Character shaping robot

Abstract

The present application relates to the cross field of artificial intelligence and mental health, clinical psychology, specifically relates to an intelligent robot system integrating dynamic cognitive computing architecture, physiological-behavior multi-modal fusion perception, XR-LLM collaborative intervention engine, through dynamic personality formation, multi-dimensional personality test and personalized intervention strategy, realizing accurate assessment, dynamic modeling, personalized intervention and adaptive growth of users. Embodiments cover education, medical treatment, enterprise team building and other scenes, the present application has remarkable novelty, creativity and practicality, and can be widely applied in service, medical treatment, education and other fields.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of artificial intelligence and mental health and clinical psychology, specifically to an intelligent robot system that integrates a dynamic cognitive computing architecture, physiological-behavioral multimodal fusion perception, and an XR-LLM collaborative intervention engine. Through dynamic personality formation, multi-dimensional personality testing, and personalized intervention strategies, it can achieve accurate user assessment, dynamic modeling, personalized intervention, and adaptive growth, and can be widely applied in service, medical, and education fields. Background Technology

[0002] Existing technological shortcomings: Traditional personality tests (such as MBTI and Big Five personality traits) rely on static questionnaires and cannot capture dynamic behavioral characteristics; The psychological intervention system lacks real-time interactive capabilities, and the intervention strategies are rigid; VR psychotherapy systems lack integrated cognitive computing models and have poor scene adaptability.

[0003] Therefore, developing a robot with personality shaping and intervention capabilities, based on our existing patented intelligent psychological counseling robot (patent number ZL2019105928955) and mental health intervention robot (CN2025111561722), has significant practical implications. Summary of the Invention

[0004] Core innovations:

[0005] Dynamic personality formation engine: Based on LLM, a robot personality model is built, and personality parameters are continuously optimized through multimodal interaction data (voice, facial expression, behavior), forming a closed loop of "initial personality setting → user interaction feedback → model iterative update"; Initial personality library: 50 basic personality templates are preset (such as "patient" and "active"), which can be quickly adapted to different scenarios through transfer learning; Dynamic adjustment algorithm: Using a reinforcement learning framework, user satisfaction is used as a reward signal to adjust personality parameters (such as "openness" and "conscientiousness" dimensions) in real time.

[0006] Multidimensional personality testing system Design a three-level testing system: Basic test: Assess user personality type through standardized dialogues (such as "How do you view failure?") (modified MBTI); Scenario testing: Simulate high-stress environments (such as negotiation, emergency rescue) and observe user behavior responses; Long-term tracking: Record user interaction data over 3 months to generate a personality change curve.

[0007] Personalized intervention strategy library Intervention strategies were developed to address 10 typical personality traits, including anxiety, depression, and impulsivity. Cognitive restructuring: Generating positive guiding statements through LLM (such as "Your decision just now was very decisive, but you could try to plan more meticulously"). Behavioral substitution: Transforming impulsive behaviors into substitute tasks (e.g., guiding "smashing objects" to "squeezing a stress ball"); Environmental adjustment: Adjust the robot's interaction mode (e.g., reduce the speech speed for introverted users and increase visual cues).

[0008] A "dynamic cognitive engine" architecture is proposed to achieve real-time evolutionary modeling of personality traits; Construct an XR-LLM collaborative intervention mechanism to support multi-round adaptive scene generation; Develop a cross-modal personality decoding algorithm that integrates 12 dimensions of data, including voice, facial expressions, and physiological signals.

[0009] Key Algorithm:

[0010] Personality trait dynamic update algorithm Input: Multimodal data stream D_t = {v_t, a_t, e_t, ...} Output: Updated personality vector CFG_{t+1} step: 1. Extracting temporal features using BiLSTM 2. Calculate the personality drift coefficient: δ_t = σ(W_δ · [CFG_t; D_t] + b_δ) 3. Dynamic updates: CFG_{t+1} = (1-δ_t)·CFG_t + δ_t·MLP(D_t).

[0011] XR scene generation algorithm Input: User's current personality state CFG_t, intervention target G Output: Optimized XR scene parameters θ Optimization goal: min L = α·L_engagement + β·L_progress + γ·L_safety in: L_engagement = -log(P(user_action|θ)) # User engagement L_progress = ||CFG_t - G||_2 # Intervention progress L_safety = Σ|θ_i - θ_default| # Safety constraint.

[0012] When the Dynamic Cognitive Engine (DCE) adopts a three-tier architecture: Input layer → Feature decoupling layer → State estimation layer │ │ │ Multimodal data → 128-dimensional personality vector → Intervention decision space (Eye movement / EMG / speech) (CFG v3.2) (MCTS search tree).

[0013] Personality trait dynamic update algorithm 2 (pseudocode) def update_cfg(D_t, CFG_t): #Temporal Attention Weighting alpha = sigmoid(W_alpha @concat(CFG_t, D_t)) # Physiological signal modulation hrv_factor = 1 - min(HRV_t / HRV_baseline, 1.0) alpha *= hrv_factor # Reduce update rate under anxiety state # Feature Decoupling Update CFG_{t+1} = (1-alpha)*CFG_t + alpha*MLP(D_t) return CFG_{t+1}.

[0014] When the Dynamic Cognitive Engine (DCE) adopts a four-layer progressive architecture: Input layer → Feature decoupling layer → State estimation layer → Decision output layer │ │ │ │ Multimodal data → 128-dimensional personality vector → Dynamic cognitive map → Intervention strategy space (Eye movement / EMG / voice / EEG) (CFG v4.0) (MCTS search tree).

[0015] Personality Trait Dynamic Update Algorithm 3 def adaptive_cfg_update(D_t, CFG_t, bio_signals): # Temporal attention weighting (physiological signal modulation) alpha = sigmoid(W_alpha @concat(CFG_t, D_t)) hrv_factor = 1 - min(HRV_t / HRV_baseline, 1.0) eda_factor = 1 + tanh(EDA_t / EDA_threshold) * 0.3 alpha *= hrv_factor * eda_factor # Reduce update rate under anxiety state # Feature decoupling update (adversarial training mechanism) CFG_{t+1} = (1-alpha)*CFG_t + alpha*MLP(D_t) CFG_{t+1} += adversarial_perturbation(CFG_{t+1}, epsilon=0.05) # Enhance robustness return CFG_{t+1}.

[0016] Module technical features

[0017] Personality trait update sequence Time axis: t-2 t-1 t t+1 t+2 CFG values: 0.72 0.75 0.78 0.76 0.79 └────┬────┘ └────┬────┘ Physiological modulation backpropagation correction Attached Figure Description

[0018] Figure 1 This is a system architecture diagram of the present invention.

[0019] Figure 2 The system architecture diagram of this invention is presented in tabular form.

[0020] Figure 3 This is a comparison table of technical parameters for some embodiments. Detailed Implementation

[0021] The principles and features of this application will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific methods and embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0022] Example 1: Character Shaping in Educational Scenarios The robot was initially programmed with a "guide" personality, interacting with an introverted student. Long-term tracking revealed that students' participation in classroom discussions was below a threshold, triggering an intervention strategy: The personality test module identified students with low scores on the "openness" dimension; The intervention strategy library invokes the "progressive questioning" method, increasing the number of open-ended questions by 10% in each conversation; Three months later, student participation increased by 40%, and the personality model was automatically updated to "guidance + encouragement".

[0023] Example 2: Emotional Intervention in Medical Rehabilitation A robot equipped with an "empathetic" personality interacted with a person suffering from anxiety disorder. In the scenario test simulating a "sudden illness" situation, the patient exhibited symptoms of hyperventilation. Intervention strategy initiated: LLM generates breathing regulation instructions ("Inhale deeply for 4 seconds, hold your breath for 2 seconds, and exhale slowly for 6 seconds"). The robot's tactile sensors detect changes in the patient's heart rate and dynamically adjust the guidance rhythm. After the intervention, the patient's anxiety scale score decreased by 25%.

[0024] Example 3: Personality Fit in Enterprise Team Building The robot, acting as a team coordinator, was initially programmed with a "neutral" personality. Multidimensional testing revealed that 60% of team members were "dominant," triggering personality adjustments. The robot's personality model incorporates a "compromise" parameter to reduce decision-making conflicts; The intervention strategy library invokes a "rotation speaking" mechanism to ensure that each member's speaking time accounts for ≥15%; Team efficiency increased by 30%, and conflicts decreased by 50%.

[0025] Example 4: Intergenerational Interaction in Home Services The robot needs to serve both elderly users (conservative type) and teenagers (open type). The dynamic personality formation engine activates the "time-based personality" mode: During interactions with elderly users: The robot adopts a "patient" personality, enlarging the font and increasing the volume; Interaction time for teenage users: Switch to "active" personality and increase popular culture topics; User satisfaction surveys show that cross-generational acceptance has increased by 60%.

[0026] Example 5: Autism Intervention in Special Education Robots are designed to create "structured personality" models for children with autism. The personality test module assesses children's attention span threshold through eye tracking; The intervention strategy library uses the "visual cue-gradual extension" method to increase attention span by 5% for each task; After 6 months of intervention, the number of children’s social responses increased threefold.

[0027] Example 6: Social Anxiety Intervention Initial assessment: Through VR social scenario testing, it was detected that users avoided eye contact more than 80% of the time. Intervention strategies: Week 1: Virtual mentor-led progressive exposure training (from one-person conversations to five-person meetings) Week 2: Introduce AI role-playing to simulate conflict scenarios, and LLM will generate real-time resolution scripts. Results: After intervention, the frequency of eye contact increased to 45%, and the DSM-5 Social Anxiety Scale score decreased by 32%.

[0028] Example 7: Adjustment of Impulsive Personality Dynamic detection: Electromyography (EMG) signals revealed that the user's decision-making time was <0.8s (normal range 1.2-2.5s). Intervention plan: The XR Maze mission includes a delayed decision-making reward mechanism. LLM generates a "Think Before You Act" prompt, combined with breathing training guidance. Results: Decision time increased to 1.5 seconds, and Barratt Impulsivity Scale score decreased by 27%.

[0029] Example 8: Cross-cultural adaptation training Scenario building: Based on the cultural characteristics of the user's destination (such as collectivist / individualistic tendencies) Intervention content: VR simulations of cultural conflict scenarios (such as differences in team decision-making styles) LLM provides a library of cultural adaptation strategies, supporting real-time translation and contextual interpretation. Assessment: Cultural Adaptation Scale (CAS) score improved by 41%.

[0030] Example 9: Cognitive Restructuring of Depression Negative thinking test: Identify negative statements (sad tone lasting >2 seconds) through voice sentiment analysis. Combined with a decrease in EEG alpha wave power (<8μV² / Hz) Intervention strategies: VR generates positive scenarios (such as sunny beaches) and simultaneously plays LLM-generated cognitive restructuring scripts. "You were thinking, 'I can't do anything right,' but the truth is you successfully completed task XX." Effect verification: BDI-II score changed from 29 to 14 (moderate to mild). EEG indicators: Alpha wave power in the left prefrontal cortex increased by 28%.

[0031] Example 10: Leadership Development Training Personality modeling: Identifying user decision-making styles (analytical / intuitive) Customized training: Setting time pressure in VR crisis management scenarios LLM provides a strategic decision-making framework and offers real-time feedback on decision quality. Improvement: The Leadership Five Factors (LFM) score increased by 36%.

[0032] Example 11: Exposure therapy for post-traumatic stress disorder (PTSD) Dynamic detection: Identifying traumatic memory trigger points using electrical conduction analysis (EDA) (signal amplitude surge >300%). Combine eye-tracking to locate the visual attention focus (area with gaze duration > 2s). Intervention strategies: VR scene exposure is tiered (starting at 5% intensity, increasing by 15% daily). LLM real-time generation system desensitization script: "What you are experiencing now is a memory replay, not a real danger." Effect verification: CAPS-5 score improved from 48 to 22 (significant clinical improvement). Physiological indicators: EDA baseline decreased by 62%, and heart rate variability (RMSSD) increased by 41%.

[0033] Example 12: Social Training for Autism Spectrum Disorder (ASD) Multimodal evaluation: Speech features: intonation flatness (F0 range < 2 semitones) Behavioral characteristic: Duration of joint attention (<1.2s) VR intervention design: Dynamic social scene generation (number of characters adjusted based on user tolerance). LLM social cue system: "Xiao Ming is looking you in the eye; he might want to share his toy with you." Quantitative results: SRS-2 Social Response Scale score decreased by 37%. The duration of joint attention increased to 2.8 seconds (close to the typical developmental level).

[0034] Example 13: Workplace Stress Management Training Pressure source identification: Neck and shoulder tension was measured using electromyography (EMG) (RMS value > 15 μV). Combined with speech rate (>4 words / second) Intervention strategies: VR-generated relaxation scene (forest waterfall) with synchronized progressive muscle relaxation instructions. LLM recommends stress management: "Break large tasks into three 20-minute sub-tasks." Quantitative results: The PSS-10 stress scale score decreased by 33%. Physiological indicators: The EMG RMS value of the neck and shoulders decreased by 42%.

[0035] Example 14: Addictive Behavior Intervention Training Desire state detection: Increased skin temperature (SKT) (>0.8℃) Combined with pupil diameter dilation (>0.5mm) Intervention strategies: VR generates aversive scenes (such as images showing lung damage caused by smoking). LLM (Lesson on the Power of Rejection) training: "When a friend offers you a cigarette, you can say, 'I'm trying a healthy lifestyle.'" Effect verification: The frequency of recording addictive behavior diaries decreased by 67%. Physiological indicators: SKT peak decreased by 51%.

[0036] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0037] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A personality-shaping intelligent robot based on a large language model, characterized in that, include: Multimodal sensing layer; The dynamic cognitive engine includes: a personality trait graph (CFG) that uses a time-weighted update mechanism modulated by physiological signals, with weight coefficients calculated in real time by electrical conductance of the skin (EDA), heart rate variability (HRV), and pupil diameter change rate; and a context-aware module that integrates a spatiotemporal Transformer encoder and a knowledge graph reasoning engine, supporting 10^7 level scene branch generation. The XR interaction layer supports: dynamically rendering intervention scenarios based on the user's personality state, with scene complexity being negatively correlated with cognitive load index (C=k*(1-σ(W·[CFG_t; L_t]))); and a multimodal feedback channel that maps at least 5 types of physiological signals (EDA / HRV / EMG / SKT / Pupil) to scene visual parameters. The core of the large language model is fine-tuned with 15 billion rounds of personality dialogue data, including a reinforcement learning dataset with DSM-5 diagnostic labels; and optimized with a dual reward function, incorporating both the Emotion Regulation Efficacy Index (ERI) and the Safety Boundary Constraint for Intervention (SBC). The intervention effectiveness evaluation includes: a multi-dimensional personality testing system, comprising three levels of evaluation modules: basic testing, scenario testing, and long-term follow-up; a personalized intervention strategy library, providing intervention strategies for at least 10 typical personality defects; and a closed-loop feedback mechanism, inputting user feedback data into the personality model for iterative optimization.

2. The system according to claim 1, characterized in that... The dynamic cognitive engine further includes: The personality drift detection submodule uses a two-stream convolutional network to process static personality features and dynamic behavior sequences respectively. Intervention safety boundary constraint, and establishment of dynamic threshold model based on user baseline physiological data.

3. The system according to claim 1, characterized in that... The XR interaction layer includes: The multi-user collaborative intervention module supports therapists to intervene in real time through digital twins; Cross-modal feedback channels map user physiological signals to scene visual parameters (such as heart rate variability → ambient light color temperature).

4. The system according to claim 1, characterized in that... The core of the large language model adopts: The three-stage training process consists of: pre-training (1.5 trillion tokens) + personality fine-tuning (12 billion rounds of dialogue) + reinforcement learning (PPO algorithm, with an environment simulator that includes a personality evolution model). The real-time knowledge distillation architecture updates the base model and the lightweight student model synchronously.

5. The system according to claim 1, characterized in that... Further includes: The personality intervention effect prediction module is based on graph neural network (GNN) modeling the user-scenario-policy ternary relationship; The privacy protection subsystem employs a federated learning framework to achieve collaborative training of multi-center data.

6. The system according to claim 1, characterized in that... The physiological signals include: Electrodermal conductance (EDA); Heart rate variability (HRV); Facial electromyography (EMG) (fEMG); Rate of change in pupil diameter; Electroencephalography (EEG); Respiratory rate (RR).

7. The system according to claim 1, characterized in that... Further includes: The federated learning module supports collaborative training across multiple data centers without leaking the original data. Differential privacy protection mechanism, noise addition intensity ε=0.3 to 0.5 (satisfying (ε,δ)-DP).

8. The system according to claim 1, characterized in that... The large language model adopts: Hybrid Expert Architecture (MoE) comprises 12 domain expert subnetworks; A dynamic routing mechanism that dynamically activates relevant experts based on user personality traits.

9. A personality testing system according to claim 1, characterized in that, This can be achieved through the following steps: Collect user data; Basic personality classification is based on at least one of the two scales: the modified MBTI scale and the Big Five personality traits. Observe user behavior responses in simulated high-stress scenarios; A personality change curve is generated by combining long-term interaction data.

Images