The parameter space of each action in is expanded to construct the action parameter space Ω={ω1,ω2,...,ω M}, each ω j The motion parameters including motion trajectory, speed, amplitude, and energy consumption are output as the motion parameter space Ω.

[0037] Optionally, the S4 specifically includes:

[0038] S41, the obtained action parameter space Ω={ω1,ω2,...,ω M As the optimization input, initialize the position X0 of the search individual of the longhorn beetle search algorithm, the initial step size δ0 and the length of the antenna d0, and initialize the direction vector set

[0039] S42, set the multi-objective fitness function F(ω j ):

[0040] F(ω j )=γ1·S(I ** ,ω j )+γ2·N(ω j )+γ3·E(ω j )+γ4·B(ω j );

[0041] Among them, S(I ** ,ω j ) is the matching degree between intention and action parameters, N(ω j ) is the naturalness of the action, E(ω j ) is energy consumption, B(ω j ) is obstacle avoidance, γ1, γ2, γ3, γ4 are weight coefficients;

[0042] S43, in each round of search iteration, respectively in the direction vector set In each direction D i Set a pair of left and right sensing sampling points on the top, the sampling point is X t +d t ·D i and X t -d t ·D i ;

[0043] S44, for each pair of sampling points, respectively, at the corresponding position, calculate the fitness value f of each pair of sampling points according to the multi-objective fitness function L,i and f R,i ,dynamically assign the weight of each direction based on the fitness change trend of the current direction sampling point and the previous historical search performance;

[0044] S45. Based on the fitness difference in each direction and the weight of the historical direction, a weighted method is used to determine the comprehensive search direction:

[0045]

[0046] in, Indicates the comprehensive search direction of the current iteration, k indicates the total number of search directions, w t,i represents the weight of the i-th direction at the t-th iteration, and sign is the sign function;

[0047] S46, updating the position of the search individual according to the current step length and the comprehensive search direction, that is, according to the comprehensive direction guidance, moving the search individual forward one step length in the direction to obtain a new search individual position;

[0048] S47, using an adaptive adjustment mechanism to dynamically adjust the length and step size of the tentacle according to the rate of fitness improvement during the current search process. When the fitness improves rapidly, the length and step size of the tentacle are appropriately increased to enhance the global search capability. When the fitness improvement slows down, the length and step size of the tentacle are gradually reduced to enhance the local fine search capability.

[0049] S48. In each iteration, update the historical weights of each direction:

[0050] w t+1,i =α·w t,i +(1-α)·|f L,i -f R,i |;

[0051] Among them, α is the memory factor, w t+1,i represents the weight of the i-th direction in the t+1-th iteration;

[0052] S49, repeat steps S43 to S48 until the preset maximum number of iterations is reached or the fitness convergence condition is met, and record the action parameter combination with the largest fitness function value as the optimal action sequence Output the final optimal action sequence

[0053] Optionally, the S5 specifically includes:

[0054] S51, sending the obtained optimal action sequence as a command to the humanoid robot control unit, driving the robot to perform various actions in sequence according to the optimal action sequence;

[0055] S52. During the execution of the action, sensor data related to each action is collected in real time, including physical feedback information such as joint angle, end effector position, action speed, and force conditions;

[0056] S53, simultaneously collecting environmental change data and external response information, including external interaction signals such as changes in environmental obstacles, changes in user instructions, and user expressions or emotional feedback;

[0057] S54, synchronously collating and normalizing all collected action execution data and external environment and user interaction data to form a standardized real-time feedback data set;

[0058] S55. Input the standardized real-time feedback data set into the intention recognition and cognitive feedback processing process to dynamically evaluate the action execution effect and the degree of intention realization. Based on the feedback evaluation results, record key feedback parameters and typical scenario data.

[0059] Optionally, the S6 specifically includes:

[0060] S61. Receive and analyze the collected real-time feedback information, and conduct a comprehensive evaluation of the action execution effect and the current environment status;

[0061] S62. Adjust the current intention confidence based on real-time feedback information, and update the cognitive reasoning results based on the actual action execution and environmental changes;

[0062] S63, determining whether the action execution meets the completion standard of the established intention, and detecting the deviation between the degree of action realization and the recognized intention;

[0063] S64. If it is determined that the action does not fully satisfy the recognition intent, or the environment has changed significantly, immediately return the latest feedback information to the intent recognition and cognitive feedback process to restart the intent recognition and action optimization process;

[0064] S65. After re-completing the intention recognition and action optimization, repeat the action and continue to collect and evaluate feedback information, forming a closed-loop process of adaptive adjustment;

[0065] S66. When the action completion degree reaches the set requirements and the recognition intention is fully realized, the final action execution result is output and the relevant data is archived.

[0066] The beneficial effects of the present invention are:

[0067] The humanoid robot motion generation method based on the cognitive feedback mechanism provided by the present invention can effectively overcome the shortcomings of the prior art, such as the single motion generation process, the lack of an adaptive closed loop, and the separation of the cognitive layer and motion optimization. The present invention organically combines multimodal perception fusion, hierarchical intention recognition, adaptive cognitive feedback, and global optimization of motion parameters, so that the humanoid robot can understand the user's complex and changeable intentions in real time, perceive dynamic changes in the environment, and dynamically correct the motion generation process based on real-time feedback, thereby realizing intelligent and autonomous motion regulation. In terms of motion optimization, the introduction of the longicorn beetle whisker search algorithm has greatly improved the globality and efficiency of motion parameter optimization. Combined with the cognitive feedback mechanism, it can also be efficiently adjusted according to the actual needs of the user and environmental feedback to ensure the naturalness and accuracy of the motion.

[0068] The present invention realizes an intelligent closed loop of cognition-action-feedback, enabling humanoid robots to demonstrate greater flexibility and autonomous adaptability when faced with complex tasks and uncertain environments, significantly improving the system's interactive intelligence level and behavioral robustness. By collecting and processing multi-dimensional feedback information on action execution in real time, the system can continuously improve and learn from itself, enhancing its intelligent evolution capabilities under long-term operation, and further improving the safety, comfort, and user satisfaction of human-machine collaboration. Compared with traditional methods, the present invention not only realizes the refinement and intelligent control of the action generation process, but also provides solid technical support for the implementation of humanoid robots in demanding application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0070] Figure 1 This is a flow chart of a method for generating humanoid robot motions based on a cognitive feedback mechanism proposed by the present invention;

[0071] Figure 2 This is a flow chart of optimizing motion parameters using a beetle whisker search algorithm in a humanoid robot motion generation method based on a cognitive feedback mechanism proposed in the present invention. DETAILED DESCRIPTION

[0072] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0073] refer to Figure 1 and Figure 2 , a humanoid robot motion generation method based on cognitive feedback mechanism, comprising the following steps:

[0074] S1. Collect multimodal perception data of the humanoid robot's environment, preprocess the multimodal perception data, and generate a standardized perception feature vector;

[0075] S2. Based on the standardized perceptual feature vector, using the intent recognition and cognitive feedback method, through multimodal deep neural network fusion analysis, identify the user's current intent and intent confidence, generate an intent label, and output a cognitive feedback signal;

[0076] S3. According to the intent label, the candidate action sequence corresponding to the identified intent is retrieved from the action database, and the action parameter space is constructed based on the current environment information;

[0077] S4. Use the beetle whisker search algorithm to globally optimize the action parameter space, take the current candidate action plan as the search individual, and optimize the action parameters based on the beetle whisker search mechanism to obtain the optimal action sequence;

[0078] S5. Send the optimal action sequence to the humanoid robot for execution, collect real-time feedback information during the action execution process, and use it as feedback input to the intention recognition and cognitive feedback link;

[0079] S6. Adjust the intention confidence and cognitive reasoning results based on real-time feedback information. If it is detected that the action does not fully meet the recognition intention or the environment has changed significantly, return to step S2 and re-execute the intention recognition and action optimization process until the action completion meets the set requirements. If the action completion meets the set requirements, output the final action execution result.

[0080] In this embodiment, the multimodal perception data specifically includes image information, voice signals, human posture data, spatial position information and environmental physical parameters of the robot's environment.

[0081] In this embodiment, the preprocessing of the multimodal perception data specifically includes denoising and feature extraction of the image information of the multimodal perception data, framing and feature encoding of the voice signal of the multimodal perception data, normalizing the posture data of the multimodal perception data, and standardizing the spatial position information and environmental physical parameters of the multimodal perception data.

[0082] In this embodiment, S2 specifically includes:

[0083] S21. Input the obtained standardized perception feature vectors into the short-term intent recognition network, the long-term intent recognition network, and the contextual intent reasoning network respectively. Each network adopts a multimodal deep neural network structure to extract feature information related to the corresponding intent type.

[0084] S22, the short-term intention recognition network outputs the short-term intention vector I s , the long-term intention recognition network outputs the long-term intention vector I l , the contextual intention inference network outputs the contextual intention vector I c ;

[0085] S23, I s , I l , I c Input adaptive fusion process, calculate the weight coefficient α based on the current environment complexity, interaction history length and user personality characteristics s , α l , α c , perform weighted fusion on the three types of intention vectors to obtain a multi-level comprehensive intention vector I fusion ;

[0086] S24, multi-level integrated intention vector I fusion Input intent classification layer, calculate the probability distribution P(I k |I fusion ) and obtain the probability values ​​corresponding to all categories;

[0087] S25, based on the probability distribution P(I k |I fusion ), a confidence gating mechanism and historical intent transfer probability are introduced to perform confidence-weighted scoring on the current candidate intent category:

[0088] S k =λ1P(I k |I fusion )+λ2Q(I k |I hist );

[0089] Among them, S k is the comprehensive score of the k-th type of intention, Q(I k |I hist ) is the historical intention transfer probability, λ1 and λ2 are weight coefficients, and the intention category with the highest score is finally selected as the current optimal intention label I * , and the corresponding score value is used as the confidence C;

[0090] S26, introduce the knowledge graph into the intention recognition process, and convert the current multi-level comprehensive intention vector I fusion , historical scene information and environmental data are used to perform semantic reasoning with the knowledge graph, and the semantically corrected intent label I is output through knowledge graph completion and verification. ** ;

[0091] S27, the obtained intention label I ** The confidence level C is input into the cognitive feedback generation process, and the real-time feedback information of the robot's current action execution, the rate of environmental change, and the user's emotional feedback signal are simultaneously collected;

[0092] S28, the cognitive feedback generation process dynamically adjusts the feedback intensity and feedback content of the cognitive feedback signal based on the collected real-time feedback data, combined with the adaptive gain adjustment mechanism and historical interaction information, to form the cognitive feedback signal R adpt ;

[0093] S29, the cognitive feedback signal R adpt It is transmitted back to the short-term intent recognition network, long-term intent recognition network and contextual intent reasoning network to correct the intent recognition results and various intent weights in real time, realize closed-loop adaptive adjustment of intent recognition and cognitive feedback, and output the corrected final intent label, confidence and cognitive feedback signal.

[0094] In this embodiment, S3 specifically includes:

[0095] S31, according to the output final intention label I ** , confidence C and cognitive feedback signal R adpt , combined with the current environment information E to build the retrieval conditions and determine the candidate action set corresponding to the intention label

[0096] S32. For each candidate action A i Extracting action description vectors The action description vector includes action type, movement amplitude, speed, energy consumption, gait, and spatial position;

[0097] S33, according to the cognitive feedback signal R adpt And environmental information E, calculate the fitness score of each candidate action with the current intention and environment, using the weighted fitness function:

[0098]

[0099] Among them, i is the fitness of the i-th candidate action, Indicates the matching degree between intention and action, Q(A i,E) represents the adaptability of the action to the environment, β1, β2, β3, β4 are weighting coefficients;

[0100] S34, normalizing the fitness scores of all candidate actions to form an action optimization vector Ψ;

[0101] S35. According to the action optimization vector Ψ, select the top K candidate actions ranked by fitness to form an action candidate sequence

[0102] S36, combining the current environment information E and the historical action execution data, The parameter space of each action in is expanded to construct the action parameter space Ω={ω1,ω2,...,ω M}, each ω j The motion parameters including motion trajectory, speed, amplitude, and energy consumption are output as the motion parameter space Ω.

[0103] In this embodiment, the S4 specifically includes:

[0104] S41, the obtained action parameter space Ω={ω1,ω2,...,ω M As the optimization input, initialize the position X0 of the search individual of the longhorn beetle search algorithm, the initial step size δ0 and the length of the antenna d0, and initialize the direction vector set

[0105] S42, set the multi-objective fitness function F(ω j ):

[0106] F(ω j )=γ1·S(I ** ,ω j )+γ2·N(ω j )+γ3·E(ω j )+γ4·B(ω j );

[0107] Among them, S(I ** ,ω j ) is the matching degree between intention and action parameters, N(ω j ) is the naturalness of the action, E(ω j ) is energy consumption, B(ω j ) is obstacle avoidance, γ1, γ2, γ3, γ4 are weight coefficients;

[0108] S43, in each round of search iteration, respectively in the direction vector set In each direction D i Set a pair of left and right sensing sampling points on the top, the sampling point is X t +d t ·Di and X t -d t ·D i ;

[0109] S44, for each pair of sampling points, respectively, at the corresponding position, calculate the fitness value f of each pair of sampling points according to the multi-objective fitness function L,i and f R,i ,dynamically assign the weight of each direction based on the fitness change trend of the current direction sampling point and the previous historical search performance;

[0110] S45. Based on the fitness difference in each direction and the weight of the historical direction, a weighted method is used to determine the comprehensive search direction:

[0111]

[0112] in, Indicates the comprehensive search direction of the current iteration, k indicates the total number of search directions, w t,i represents the weight of the i-th direction at the t-th iteration, and sign is the sign function;

[0113] S46, updating the position of the search individual according to the current step length and the comprehensive search direction, that is, according to the comprehensive direction guidance, moving the search individual forward one step length in the direction to obtain a new search individual position;

[0114] S47, using an adaptive adjustment mechanism to dynamically adjust the length and step size of the tentacle according to the rate of fitness improvement during the current search process. When the fitness improves rapidly, the length and step size of the tentacle are appropriately increased to enhance the global search capability. When the fitness improvement slows down, the length and step size of the tentacle are gradually reduced to enhance the local fine search capability.

[0115] S48. In each iteration, update the historical weights of each direction:

[0116] w t+1,i =α·w t,i +(1-α)·|f L,i -f R,i |;

[0117] Among them, α is the memory factor, w t+1,i represents the weight of the i-th direction in the t+1-th iteration;

[0118] S49, repeat steps S43 to S48 until the preset maximum number of iterations is reached or the fitness convergence condition is met, and record the action parameter combination with the largest fitness function value as the optimal action sequence Output the final optimal action sequence

[0119] In this embodiment, the S5 specifically includes:

[0120] S51, sending the obtained optimal action sequence as a command to the humanoid robot control unit, driving the robot to perform various actions in sequence according to the optimal action sequence;

[0121] S52. During the execution of the action, sensor data related to each action is collected in real time, including physical feedback information such as joint angle, end effector position, action speed, and force conditions;

[0122] S53, simultaneously collecting environmental change data and external response information, including external interaction signals such as changes in environmental obstacles, changes in user instructions, and user expressions or emotional feedback;

[0123] S54, synchronously collating and normalizing all collected action execution data and external environment and user interaction data to form a standardized real-time feedback data set;

[0124] S55. Input the standardized real-time feedback data set into the intention recognition and cognitive feedback processing process to dynamically evaluate the action execution effect and the degree of intention realization. Based on the feedback evaluation results, record key feedback parameters and typical scenario data.

[0125] In this embodiment, S6 specifically includes:

[0126] S61. Receive and analyze the collected real-time feedback information, and conduct a comprehensive evaluation of the action execution effect and the current environment status;

[0127] S62. Adjust the current intention confidence based on real-time feedback information, and update the cognitive reasoning results based on the actual action execution and environmental changes;

[0128] S63, determining whether the action execution meets the completion standard of the established intention, and detecting the deviation between the degree of action realization and the recognized intention;

[0129] S64. If it is determined that the action does not fully satisfy the recognition intent, or the environment has changed significantly, immediately return the latest feedback information to the intent recognition and cognitive feedback process to restart the intent recognition and action optimization process;

[0130] S65. After re-completing the intention recognition and action optimization, repeat the action and continue to collect and evaluate feedback information, forming a closed-loop process of adaptive adjustment;

[0131] S66. When the action completion degree reaches the set requirements and the recognition intention is fully realized, the final action execution result is output and the relevant data is archived.

[0132] Example 1:

[0133] In order to verify the feasibility of the present invention in implementation, the present invention was applied to the Rehabilitation Medicine Center of a Class III hospital. In the daily training area of ​​the Rehabilitation Medicine Center, patients often need long-term and repeated limb rehabilitation training such as standing and walking after stroke, brain trauma or orthopedic surgery. In the past, the center mainly used traditional humanoid robots to assist patients in standing training. These robots mainly used action templates and simple perception decisions. When faced with the dynamic needs of patients, environmental changes or emergencies, they often showed stiff movements and slow reactions. For example, if a patient suddenly feels slightly unwell, needs to slow down the pace while standing, or encounters movement obstacles in the training space, traditional robots are often unable to flexibly adjust the auxiliary movements, resulting in a decrease in patient rehabilitation compliance and a high rate of medical safety alerts.

[0134] Starting from March 2025, the center will pilot the introduction of an intelligent humanoid robot based on the present invention, focusing on rehabilitation-assisted standing training for patients with moderate to severe hemiplegia. Before the training begins, the robot collects data such as the patient's voice commands, limb movements, facial expressions, spatial positions, and ambient temperature and humidity through a multimodal perception system. The system pre-processes this data and inputs it into a hierarchical intention recognition network to achieve a comprehensive judgment of the patient's immediate needs, rehabilitation goals, and environmental dynamics. For example, when a patient whispers "slow down" and shows a nervous expression while trying to stand up, the robot can comprehensively identify that the patient needs to slow down and increase auxiliary force in a short period of time, and dynamically adjust the action parameters and execution strategies accordingly.

[0135] During the motion generation and optimization phase, the robot extracts multiple sets of candidate auxiliary motion solutions from the database, and combines the beetle whisker search algorithm and cognitive feedback mechanism to perform global optimization and adaptive parameter fine-tuning. During the process, the system collects the patient's micro-expression changes, movement deviations, and temporary instructions from medical staff in real time, and transmits feedback information back to the intention recognition and optimization link, continuously adjusting parameters such as movement speed and support angle. During a complete training cycle, the robot can respond to changes in the patient's status in less than 0.5 seconds, and autonomously adjust the support mode multiple times to help patients safely complete standing training.

[0136] In actual applications, the humanoid robot of the present invention significantly outperforms traditional robots in terms of intelligent adaptability, responsiveness, and patient interaction experience. Comparative data statistics from March to June 2025 showed that a total of 286 standing training sessions were conducted using the robot of the present invention. The success rate of standing for patients increased from 66.2% for traditional robots to 92.8%; the number of automatic adjustments to movement parameters for a single training session increased from 1.3 to 5.4 times, and the average response time of the robot was shortened from 1.21 seconds to 0.41 seconds; patient satisfaction scores increased from 6.9 to 9.5, and the trigger rate of medical safety alarms decreased by 73%. No safety incidents caused by misoperation of the robot occurred during the training period.

[0137] Table 1 Comparison of robot-assisted rehabilitation training effects

[0138]

[0139] As can be seen from the data in Table 1, the humanoid robot based on the cognitive feedback mechanism of the present invention exhibits significant comprehensive advantages in rehabilitation-assisted training scenarios. The success rate of standing training increased from 66.2% of the traditional humanoid robot to 92.8%, an increase of 26.6%, which fully demonstrates that the robot of the present invention has higher effectiveness in understanding patient intentions and dynamic self-adaptation. The number of automatic adjustment of motion parameters increased significantly from 1.3 times of the traditional robot to 5.4 times, an increase of 315%. This reflects that the robot of the present invention actively and repeatedly optimizes its own movements according to patient feedback and environmental changes during training, making the assisted behavior more personalized and flexible.

[0140] The average response time of the robot of the present invention has been shortened to 0.41 seconds, which is only one-third of that of traditional robots. This has improved the system's ability to quickly respond to emergencies and changes in patient needs, and further enhanced the real-time and safety of the interaction. The patient satisfaction score has increased from 6.9 to 9.5, indicating that patients generally recognize the comfort, collaborative experience and intelligence level of the new system, which is significantly higher than that of traditional robots. The trigger rate of medical safety alarms has dropped from 3.4% to 0.9%, effectively reducing the safety risks caused by the robot's untimely response or improper actions, and ensuring the smooth progress of the rehabilitation process. The number of major safety incidents during training has dropped from 1 to 0, reflecting the system's ability to continuously optimize actions and intelligent risk avoidance. Although the total number of statistical samples is basically the same, many core indicators show the outstanding advantages and promotion value of the method of the present invention in practical applications, which can bring a higher level of intelligent service experience to the medical rehabilitation industry.

[0141] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for generating humanoid robot motions based on cognitive feedback mechanism, characterized in that: The steps include: S1. Collect multimodal perception data of the humanoid robot's environment, preprocess the multimodal perception data, and generate a standardized perception feature vector; S2. Based on the standardized perceptual feature vector, using the intent recognition and cognitive feedback method, through multimodal deep neural network fusion analysis, identify the user's current intent and intent confidence, generate an intent label, and output a cognitive feedback signal; S3. According to the intent label, the candidate action sequence corresponding to the identified intent is retrieved from the action database, and the action parameter space is constructed based on the current environment information; S4. Use the beetle whisker search algorithm to globally optimize the action parameter space, take the current candidate action plan as the search individual, and optimize the action parameters based on the beetle whisker search mechanism to obtain the optimal action sequence; S5. Send the optimal action sequence to the humanoid robot for execution, collect real-time feedback information during the action execution process, and use it as feedback input to the intention recognition and cognitive feedback link; S6. Adjust the intention confidence and cognitive reasoning results based on real-time feedback information. If it is detected that the action does not fully meet the recognition intention or the environment has changed significantly, return to step S2 and re-execute the intention recognition and action optimization process until the action completion meets the set requirements. If the action completion meets the set requirements, output the final action execution result.

2. The method for generating humanoid robot motions based on cognitive feedback mechanism according to claim 1, characterized in that: The multimodal perception data specifically includes image information, voice signals, human posture data, spatial position information and environmental physical parameters of the robot's environment.

3. The method for generating humanoid robot motions based on cognitive feedback mechanism according to claim 1, characterized in that: The preprocessing of the multimodal perception data specifically includes denoising and feature extraction of the image information of the multimodal perception data, framing and feature encoding of the voice signal of the multimodal perception data, normalizing the posture data of the multimodal perception data, and standardizing the spatial position information and environmental physical parameters of the multimodal perception data.

4. The method for generating humanoid robot motions based on cognitive feedback mechanism according to claim 1, characterized in that: The S2 specifically includes: S21. Input the obtained standardized perception feature vectors into the short-term intent recognition network, the long-term intent recognition network, and the contextual intent reasoning network respectively. Each network adopts a multimodal deep neural network structure to extract feature information related to the corresponding intent type. S22, the short-term intention recognition network outputs the short-term intention vector I s , the long-term intention recognition network outputs the long-term intention vector I l , the contextual intention inference network outputs the contextual intention vector I c ; S23, I s , I l , I c Input adaptive fusion process, calculate the weight coefficient α based on the current environment complexity, interaction history length and user personality characteristics s , α l , α c , perform weighted fusion on the three types of intention vectors to obtain a multi-level comprehensive intention vector I fusion ; S24, multi-level integrated intention vector I fusion Input intent classification layer, calculate the probability distribution P(I k |I fusion ) and obtain the probability values ​​corresponding to all categories; S25, based on the probability distribution P(I k |I fusion ), a confidence gating mechanism and historical intent transfer probability are introduced to perform confidence-weighted scoring on the current candidate intent category: S k =λ1P(I k |I fusion )+λ2Q(I k |I hist ); Among them, S k is the comprehensive score of the k-th type of intention, Q(I k |I hist ) is the historical intention transfer probability, λ1 and λ2 are weight coefficients, and the intention category with the highest score is finally selected as the current optimal intention label I * , and the corresponding score value is used as the confidence C; S26, introduce the knowledge graph into the intention recognition process, and convert the current multi-level comprehensive intention vector I fusion , historical scene information and environmental data are used to perform semantic reasoning with the knowledge graph, and the semantically corrected intent label I is output through knowledge graph completion and verification. ** ; S27, the obtained intention label I ** The confidence level C is input into the cognitive feedback generation process, and the real-time feedback information of the robot's current action execution, the rate of environmental change, and the user's emotional feedback signal are simultaneously collected; S28, the cognitive feedback generation process dynamically adjusts the feedback intensity and feedback content of the cognitive feedback signal based on the collected real-time feedback data, combined with the adaptive gain adjustment mechanism and historical interaction information, to form the cognitive feedback signal R adpt ; S29, the cognitive feedback signal R adpt It is transmitted back to the short-term intent recognition network, long-term intent recognition network and contextual intent reasoning network to correct the intent recognition results and various intent weights in real time, realize closed-loop adaptive adjustment of intent recognition and cognitive feedback, and output the corrected final intent label, confidence and cognitive feedback signal.

5. The method for generating humanoid robot motions based on cognitive feedback mechanism according to claim 1, characterized in that: The S3 specifically includes: S31, according to the output final intention label I ** , confidence C and cognitive feedback signal R adpt , combined with the current environment information E to build the retrieval conditions and determine the candidate action set corresponding to the intention label S32. For each candidate action A i Extracting action description vectors The action description vector includes action type, movement amplitude, speed, energy consumption, gait, and spatial position; S33, according to the cognitive feedback signal R adpt And environmental information E, calculate the fitness score of each candidate action with the current intention and environment, using the weighted fitness function: Among them, i is the fitness of the i-th candidate action, Indicates the matching degree between intention and action, Q(A i ,E) represents the adaptability of the action to the environment, β1, β2, β3, β4 are weighting coefficients; S34, normalizing the fitness scores of all candidate actions to form an action optimization vector Ψ; S35. According to the action optimization vector Ψ, select the top K candidate actions ranked by fitness to form an action candidate sequence S36, combining the current environment information E and the historical action execution data, The parameter space of each action in is expanded to construct the action parameter space Ω={ω1,ω2,...,ω M }, each ω j The motion parameters including motion trajectory, speed, amplitude, and energy consumption are output as the motion parameter space Ω.

6. The method for generating humanoid robot motions based on cognitive feedback mechanism according to claim 1, characterized in that: The S4 specifically includes: S41, the obtained action parameter space Ω={ω1,ω2,...,ω M As the optimization input, initialize the position X0 of the search individual of the longhorn beetle search algorithm, the initial step size δ0 and the length of the antenna d0, and initialize the direction vector set S42, set the multi-objective fitness function F(ω j ): F(ω j )=γ1·S(I ** ,oh j )+γ2·N(ω j )+γ3·E(ω j )+γ4·B(ω j ); Among them, S(I ** ,ω j ) is the matching degree between intention and action parameters, N(ω j ) is the naturalness of the action, E(ω j ) is energy consumption, B(ω j ) is obstacle avoidance, γ1, γ2, γ3, γ4 are weight coefficients; S43, in each round of search iteration, respectively in the direction vector set In each direction D i Set a pair of left and right sensing sampling points on the top, the sampling point is X t +d t ·D i and X t -d t ·D i ; S44, for each pair of sampling points, respectively, at the corresponding position, calculate the fitness value f of each pair of sampling points according to the multi-objective fitness function L,i and f R,i ,dynamically assign the weight of each direction based on the fitness change trend of the current direction sampling point and the previous historical search performance; S45. Based on the fitness difference in each direction and the weight of the historical direction, a weighted method is used to determine the comprehensive search direction: in, Indicates the comprehensive search direction of the current iteration, k indicates the total number of search directions, w t,i represents the weight of the i-th direction at the t-th iteration, and sign is the sign function; S46, updating the position of the search individual according to the current step length and the comprehensive search direction, that is, according to the comprehensive direction guidance, moving the search individual forward one step length in the direction to obtain a new search individual position; S47, using an adaptive adjustment mechanism to dynamically adjust the length and step size of the tentacle according to the rate of fitness improvement during the current search process. When the fitness improves rapidly, the length and step size of the tentacle are appropriately increased to enhance the global search capability. When the fitness improvement slows down, the length and step size of the tentacle are gradually reduced to enhance the local fine search capability. S48. In each iteration, update the historical weights of each direction: w t+1,i =α·w t,i +(1-a)·|f L,i -f R,i |; Among them, α is the memory factor, w t+1,i represents the weight of the i-th direction in the t+1-th iteration; S49, repeat steps S43 to S48 until the preset maximum number of iterations is reached or the fitness convergence condition is met, and record the action parameter combination with the largest fitness function value as the optimal action sequence Output the final optimal action sequence 7. The method for generating humanoid robot motions based on cognitive feedback mechanism according to claim 1, characterized in that: The S5 specifically includes: S51, sending the obtained optimal action sequence as a command to the humanoid robot control unit, driving the robot to perform various actions in sequence according to the optimal action sequence; S52. During the execution of the action, sensor data related to each action is collected in real time, including physical feedback information such as joint angle, end effector position, action speed, and force conditions; S53, simultaneously collecting environmental change data and external response information, including external interaction signals such as changes in environmental obstacles, changes in user instructions, and user expressions or emotional feedback; S54, synchronously collating and normalizing all collected action execution data and external environment and user interaction data to form a standardized real-time feedback data set; S55. Input the standardized real-time feedback data set into the intention recognition and cognitive feedback processing process to dynamically evaluate the action execution effect and the degree of intention realization. Based on the feedback evaluation results, record key feedback parameters and typical scenario data.

8. The method for generating humanoid robot motions based on cognitive feedback mechanism according to claim 1, characterized in that: The S6 specifically includes: S61. Receive and analyze the collected real-time feedback information, and conduct a comprehensive evaluation of the action execution effect and the current environment status; S62. Adjust the current intention confidence based on real-time feedback information, and update the cognitive reasoning results based on the actual action execution and environmental changes; S63, determining whether the action execution meets the completion standard of the established intention, and detecting the deviation between the degree of action realization and the recognized intention; S64. If it is determined that the action does not fully satisfy the recognition intent, or the environment has changed significantly, immediately return the latest feedback information to the intent recognition and cognitive feedback process to restart the intent recognition and action optimization process; S65. After re-completing the intention recognition and action optimization, repeat the action and continue to collect and evaluate feedback information, forming a closed-loop process of adaptive adjustment; S66. When the action completion degree reaches the set requirements and the recognition intention is fully realized, the final action execution result is output and the relevant data is archived.

Images