An autonomous decision-making method and system for embodied intelligent robots

Abstract

This invention relates to the field of embodied intelligence, and provides an autonomous decision-making method and system for embodied intelligent robots. The autonomous decision-making method includes the following steps: using an active perception mechanism to collect environmental and interaction data in real time and converting this data into embodied semantic codes that fuse scene and embodiment; inputting the embodied semantic codes into a triage decision layer to output a structured semantic summary and an execution action, the structured semantic summary including scene parameters, user needs, and confidence levels; responding to a confidence level within the middle confidence range, determining and executing a non-intrusive trial action instruction to verify user needs based on the execution action and scene parameters; collecting user feedback data during the execution of the non-intrusive trial action instruction to determine the user feedback result; and determining a decision result based on the user feedback result, and adjusting the range of the dynamic confidence range and the correlation weight between the execution action and user needs in the triage decision layer.

Description

Technical Field

[0001] This invention relates to the field of embodied intelligence, and more particularly to an autonomous decision-making method and system for embodied intelligent robots. Background Technology

[0002] Existing technologies for autonomous decision-making and demand prediction in embodied intelligent robots can be mainly divided into three categories: instruction / rule-triggered schemes, perception-mapping schemes, and schemes that directly connect visual language models with large-scale language models for reasoning.

[0003] Command / rule-triggered solutions rely on explicit user commands or preset rules (such as pressure sensors triggering alarms or schedule-driven cleaning), using rule engines or thresholds to determine the action to be executed. However, command / rule-triggered solutions can only handle scenarios with clear triggering conditions and lack an autonomous closed loop of perception, decision-making, and feedback. Therefore, this solution cannot address ambiguous needs.

[0004] Perceptual mapping solutions establish a mapping table between actions and needs through manual design or supervised learning, and then determine needs by combining the perceived action information to make decisions. However, the mapping relationship of perceptual mapping solutions is fixed, making them unsuitable for ambiguous scenarios, and these solutions typically acquire external information through passive perception.

[0005] The direct inference approach between a visual language model (VLM) and a large language model (LLM) is currently the mainstream technical solution. After generating a natural language description of the scene using a VLM, it is directly input into a large language model (LLM) for user intent reasoning to obtain decision instructions. This approach relies on the generalization ability of the large model to handle open scenarios, but it suffers from VLM description redundancy and high LLM inference latency (hundreds of milliseconds). When dealing with ambiguous scenarios that do not require deep reasoning, redundancy and latency prevent the robot from making timely and effective decisions, resulting in a waste of computational resources.

[0006] Therefore, existing technologies only distinguish between direct execution under high confidence and deep reasoning under low confidence, lacking technical solutions for handling ambiguous scenarios. The aforementioned existing technologies suffer from the following technical problems when facing ambiguous scenarios: First, in such scenarios, existing technologies may choose to execute decisions directly, leading to misoperation; they may rely on lengthy voice confirmations, interfering with user behavior; or they may employ deep reasoning, but suffer from excessive computational load, long delays, and inaccurate decision instructions. Second, existing technologies cannot effectively form a closed loop between the decision-making process and user feedback in ambiguous scenarios, nor can they optimize decision-making capabilities based on actual user reactions to achieve the evolution of autonomous decision-making capabilities. Third, the feedback collection of existing technologies highly depends on explicit user interactions (such as voice responses and touchscreen confirmations), lacking mechanisms for trial and verification.

[0007] In order to overcome the above-mentioned defects of the existing technology, there is an urgent need in the field for an autonomous decision-making method for embodied intelligent robots that can construct a mechanism that integrates confidence diversion with non-intrusive trial and feedback mechanisms, so as to achieve low-intervention and evolvable autonomous decision-making of embodied intelligent robots while ensuring real-time performance and privacy. Summary of the Invention

[0008] The following provides a brief overview of one or more aspects to offer a basic understanding of them. This overview is not an exhaustive summary of all conceived aspects, nor is it intended to identify key or decisive elements of all aspects, nor to define the scope of any or all aspects. Its sole purpose is to present some concepts of one or more aspects in a simplified form as a prelude to the more detailed descriptions that follow.

[0009] To overcome the aforementioned deficiencies in the existing technology, this invention provides an autonomous decision-making method and system for embodied intelligent robots. It can construct a mechanism that integrates confidence diversion with non-intrusive trial and feedback mechanisms, enabling low-intervention and evolvable autonomous decision-making for embodied intelligent robots while ensuring real-time performance and privacy.

[0010] Specifically, the autonomous decision-making method for an embodied intelligent robot provided by the first aspect of the present invention includes the following steps: acquiring environmental and interaction data in real time using an active perception mechanism and converting the environmental and interaction data into an embodied semantic code that integrates scene and embodiment; inputting the embodied semantic code into a triage decision layer to output a structured semantic summary and an execution action, wherein the structured semantic summary includes scene parameters, user needs, and confidence levels, the confidence level being located within a dynamic confidence level interval, the dynamic confidence level interval including a low confidence level interval, a medium confidence level interval, and a high confidence level interval; responding to the confidence level being located within the medium confidence level interval, determining and executing a non-intrusive probing action instruction for verifying the user needs based on the execution action and the scene parameters; acquiring user feedback data during the execution of the non-intrusive probing action instruction to determine a user feedback result; and determining a decision result based on the user feedback result, and adjusting the range of the dynamic confidence level interval and the correlation weight between the execution action and the user needs in the triage decision layer, the correlation weight affecting the magnitude of the confidence level output by the triage decision layer.

[0011] Furthermore, in some embodiments of the present invention, the active perception mechanism includes dynamic perspective optimization, text context sampling, and multimodal coupling representation.

[0012] Furthermore, in some embodiments of the present invention, the triage decision layer is a lightweight world model optimized by RLHF. The pre-training process of the lightweight world model optimized by RLHF includes the following steps: determining a semantic sequence based on historical environment and interaction data, the semantic sequence including user actions, perceptual states, and semantic events; and inputting the semantic sequence into the lightweight world model to output a structured semantic summary and an execution action. The reward signal of the lightweight world model covers the rationality of the response of the execution action, the timing of intervention, and the degree of interference. The value function of the lightweight world model is optimized by RLHF.

[0013] Furthermore, in some embodiments of the present invention, the step of determining and executing a non-invasive probing action instruction for verifying the user's needs based on the execution action and the scene parameters includes: adjusting the action type and / or action parameters of the execution action in combination with the reasonableness of the response, the timing of intervention, and the degree of interference under the scene parameters to generate a non-invasive probing action instruction; or pre-setting the action type and / or preset action parameter range of the execution action mapping based on the reasonableness of the response, the timing of intervention, and the degree of interference under the scene parameters, and generating a non-invasive probing action instruction based on the preset action type and / or preset action parameter range of the execution action mapping.

[0014] Furthermore, in some embodiments of the present invention, the method further includes the steps of: executing the execution action in response to the confidence level being in the high confidence interval; collecting user feedback data during the execution of the execution action to determine the user feedback result; adjusting the range of the dynamic confidence interval and the association weight between the execution action and the user need in the triage decision layer based on the user feedback result; or, in response to the confidence level being in the low confidence interval, calling a large language model to perform semantic reasoning on the structured semantic summary to obtain a decision instruction, and executing the decision instruction; collecting user feedback data during the execution of the decision instruction to determine the user feedback result; and adjusting the range of the dynamic confidence interval and the association weight between the execution action and the user need in the triage decision layer based on the user feedback result.

[0015] Furthermore, in some embodiments of the present invention, the user feedback result includes acceptance, rejection, and no feedback. The step of adjusting the range of the dynamic confidence interval and the association weight between the execution action and the user need in the triage decision layer based on the user feedback result includes: in response to the user feedback result being acceptance, lowering the high confidence interval corresponding to the scenario parameter to increase the association weight between the execution action and the user need in the triage decision layer; in response to the user feedback result being rejection, raising the medium confidence interval corresponding to the scenario parameter to reduce the association weight between the execution action and the user need in the triage decision layer by a preset first threshold; or in response to the user feedback result being no feedback, adding a delayed trial marker to the scenario parameter to reduce the association weight between the execution action and the user need in the triage decision layer by a preset second threshold, wherein the second threshold is less than the first threshold.

[0016] Furthermore, in some embodiments of the present invention, the step of recording and storing a time-series interpretable decision log is also included, wherein the time-series interpretable decision log includes the input and output data of the triage decision layer, the confidence triage path selection result, the non-intrusive trial action instruction and / or the user feedback result.

[0017] Furthermore, the autonomous decision-making system for an embodied intelligent robot provided according to the second aspect of the present invention includes a memory and a processor. The memory stores computer instructions. The processor is connected to the memory and configured to execute the computer instructions stored in the memory to implement the autonomous decision-making method for an embodied intelligent robot provided in any of the above embodiments.

[0018] Furthermore, the embodied intelligent robot provided according to the third aspect of the present invention includes a perception module, an execution module, and an autonomous decision-making system for the embodied intelligent robot provided in the second aspect of the present invention.

[0019] Furthermore, the computer-readable storage medium provided according to the fourth aspect of the present invention stores computer instructions. When the computer instructions are executed by a processor, the autonomous decision-making method for an embodied intelligent robot provided in the first aspect of the present invention is implemented.

[0020] The beneficial effects of this invention are as follows: The autonomous decision-making system provided by this invention integrates confidence level diversion with non-intrusive trial and feedback mechanisms. Under the premise of ensuring real-time performance and privacy, it overcomes the defects of existing solutions such as confidence level misoperation, interactive interference, and disconnect between decision-making and feedback. It realizes autonomous decision-making of embodied intelligent robots with low intervention and evolution, and improves the accuracy and naturalness of decision-making. Attached Figure Description

[0021] The above-described features and advantages of the present invention will be better understood after reading the following detailed description of embodiments of the present disclosure in conjunction with the accompanying drawings. In the drawings, components are not necessarily drawn to scale, and components having similar related characteristics or features may have the same or similar reference numerals.

[0022] Figure 1 A schematic diagram of an autonomous decision-making system for an embodied intelligent robot, according to some embodiments of the present invention, is shown.

[0023] Figure 2 A flowchart of an autonomous decision-making method for an embodied intelligent robot, according to some embodiments of the present invention, is shown;

[0024] Figure 3 A flowchart illustrating a complete closed loop of an autonomous decision-making method for an embodied intelligent robot, including perception, prediction, triage, and feedback, is shown according to some embodiments of the present invention.

[0025] Figure label:

[0026] 100: Autonomous decision-making systems for embodied intelligent robots;

[0027] 110: Memory;

[0028] 111: Computer-readable storage medium;

[0029] 120: Processor;

[0030] S210~S250: Steps. Detailed Implementation

[0031] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Although the description of the present invention will be presented in conjunction with preferred embodiments, this does not mean that the features of the invention are limited to these embodiments. On the contrary, the purpose of describing the invention in conjunction with embodiments is to cover other options or modifications that may be derived based on the claims of the present invention. To provide a thorough understanding of the invention, many specific details will be included in the following description. The invention may also be implemented without using these details. Furthermore, to avoid confusion or obscuring the focus of the invention, some specific details will be omitted in the description.

[0032] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; 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; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0033] Furthermore, the terms "upper," "lower," "left," "right," "top," "bottom," "horizontal," and "vertical" used in the following description should be understood as the orientations shown in the relevant paragraphs and accompanying drawings. These relative terms are for illustrative purposes only and do not imply that the described apparatus must be manufactured or operated in a specific orientation, and therefore should not be construed as limiting the invention.

[0034] It is understood that although terms such as "first," "second," and "third" may be used herein to describe various components, regions, layers, and / or parts, these components, regions, layers, and / or parts should not be limited by these terms, and these terms are only used to distinguish different components, regions, layers, and / or parts. Therefore, the first components, regions, layers, and / or parts discussed below may be referred to as second components, regions, layers, and / or parts without departing from some embodiments of the present invention.

[0035] As mentioned above, existing technologies only distinguish between direct execution under high confidence and deep reasoning under low confidence, lacking technical solutions for handling ambiguous scenarios. These existing technologies suffer from the following technical problems when facing ambiguous scenarios: First, in such scenarios, existing technologies may choose to execute decisions directly, leading to misoperation; they may rely on lengthy voice confirmations, interfering with user behavior; or they may employ deep reasoning, but suffer from excessive computational load, long delays, and inaccurate decision instructions. Second, existing technologies cannot effectively create a closed loop between the decision-making process and user feedback in ambiguous scenarios, nor can they optimize decision-making capabilities based on actual user reactions to achieve the evolution of autonomous decision-making abilities. Third, the feedback collection of existing technologies highly depends on explicit user interactions (such as voice responses and touchscreen confirmations), lacking mechanisms for trial and verification.

[0036] To overcome the aforementioned deficiencies in the existing technology, this invention provides an autonomous decision-making method and system for embodied intelligent robots. It can construct a mechanism that integrates confidence diversion with non-intrusive trial and feedback mechanisms, enabling low-intervention and evolvable autonomous decision-making for embodied intelligent robots while ensuring real-time performance and privacy.

[0037] In some embodiments, the autonomous decision-making method for embodied intelligent robots provided in the first aspect of the present invention can be implemented via the autonomous decision-making system for embodied intelligent robots provided in the second aspect of the present invention.

[0038] Please refer to Figure 1 , Figure 1 A schematic diagram of an autonomous decision-making system for an embodied intelligent robot, according to some embodiments of the present invention, is shown.

[0039] like Figure 1 As shown, the autonomous decision-making system 100 for an embodied intelligent robot may be configured with a memory 110 and a processor 120. The memory 110 includes, but is not limited to, the computer-readable storage medium 111 described in the fourth aspect of the present invention, which stores computer instructions thereon. The processor 120 is connected to the memory 110 and configured to execute the computer instructions stored in the memory 110 to implement the autonomous decision-making method for an embodied intelligent robot provided in the first aspect of the present invention.

[0040] In some embodiments, the autonomous decision-making system 100 for an embodied intelligent robot can be configured in the embodied intelligent robot provided in the third aspect of the present invention, as part of the edge computing module of the embodied intelligent robot. As the executing entity, the embodied intelligent robot, with the special authorization of the user, actively perceives environmental and interaction data as well as user feedback data, thereby enabling the embodied intelligent robot to execute the steps of the autonomous decision-making method for the embodied intelligent robot based on the autonomous decision-making system 100.

[0041] The following will first describe the working principle of the aforementioned autonomous decision-making system and embodied intelligent robot in conjunction with some embodiments of autonomous decision-making methods. Those skilled in the art will understand that these embodiments of autonomous decision-making methods are merely non-limiting implementations provided by the present invention, intended to clearly demonstrate the main concepts of the invention and provide specific solutions convenient for public implementation, rather than limiting all functions or all working methods of the autonomous decision-making system and embodied intelligent robot. Similarly, the autonomous decision-making system and embodied intelligent robot are also merely non-limiting implementations provided by the present invention, and do not constitute a limitation on the executing subject and execution order of the steps in these autonomous decision-making methods.

[0042] Please refer to Figure 2 and Figure 3 , Figure 2 A flowchart of an autonomous decision-making method for an embodied intelligent robot, according to some embodiments of the present invention, is shown. Figure 3A flowchart illustrating a complete closed loop of an autonomous decision-making method for an embodied intelligent robot, including perception, prediction, triage, and feedback, is shown according to some embodiments of the present invention.

[0043] like Figure 2 and Figure 3 As shown, after the embodied intelligent robot enters the work scene, the embodied intelligent robot can perform step S210: using an active perception mechanism to collect environmental and interaction data in real time and convert the environmental and interaction data into embodied semantic codes that integrate the scene and the embodied.

[0044] The robot uses visual sensors (depth cameras), auditory sensors (microphone arrays), force / tactile sensors (six-dimensional force sensors), and body sensors (IMU, joint encoders) to actively collect environmental and interaction data in real time to achieve multimodal active perception and embodied data collection.

[0045] Unlike traditional "passive observation" perception strategies, autonomous decision-making systems employ an active perception mechanism to obtain higher-quality data, thereby improving the accuracy of subsequent semantic summaries and their confidence levels.

[0046] The autonomous decision-making system achieves an active perception mechanism through dynamic perspective optimization, text context sampling, and multimodal coupled representation.

[0047] Dynamic perspective optimization can specifically include: acquiring collected visual signals; in response to occlusion of the visual signals, the embodied intelligent robot plans a movement path based on environmental data and autonomously moves to the preferred observation point according to the movement path. In this way, when the visual signals are occluded or ambiguous, such as when a user's gesture is blocked by a bookshelf, the robot can autonomously move to a better observation posture, such as moving laterally to the side of the bookshelf, thereby ensuring the integrity of the robot's perception data.

[0048] Behavioral context sampling can specifically include recording the embodied intelligent robot's own movement intentions, speed, and path changes. While collecting external environmental and interaction data, the autonomous decision-making system can simultaneously record the robot's own movement intentions, speed, and path changes as part of the raw data input to the subsequent model, enabling the model to understand "why the robot is observing" and providing contextual support for the robot's autonomous decision-making.

[0049] Multimodal coupling representation can specifically include: fusing environmental and interaction data to form embodied semantic encoding that integrates scene and embodiment. Environmental and interaction data can include visual image features, sound signals, joint states, and force responses. Joint states can refer to the posture of the robotic arm of an embodied intelligent robot.

[0050] Subsequently, the embodied intelligent robot can execute step S220: inputting the embodied semantic code into the triage decision layer to output a structured semantic summary and an action to be performed. The structured semantic summary includes scene parameters, user needs, and confidence levels. Here, the output confidence level can be located within a dynamic confidence level range, which can include a low confidence level range, a medium confidence level range, and a high confidence level range.

[0051] In some embodiments, the autonomous decision-making system can acquire embodied semantic encoding obtained from the multimodal coupled representation of the active perception mechanism, input the embodied semantic encoding into a trained prediction model in the triage decision layer to output a structured semantic summary including user needs, and make preliminary autonomous decisions based on candidate user needs. User needs may include user actions, the object of user actions, the risk of user actions, or the intention of user actions; that is, user needs can be represented based on the structure of "user actions, the object of user actions, the risk of user actions, or the intention of user actions".

[0052] As the "pre-conscious decision-making layer" of the embodied intelligent robot, the diversion decision layer is responsible for real-time compression and intent inference of multimodal perception information, and outputs structured semantic summaries and corresponding confidence scores. The output confidence scores can provide a basis for judgment for the diversion mechanism.

[0053] The determination of the execution action is closely related to the user requirements output by the structured semantic summary. The triage decision layer can obtain the scene parameters indicated by the current structured semantic summary and the robot's execution actions associated with the user requirements based on preset dynamic association weights. The preset dynamic association weights between the execution actions and user requirements can further influence the confidence level of the output of the triage decision layer when making predictions.

[0054] In some embodiments, the triage decision layer can be a lightweight world model optimized by RLHF (Reinforcement Learning from Human Feedback). The RLHF-optimized lightweight world model can make predictions based on embodied semantic encodings of scene structured semantics (such as objects, actions, spatial relationships, and risk states) extracted from multimodal perception data, so as to output high-value structured semantic summaries and corresponding confidence scores.

[0055] In some embodiments, the pre-training process of the RLHF-optimized lightweight world model may include: determining semantic sequences based on historical environment and interaction data, the input sequences including user actions, perceptual states, and semantic events; inputting the semantic sequences into the lightweight world model to output a structured semantic summary including confidence and an execution action, the reward signal of the lightweight world model covering the rationality of the execution action response, the timing of intervention, and the degree of interference, and the value function of the lightweight world model being optimized by RLHF.

[0056] In the pre-training process of the lightweight world model, replacing all video frames with low-dimensional semantic sequences determined by historical environment and interaction data can reduce the amount of data and solve the problem of high training costs in traditional video training. Reward signals can be generated by human annotation or expert demonstrations, covering dimensions such as the rationality of the robot's reaction, the timing of intervention, and the degree of interference. Combining RLHF with optimized value functions of the lightweight world model enables it to accurately assess confidence in fuzzy situations without relying on massive amounts of visual data. Through the pre-training method provided by this invention, the lightweight world model can reduce its reliance on visual data by 90% and lower training costs by 60%, while achieving millisecond-level prediction. Compared to existing VLM and LLM combined schemes, this reduces inference latency by 80%.

[0057] To further improve the reliability of confidence assessment, the lightweight world model optimized by RLHF can draw on the uncertainty quantification methods in architectures such as PEVA and DeepConf. By analyzing internal token entropy, attention distribution, and tail confidence signals, the model confidence can be quantified, and the certainty of the output results can be verified.

[0058] The pre-trained, RLHF-optimized lightweight world model can output structured embeddings in a refined format, such as "user action - object of user action - risk of user action - confidence level," without relying on natural language generation. In one example, when an embodied intelligent robot forms an embodied semantic code representing "user crouching, touching a book, unstable posture" based on an active perception mechanism, this embodied semantic code is input into the RLHF-optimized lightweight world model, which outputs a structured semantic summary: "object retrieval - book - fall risk - 0.82." Furthermore, the lightweight world model can obtain the action associated with the user's needs indicated by this structured semantic summary. For example, the action associated with the structured semantic summary "object retrieval - book - fall risk - 0.82" could be "provide support."

[0059] In this way, the lightweight world model optimized by RLHF can extract core elements, including scene parameters and user needs, from the collected multimodal perception data. These extracted core elements serve as the input basis for the lightweight world model's confidence assessment, as well as subsequent non-intrusive feedback verification and LLM inference, avoiding redundant information interference in these processes and ensuring decision-making efficiency. Furthermore, by utilizing this extracted and filtered semantic information, the amount of privacy data transmitted during subsequent LLM inference can be reduced by 95%, guaranteeing user privacy and security.

[0060] Furthermore, the lightweight world model optimized by RLHF has the advantages of low computational cost (parameter size controlled within 2B) and low latency (millisecond-level response). The latency of the entire inference process is controlled within 100ms, ensuring real-time response and thus supporting real-time "instinctive prediction," providing a data foundation for subsequent confidence-based splitting.

[0061] The confidence level corresponding to the structured semantic summary output by the diversion decision layer can be used to characterize whether the prediction of the diversion decision layer is reliable.

[0062] like Figure 3 As shown, the autonomous decision-making system can determine whether the predicted user demand is accurate based on the dynamic confidence interval where the confidence level is located, and further decide whether the embodied intelligent robot should execute the execution action output by the diversion decision layer based on the dynamic confidence interval where the confidence level is located.

[0063] exist Figure 3 In the illustrated embodiment, the dynamic confidence interval may include a low confidence interval, a medium confidence interval, and a high confidence interval.

[0064] When the confidence level is in the low confidence level range, the autonomous decision-making system executes the deep reasoning path; when the confidence level is in the medium confidence level range, it executes the non-intrusive feedback verification path; and when the confidence level is in the high confidence level range, it executes the direct execution path.

[0065] exist Figure 3 In the illustrated embodiment, the low confidence interval can be less than 0.6, the medium confidence interval can be greater than or equal to 0.6 and less than or equal to 0.9, and the high confidence interval can be greater than 0.9. The range of the confidence interval is dynamic, as described below.

[0066] When the confidence level is in the low confidence range, it indicates that the triage decision layer cannot accurately determine the user's needs in the current scenario and the actions that the embodied intelligent robot can perform based on those needs. The user's needs are highly ambiguous, thus requiring deep reasoning. The autonomous decision-making system can execute a deep reasoning path, calling a large language model to perform semantic reasoning on the structured semantic summary to obtain decision instructions, and then execute those instructions.

[0067] In one example, the autonomous decision-making system of an embodied intelligent robot can be configured with an LLM in an edge computing unit or connected to an LLM in an external cloud. When the confidence level is in the low confidence range, the autonomous decision-making system determines that the user's needs are highly ambiguous and sends the structured semantic summary output from the triage decision layer to the LLM. The LLM is then responsible for high-level causal and socio-semantic reasoning and outputs explicit decision instructions. After receiving the decision instructions returned by the LLM, the autonomous decision-making system can directly proceed to the corresponding execution path.

[0068] Correspondingly, when the confidence level is in the high confidence range, it indicates that the triage decision layer judges that the user needs in the current scenario are clear, and the execution actions associated with these clear user needs are accurate. The robot can directly execute the execution actions output by the triage decision layer.

[0069] In one example, when a user repeatedly points to the water cup without any other ambiguous actions, the autonomous decision-making system can determine that the user's need is clear and output a predicted structured semantic summary: "Instruction - Water Cup - Want Water Cup - 0.95". Since this confidence level is in the high-confidence range, the autonomous decision-making system can directly invoke the Hardware Abstraction Layer (HAL) to drive the embodied intelligent robot's actuators to perform the associated actions of delivering the water cup and obstacle avoidance. Here, the action parameters for instructing the delivery of the water cup and initiating obstacle avoidance based on obstacle avoidance risk can be set based on historical optimization results, without the need for additional feedback verification. The Hardware Abstraction Layer (HAL) is connected to the embodied intelligent robot's actuators (such as robotic arms or mobile chassis), which can convert the execution actions into standardized hardware control signals and adapt to different hardware drive protocols to achieve hardware-software decoupling.

[0070] When the confidence level is in the medium confidence range, it indicates that the triage decision layer judges the current scenario to be ambiguous. In this ambiguous scenario, user needs may be ambiguous, but no deep reasoning is required. The autonomous decision-making system can initiate a non-intrusive feedback mechanism to execute a non-intrusive feedback verification path.

[0071] The non-intrusive feedback mechanism is a human-computer interaction and feedback collection mechanism designed by the autonomous decision-making method provided in this invention for ambiguous scenarios with medium confidence levels. Through the non-intrusive feedback mechanism, the autonomous decision-making system can form a logical chain of trial and error, feedback verification, and decision adjustment.

[0072] like Figure 2 As shown, the embodied intelligent robot can perform step S230: in response to the confidence level being in the medium confidence level range, it determines and executes non-intrusive trial action instructions for verifying user needs based on the action to be performed and scene parameters.

[0073] Non-intrusive feedback mechanisms can include the generation of non-intrusive probing action commands and the collection of natural user responses.

[0074] The generation of non-invasive probing action commands refers to the process by which the robot verifies the user demand assumptions output by the triage decision layer through light physical / visual movements (such as gently moving objects, lightly pushing tests, and posture preparation) without interfering with user behavior, intruding into safe spaces, or relying on voice interruptions. The collection of natural user reactions involves using sensors to capture the user's natural responses to the non-invasive probing actions, and using the user feedback results to form key signals for the autonomous decision-making system to optimize its decisions.

[0075] In some embodiments, the generation of non-intrusive probing action commands can be generated by the autonomous decision-making system adjusting the predicted action type and / or action parameters based on the scene parameters, intervention timing, and interference level in the core information output by the triage decision layer. For example, when the triage decision layer determines that the user's need is to wipe the table and the associated action is to move a rag, since the confidence level is in the medium confidence range, meaning that this user need assumption may be ambiguous, the autonomous decision-making system can adjust the action "move a rag" based on the scene parameters, combining the reaction rationality index, intervention timing index, and interference level index. For example, the action parameters of the action "move a rag" can be adjusted to have a force of less than 5N, a movement distance of less than 0.5 meters, and a speed of less than 0.1m / s.

[0076] In some embodiments, the generation of non-invasive probing action instructions may also be based on the reasonableness of the response under scene parameters, the timing of intervention and the degree of interference, the preset action type of the execution action mapping and / or the preset action parameter range, and then the non-invasive probing action instructions may be generated based on the preset action type of the execution action mapping and / or the preset action parameter range.

[0077] Specifically, the generation of non-intrusive probing action commands can be achieved by adjusting the action parameters of the executed action to a preset range based on the reasonableness of the response, the timing of intervention, and the degree of interference in the corresponding scenario. For example, setting the force, distance, and speed can ensure that the probing does not interfere with the user. Alternatively, the generation of non-intrusive probing action commands can also be achieved by mapping the probing action and its action parameters to the action type of the executed action based on a preset association database. For example, when the user's need is to retrieve an object and the associated execution action is delivery, the probing action "gentle push" can be associated with the execution action "delivery".

[0078] In this way, the autonomous decision-making system can generate light physical / visual movements that enable the robot to perform without interfering with user behavior, intruding into safe spaces, or relying on voice interruptions. These non-intrusive trial movements are used to verify the user demand assumptions output by the triage decision layer.

[0079] like Figure 2 As shown, the embodied intelligent robot can perform step S240: collecting user feedback data when non-invasive trial action commands are executed to determine the user feedback results.

[0080] When an embodied intelligent robot executes non-invasive trial action commands, the autonomous decision-making system can collect user feedback data during the execution of the non-invasive trial action commands to collect the user's natural reactions. Based on the collected user natural reactions, the system determines the user feedback results, which serve as the direct basis for decision adjustments.

[0081] User feedback can be either accepted, rejected, or no feedback. In one example, when the robot collects video data of a user reaching out to touch an object and interprets and understands it using a lightweight world model, it can determine that the user feedback is accepted; or, when the robot determines that the user has taken a physical avoidance action based on the collected user feedback data, it can determine that the user feedback is rejected; or, when the robot determines that the user continues the original behavior based on the collected user feedback data, it can determine that the user feedback is no feedback.

[0082] Then, the embodied intelligent robot can perform step S250: determine the decision result based on the user feedback, and adjust the range of the dynamic confidence interval and the correlation weight between the action to be performed and the user's needs in the triage decision layer. Here, the correlation weight affects the confidence level of the output of the triage decision layer.

[0083] The autonomous decision-making system can determine the decision outcome based on user feedback. Specifically, when the user feedback is "accept," it indicates that the triage decision-making layer's assumptions about the user's needs are accurate, and the embodied intelligent robot can execute the actions output by the triage decision-making layer. When the user feedback is "reject" or there is no feedback, it indicates that the triage decision-making layer's assumptions about the user's needs are questionable, and the embodied intelligent robot can stop executing non-intrusive trial-and-error commands to reduce interference with the user.

[0084] Furthermore, the autonomous decision-making system can adjust the range of the confidence interval and the correlation weight between the actions executed and user needs in the triage decision-making layer based on user feedback.

[0085] The confidence interval range in the autonomous decision-making system is dynamically calibrated. The confidence interval range can be adaptively adjusted based on user feedback, thereby enabling dynamic traffic distribution in the autonomous decision-making system. This allows it to adapt to different user behavior patterns and response styles, replacing the limitations of traditional single decision paths or fixed threshold judgments.

[0086] Furthermore, the autonomous decision-making system can update the correlation weights between user needs and actions in the lightweight world model based on user feedback. By updating the correlation weights between user needs and actions in the lightweight world model, the model can adjust the confidence level of its output during prediction.

[0087] Furthermore, the autonomous decision-making system can adapt the parameters and types of non-intrusive trial actions based on user feedback.

[0088] Therefore, the autonomous decision-making system ensures that its autonomous decision-making capability can continue to iterate through feedback-driven autonomous decision optimization, so as to better adapt to the dynamic and personalized needs of users.

[0089] In some embodiments, when the user feedback is "accept," the autonomous decision-making system can lower the high-confidence interval corresponding to the scenario parameters, increase the correlation weight between the executed action and the user's needs in the triage decision layer, and optimize the action parameters of the non-intrusive probing action command. In one example, the autonomous decision-making system can lower the high-confidence interval corresponding to the scenario parameters by 5%, for example, from 0.9 to 0.85, thereby prioritizing triage to the direct execution path in the next similar scenario. Furthermore, it increases the correlation weight between the current executed action and the user's needs by 10% and optimizes the action parameters of the non-intrusive probing action command (such as delivery height and force), using these as the default parameters for the non-intrusive probing action command in this scenario.

[0090] When the user feedback is a rejection, the autonomous decision-making system can increase the medium confidence interval corresponding to the scene parameters to reduce the correlation weight between the executed action and the user's needs in the triage decision layer by setting a first threshold, and adjust the action parameters of the non-intrusive trial action command according to the degree of interference. In one example, the autonomous decision-making system can increase the medium confidence interval corresponding to the scene parameters by 5%, for example, from 0.6 to 0.65, to reduce the frequency of similar trials. Furthermore, it can reduce the correlation weight between the current executed action and the user's needs by 8%, and adjust the action parameters of the non-intrusive trial action command (e.g., reducing the robotic arm height by 15cm and increasing the movement distance by 0.2 meters) to reduce interference with the user.

[0091] When the user feedback result is no feedback, the autonomous decision-making system can add a delayed trial flag to the scene parameters. This allows for a preset second threshold to reduce the correlation weight between the executed action and the user's needs in the triage decision layer and to simplify the action type of the non-intrusive trial action command. This second threshold is lower than the first threshold to fine-tune the weights and avoid over-prediction. In one example, if the user continues the original behavior for 5 seconds without response, the autonomous decision-making system can determine that the user feedback result is no feedback. Instead of changing the confidence threshold ranges, the autonomous decision-making system can add a delayed trial flag to the scene parameters, marking the scene as a low-priority trial. The next time the same scene is encountered, the autonomous decision-making system can choose to delay for 2 seconds before executing the non-intrusive trial action command based on this delayed trial flag. The autonomous decision-making system can fine-tune the correlation weight between the current executed action and the user's needs, for example, by reducing it by 3%, to avoid over-prediction. Furthermore, the autonomous decision-making system can simplify the action type of the non-intrusive trial action command to further reduce ineffective intervention, for example, simplifying the action type of "moving an object" to "in-place gentle push prompt".

[0092] Preferably, in Figure 3 In the illustrated embodiment, the autonomous decision-making system can adjust the confidence interval and association weights not only based on user feedback results during the execution of non-intrusive trial action instructions, but also based on user feedback results under the direct execution path and user feedback results under the deep inference path.

[0093] Specifically, in response to a high confidence level, the autonomous decision-making system collects user feedback data during the execution of actions via the direct execution path to determine the user feedback result. Then, based on the user feedback result, it adjusts the range of the dynamic confidence interval and the correlation weight between the executed actions and user needs in the triage decision layer. In response to a low confidence level, the autonomous decision-making system collects user feedback data during the execution of decision instructions derived from large language model reasoning via the deep reasoning path to determine the user feedback result. Again, based on the user feedback result, it adjusts the range of the dynamic confidence interval and the correlation weight between the executed actions and user needs in the triage decision layer.

[0094] In addition, such as Figure 3 As shown, the autonomous decision-making system can also achieve periodic model optimization through autonomous knowledge distillation. The system can record user rejection use cases, encapsulate the current decision scenario, user rejection response, and environmental context into knowledge samples, and submit them to a larger model for specific reasoning. The larger model, acting as a "teacher," generates decision explanations and alternatives, which the autonomous decision-making system records as higher-order knowledge and uses as correction samples. Simultaneously, the system can incorporate use cases that accept feedback, enabling targeted optimization and distillation of the lightweight world model, thereby achieving a fundamental policy upgrade.

[0095] Furthermore, the autonomous decision-making system can record a time-series interpretable decision log throughout the entire process, stored in a local encrypted unit. The content of the time-series interpretable decision log can include the input (e.g., embodied semantic encoding) and structured semantic summary and confidence basis of the lightweight world model of the triage decision layer; the confidence-based triage path selection results; the trajectory and action parameters of non-intrusive trial actions; and user feedback results and corresponding autonomous decision parameter adjustment records. This time-series interpretable decision log supports subsequent auditing, fault debugging, and model retraining, ensuring the transparency and traceability of the autonomous decision-making process and enhancing user trust.

[0096] In summary, the autonomous decision-making system provided by this invention integrates confidence level diversion with a non-intrusive trial and feedback mechanism. While ensuring real-time performance and privacy, it overcomes the shortcomings of existing solutions such as mid-confidence level misoperation, interactive interference, and disconnect between decision-making and feedback. It enables low-intervention, evolvable embodied intelligent robots to make autonomous decisions and improves the accuracy and naturalness of decision-making.

[0097] Non-invasive trial-and-error and feedback mechanisms enable embodied intelligent robots to better handle ambiguous scenarios, reducing the interaction interference rate from 35% in existing solutions to below 14% and the misoperation rate to below 1%. The feedback-driven confidence and motion parameter optimization provided by this invention improves the autonomous decision-making accuracy of embodied intelligent robots from 82% to over 95% and better enables user personalization.

[0098] The lightweight world model provided by this invention can achieve millisecond-level semantic filtering and structured embedding output, keeping the response latency within 100ms, meeting users' immediate needs in real time. Furthermore, in low-confidence scenarios, only the structured semantic summary output by the lightweight world model is uploaded (the data volume is 1 / 20 of that of the existing VLM description), which greatly reduces the risk of privacy data leakage.

[0099] In addition, the present invention enables the continuous evolution of autonomous decision-making capabilities. The feedback-driven optimization mechanism allows the autonomous decision-making system to adapt to the behavior patterns of specific users after 3 to 5 interactions. In long-term use, the confidence assessment error is reduced by 8%-12%, and the accuracy of non-intrusive action adaptation is improved by 20%, achieving autonomous evolution in terms of accuracy and comprehension.

[0100] Furthermore, this invention can utilize logs to fully record data from the entire process of perception, prediction, triage, feedback, and optimization, making decision-making transparent, enhancing user trust, and providing traceable evidence such as action parameters and trajectories for safety incidents, thus meeting safety compliance requirements.

[0101] The following are several non-limiting preferred embodiments, which are used to further elaborate on the autonomous decision-making method for embodied intelligent robots proposed in this invention.

[0102] In the preferred embodiment, the autonomous decision-making system can be composed of multiple functional modules, and the above-mentioned autonomous decision-making method can be realized through the architecture formed by these multiple functional modules.

[0103] Autonomous decision-making systems can include a perception layer. The perception layer includes, but is not limited to, depth cameras, microphone arrays, six-dimensional force sensors, LiDAR, IMUs, and joint encoders. The perception layer can employ active perception mechanisms to acquire multimodal environmental and ontological data and output embodied semantic codes to edge computing units.

[0104] The autonomous decision-making system may include an RLHF optimization module, which can be deployed at the edge of the embodied intelligent robot or in the cloud. This RLHF optimization module is responsible for the pre-training and iterative updating of the lightweight world model in the edge computing unit. In actual operation, the low-dimensional semantic sequence extracted from the perception layer, including user actions, perceptual states, and semantic events, along with human-annotated reward signals, is input into the RLHF optimization module. The module optimizes the value function of the lightweight world model through RLHF and outputs the optimized parameters of the lightweight world model to the edge computing unit.

[0105] The autonomous decision-making system may include an edge computing unit, integrated into the embodied intelligent robot body, which is the core execution unit for confidence-based traffic splitting and non-intrusive feedback fusion. This edge computing unit incorporates a lightweight world model, an active perception and control module, a confidence-based traffic splitting module, a non-intrusive feedback generation module, and a feedback acquisition module. Through these integrated functional modules, the edge computing unit can achieve active perception scheduling, instinctive prediction, traffic splitting decision-making, trial action generation, and real-time feedback recognition to ensure low-latency autonomous decision-making.

[0106] The autonomous decision-making system may include an intelligent platform, which can be deployed at the edge or in the cloud. The intelligent platform includes LLM and big data analytics modules. This intelligent platform can receive structured semantic summaries of low-confidence scenarios output by the lightweight world model in the edge computing unit, thereby performing deep reasoning. Preferably, the intelligent platform can also store robot interaction data to provide training data for the RLHF optimization module. More preferably, the intelligent platform can also periodically push model iteration update packages.

[0107] An autonomous decision-making system may include a feedback optimization unit. This unit determines the user feedback result based on user feedback data and further performs confidence threshold calibration, weight updates between the actions performed in the lightweight world model and user needs, and adjustments to the action parameters of non-intrusive trial action instructions based on the user feedback result. The feedback optimization unit is the core module for realizing the evolution of the autonomous decision-making capability of the autonomous decision-making system.

[0108] An autonomous decision-making system may include a hardware abstraction layer. This hardware abstraction layer can connect edge computing units and the actuators of embodied intelligent robots. Through this hardware abstraction layer, execution commands, decision commands, or non-intrusive trial commands output by the edge computing units can be converted into standardized hardware control signals.

[0109] The autonomous decision-making system may include an interpretable decision log unit. This interpretable decision log unit is a locally encrypted storage module that records data from the entire autonomous decision-making process in chronological order to support subsequent queries, audits, and retraining data export, ensuring that decisions are transparent and traceable.

[0110] In a preferred embodiment, the embodied intelligent robot provided by this invention can be used to assist the elderly in a home setting. The embodied intelligent robot can navigate within this home environment and collect environmental and interaction data in real time through an active perception mechanism to make real-time decisions.

[0111] In Example 2, during its navigation, the embodied intelligent robot's depth camera initially detects an obstruction from the sofa armrest (only the upper body of the elderly person is visible). The robot autonomously moves laterally to the side of the sofa to achieve dynamic perspective optimization. Simultaneously, its movement path (from beside the coffee table to the side of the sofa) and speed (0.1 m / s) are recorded to achieve behavioral context sampling. Under dynamic perspective optimization, the robot's visual sensors collect and identify the elderly person's forward lean and hand on the sofa armrest; the robot's force sensors collect and identify the increased pressure on the armrest (from 50 N to 120 N); and the IMU sensor collects and identifies multimodal data indicating that the robot's body tilt angle exceeds a threshold. The robot utilizes multimodal coupling representation to fuse the collected and identified data into embodied semantic encoding.

[0112] The embodied semantic encoding is input into a pre-trained lightweight world model for processing, outputting a structured semantic summary: scene parameters (elderly person), user action (hand supporting the sofa), user's intention (user needs to get up), risk of the user's action (potential imbalance), and confidence score (0.82). Based on this structured semantic summary, the lightweight world model can determine the robot's action with the highest association weight (using a robotic arm for support). Preferably, the lightweight world model can further confirm its confidence by analyzing its internal token entropy (entropy value 0.3, below the ambiguity threshold of 0.5).

[0113] By combining the confidence value in the structured semantic summary and the entropy value in the model analysis, the current ambiguous scenario can be determined to be at a medium confidence level, and the autonomous decision-making system can trigger a non-intrusive feedback verification path.

[0114] The autonomous decision-making system determines and executes non-invasive trial action instructions to verify user needs based on the execution actions and scenario parameters. The system generates non-invasive trial action instructions by combining scenario parameters (elderly person) and the execution actions (using a robotic arm for support). It adjusts the execution actions while considering the rationality of the response, the timing of intervention, and the degree of interference under the scenario parameters. In this second embodiment, the autonomous decision-making system can determine the safe distance and the safe force and speed of the robotic arm in the elderly scenario based on the rationality of the response, the timing of intervention, and the degree of interference. Then, while meeting the safe distance and the safe force and speed of the robotic arm in the elderly scenario, it adjusts the action parameters of the execution actions to generate non-invasive trial action instructions: the robot moves to 0.4 meters in front of and to the side of the elderly person (meeting the safe distance); the robot's robotic arm slowly rises from the side of the chassis at a speed of 0.1 m / s to the height of the elderly person's elbow (80 cm) and maintains a bent, ready-to-support posture (meeting the safe force and speed of the robotic arm). Preferably, the generation of these non-invasive trial action instructions can be further adjusted based on historical feedback data (e.g., user feedback in the previous three similar scenarios indicated acceptance of the trial).

[0115] The autonomous decision-making system can continuously monitor the user's natural reactions and determine the user feedback result based on the collected user feedback data. In this second embodiment, the autonomous decision-making system can collect and identify the elderly person's elbow moving towards the robotic arm and the force sensor detecting a slight contact signal, thereby determining that the user feedback result is acceptance.

[0116] Based on the accepted results, the autonomous decision-making system calibrates the confidence threshold, lowering the high-confidence threshold for the elderly scenario from 0.9 to 0.85. It also updates the association weight between user needs (the user needs to get up) and the action performed (using the robotic arm for support), increasing this weight from 0.7 to 0.77. Furthermore, it optimizes the non-invasive trial action command; for example, in the next trial in the same scenario, the robot's initial arm height can be directly set to 80cm, thus shortening the action preparation time. The autonomous decision-making system can then update the model based on the adjustments and record them in the interpretable decision log.

[0117] Furthermore, when the elderly person gets up from the sofa again, the autonomous decision-making system acquires environmental and interaction data and generates embodied semantic codes based on the active perception mechanism. After inputting the embodied semantic codes into the lightweight world model, due to the increased association weights, the confidence level of the lightweight world model's output is 0.86. When the confidence threshold is lowered to 0.85, the triage decision layer can directly determine it as high confidence, triggering the direct execution path. The robot does not need to probe and directly executes the action supported by the robotic arm, reducing the response latency to 80ms, verifying the effectiveness of the autonomous decision-making optimization.

[0118] Similarly, in the preferred embodiment three, the embodied intelligent robot provided by the present invention can be applied to provide assistance to children in a home setting.

[0119] A 5-year-old child is playing in the living room, repeatedly glancing at a box of building blocks (40cm high) on the coffee table, reaching out towards the table but not touching the objects. The autonomous decision-making system can achieve accurate autonomous decision-making in moderately confident scenarios through proactive perception and non-intrusive feedback.

[0120] Specifically, the autonomous decision-making system can achieve dynamic perspective optimization through proactive perception and autonomous movement, and collect embodied semantic codes representing "child's gaze focusing on the block box, reaching out (visual), and the distance between the child and the coffee table 0.6 meters (LiDAR)". The lightweight world model then outputs a structured semantic summary and an execution action. The structured semantic summary can be: "user action (gaze focusing, reaching out), object of user action (block box), risk of user action (none), confidence level (0.75)", and the associated execution action can be "delivering an item".

[0121] Since a confidence level of 0.75 falls within the medium confidence range of 0.6 to 0.9, the autonomous decision-making system can trigger a non-intrusive feedback verification pathway.

[0122] The autonomous decision-making system determines and executes non-invasive probing action instructions to verify user needs based on the execution actions and scene parameters. The system combines scene parameters (child) and the execution action (delivering an item) to generate non-invasive probing action instructions, adjusting the execution actions while considering the reasonableness of the response, the timing of intervention, and the degree of interference under the child's scene parameters. In this third embodiment, the autonomous decision-making system can determine the safe distance, safe force, and safe height under the child's scene based on the reasonableness of the response, the timing of intervention, and the degree of interference. Then, while adhering to the safe distance, safe force, and safe height under the child's scene, it adjusts the action parameters of the execution action to generate non-invasive probing action instructions: the robotic arm gently grasps the block box with a safe force and slowly delivers it to a position 0.2 meters in front of the child (height 18cm) based on the safe distance and safe height, at a speed of 0.05m / s.

[0123] The autonomous decision-making system can continuously monitor the user's natural reactions and determine the user feedback result based on the collected user feedback data. In this third embodiment, the autonomous decision-making system can collect and identify the child actively grasping the building block box, thereby determining that the user feedback result is acceptance.

[0124] Based on the accepted results, the autonomous decision-making system calibrated the confidence threshold, lowering the high-confidence threshold for the child scenario from 0.9 to 0.88, and updated the association weight between user needs (focused gaze and outstretched hand) and the executed action (delivering an item), increasing the association weight from 0.65 to 0.72. Furthermore, it optimized the non-invasive probing action instruction; for example, based on user feedback data, it fine-tuned the height of the action to 17cm to better fit the child's arm length.

[0125] When the same scenario reappears, the autonomous decision-making system acquires environmental and interaction data and generates embodied semantic codes based on an active perception mechanism. After inputting the embodied semantic codes into the lightweight world model, due to the increased association weights, the confidence level of the lightweight world model's output is 0.89. The triage decision layer can directly determine this as a high-confidence scenario, triggering the direct execution path. The robot can then directly execute the action of delivering the item without trial and error, reducing the response latency to 90ms.

[0126] As can be seen from Embodiments 2 and 3 above, the autonomous decision-making method for embodied intelligent robots provided by this invention can not only effectively handle ambiguous scenarios, but also be personalized according to user groups, making the autonomous decision-making capability adaptable to the user group. The lightweight world model achieves millisecond-level real-time prediction and structured data transmission, meeting the user's immediate needs. Compared with the existing VLM combined with LLM scheme, the latency of this invention can be reduced by 67%. Furthermore, feedback-driven autonomous decision optimization can improve the efficiency of decision evolution. After 3 to 5 interactions, the confidence assessment error of the autonomous decision-making method provided by this invention has been reduced to 8% to 12%, and the response time of non-intrusive actions has been shortened by 15% to 20%, achieving personalized adaptation and evolution for users. In addition, the autonomous decision-making method provided by this invention replaces voice confirmation or direct execution with a non-intrusive feedback mechanism, which can significantly reduce interference to users in ambiguous scenarios (for example, the interaction interference rate in the home scenario is reduced from 35% to below 14%), optimizing the user experience.

[0127] In summary, this invention utilizes an optimized lightweight world model as the triage decision layer, providing accurate confidence scores and structured semantic summaries, which form the basis for subsequent confidence-based triage. It employs confidence-based triage with integrated non-intrusive feedback as its core framework, achieving low-intervention decision-making through a three-level path and feedback loop. It leverages feedback-driven autonomous decision optimization to evolve autonomous decision-making capabilities, ensuring continuous adaptation to different scenarios. Finally, it utilizes interpretable logs to record time-series data, addressing the pain point of untraceable embodied intelligent decision-making.

[0128] This technology combination effectively breaks through the industry bottlenecks of embodied intelligent robots, such as "difficulty in handling medium confidence, disconnect between decision-making and feedback, and weak evolutionary ability," providing a feasible path for robots to upgrade from "passive response" to "active and autonomous decision-making intelligent agents." It can be widely applied in scenarios such as elderly care, childcare, and daily housework.

[0129] Although the methods described above are illustrated and depicted as a series of actions for the sake of simplicity, it should be understood and appreciated that these methods are not limited by the order of the actions, as some actions may occur in a different order and / or concurrently with other actions from the illustrations and descriptions herein or not illustrated and described herein but which may be understood by those skilled in the art, according to one or more embodiments.

[0130] Those skilled in the art will understand that information, signals, and data can be represented using any of a variety of different techniques and skills. For example, the data, instructions, commands, information, signals, bits, symbols, and chips described throughout the above description can be represented by voltage, current, electromagnetic waves, magnetic fields or magnetic particles, light fields or optical particles, or any combination thereof.

[0131] Those skilled in the art will further appreciate that the various illustrative logic blocks, modules, circuits, and algorithm steps described in conjunction with the embodiments disclosed herein can be implemented as electronic hardware, computer software, or a combination of both. To clearly illustrate this interchangeability between hardware and software, the various illustrative components, blocks, modules, circuits, and steps are described above in a generalized manner in terms of their functionality. Whether such functionality is implemented as hardware or software depends on the specific application and the design constraints imposed on the overall system. Those skilled in the art may implement the described functionality in different ways for each specific application, but such implementation decisions should not be construed as departing from the scope of the invention.

[0132] The various illustrative logic modules and circuits described in conjunction with the embodiments disclosed herein may be implemented or performed using a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. The general-purpose processor may be a microprocessor, but in alternatives, it may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors cooperating with a DSP core, or any other such configuration.

[0133] The prior description of this disclosure is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to this disclosure will be apparent to those skilled in the art, and the general principles defined herein may be applied to other variations without departing from the spirit or scope of this disclosure. Therefore, this disclosure is not intended to be limited to the examples and designs described herein, but should be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An autonomous decision-making method for embodied intelligent robots, characterized in that, Including the following steps: An active perception mechanism is used to collect environmental and interaction data in real time and transform the environmental and interaction data into embodied semantic codes that integrate scene and embodiment. The embodied semantic encoding is input into the triage decision layer to output a structured semantic summary and an execution action. The structured semantic summary includes scene parameters, user needs, and confidence levels. The confidence level is located in a dynamic confidence interval, which includes a low confidence interval, a medium confidence interval, and a high confidence interval. The triage decision layer is a lightweight world model optimized by RLHF. The pre-training process of the lightweight world model optimized by RLHF includes the following steps: determining a semantic sequence based on historical environment and interaction data. The semantic sequence includes user actions, perceptual states, and semantic events. The semantic sequence is input into the lightweight world model to output a structured semantic summary and an action. The reward signal of the lightweight world model covers the rationality of the action, the timing of intervention, and the degree of interference. The value function of the lightweight world model is optimized by RLHF. In response to the confidence level being within the medium confidence level range, a non-intrusive probing action instruction for verifying the user's needs is determined and executed based on the action to be performed and the scenario parameters; Collect user feedback data when the non-invasive probing action command is executed to determine the user feedback result; as well as The decision result is determined based on the user feedback, and the range of the dynamic confidence interval and the correlation weight between the execution action and the user demand in the triage decision layer are adjusted. The correlation weight affects the confidence level output by the triage decision layer.

2. The autonomous decision-making method for embodied intelligent robots as described in claim 1, characterized in that, The active perception mechanism includes dynamic perspective optimization, text context sampling, and multimodal coupled representation.

3. The autonomous decision-making method for embodied intelligent robots as described in claim 1, characterized in that, The step of determining and executing a non-intrusive probing action instruction for verifying the user's needs based on the execution action and the scene parameters includes: The action type and / or action parameters of the executed action are adjusted based on the rationality of the response under the aforementioned scenario parameters, the timing of intervention, and the degree of interference, in order to generate non-invasive probing action instructions; or Based on the rationality of the response, the timing of intervention, and the degree of interference under the scenario parameters, the action type and / or the range of preset action parameters of the execution action mapping are preset, and non-intrusive probing action instructions are generated based on the preset action type and / or the range of preset action parameters of the execution action mapping.

4. The autonomous decision-making method for embodied intelligent robots as described in claim 1, characterized in that, It also includes the following steps: In response to the confidence level being within the high confidence interval, the execution action is performed; Collect user feedback data during the execution of the action to determine the user feedback result; Based on the user feedback results, the range of the dynamic confidence interval and the correlation weight between the execution action and the user demand in the triage decision layer are adjusted; or In response to the confidence level being in the low confidence interval, a large language model is invoked to perform semantic reasoning on the structured semantic summary to obtain a decision instruction, and the decision instruction is executed. Collect user feedback data during the execution of the decision instruction to determine the user feedback result; Based on the user feedback results, the range of the dynamic confidence interval and the correlation weight between the execution action and the user demand in the triage decision layer are adjusted.

5. The autonomous decision-making method for embodied intelligent robots as described in claim 4, characterized in that, The user feedback results include acceptance, rejection, and no feedback. The step of adjusting the range of the dynamic confidence interval and the correlation weight between the execution action and the user need in the triage decision layer based on the user feedback results includes: In response to the user feedback result being accepted, the high confidence interval corresponding to the scenario parameter is lowered to increase the correlation weight between the execution action and the user's needs in the triage decision layer; In response to the user feedback result being a rejection, the medium confidence interval corresponding to the scenario parameter is increased, and the correlation weight between the execution action and the user's demand in the triage decision layer is reduced by a preset first threshold; or In response to the user feedback result being no feedback, a delay trial flag is added to the scenario parameters to reduce the correlation weight between the execution action and the user's needs in the triage decision layer by setting a second threshold, wherein the second threshold is less than the first threshold.

6. The autonomous decision-making method for an embodied intelligent robot as described in claim 1, characterized in that, It also includes the following steps: Record and store time-series interpretable decision logs, which include the input and output data of the triage decision layer, the confidence-based triage path selection results, the non-intrusive trial action instructions, and / or the user feedback results.

7. An autonomous decision-making system for embodied intelligent robots, characterized in that, include: Memory, on which computer instructions are stored; and A processor, connected to the memory, and configured to execute computer instructions stored in the memory to implement the autonomous decision-making method for an embodied intelligent robot as described in any one of claims 1 to 6.

8. A embodied intelligent robot, characterized in that, It includes a perception module, an execution module, and an autonomous decision-making system for an embodied intelligent robot as described in claim 7.

9. A computer-readable storage medium storing computer instructions thereon, characterized in that, When the computer instructions are executed by the processor, the autonomous decision-making method for an embodied intelligent robot as described in any one of claims 1 to 6 is implemented.

Images