Automobile robot and autonomous decision control method thereof

Abstract

The application relates to the technical field of vehicle control, in particular to an automobile robot and an autonomous decision control method thereof, wherein the automobile robot comprises: a vehicle-mounted perception module configured to acquire environment information and user information after the vehicle is started; a demand understanding module configured to perform fusion analysis on the environment information and the user information to obtain a user intention; an autonomous behavior decision module configured to generate an autonomous decision result comprising a driving action sequence, a driving speed and a parking position within a preset safety compliance framework according to the environment information and the user intention through a pre-constructed multi-target optimization decision model; a planning control module configured to generate a motion trajectory based on the driving action sequence and the parking position, combine the motion trajectory with the driving speed to obtain chassis control instructions and send the chassis control instructions to an actuator module; and the actuator module configured to receive and execute the chassis control instructions.

Description

Technical Field

[0001] This application relates to the field of vehicle control technology, and in particular to an automotive robot and its autonomous decision-making and control method. Background Technology

[0002] With the development of autonomous driving technology, autonomous vehicles and intelligent cars have gradually acquired standardized driving functions such as following other vehicles, lane keeping, and automatic parking. These functions are generally implemented using a command-driven mode.

[0003] Before departure or during the journey, users pre-set their destination through the in-vehicle interface. The vehicle then plans a navigation route based on this destination and performs standardized driving operations such as following the vehicle in front, lane keeping, and automatic parking along the navigation route. Throughout the journey, the vehicle's path is pre-defined by the navigation route, the driving speed is usually manually set by the user or maintained according to the cruise control system's preset value, and the parking location also needs to be pre-specified by the user in the navigation system. All driving behaviors of the vehicle are triggered and bounded by user commands.

[0004] Because the existing command-driven mode requires users to issue fine-grained control commands one by one, the vehicle can only respond to explicit commands. Furthermore, driving speed strategies largely rely on preset cruise values ​​or fixed following logic, making it difficult to dynamically adjust the optimal driving rhythm based on real-time road conditions and user status. Since parking locations require users to specify parking lots or spaces in advance, the vehicle lacks the ability to autonomously search for and assess nearby parking resources. And because all actions require human intervention, the vehicle cannot proactively provide service without a command trigger.

[0005] In conclusion, existing intelligent vehicles are essentially still passive automated transportation tools, which are unable to meet users' growing demand for intelligent and humanized travel services. Summary of the Invention

[0006] (a) Technical problems to be solved In view of the above-mentioned shortcomings and deficiencies of the prior art, this application provides an automotive robot and its autonomous decision-making and control method, which solves the technical problems of existing automotive robots lacking autonomous decision-making capabilities and failing to achieve the closed-loop operation characteristics of robots.

[0007] (II) Technical Solution To achieve the above objectives, the main technical solutions adopted in this application include: In a first aspect, embodiments of this application provide an automotive robot, comprising: an onboard perception module for acquiring environmental information and user information after the vehicle is started; a demand understanding module for fusing and analyzing the environmental information and user information to obtain user intent; an autonomous behavior decision-making module for generating autonomous decision-making results within a preset safety and compliance framework based on the environmental information and user intent through a pre-built multi-objective optimization decision-making model; the autonomous decision-making results including a driving action sequence, driving speed, and parking position; a planning and control module for generating a motion trajectory that conforms to the user intent based on the driving action sequence and parking position, combining the motion trajectory with the driving speed to obtain chassis control commands, and sending them to an execution mechanism module; and an execution mechanism module for receiving and executing the chassis control commands.

[0008] Optionally, in some embodiments of this application, the demand understanding module includes a trained hierarchical dynamic Bayesian network model; the user information is multimodal interaction perception data, including explicit instruction input and / or implicit state information; in the demand understanding module, the fusion analysis of environmental information and user information to obtain user intent includes: performing semantic abstraction processing on the environmental information to generate a context modulation factor of the current interaction scenario semantics; inputting the multimodal interaction perception data as observation evidence, together with the context modulation factor, into the hierarchical dynamic Bayesian network model to obtain the posterior probability distribution of the user's mental state; the user's mental state includes physiological load, emotional valence, and time urgency; and generating the user intent based on the posterior probability distribution.

[0009] Optionally, in some embodiments of this application, generating the user intent based on the posterior probability distribution includes: obtaining an interaction strategy weight factor; the interaction strategy weight factor is determined by performing a confidence assessment on the posterior probability distribution and / or an intermodal consistency assessment on the multimodal interaction perception data; according to the posterior probability distribution and the interaction strategy weight factor, obtaining the expected utility value of each candidate demand parsing action through a preset utility function; the utility function represents the expected utility value of performing different candidate demand parsing actions under different mental states; selecting the parsing result corresponding to the candidate demand parsing action with the highest expected utility value as the user intent, and normalizing the highest expected utility value as the confidence level of the user intent.

[0010] Optionally, in some embodiments of this application, the hierarchical dynamic Bayesian network model includes an observation layer, a mental state layer, and an interaction expectation layer; the observation layer includes observation variable nodes corresponding to the interactive perception data of each modality, and is connected to the mental state layer through an observation likelihood function; the mental state layer includes latent variable nodes for physiological load, emotional valence, and time urgency, and the latent variable nodes are connected by directed edges to represent causal dependencies, which are used to output the posterior probability distribution; the interaction expectation layer includes latent variable nodes for interaction mode preference, and takes the posterior probability distribution and the context modulation factor as input to output a probability distribution representing the user's acceptance of the candidate demand parsing action in the current scenario; wherein, the acceptance probability distribution is used to correct each expected utility value.

[0011] Optionally, in some embodiments of this application, the autonomous behavior decision-making module generates an autonomous decision result within a preset safety and compliance framework based on the environmental information and user intent using a pre-constructed multi-objective optimization decision model. This includes: determining a safety margin modulation coefficient based on the confidence level of the user intent, and adaptively modulating the safety boundary of the preset safety and compliance framework; wherein, the lower the confidence level of the user intent, the larger the safety margin modulation coefficient, and the more conservative the safety boundary; determining a decision target weight vector based on the user intent and the environmental information; using the modulated safety boundary as a constraint and the decision target weight vector as the objective function, solving the multi-objective optimization decision model to generate multiple candidate decision schemes; the candidate decision schemes at least include candidate driving action sequences, candidate driving speeds, and candidate parking positions; performing multi-objective evaluation on the multiple candidate decision schemes, and selecting the candidate decision scheme with the highest comprehensive evaluation value as the autonomous decision result output.

[0012] Optionally, in some embodiments of this application, determining the decision target weight vector based on the user intent and the environmental information includes: matching a basic weight vector in a pre-configured intent-weight mapping table according to the discrete behavioral labels of the user intent; the intent-weight mapping table represents the correspondence between different candidate intent types and at least one of timeliness weight, comfort weight, safety weight, and energy consumption weight; extracting the scene complexity index from the environmental information; dynamically correcting the basic weight vector using the scene complexity index to obtain the decision target weight vector; wherein, when the scene complexity index is higher than a preset threshold, the safety weight is increased and the comfort weight is decreased.

[0013] Optionally, in some embodiments of this application, the multi-objective optimization decision model is configured with a multi-objective trade-off function; the multi-objective trade-off function is used to evaluate candidate decision schemes from multiple dimensions, the multiple dimensions including at least two of timeliness, comfort, safety, and energy consumption; the multi-objective optimization decision model uses the modulated safety boundary as a constraint condition, and the decision objective weight vector as the weight configuration corresponding to each dimension in the multi-objective trade-off function to solve the weighted optimal solution of the multi-objective trade-off function; the candidate decision scheme corresponding to the weighted optimal solution is output as the candidate decision scheme with the highest comprehensive evaluation value.

[0014] Optionally, in some embodiments of this application, the step of generating a motion trajectory that conforms to the user's intention based on the driving action sequence and the parking position, and combining the motion trajectory with the driving speed to obtain chassis control commands, includes: determining the macroscopic driving path topology of the vehicle according to the driving action sequence, and using the parking position as the endpoint constraint of the macroscopic driving path topology; generating a continuous reference path that satisfies the vehicle's kinematic constraints based on the macroscopic driving path topology using a curve smoothing algorithm; performing spatiotemporal fusion of the continuous reference path and the driving speed to generate a spatiotemporal motion trajectory that includes position, heading angle, and target speed; and calculating the chassis control commands through a trajectory tracking controller based on the deviation between the spatiotemporal motion trajectory and the vehicle's current actual state.

[0015] Optionally, in some embodiments of this application, the car robot further includes a closed-loop correction module; the closed-loop correction module is used to continuously acquire real-time perception data from the vehicle perception module and monitor environmental changes and user state changes after the autonomous behavior decision-making module outputs the autonomous decision-making result and the execution mechanism module begins execution; when the amount of change in the environmental information exceeds a preset environmental change threshold, and / or the amount of change in the user information exceeds a preset user state change threshold, the closed-loop correction module sends a replanning trigger signal to the autonomous behavior decision-making module; the autonomous behavior decision-making module responds to the replanning trigger signal, makes a new autonomous decision based on the updated environmental information and user intent, generates an updated autonomous decision-making result, and the updated decision result is executed by the planning control module and the execution mechanism module; the above continuous monitoring and dynamic updating process is executed cyclically until the target parking location is reached or the user terminates the travel task.

[0016] Secondly, embodiments of this application provide an autonomous decision-making and control method for an automotive robot.

[0017] This application proposes an autonomous decision-making and control method for an automotive robot, comprising: acquiring environmental information and user information after vehicle startup; performing fusion analysis on the environmental information and user information to obtain user intent; generating autonomous decision results within a preset safety and compliance framework based on the environmental information and user intent through a pre-constructed multi-objective optimization decision model; the autonomous decision results including a driving action sequence, driving speed, and parking position; generating a motion trajectory conforming to the user intent based on the driving action sequence and parking position; combining the motion trajectory with the driving speed to obtain chassis control commands; and executing the chassis control commands.

[0018] (III) Beneficial Effects The beneficial effects of this application are as follows: The automotive robot and its autonomous decision-making and control method of this application, by using an onboard perception module to acquire real-time environmental information and user information, and then using a demand understanding module to fuse and analyze the two to generate user intent, and then using an autonomous behavior decision-making module to generate an autonomous decision result including driving action sequence, driving speed and parking position based on the environmental information and user intent within a preset safety and compliance framework, and calling a multi-objective optimization decision model, generates an autonomous decision result. Then, a planning and control module generates a motion trajectory based on the driving action sequence and parking position, and combines the motion trajectory with the driving speed to obtain chassis control commands. Finally, the actuator module executes the chassis control commands. Compared with the prior art, it can solve the technical problems of existing automotive robots lacking autonomous decision-making ability and failing to achieve the closed-loop operation characteristics of robots. It enables automotive robots to get rid of dependence on users' detailed instructions during operation and achieve autonomous operation from environmental perception, user intent understanding to behavior decision-making and motion control. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the structure of an automotive robot provided in an embodiment of this application; Figure 2 A schematic diagram of the structure of the hierarchical dynamic Bayesian network model provided in the embodiments of this application; Figure 3 This is a flowchart illustrating an autonomous decision-making and control method for an automotive robot, as provided in an embodiment of this application.

[0020] [Explanation of Labels in the Attached Image] 100: Vehicle-mounted sensing module; 200: Requirements Understanding Module; 300: Autonomous Behavior Decision-Making Module; 400: Planning and Control Module; 500: Actuator module; 21: Observation layer; 22: Mental state layer; 23: Interaction expectation layer; 22a: Latent variable node of time urgency; 22b: Latent variable node of physiological load; 22c: Latent variable node of emotional valence. Detailed Implementation

[0021] To better understand the above technical solutions, exemplary embodiments of this application will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application can be understood more clearly and thoroughly, and that the scope of this application can be fully conveyed to those skilled in the art.

[0022] Figure 1 This is a schematic diagram of the structure of a car robot provided in an embodiment of this application. The car robot can be mounted within a car control system, and after the user starts the vehicle, it actively interprets the user's intent, generates autonomous decisions, and executes them, such as... Figure 1 As shown, the car robot includes: The vehicle-mounted sensing module 100 is used to acquire environmental and user information after the vehicle is started.

[0023] The requirement understanding module 200 is used to integrate and analyze environmental and user information to obtain user intent.

[0024] The autonomous behavior decision-making module 300 is used to generate autonomous decision-making results within a preset safety and compliance framework by using a pre-built multi-objective optimization decision-making model based on environmental information and user intent. The autonomous decision-making results include driving action sequence, driving speed, and parking position.

[0025] The planning and control module 400 is used to generate a motion trajectory that conforms to the user's intention based on the driving action sequence and parking position, and to combine the motion trajectory with the driving speed to obtain chassis control commands, which are then sent to the actuator module 500.

[0026] The actuator module 500 is used to receive and execute chassis control commands.

[0027] The core of this car robot lies in building a fully autonomous operating architecture that integrates perception, understanding, decision-making, control, and execution. This allows the vehicle to no longer rely on users to issue detailed control commands one by one, but instead to autonomously complete travel behaviors based on user needs, giving the vehicle proactive judgment and flexible response capabilities.

[0028] It should be noted that the autonomous decision-making results of the car robot in this application embodiment cover three dimensions: driving action sequence, driving speed, and parking position, corresponding to the three core travel decision questions: what action to take, at what speed to drive, and where to park. Since the decisions in these three dimensions are all generated autonomously by the system rather than specified by the user, the vehicle can autonomously plan and execute a complete travel service process even when the user has no instructions or only provides a vague travel goal.

[0029] The vehicle-mounted perception module 100 may include an environmental perception subunit and a user perception subunit. The environmental perception subunit is used to collect real-time information about the environment inside and outside the vehicle, while the user perception subunit is used to collect multimodal interaction perception data of the user inside the vehicle.

[0030] Specifically, the environmental perception subunit may include sensing devices such as LiDAR, millimeter-wave radar, surround-view cameras, and forward-looking cameras to perceive road users, traffic signs, traffic congestion levels, available parking lots and vacant parking spaces, road geometry, weather conditions, and user comfort information such as in-vehicle temperature and air pressure. The real-time environmental information output by the environmental perception subunit is structured data, including target category, location, speed, confidence level, and scene semantic tags. Road users include vehicles, pedestrians, and non-motorized vehicles; traffic signs include traffic lights, speed limit signs, and no-parking signs; and weather conditions include rain, snow, and fog.

[0031] The user perception subunit may include devices such as in-vehicle cameras, microphone arrays, seat pressure sensors, and / or steering wheel biosensors to collect multimodal interactive perception data, including facial expressions, voice commands, body posture, and physiological signals. In addition, the user perception subunit can also obtain user schedule information, navigation history, music preferences, and other data through the in-vehicle infotainment system to assist in understanding user needs.

[0032] Optionally, in some embodiments of this application, the car robot further includes a user preference knowledge base. The user preference knowledge base stores decision feedback data from the user's historical travels, including records of the user's acceptance of each autonomous decision result, records of manual corrections, and records of explicit evaluations. The autonomous behavior decision-making module is further configured to: collect user feedback data on each autonomous decision result after it is output, and write it into the user preference knowledge base; incrementally calibrate the decision target weight baseline in the multi-objective optimization decision model based on the historical feedback data in the user preference knowledge base to obtain a personalized weight baseline; determine the decision target weight vector for the current decision based on the personalized weight baseline and the current user intent; wherein, when the number of feedback samples for a certain type of intent reaches a preset sample threshold, a model parameter update is triggered; when the preset sample threshold is not reached, the personalized weight baseline is determined by a weighted fusion of the preset weight baseline and a small amount of feedback data.

[0033] Optionally, in some embodiments of this application, the demand understanding module includes a trained hierarchical dynamic Bayesian network model; the user information is multimodal interaction perception data, including explicit command input and / or implicit state information.

[0034] In the requirements understanding module, environmental information and user information are fused and analyzed to obtain user intent, including: The environmental information is semantically abstracted to generate a context modulation factor for the semantics of the current interaction scene.

[0035] Semantic abstraction processing refers to extracting semantic features from sensor data collected by the environmental perception submodule to form a structured description of the current interaction scenario.

[0036] The structured environmental information output by the environmental perception module is subjected to high-level semantic abstraction to obtain semantic labels for the current scene. Structured environmental information includes the current road segment type, traffic congestion level, surrounding facility types, and weather conditions. Scene semantic labels include, for example, congested main roads during weekday morning rush hour, empty parking lots in shopping malls on weekend afternoons, and smooth highways at night. These scene semantic labels serve as context modulation factors to adjust the prior probability distribution of user intent reasoning. For example, in a weekday morning rush hour scenario, the prior probability of a user's intention to be in a hurry is higher than in a normal scenario; in a shopping mall parking lot scenario, the prior probability of a user's intention to find a convenient parking location is higher.

[0037] Different external scenarios can have varying impacts on a user's mental state. For example, in a highway cruising scenario, a user's physiological load is typically low, and their emotional valence is generally positive; while in a traffic jam scenario, a user's physiological load may increase, their sense of time urgency may intensify, and their emotional valence may become more negative. Context modulation factors, as prior information input to the subsequent hierarchical dynamic Bayesian network model, are used to modulate the model's inference process regarding mental states.

[0038] Multimodal interactive perception data is used as observational evidence and input into a hierarchical dynamic Bayesian network model along with context modulation factors to obtain the posterior probability distribution of the user's mental state; the user's mental state includes physiological load, emotional valence, and sense of time urgency.

[0039] The hierarchical dynamic Bayesian network model takes multimodal observation evidence and contextual modulation factors as input and calculates the posterior probability distribution of the user's mental state through Bayesian inference. This posterior probability distribution describes the probability of the user being in various combinations of mental states given the current observation evidence and scene context.

[0040] It should be noted that the dimensions of user mental state are defined as follows: Physiological load: Characterizes a user's current level of physical fatigue and cognitive resource utilization. Physiological load can be inferred comprehensively from physiological and behavioral indicators such as the low-frequency / high-frequency ratio of heart rate variability, the rate of change in skin conductance response, and blink frequency. Physiological load is discretized into multiple levels, such as low load, medium load, and high load.

[0041] Emotional valence: Represents the degree of a user's current positive or negative emotion. Emotional valence can be inferred comprehensively through indicators such as facial expression features and paralinguistic features of speech. Emotional valence is discretized into multiple levels, such as positive, neutral, and negative.

[0042] Time urgency: Represents a user's current level of time pressure. Time urgency can be inferred from indicators such as the user's explicit instructions, speech rate, forward leaning angle, and remaining time to the destination. Time urgency is discretized into multiple levels, such as low urgency, medium urgency, and high urgency.

[0043] User intent is generated based on the posterior probability distribution.

[0044] This application's embodiments generate context modulation factors by semantically abstracting real-time environmental information and inputting these factors as prior information into a hierarchical dynamic Bayesian network model. This enables the system to fully integrate the contextual information of the current interaction scenario when inferring the user's mental state, thereby overcoming the limitation of traditional autonomous driving systems that can only passively respond based on a single user command. Since different external scenarios significantly influence the user's physiological load, emotional valence, and sense of time urgency, using scenario semantic information as the basis for adjusting prior probabilities allows the inference of mental states to better reflect the user's actual psychological performance in different scenarios, improving the accuracy and robustness of the inference.

[0045] Meanwhile, by inputting multimodal interactive perception data as observational evidence and contextual modulation factors into a hierarchical dynamic Bayesian network model, and utilizing Bayesian inference to calculate the posterior probability distribution of the user's mental state, heterogeneous information from different modalities can be effectively integrated. Even when explicit instructions are incomplete or the user has not clearly expressed their needs, it can still accurately estimate the user's implicit needs based on joint inference from observational evidence and scene context. Furthermore, by comprehensively inferring and quantifying the three dimensions of physiological load, emotional valence, and time urgency, detailed user state information can be provided for subsequent autonomous behavior decisions. This allows the decision-making module to plan behavior based on the user's current real physical and mental state rather than solely relying on the content of the instructions, thereby improving the initiative and humanization level of automotive robot services.

[0046] Optionally, in some embodiments of this application, user intent is generated based on a posterior probability distribution, including: Obtain the interaction strategy weight factor; the interaction strategy weight factor is determined by evaluating the confidence of the posterior probability distribution and / or evaluating the intermodal consistency of the multimodal interaction perception data.

[0047] Confidence assessment measures the concentration of the posterior probability distribution. When the posterior probability distribution is highly concentrated around a particular combination of mental states, it indicates that the model's judgment of the user's current state has high certainty, resulting in high confidence. Conversely, when the posterior probability distribution is relatively dispersed, with multiple combinations of mental states having similar probabilities, it indicates greater uncertainty in the judgment, resulting in lower confidence. In some embodiments of this application, confidence can be measured using statistics such as the entropy of the posterior probability distribution, the maximum a posteriori probability, or the variance of the probability distribution.

[0048] Intermodal consistency assessment is used to measure the degree of consistency between perceptual data from different modalities. For example, when facial expressions indicate that a user is in a negative mood, but speech paralinguistic features show that the user's speech rate is stable and the pitch is normal, the judgments of the two modalities are contradictory, and the intermodal consistency is low. In some embodiments of this application, intermodal consistency can be measured by indices such as Kullback-Leibler divergence or cosine similarity between the posterior distributions of mental states independently inferred from each modality.

[0049] The value of the interaction strategy weight factor is positively correlated with the confidence assessment result and / or the intermodal consistency assessment result. That is, the higher the confidence and / or the higher the intermodal consistency, the larger the interaction strategy weight factor, indicating that the model's judgment of user intent is more reliable; conversely, the smaller the interaction strategy weight factor, the lower the reliability of the judgment, and subsequent decisions should tend to be conservative.

[0050] Based on the posterior probability distribution and the interaction strategy weight factor, the expected utility value of each candidate demand parsing action is obtained through a preset utility function; the utility function represents the expected utility value of performing different candidate demand parsing actions under different mental states.

[0051] The candidate demand parsing action refers to the possible ways the system interprets user intent, including: for high time urgency, the user intent is interpreted as "to reach the destination quickly"; for high physiological load or low time urgency, the user intent is interpreted as "to reach the destination comfortably and smoothly"; for negative emotions or high physiological load, the user intent is interpreted as "to reach the destination safely and cautiously"; for low time urgency and positive emotions, the user intent is interpreted as "to reach the destination energy-efficiently"; and for low urgency and positive emotions, the user intent is interpreted as "to find points of interest along the way".

[0052] The formula for calculating the expected utility value E[U(a,s)] is: E[U(a,s)]=W int ×Σ s [P(s|e,c)×U(a,s)]; Among them, W int , where is the interaction strategy weight factor, with a value range of [0,1]; 'a' is the encoding identifier of the candidate demand parsing action; 's' is the mental state vector, representing the combined state of physiological load, emotional valence, and time urgency; 'e' is the multimodal observation evidence fusion vector; 'c' is the context modulation factor; 'P(s|e,c)' is the posterior probability distribution of mental state 's' given 'e' and 'c'; 'U(a,s)' is the utility value of the utility function executed under mental state vector 's'; and the utility evaluation values ​​of each candidate action are normalized to the same dimension interval.

[0053] The utility function U(a,s) is pre-defined and stored in the vehicle computing platform in the form of a utility matrix. Based on the utility function, when the user's mental state is "high time urgency + neutral emotion + low physiological load," the utility value for performing the analytical action of "quickly reaching the destination" is higher; when the user's mental state is "low time urgency + negative emotion + high physiological load," the utility value for performing the analytical action of "comfortably and smoothly reaching the destination" is higher.

[0054] The parsing result corresponding to the candidate demand parsing action with the highest expected utility value is selected as the user intent, and the highest expected utility value is normalized and used as the confidence level of the user intent.

[0055] Specifically, the expected utility values ​​of all candidate demand parsing actions are compared, and the action with the highest expected utility value is selected. The parsing result corresponding to this action is the final determined user intent. For example, if "arriving at the destination comfortably and smoothly" has the highest expected utility value, then the user intent is determined to be "comfortable driving".

[0056] Simultaneously, the highest expected utility value is normalized and mapped to the [0,1] interval, serving as the confidence level for the user's intent. Normalization can employ methods such as the Softmax function or min-max normalization. This confidence level will be used for adaptive modulation of the safety boundary in the subsequent autonomous behavior decision-making module.

[0057] This application embodiment obtains an interaction strategy weight factor and uses this weight factor to modulate the expected utility calculation of the utility function. This enables the intent inference process to simultaneously consider the certainty of the model's judgment of the user's state and the consistency level between the observation data of each modality. This effectively suppresses the risk of misjudgment that may occur when evidence is insufficient or there are conflicts between modalities, thus improving the reliability of the intent recognition results. Since the interaction strategy weight factor is positively correlated with the confidence assessment results of the posterior probability distribution and / or the consistency assessment results between modalities, the intent inference results have high credibility when the model's judgment of the mental state is relatively certain and the evidence from each modality corroborates each other. However, when the judgment is uncertain or there are contradictions between modalities, the weight factor automatically compresses the expected utility value, guiding subsequent decisions to be more conservative and reducing the risk of unreasonable or dangerous driving behavior caused by excessive reliance on unreliable intents.

[0058] Furthermore, by normalizing the highest expected utility value as the confidence level of the user's intent and outputting it to the autonomous behavior decision-making module, a quantifiable reliability index is provided for the adaptive modulation of the safety boundary in the subsequent decision-making process. This enables the entire system to form a complete information transmission link from state perception to intent understanding and then to behavior decision-making, realizing a deep understanding and reliable expression of user needs, and laying the decision-making foundation for providing safe, comfortable and personalized autonomous services for automotive robots.

[0059] Figure 2 This is a schematic diagram of the structure of a hierarchical dynamic Bayesian network model provided in an embodiment of this application. Optionally, as... Figure 2 As shown, in some embodiments of this application, the hierarchical dynamic Bayesian network model includes an observation layer 21, a mental state layer 22, and an interaction expectation layer 23. This three-layer structure follows the general paradigm of hierarchical Bayesian modeling, decomposing the complex joint probability distribution into a hierarchical product of conditional probabilities, thereby achieving structured probabilistic inference of the user's mental state and interaction expectations.

[0060] The observation layer 21 includes observation variable nodes corresponding to the interactive perception data of each modality, and is connected to the mental state layer 22 through the observation likelihood function.

[0061] Specifically, each modal data in the multimodal interactive perception data serves as observational evidence for the corresponding observation variable nodes in the hierarchical dynamic Bayesian network model. Each observation variable node is connected to all latent variable nodes in the mental state layer 22 through an observation likelihood function.

[0062] For example, the observation layer 21 contains N observation variable nodes, denoted as the first modality observation variable node to the nth modality observation variable node, where N is an integer greater than or equal to 2. The observation variable nodes may include: Speech modality observation variable nodes are used to receive speech paralinguistic features, such as speech rate, pitch variation rate, and energy envelope.

[0063] Visual modal observation variable nodes are used to receive facial expression features, such as the degree of eyebrow furrowing, the curvature of the corners of the mouth, and the frequency of blinking.

[0064] Physiological modality observation variable nodes are used to receive physiological signals, such as HRV (Heart Rate Variability) and GSR (Galvanic Skin Response).

[0065] The posture modal observation variable node is used to receive body posture features, such as the forward tilt angle of sitting posture and the direction of head turning.

[0066] The observation likelihood function describes the conditional probability distribution of modal observation data under a given mental state, i.e., P(observation data|mental state). Under different mental states, the observation data of each modality exhibits different probability distribution characteristics.

[0067] In some embodiments of this application, the local observation likelihood functions of each modality adopt different parameterized distribution forms to adapt to the statistical characteristics of different modal data: N heterogeneous modalities (such as speech paralinguistic features, facial expression AU encoding, HRV frequency domain features, seat pressure distribution, etc.) are mapped to a unified 3D discrete mental state space. A local observation likelihood function with a specific parameterization is designed for each modality (e.g., Gaussian mixture for speech modality, Gamma distribution for physiological modality, and multinomial distribution for visual modality). This specific parameterization is calibrated based on human factors engineering experimental data in driving scenarios.

[0068] For example, for the speech modality, a Gaussian mixture model is used for the local observation likelihood function. For the physiological modality, since the low-frequency / high-frequency ratio of HRV exhibits a positively skewed distribution, the local observation likelihood function adopts a log-normal distribution. The parameter sets of the local observation likelihood functions for each modality are learned from labeled data through offline training. This conditional independence assumption simplifies the computational complexity of the observation likelihood function while providing sufficiently accurate inference results in practical applications.

[0069] The mental state layer 22 includes latent variable nodes for physiological load 22b, latent variable nodes for emotional valence 22c, and latent variable nodes for time urgency 22a. The causal dependencies between the latent variable nodes are represented by directed edges, which are used to output the posterior probability distribution.

[0070] Causal dependencies are connected by directed edges, and conditional probability tables are used to quantify the strength of the dependencies between latent variables. The parameters in the conditional probability tables are learned from historical driving data through a training process.

[0071] Within the framework of dynamic Bayesian networks, the latent variable nodes in mental state layer 22 exhibit state transition relationships across different time slices. That is, the latent variable state at time t depends on the latent variable state at time t-1, satisfying the first-order Markov property assumption. The state transition probability matrix is ​​also learned from time-series labeled data through the training process.

[0072] The context modulation factor, as an external input, influences the prior probability distribution of each latent variable node in the mental state layer 22. For example, when the context modulation factor indicates that the current scenario is "traffic jam and following in a car", the prior probability distribution of physiological load and time urgency will shift towards higher levels.

[0073] The interaction expectation layer 23 includes latent variable nodes for interaction mode preferences. It takes the posterior probability distribution and context modulation factor as input and outputs a probability distribution representing the user's acceptance of candidate demand parsing actions in the current scenario.

[0074] Specifically, the latent variable node representing the interaction mode preference characterizes the degree of user preference for different interaction methods (such as proactive suggestions, passive execution, and confirmatory inquiries). The value of this node is influenced by both the user's current mental state and the current context.

[0075] For example, when users are in a state of high physiological load and high time urgency, they prefer passive execution, that is, the system directly executes the optimal decision without user confirmation; when users are in a state of low physiological load and positive emotions, they may be more receptive to proactive suggestions, that is, the system provides multiple options for users to choose from.

[0076] The acceptance probability distribution is used to adjust the expected utility value of each candidate demand parsing action.

[0077] Specifically, the acceptance probability distribution is used as a correction coefficient to adjust the expected utility values ​​in a weighted manner. If a user has a high acceptance of a certain parsing action, the expected utility value of that action will increase accordingly, and vice versa. This allows for dynamic adjustment of the interaction strategy based on the user's real-time status, thereby improving the user experience.

[0078] The structure of the hierarchical dynamic Bayesian network model is predefined based on psychology and human-computer interaction theory, while the parameters in the model, such as the conditional probability table, state transition probability matrix, and observation likelihood function parameters, are learned from historical driving data through the following training methods.

[0079] The hierarchical dynamic Bayesian network model structure provided in this application constructs a three-layer Bayesian hierarchical architecture consisting of an observation layer, a mental state layer, and an interaction expectation layer. This architecture integrates multimodal heterogeneous observation data, temporal state evolution patterns, and scene context information into a probabilistic inference framework, achieving structured inference and interpretable modeling of user mental states. Specifically, the observation layer designs local observation likelihood functions with different parameterizations for various modalities such as speech, vision, physiological, and posture data, and calibrates them using human factors experiment data from driving scenarios, effectively improving the accuracy and robustness of multimodal fusion perception. The mental state layer represents the causal dependence between physiological load, emotional valence, and time urgency through directed edges, and introduces a state transition mechanism that satisfies the first-order Markov property assumption. Combined with context modulation factors, it adaptively modulates the prior distribution of the scene, achieving continuous, stable, and scene-sensitive tracking and inference of user states. The interaction expectation layer further takes the posterior mental state and scene context as input, outputting the probability distribution of user acceptance of candidate parsed actions, and uses this as a correction coefficient to dynamically optimize the interaction strategy. Meanwhile, the model structure is predefined based on psychology and human-computer interaction theory, and the parameters are obtained through offline learning of historical driving data, which enables the car robot to have both theoretical interpretability and individual adaptability, providing a reliable and efficient intention understanding model for the car robot.

[0080] The training process is completed offline before vehicle deployment and includes the following steps: First, a training dataset is constructed. The training dataset includes a large amount of time-series sample data. Each sample contains: multimodal interaction perception data, including multimodal observation data sequences of speech, vision, physiological, and posture, corresponding scene context information, and manually labeled user mental state and user intent labels. The training dataset can be obtained through real-vehicle road tests, driving simulator experiments, and digital twin simulation data collection.

[0081] Next, parameter learning is performed. In some embodiments of this application, the expectation-maximization algorithm is used for parameter learning. The expectation-maximization algorithm iteratively optimizes the likelihood values ​​of the model parameters by alternately executing the E-step (expectation step) and the M-step (maximization step): in the E-step, the posterior distribution of the latent variables is estimated based on the current parameters; in the M-step, the model parameters are updated based on the posterior distribution of the latent variables to maximize the expected complete log-likelihood of the data. Iteration continues until the change in the complete log-likelihood is less than a preset convergence threshold or the maximum number of iterations is reached.

[0082] In other embodiments of this application, Bayesian inference methods such as maximum likelihood estimation, variational inference, and Monte Carlo Markov chains can also be used for parameter learning, and this application does not limit this.

[0083] After training, the converged model parameters are exported and deployed to the in-vehicle computing platform. During vehicle operation, the model only performs online probabilistic inference, without performing any online training or parameter update operations, ensuring the determinism and verifiability of the inference results.

[0084] Optionally, in some embodiments of this application, the autonomous behavior decision-making module generates autonomous decision-making results within a preset security and compliance framework based on environmental information and user intent through a pre-built multi-objective optimization decision-making model, including: The security margin modulation coefficient is determined based on the confidence level of the user's intent, and the security boundary of the preset security compliance framework is adaptively modulated; the lower the confidence level of the user's intent, the larger the security margin modulation coefficient, and the more conservative the security boundary.

[0085] Optionally, in some embodiments of this application, the preset security compliance framework includes a hard constraint layer and a soft constraint layer; The hard constraint layer is configured with a first safety boundary based on traffic rules and vehicle kinematics boundaries. This first safety boundary is a non-adjustable safety constraint. For example, the maximum deceleration does not exceed -8 m / s², and the lane deviation does not exceed 0.3 m. This boundary is a rigid, non-adjustable constraint that must not be violated by any planned trajectory.

[0086] The soft constraint layer is configured with a second safety boundary based on the safety margin, and the tightness of the second safety boundary is adaptively adjusted according to the decision confidence.

[0087] Adaptive modulation of the security boundary of the preset security compliance framework includes: adjusting the relaxation level of the second security boundary according to the confidence level; the higher the confidence level, the more relaxed the second security boundary, and the lower the confidence level, the more stringent the second security boundary.

[0088] When the confidence level is higher than the preset confidence threshold (e.g., 0.85), it indicates that the current interaction intent recognition result is reliable, and the system appropriately relaxes the second security boundary to improve passage efficiency and interaction naturalness; when the decision confidence level is lower than the threshold, it indicates that there is uncertainty in intent recognition, and the system automatically tightens the second security boundary to increase security redundancy and reduce the risk of misjudgment.

[0089] For example, a set of bottom-line constraints is pre-stored within the pre-defined security compliance framework, which may include: Following distance constraints: minimum safe distance from the vehicle in front, vehicle speed, and road surface adhesion coefficient.

[0090] Speed ​​limit constraints: Road speed limits, minimum speed limits.

[0091] Lateral acceleration constraint: The lateral acceleration during cornering shall not exceed the occupant comfort threshold and the vehicle stability threshold.

[0092] Longitudinal acceleration / deceleration constraint: The longitudinal acceleration during acceleration and braking shall not exceed the safety threshold.

[0093] Collision time constraint: The collision time with surrounding obstacles is not less than the safety threshold.

[0094] Traffic regulations include: traffic light rules, right-of-way rules, lane-changing rules, etc.

[0095] The safety margin modulation factor is a coefficient greater than or equal to 1, used to further increase the safety margin beyond the basic safety boundary. Its calculation formula is as follows: K safe =1+α×(1-C intent ); Among them, K safe C is the safety margin modulation coefficient. intent The confidence level of the user's intent, with a value range of [0,1], and α is a preset modulation gain parameter (α>0).

[0096] When the confidence level of the user's intent is high, i.e. close to 1, K safe When the confidence level is close to 1, the safety boundary remains at a basic level; when the confidence level of the user's intent is low, i.e. close to 0, K... safe As the value approaches 1+α, the safety margin expands significantly, and the system adopts a more conservative decision-making strategy.

[0097] Therefore, when the system is not confident enough in judging the user's intention, the decision-making risk is reduced by expanding the safety boundary, so that even in the case of a deviation in the understanding of the intention, the driving safety of the vehicle can still be fully guaranteed.

[0098] Based on user intent and environmental information, determine the decision objective weight vector.

[0099] This application provides a safety boundary adaptive modulation scheme. By introducing a safety margin modulation coefficient and dynamically adjusting the safety boundary of a preset safety compliance framework based on the confidence level of the user's intent, the system can automatically expand the safety margin and adopt a more conservative decision-making strategy when the uncertainty of intent understanding is high. This effectively reduces the driving risks that may be caused by intent inference bias. At the same time, when the confidence level is high, the basic safety boundary is maintained to avoid unnecessary losses in driving efficiency due to excessive conservatism, thus achieving a dynamic balance between safety and traffic efficiency. This mechanism does not rely on prior assumptions about specific scenarios and can be applied to various driving conditions. It provides a structured risk buffer mechanism for the safe and reliable operation of automotive robots in complex open environments and provides adaptive and adjustable safety constraints for subsequent multi-objective optimization decisions.

[0100] Optionally, in some embodiments of this application, determining the decision target weight vector based on user intent and environmental information includes: Based on the discrete behavioral labels of user intent, match the base weight vector in a pre-configured intent-weight mapping table.

[0101] The intent-weight mapping table represents the correspondence between different candidate intent types and at least one of the following weights: timeliness weight, comfort weight, safety weight, and energy consumption weight. The intent-weight mapping table is pre-configured in the in-vehicle computing platform in the form of a lookup table.

[0102] For example, the intent-weight mapping table can be shown in Table 1 below:

[0103] Table 1 Intent-Weight Mapping Table The sum of the weight values ​​in the table above is 1, representing the relative importance of each dimension in the decision-making objective. These weight values ​​are exemplary and can be adjusted based on calibration experiments and expert experience in practical applications.

[0104] The intent-weight mapping table represents the correspondence between different candidate intent types and at least one of the following weights: timeliness weight, comfort weight, safety weight, and energy consumption weight.

[0105] Extract scene complexity indicators from environmental information.

[0106] The scenario complexity metric is used to quantify the complexity of the current driving scenario. For example, its calculation can take into account the following factors: Dynamic obstacle density: The number of dynamic obstacles (vehicles, pedestrians, non-motorized vehicles) within a unit of sensing range.

[0107] Uncertainty in the motion of dynamic obstacles: the variance in the prediction of the motion trajectory of surrounding obstacles.

[0108] Road topology complexity: number of lanes, number of intersections, rate of change of road curvature, etc.

[0109] Traffic light density: The number of traffic lights per unit distance.

[0110] Visibility conditions: Visibility levels related to weather and lighting conditions.

[0111] The scene complexity index is obtained by weighted summation or nonlinear fusion of the above factors, and its value range is [0,1]. The larger the value, the more complex the scene.

[0112] The basic weight vector is dynamically adjusted using a scene complexity index to obtain a decision target weight vector. Specifically, when the scene complexity index exceeds a preset threshold, the safety weight is increased while the comfort weight is decreased. After adjustment, each weight is normalized to ensure that the sum of the weights is 1.

[0113] Therefore, regardless of the user's preferred driving style, the system will automatically prioritize safety and de-prioritize comfort when the complexity of the scenario increases, ensuring driving safety in complex scenarios.

[0114] Using the modulated safety boundary as a constraint and the decision target weight vector as the objective function, a multi-objective optimization decision model is used to solve the problem and generate multiple candidate decision schemes. Each candidate decision scheme includes at least a candidate driving action sequence, a candidate driving speed, and a candidate parking position.

[0115] The multi-objective optimization decision model is a pre-built mathematical optimization solution framework. Before deployment, the mathematical expression of the multi-objective trade-off function, the inequality form of the safety and compliance constraints, and the numerical solution algorithm are predefined. During vehicle operation, the model constructs and solves a constrained optimization problem online based on the real-time acquired decision objective weight vector and the modulated safety boundary, thereby generating the optimal decision scheme that meets the requirements of the current scenario.

[0116] In some embodiments of this application, the comfort evaluation function f comfort The calculation of (x) comprehensively considers the root mean square or peak values ​​of kinematic parameters such as longitudinal acceleration, lateral acceleration, and jerk; the smaller the value, the less comfortable the ride. Safety evaluation function f safety The calculation of (x) takes into account indicators such as minimum following distance, minimum collision time, and minimum safe distance.

[0117] Multi-objective optimization decision models can be solved using weighted summation, ε-constraints, non-dominated sorting genetic algorithms, or other multi-objective optimization algorithms. In some embodiments of this application, the weighted summation method is used to transform the multi-objective problem into a single-objective problem. This constrained optimization problem can be numerically solved using nonlinear programming solvers such as quadratic programming solvers. The solver outputs one or more sets of optimal or suboptimal solutions that satisfy the constraints, with each set of solutions corresponding to a candidate decision scheme.

[0118] Multiple candidate decision-making schemes are evaluated for multiple objectives, and the candidate decision-making scheme with the highest comprehensive evaluation value is selected as the output of the autonomous decision-making result.

[0119] The comprehensive evaluation value is calculated using the same weighted summation method as the objective function. The candidate decision scheme with the highest score is selected as the final autonomous decision result.

[0120] For example, in one specific embodiment of this application, the autonomous decision-making result includes: driving action sequence "maintain the current lane for 3 kilometers → turn right into the target road → turn right at the entrance of the target parking lot → park in parking space No. 15 on B2 floor", driving speed "target speed curve", and parking location "specific parking space coordinates of the target parking lot".

[0121] This application embodiment achieves coordinated optimization of decision weights across both intent-driven and scenario-adaptive dimensions by integrating discrete behavioral labels of user intents with real-time scenario complexity metrics. Specifically, a pre-configured intent-weight mapping table is used to quickly map user intents to a basic weight vector, ensuring differentiated configuration of decision targets under different intent types, enabling vehicle behavior to accurately respond to users' explicit demands. Simultaneously, by extracting a scenario complexity metric that comprehensively quantifies multiple factors, the basic weight vector is dynamically adjusted. This allows the system to automatically increase safety weights and decrease comfort weights when scenario complexity increases, proactively sacrificing some comfort in complex environments to achieve higher safety redundancy. This avoids the problem of fixed weight configurations being too conservative in simple scenarios or too aggressive in complex scenarios. By decoupling user intent from scenario complexity and then fusing the decision-making process, the predictability of intent-driven behavior is maintained, while the system is given the ability to autonomously adjust the conservatism of its decisions based on the level of environmental risk. The modified weight vector is normalized to ensure that the sum of the weights of each dimension is 1, making the multi-objective trade-off function comparable and numerically stable in different scenarios. It can be applied to various driving environments without relying on prior assumptions of specific scenarios, providing an interpretable decision-making basis for car robots to achieve safe, efficient and personalized autonomous driving behavior.

[0122] Optionally, in some embodiments of this application, the self-service behavior decision module is further used to determine the parking evaluation dimensions and the initial weights of each dimension based on the user's intent; the parking evaluation dimensions include at least three of the following: walking distance dimension, safety dimension, parking fee dimension, accessibility dimension, and environmental condition dimension.

[0123] Acquire attribute data for multiple candidate parking spaces within the target area. Attribute data includes parking space location, parking space type, charging standard, access width, sun and rain shelter conditions, and surrounding pedestrian and vehicle traffic density.

[0124] The attribute data comes from high-precision maps, parking management system interfaces, and real-time observations from vehicle-mounted perception modules. For each candidate parking space, its attribute data is mapped to normalized scores for each evaluation dimension: the walking distance dimension uses inverse linear normalization, with shorter distances resulting in higher scores; the safety dimension is weighted based on the density of surrounding pedestrian and vehicle traffic, lighting conditions, and surveillance coverage; the parking fee dimension is normalized based on the ratio of the fee rate to the user's preset budget limit; the ease of entry and exit dimension is obtained by looking up tables based on passage width, turning radius, and parking space type; and the environmental conditions dimension is quantified based on the completeness of sunshade and rain shelter facilities and ground flatness.

[0125] Based on the scene context parameters in the current environmental information, the initial weights of each dimension are dynamically adjusted to obtain the parking evaluation weight vector.

[0126] Contextual parameters include weather conditions, time of day, regional congestion index, and real-time parking lot saturation. For example, when rain or high temperatures are detected, the weight of the environmental conditions dimension is automatically increased; when the parking lot saturation exceeds 90%, the weight of the walking distance dimension is automatically decreased to avoid long search times and detours caused by chasing nearby parking spaces.

[0127] The attribute data of each candidate berth is mapped to the normalized score of each evaluation dimension.

[0128] Based on the parking evaluation weight vector and the normalized scores of each dimension, the comprehensive score of each candidate parking space is calculated, and the candidate parking space with the highest comprehensive score is selected as the parking location output.

[0129] It directly triggers weight modulation rules through the semantic features of user intent. When the user intent includes the semantics of temporary parking, it automatically increases the weights of the dimensions of ease of entry and exit and safety. When the user intent includes the semantics of long-term parking, it automatically increases the weights of the dimensions of environmental conditions and parking fees.

[0130] The autonomous behavior decision-making module also includes a social interaction reasoning subunit; The social interaction reasoning subunit is used to extract behavioral characteristics of surrounding traffic participants from environmental information. Surrounding traffic participants include other vehicles, pedestrians, and non-motorized vehicles, and their behavioral characteristics include position, speed, heading angle, and historical movement trajectory.

[0131] By using kinematic geometry analysis and traffic rule state machines, we can infer the intentions of each surrounding traffic participant and obtain the probability distribution representing their behavioral intentions and the corresponding predicted trajectories.

[0132] For example, based on its current position, velocity vector, rate of change of heading angle, and lateral offset from the lane boundary, the geometric likelihood of its discrete behaviors such as lane changing, turning, going straight, or stopping is calculated. Combined with the topology of its lane, traffic light status, turn signal, and right-of-way markers, a predefined finite state machine is used to determine its rule legality under the current traffic rules. The geometric likelihood and rule legality are then fused using Bayesian methods to obtain a probability distribution representing the traffic participant's behavioral intention. Multiple predicted trajectories and their associated probabilities are then extrapolated based on the kinematic constraints corresponding to each intention branch. Based on the predicted behavioral intentions and trajectories of surrounding traffic participants, a spatiotemporal competition relationship between the vehicle and surrounding traffic participants is constructed. The spatiotemporal competition relationship includes the spatial overlap and temporal arrival overlap between the vehicle's driving path and the predicted trajectories of surrounding traffic participants.

[0133] Specifically, spatial overlap is calculated by geometrically intersecting the future travel path envelope corresponding to the vehicle's candidate decision with the confidence ellipsoids of the predicted trajectories of each traffic participant; temporal arrival overlap is determined by comparing the intersection length of the time windows of the expected arrival spatial overlap areas of the vehicle and each traffic participant. Together, they quantify the potential conflict between the vehicle's decision and the expected behaviors of other traffic participants.

[0134] When solving for candidate decision-making schemes, the interaction risk value and social acceptability assessment value of each candidate decision-making scheme are calculated based on the spatiotemporal competition relationship, and used as an additional constraint dimension in the selection of candidate decision-making schemes. The interaction risk value is calculated based on the minimum distance and collision time between the candidate decision-making scheme of this vehicle and the predicted trajectory of surrounding traffic participants; the social acceptability assessment value is determined based on whether the candidate decision-making scheme of this vehicle conforms to the preset traffic yielding rules, right-of-way priority rules, and driving etiquette norms.

[0135] Candidate decision-making solutions that have been filtered by interaction risk value and social acceptability assessment value, meet the preset interaction safety threshold, and align with the user's intention optimization goals are selected as the autonomous decision-making output. If multiple candidate solutions all meet the safety threshold, the solution with the highest social acceptability assessment value is selected first. When the social acceptability assessment values ​​of all candidate solutions are below the preset lower limit but there is a solution that meets the safety threshold, a conservative degradation strategy is triggered, and a defensive decision with minimizing interaction risk as the primary objective is output.

[0136] Optionally, in some embodiments of this application, the multi-objective optimization decision model is configured with a multi-objective trade-off function.

[0137] A multi-objective trade-off function is used to evaluate candidate decision schemes from multiple dimensions, including at least two of the following dimensions: timeliness, comfort, safety, and energy consumption.

[0138] The evaluation functions for each dimension are defined as follows: (1) Timeliness dimension The timeliness evaluation function characterizes the traffic efficiency of candidate decision-making schemes. Its calculation considers the following factors: the deviation between the estimated arrival time and the user's expected arrival time; the ratio of average driving speed to the road speed limit; and the number of unnecessary stops and waiting times along the way. A higher function value indicates higher traffic efficiency.

[0139] (2) Comfort dimension The comfort evaluation function characterizes the ride comfort level of candidate decision options. Its calculation considers the following factors: the root mean square value of the absolute value of longitudinal acceleration; the root mean square value of the absolute value of lateral acceleration; the peak and root mean square values ​​of jerk; and the rate of change of steering angle. A larger value for this function indicates a more comfortable ride. In some embodiments of this application, the penalty coefficient for jerk in the comfort evaluation function is determined through offline calibration tests to ensure that the evaluation results are consistent with the occupant's subjective comfort experience.

[0140] (3) Security dimension The safety evaluation function characterizes the driving safety level of candidate decision schemes. Its calculation considers the following factors: the sum of the reciprocals of the minimum collision times with surrounding obstacles during driving; the ratio of following distance to safe distance; and the interaction risk with vehicles in adjacent lanes during lane changes. A higher function value indicates greater driving safety.

[0141] (4) Energy consumption dimension The energy efficiency evaluation function characterizes the energy utilization efficiency of candidate decision-making schemes. For electric vehicles, the calculation considers the following factors: expected energy consumption; braking energy recovery efficiency; and energy consumption of air conditioning and auxiliary systems. The larger the value of this function, the higher the energy utilization efficiency.

[0142] The multi-objective optimization decision model uses the modulated safety boundary as a constraint and the decision objective weight vector as the weight configuration of each dimension in the multi-objective trade-off function to solve the weighted optimal solution of the multi-objective trade-off function. The candidate decision scheme corresponding to the weighted optimal solution is output as the candidate decision scheme with the highest comprehensive evaluation value.

[0143] Optionally, in some embodiments of this application, a motion trajectory conforming to the user's intent is generated based on the driving action sequence and parking position, and the motion trajectory is combined with the driving speed to obtain chassis control commands, including: Based on the sequence of driving actions, the macroscopic driving path topology of the vehicle is determined, and the parking position is used as the endpoint constraint of the macroscopic driving path topology.

[0144] Macroscopic driving path topology refers to the topological structure composed of a series of path nodes and connecting road segments, describing the macroscopic driving route of a vehicle from its current location to its target parking location. Path nodes include the current location point, the execution points of each driving action (such as lane change starting point, turning point), and the target parking location point.

[0145] Based on the macroscopic driving path topology, a continuous reference path that satisfies vehicle kinematic constraints is generated using a curve smoothing algorithm.

[0146] Curve smoothing algorithms are used to smooth polyline paths in macroscopic driving path topologies into continuous, differentiable curved paths, while ensuring that the path meets the vehicle's minimum turning radius constraint. In some embodiments of this application, the curve smoothing algorithm employs methods such as B-spline curves, Bézier curves, or fifth-order polynomial interpolation, and this application does not limit the methods used.

[0147] By spatiotemporally fusing continuous reference paths with driving speeds, a spatiotemporal motion trajectory containing position, heading angle, and target speed is generated.

[0148] Spatiotemporal fusion is the process of matching a reference path in the spatial dimension with a velocity curve in the temporal dimension. Specifically, based on the driving speed output by the autonomous behavior decision-making module, speed is allocated along a continuous reference path to generate a continuous position-heading angle-velocity sequence in the temporal dimension, i.e., the spatiotemporal motion trajectory.

[0149] Based on the deviation between the spatiotemporal motion trajectory and the vehicle's current actual state, chassis control commands are calculated by the trajectory tracking controller.

[0150] The task of the trajectory tracking controller is to calculate the control input so that the actual motion trajectory of the vehicle tracks the reference spatiotemporal motion trajectory as closely as possible. In some embodiments of this application, the trajectory tracking controller employs the MPC (Model Predictive Control) algorithm. Within each control cycle, the MPC controller optimizes the control input sequence in the prediction time domain based on the vehicle's kinematics or dynamics model, minimizing the tracking error between the predicted trajectory and the reference trajectory while simultaneously satisfying the physical constraints of the actuators (such as steering wheel angle limits and accelerator / brake pedal travel limits).

[0151] Chassis control commands include steering wheel angle commands, throttle opening commands, brake pressure commands, gear commands, etc., which are used to control the vehicle's steering system, drive system, and braking system.

[0152] The actuator module includes chassis actuators such as the steer-by-wire system, drive-by-wire system, brake-by-wire system, and automatic transmission. Each actuator precisely controls the vehicle's steering angle, driving torque, braking torque, and gear position according to chassis control commands sent by the planning and control module, thereby enabling autonomous vehicle operation.

[0153] By progressively transforming the driving action sequence and parking position output by the decision module into a smooth spatiotemporal motion trajectory and ultimately generating chassis control commands, a complete pathway from macroscopic behavioral intent to low-level execution response is constructed. Specifically, firstly, the macroscopic driving path topology is determined based on the driving action sequence, with the parking position as the endpoint constraint, providing a structured geometric framework and logical boundary for subsequent path generation. Based on this, a curve smoothing algorithm is used to generate a continuous reference path that satisfies vehicle kinematic constraints, eliminating unexecutable curvature abrupt changes in the polyline path. Furthermore, by spatiotemporally fusing the continuous reference path with the driving speed, a spatiotemporal motion trajectory including position, heading angle, and target speed is generated, achieving a unified expression of spatial path and time dimension, avoiding execution conflicts caused by the separation of path planning and speed planning. Simultaneously, based on the deviation between the spatiotemporal motion trajectory and the actual vehicle state, the trajectory tracking controller continuously optimizes the control input, minimizing tracking errors in the prediction time domain while satisfying the physical constraints of the actuator. This ensures stable tracking even with model errors and external disturbances, providing an engineering-featured underlying technology support for the stable, accurate, and safe execution of autonomous decision-making results by automotive robots in complex dynamic environments.

[0154] Optionally, in some embodiments of this application, the car robot further includes a closed-loop correction module.

[0155] The closed-loop correction module is used to continuously acquire real-time perception data from the vehicle perception module after the autonomous behavior decision-making module outputs the autonomous decision-making result and the execution mechanism module starts to execute it, and to monitor environmental changes and user status changes.

[0156] Specifically, the closed-loop correction module continuously acquires real-time environmental and user information output by the vehicle perception module at a preset monitoring cycle (e.g., 100ms), and calculates the environmental and user state changes relative to the last decision time.

[0157] For example, the calculation of environmental changes can take into account the degree of change of the following factors: newly added or disappeared dynamic obstacles; deviations in the movement trajectories of surrounding obstacles exceeding the prediction range; changes in traffic light status; changes in road conditions (such as the appearance of construction areas ahead, traffic accidents, etc.).

[0158] The calculation of changes in user state can take into account the degree of change of the following factors: the confidence of the user's intent decreases by more than a threshold; the user's mental state changes significantly (e.g., from low urgency to high urgency); the user issues new explicit instructions.

[0159] When the detected change in environmental information exceeds the preset environmental change threshold, and / or the change in user information exceeds the preset user status change threshold, the closed-loop correction module sends a replanning trigger signal to the autonomous behavior decision module.

[0160] The autonomous behavior decision-making module responds to the replanning trigger signal, makes a new autonomous decision based on the updated environmental information and user intent, generates an updated autonomous decision result, and executes the updated decision result by the planning control module and the execution mechanism module.

[0161] The above continuous monitoring and dynamic updating process is executed cyclically until the target parking location is reached or the user terminates the trip.

[0162] The introduction of the closed-loop correction module in this application embodiment enables the car robot to respond in real time to changes in the environment and user status during driving, ensuring that the decision results are always consistent with the current actual situation, thereby improving the driving stability and adaptability of the car robot.

[0163] Figure 3 This is a flowchart illustrating an autonomous decision-making and control method for an automotive robot, provided as an embodiment of this application. Figure 3 As shown, the method includes: S100: After the vehicle is started, acquire environmental and user information.

[0164] S200: Perform integrated analysis of environmental and user information to obtain user intent.

[0165] S300 generates autonomous decision-making results within a preset safety and compliance framework by using a pre-built multi-objective optimization decision-making model based on environmental information and user intent; the autonomous decision-making results include driving action sequence, driving speed, and parking position.

[0166] S400 generates a motion trajectory that matches the user's intent based on the driving action sequence and parking position, and combines the motion trajectory with the driving speed to obtain chassis control commands.

[0167] S500, execute chassis control commands.

[0168] To facilitate understanding of the practical application scenarios of the automotive robot and its autonomous decision-making and control method provided by this invention, an exemplary specific embodiment is given below.

[0169] For example, if a user plans to drive to a shopping mall on a weekend afternoon, the car robot begins to acquire environmental and user information after the vehicle is started.

[0170] The vehicle's current location is detected by the environmental perception submodule of the vehicle perception module, indicating that the traffic density is moderate and the weather is sunny.

[0171] The following multimodal interaction perception data was collected through the user perception submodule: Speech modality: Speech rate is moderate to slow, tone is steady, and energy envelope is flat; Visual modality: Facial expression shows mild fatigue (frequent blinking, slightly drooping corners of the mouth); Physiological modality: The ratio of low-frequency power to high-frequency power in heart rate variability is high (suggesting mild fatigue); Postural modality: Sitting posture with the head slightly tilted forward and the back slightly leaning back.

[0172] The requirement understanding module performs semantic abstraction on real-time environmental information and generates context modulation factors: the scene type is "urban road approaching the destination", the traffic density is "medium", the weather is "sunny", and the distance to the destination is about 800 meters.

[0173] By inputting multimodal interaction perception data and context modulation factors into a hierarchical dynamic Bayesian network model, the posterior probability distribution of the user's mental state is obtained: Physiological load: high load probability 0.65, medium load probability 0.25, low load probability 0.10; Emotional valence: Neutral probability 0.55, negative probability 0.30, positive probability 0.15; Sense of urgency: Low urgency probability 0.70, medium urgency probability 0.20, high urgency probability 0.10.

[0174] The entropy of the posterior probability distribution is low, with a confidence level of 0.78. The inference results across modalities show high consistency (KL divergence is less than the threshold), with an intermodal consistency assessment of 0.85. The overall interaction strategy weight factor is 0.82.

[0175] The expected utility value of each candidate demand action was calculated using a utility function. The expected utility value for arriving at the destination comfortably and smoothly is 0.72, for arriving safely and cautiously is 0.58, for arriving economically and energy-efficiently is 0.45, and for arriving quickly is 0.20. The highest expected utility value, "arriving at the destination comfortably and smoothly," was selected as the user intent, with the corresponding discrete behavior label being "comfortable driving," and a normalized confidence score of 0.78.

[0176] Make autonomous behavioral decisions. Safety margin modulation coefficient K safe =1+0.5×(1-0.78)=1.11, the safety boundary is expanded by 11%. Based on the user's intent "comfortable driving", the basic weight vector is matched in the intent-weight mapping table: timeliness 0.10, comfort 0.45, safety 0.30, energy consumption 0.15.

[0177] The extracted scenario complexity index SC = 0.45 (below the preset threshold of 0.6), so no weight adjustment is required. The final decision objective weight vector is: timeliness 0.10, comfort 0.45, safety 0.30, and energy consumption 0.15.

[0178] Using the expanded safety boundary as a constraint and the decision objective weight vector as the objective function, a multi-objective optimization decision model is invoked to solve the problem and generate candidate decision schemes. The scheme with the highest comprehensive evaluation value is: turn right 200 meters ahead into the mall parking lot entrance, drive at a speed of 25 km / h to the B1 level, and park in the nearest available parking space (numbered B1-088) to the elevator entrance. This scheme has the highest comfort evaluation value, with a root mean square longitudinal acceleration value below 0.5 m / s² and a peak jerk value below 0.8 m / s³.

[0179] The planning and control module generates a macroscopic driving path topology based on the driving action sequence, smooths it using B-spline curves to generate a continuous reference path, and performs spatiotemporal fusion with a target speed of 25 km / h to generate a spatiotemporal motion trajectory. The MPC controller calculates chassis control commands and sends them to the steer-by-wire, drive, and braking systems for execution.

[0180] In the above embodiment, while the vehicle is in motion, a closed-loop correction module continuously monitors environmental changes. If the B1 level of the parking lot ahead is temporarily closed, and the detected environmental change exceeds a threshold, a replanning is triggered. Based on the updated environmental information—B1 level is closed, requiring a move to B2 level, and the user's intention remains "comfort driving"—the autonomous behavior decision module regenerates a decision plan: proceed to B2 level and park in the nearest available parking space. The planning control module and the execution mechanism module then execute the updated decision.

[0181] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0182] In this application, unless otherwise expressly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0183] In this application, unless otherwise expressly specified and limited, "above" or "below" the second feature can mean that the first and second features are in direct contact, or that they are in indirect contact through an intermediate medium. Furthermore, "above," "on top of," and "over" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0184] In the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0185] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A car robot, characterized in that, include: The vehicle-mounted sensing module is used to acquire environmental and user information after the vehicle is started. The requirements understanding module is used to integrate and analyze environmental and user information to obtain user intent; The autonomous behavior decision-making module is used to generate autonomous decision-making results within a preset security and compliance framework by using a pre-built multi-objective optimization decision-making model based on the environmental information and user intent. The autonomous decision-making results include the sequence of driving actions, driving speed, and parking position; The planning and control module is used to generate a motion trajectory that conforms to the user's intention based on the driving action sequence and parking position, and to combine the motion trajectory with the driving speed to obtain chassis control commands, which are then sent to the actuator module. The actuator module is used to receive and execute the chassis control commands.

2. The automotive robot according to claim 1, characterized in that, The demand understanding module includes a trained hierarchical dynamic Bayesian network model; the user information is multimodal interaction perception data, including explicit command input and / or implicit state information; The requirement understanding module integrates and analyzes environmental and user information to obtain user intent, including: The environmental information is semantically abstracted to generate a context modulation factor for the semantics of the current interaction scene; The multimodal interaction perception data is used as observational evidence and, together with the context modulation factor, is input into the hierarchical dynamic Bayesian network model to obtain the posterior probability distribution of the user's mental state; the user's mental state includes physiological load, emotional valence, and sense of time urgency. The user intent is generated based on the posterior probability distribution.

3. The automotive robot according to claim 2, characterized in that, The step of generating the user intent based on the posterior probability distribution includes: Obtain the interaction strategy weight factor; the interaction strategy weight factor is determined by performing confidence evaluation on the posterior probability distribution and / or performing intermodal consistency evaluation on the multimodal interaction perception data. Based on the posterior probability distribution and the interaction strategy weight factor, the expected utility value of each candidate demand parsing action is obtained through a preset utility function; the utility function represents the expected utility value of performing different candidate demand parsing actions under different mental states. The parsing result corresponding to the candidate demand parsing action with the highest expected utility value is selected as the user intent, and the highest expected utility value is normalized and used as the confidence level of the user intent.

4. The automotive robot according to claim 3, characterized in that, The hierarchical dynamic Bayesian network model includes an observation layer, a mental state layer, and an interaction expectation layer. The observation layer includes observation variable nodes corresponding to the interactive perception data of each modality, and is connected to the mental state layer through the observation likelihood function. The mental state layer includes latent variable nodes for physiological load, latent variable nodes for emotional valence, and latent variable nodes for time urgency. The causal dependencies between each latent variable node are represented by directed edges, which are used to output the posterior probability distribution. The interaction expectation layer includes latent variable nodes for interaction mode preferences. It takes the posterior probability distribution and the context modulation factor as input and outputs a probability distribution representing the user's acceptance of candidate demand parsing actions in the current scenario. The acceptance probability distribution is used to correct each expected utility value.

5. The automotive robot according to claim 3, characterized in that, In the autonomous behavior decision-making module, based on the environmental information and user intent, an autonomous decision-making result is generated within a preset security and compliance framework using a pre-built multi-objective optimization decision-making model, including: The security margin modulation coefficient is determined based on the confidence level of the user intent, and the security boundary of the preset security compliance framework is adaptively modulated; wherein, the lower the confidence level of the user intent, the larger the security margin modulation coefficient, and the more conservative the security boundary. Based on the user intent and the environmental information, determine the decision target weight vector; Using the modulated safety boundary as a constraint and the decision target weight vector as the objective function, the multi-objective optimization decision model is used to solve the problem and generate multiple candidate decision schemes; the candidate decision schemes include at least candidate driving action sequences, candidate driving speeds, and candidate parking positions. A multi-objective evaluation is performed on the multiple candidate decision schemes, and the candidate decision scheme with the highest comprehensive evaluation value is selected as the output of the autonomous decision result.

6. The automotive robot according to claim 5, characterized in that, The step of determining the decision target weight vector based on the user intent and the environmental information includes: Based on the discrete behavioral labels of the user intent, a basic weight vector is matched in a pre-configured intent-weight mapping table; the intent-weight mapping table represents the correspondence between different candidate intent types and at least one of timeliness weight, comfort weight, safety weight and energy consumption weight. Extract the scene complexity index from the environmental information; The basic weight vector is dynamically adjusted using the scene complexity index to obtain the decision target weight vector; wherein, when the scene complexity index is higher than a preset threshold, the safety weight is increased and the comfort weight is decreased.

7. The automotive robot according to claim 5, characterized in that, The multi-objective optimization decision model is configured with a multi-objective trade-off function; The multi-objective trade-off function is used to evaluate candidate decision schemes from multiple dimensions, including at least two of the following dimensions: timeliness, comfort, safety, and energy consumption. The multi-objective optimization decision model uses the modulated safety boundary as a constraint and the decision objective weight vector as the weight configuration of each dimension in the multi-objective trade-off function to solve the weighted optimal solution of the multi-objective trade-off function. The candidate decision scheme corresponding to the weighted optimal solution is output as the candidate decision scheme with the highest comprehensive evaluation value.

8. The automotive robot according to claim 1, characterized in that, The process of generating a motion trajectory that matches the user's intent based on the driving action sequence and parking position, and combining the motion trajectory with the driving speed to obtain chassis control commands, includes: Based on the driving action sequence, the macroscopic driving path topology of the vehicle is determined, and the parking position is used as the endpoint constraint of the macroscopic driving path topology. Based on the aforementioned macroscopic driving path topology, a continuous reference path that satisfies vehicle kinematic constraints is generated using a curve smoothing algorithm. The continuous reference path and the driving speed are spatiotemporally fused to generate a spatiotemporal motion trajectory that includes position, heading angle and target speed; Based on the deviation between the spatiotemporal motion trajectory and the vehicle's current actual state, the chassis control command is calculated by the trajectory tracking controller.

9. The automotive robot according to claim 1, characterized in that, It also includes a closed-loop correction module; The closed-loop correction module is used to continuously acquire real-time perception data from the vehicle perception module and monitor environmental changes and user status changes after the autonomous behavior decision-making module outputs the autonomous decision-making result and the execution mechanism module starts to execute it. When the detected change in the environmental information exceeds a preset environmental change threshold, and / or the change in the user information exceeds a preset user state change threshold, the closed-loop correction module sends a replanning trigger signal to the autonomous behavior decision module. The autonomous behavior decision-making module responds to the replanning trigger signal, makes a new autonomous decision based on the updated environmental information and user intent, generates an updated autonomous decision result, and executes the updated decision result by the planning control module and the execution mechanism module. The above continuous monitoring and dynamic updating process is executed cyclically until the target parking location is reached or the user terminates the trip.

10. An autonomous decision-making and control method for an automotive robot, characterized in that, include: After the vehicle is started, environmental and user information is acquired; By fusing and analyzing environmental and user information, user intent can be obtained. Based on the environmental information and user intent, an autonomous decision-making result is generated within a preset safety and compliance framework through a pre-built multi-objective optimization decision-making model; the autonomous decision-making result includes driving action sequence, driving speed, and parking position. Based on the driving action sequence and parking position, a motion trajectory that matches the user's intention is generated, and the motion trajectory is combined with the driving speed to obtain chassis control commands; Execute the chassis control commands.

Images