Community service robot intelligent decision-making system based on large model and behavior tree

Abstract

The invention belongs to the technical field of service robots and artificial intelligence, discloses a community service robot intelligent decision-making system based on a large model and a behavior tree, and solves the problems that a traditional community service robot is weak in complex instruction understanding, rigid in decision-making logic, unreliable in execution and free of autonomous learning ability. The system comprises an instruction input module, an intention analysis and task arrangement module, a behavior tree construction module, a behavior execution module, an autonomous learning module and a knowledge base module, a complex instruction is collaboratively disassembled into a structured sub-action sequence through a cloud edge end model, a behavior tree is automatically constructed and executed based on an ROS system, autonomous learning is achieved by relying on a double-agent evaluation mechanism, and the knowledge base module is established. And the knowledge base stores community scenes and robot capability data. According to the invention, the intelligence and service reliability of the robot in a smart community are significantly improved, and the method is suitable for scenes such as reception guidance, article delivery and safety inspection.

Description

Technical Field

[0001] This invention relates to the fields of service robots and artificial intelligence decision-making technology, specifically to an intelligent decision-making system for community service robots that integrates Large Language Model (LLM) and Behavior Tree (BT), applicable to scenarios in smart communities that require understanding and reliably executing complex instructions, such as reception and guidance, item delivery, and security inspection. Background Technology

[0002] With the advancement of smart community construction, service robots have become an important tool for improving community management efficiency and residents' quality of life. However, the existing intelligent decision-making systems of community service robots have significant limitations in practical applications: First, the system has a weak ability to parse complex instructions. Traditional systems often rely on predefined, simple instruction sets and cannot understand complex task instructions given by users in natural language, which contain multiple steps and ambiguous semantics. This results in insufficient intelligence in robots and a poor user experience.

[0003] Secondly, the decision-making logic is rigid and lacks scalability. Most systems use decision-making logic based on fixed rules or finite state machines, requiring developers to manually write code for each new task. When community service requirements change or new task types are added, the system struggles to adapt quickly, resulting in high maintenance and update costs.

[0004] Secondly, task execution reliability is insufficient. Complex tasks typically consist of multiple dependent sub-actions. Existing systems lack standardized descriptions and unified scheduling of execution logic, making them prone to task interruption or failure due to incorrect action sequence, chaotic state management, or lack of exception handling.

[0005] Finally, it lacks self-learning and optimization capabilities. The system cannot learn from historical execution experience, and its decision-making capabilities are static and rigid. Faced with constantly emerging new scenarios and instructions within the community, it relies entirely on manual upgrades, cannot achieve autonomous evolution, and struggles to meet the dynamic development needs of smart communities.

[0006] Therefore, there is an urgent need for an intelligent decision-making system for community service robots that can understand complex instructions, has flexible and reliable execution logic, and can continuously optimize through autonomous learning. Summary of the Invention

[0007] To address the problems of traditional community service robot systems in understanding complex instructions, decision-making logic flexibility, task execution reliability, and autonomous learning ability, this invention proposes an intelligent decision-making system for community service robots based on a large model and behavior tree.

[0008] The technical solution of the present invention includes: (1) Command input module, used to receive and preprocess service commands input by users via voice or text; (2) The intent parsing and task orchestration module is connected to the instruction input module. It is used to determine the instruction type. For control instructions, it calls the large language model and combines the knowledge base and example library to parse and orchestrate the user instructions into a structured sequence of sub-actions. (3) Behavior tree construction module, connected to intent parsing and task orchestration module, used to automatically construct the sub-action sequence into a behavior tree composed of control nodes, execution nodes and decoration nodes; (4) Behavior execution module, which is connected to behavior tree construction module, schedules and executes the behavior tree based on robot operating system, and controls robot entity to complete corresponding actions; (5) Self-learning module, which is connected to the intent parsing and task orchestration module and behavior execution module, is used to collect task execution samples and automatically expand the effective samples to the example library through the dual agent evaluation mechanism; (6) Knowledge base module, connected to intent parsing and task orchestration module, is used to store community scene environment information and robot hardware capability information.

[0009] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By integrating large language models and behavior trees, the system can accurately understand the complex, multi-step instructions expressed in natural language and generate structured, reliable execution logic.

[0010] 2. By using behavior trees to visualize and structure the task process, the success rate and robustness of complex task execution are significantly improved through clear node logic, state management and built-in retry mechanism.

[0011] 3. Through the dual-agent evaluation mechanism of the self-learning module, the system can automatically learn from successful execution experiences and optimize the example library, enabling decision-making capabilities to continuously adapt to the dynamic changes in community services. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the principle of the present invention; Figure 2 This is a flowchart illustrating the retrieval process in this invention. Figure 3 A schematic diagram of the behavior tree structure constructed for a specific task in this invention; Figure 4 This is a flowchart of the self-learning module in this invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0014] like Figure 1 As shown, the workflow of the community service robot intelligent decision-making system based on large models and behavior trees includes the following steps: (1) Users input service commands via voice or text, and the system classifies the intent (chat, policy consultation, robot control) using a large language model fine-tuned at the edge.

[0015] (2) If the intention is identified as robot control, the cloud-based large language model is invoked, and the natural language instruction is parsed into a structured sequence of sub-actions (JSON format) by combining the behavior library and the example library.

[0016] (3) The system automatically parses the JSON action sequence and builds it into a behavior tree. The behavior tree executor schedules the APIs of each module of the robot to complete the action execution.

[0017] (4) The system automatically filters high-quality execution samples through a dual-agent evaluation mechanism (diversity + correctness), expands the example library, and achieves continuous learning.

[0018] (5) Throughout the process, the system integrates multimodal information such as voice, facial expressions, and sensor data to achieve humanized interaction and real-time status monitoring.

[0019] The detailed steps are as follows: Step (1): Build a sample library for community service robot task planning.

[0020] This invention constructs an example library to store standard examples of robot motion choreography. The example library uses JSON format for storage, facilitating data retrieval and expansion. The core function of the example library is to provide reference examples for large language models, serving as input for few-shot learning, thereby improving the adaptability of large language models to diverse tasks. Simultaneously, it standardizes the output format of large language models, ensuring that the generated JSON data conforms to the structural requirements of behavior trees, facilitating subsequent parsing and generation of executable behavior trees. Each example includes the user-input natural language command, the reasoning process of the large language model, and the final generated JSON format action sequence. Each action node in the JSON data includes a node name and type, along with input parameter information. For example, the input parameter for the `navigation` action node is "balcony," while the `navigation_xyzw` action node, which navigates to a certain coordinate, contains four input parameters, each corresponding to the coordinate value of the target position. The value of the `x` parameter is `${GetCurrentPosition / output / x}`, indicating that this value comes from the output of the `GetCurrentPosition` action node, rather than a specific numerical value.

[0021] Step (2): Similar example retrieval Since the example library contains a large number of examples, directly inputting all examples into a large language model would not only lead to redundancy but also significantly increase inference time. Therefore, an efficient similar example retrieval algorithm is needed to select few examples similar to the current user input from the example library, improving the quality of large model generation while maintaining inference efficiency. This invention designs a similar example retrieval method based on a text vector database; the specific process is as follows: Figure 2 First, the text embedding model client calls the cloud-based text embedding model to convert all user input text in the example library into vectors and store them in the FAISS vector database. Then, the same vectorization process is applied to the user's current input command. The FAISS database is retrieved, and the similarity coefficient between the current user input text vector and the vectors in the example vector database is calculated to obtain the Top-K similar user inputs. Finally, the corresponding examples are retrieved from the example library to obtain the Few-shot examples.

[0022] Step (3): Robot task execution and dynamic adjustment module based on behavior tree.

[0023] The behavior tree construction and execution module transforms structured action sequences into executable behavior trees, enabling adaptive execution across diverse task scenarios. The module's structure is as follows: Figure 3 As shown, it consists of two parts: a behavior tree parser and a behavior tree execution engine. The behavior tree parser takes a sequence of actions in JSON format as input and constructs a mapping from task description to behavior tree nodes; the behavior tree execution engine schedules the APIs of various robot modules based on the tree structure generated by the parser, and monitors and dynamically adjusts task execution.

[0024] Specifically, the behavior tree parser uses a recursive descent algorithm to convert JSON action sequences into a behavior tree composed of control nodes (Sequence, Selector, Parallel), execution nodes (Action), and decorator nodes (Retry, Inverter, etc.). For example, for an action sequence containing "navigate to point A", "play voice message", and "return to original position", the parser will generate a behavior tree with Sequence as the root node and three Actions as child nodes.

[0025] To ensure more stable and reliable task execution, the system introduces a behavior tree blackboard as a globally shared data area, enabling parameter passing and state sharing between nodes. The mechanism is as follows: during execution, a node can write its output parameters to the blackboard, and subsequent nodes can read these parameters from the blackboard as input. For example, the "Get Current Position" node writes coordinates to the blackboard, and the subsequent "Return to Original Position" node reads these coordinates from the blackboard as the navigation target.

[0026] The behavior tree execution engine operates using a multi-process architecture, ensuring that the execution of the behavior tree does not block other tasks (such as listening for new instructions). The engine has a built-in state monitoring and node retry mechanism: if an action node fails to execute, the system will attempt to recover according to a preset strategy (such as retrying 3 times); if it ultimately fails, the task will be terminated and the user will be notified.

[0027] Step (4): Use a cloud-edge-device collaborative robot intelligent decision-making system to achieve closed-loop execution of complex tasks.

[0028] like Figure 4 As shown, this invention is a typical cloud-edge-device collaborative intelligent decision-making system to ensure efficient understanding, planning, and execution of complex tasks. A lightweight edge model is used as the backbone unit for real-time intent perception, while user command features generated in the early stages serve as query input for the large cloud model's planning. By constructing a deep reasoning mechanism of "command semantics - planning logic," precise alignment between user intent and robot capabilities is achieved, ensuring highly executable planning results. The robot, based on the structured planning results delivered from the cloud, utilizes a behavior tree execution engine to reliably execute the task. This system uses real-time perception data and multimodal interaction information as input conditions. Through dynamic task decomposition and adaptive adjustment of the behavior tree, it drives the robot to execute customized services for specific scenarios and task requirements, achieving closed-loop optimization of decision-making and execution guided by scenario perception.

[0029] Step (5): Comparative experiment and result analysis.

[0030] To verify the effectiveness of the method of this invention, we designed a comparative experiment to compare the performance of the behavior tree planning method based on Large Language Model (LLM) with that of traditional task planning methods. The experimental results are shown in Table 1, and are analyzed from the aspects of task completion rate and response time.

[0031] Regarding task completion rate, traditional methods perform stably with a 100% completion rate for simple tasks, but fail to complete complex tasks (achieving a 0% completion rate). The method proposed in this invention, however, also achieves a 100% completion rate for simple tasks, and even with 3-shot and 5-shot settings, the completion rate for complex tasks remains at 100%, significantly outperforming traditional methods. This demonstrates that LLM-based behavior tree planning possesses strong task parsing and action orchestration capabilities.

[0032] Regarding response time, traditional methods have a low average latency (approximately 0.5 seconds) due to their reliance on predefined rules. The method of this invention involves cloud-based LLM inference and dynamic generation of behavior trees, resulting in increased latency (approximately 5.3 seconds for complex tasks under 3-shot conditions), but it remains within an acceptable range and can meet the needs of daily applications.

[0033] Furthermore, this invention exhibits better generalization ability when facing new tasks, and can quickly adapt through task decomposition and dynamic generation of behavior trees, while traditional methods require redefining rules and have weaker scalability. Experimental results show that the LLM-based behavior tree planning method has significant advantages in task completion rate and generalization ability, providing an effective intelligent solution for robot motion planning.

[0034] Table 1 Comparison Experiment of Motion Planning Execution

[0035] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A community service robot intelligent decision-making system based on a large model and behavior tree, characterized by: (1) Instruction input module, used to receive and preprocess voice or text service instructions input by the user; (2) The intent parsing and task orchestration module is connected to the instruction input module. It is used to determine the instruction type and, when it is confirmed to be a control intent, to call the large language model and combine it with the knowledge base and example library to parse and decompose the complex instruction into a structured sub-action sequence. (3) Behavior tree construction module, connected to the intent parsing and task orchestration module, used to parse the sub-action sequence and automatically construct a behavior tree consisting of control nodes, execution nodes and decoration nodes; (4) Behavior execution module, connected to the behavior tree construction module, used to schedule and execute the behavior tree, thereby controlling the robot to complete the corresponding actions; (5) Self-learning module, connected to the intent parsing and task orchestration module and behavior execution module, is used to collect task execution samples and automatically expand the effective samples to the example library through an evaluation mechanism; (6) Knowledge base module, connected to the intent parsing and task orchestration module, used to store community scene environment information and robot hardware capability information.

2. The intelligent decision-making system for community service robots based on large models and behavior trees according to claim 1, characterized in that, The intent parsing and task orchestration module includes: an edge-based lightweight model unit for quickly classifying input commands into intents; a cloud-based large language model unit for deep parsing and task planning of control commands; a behavior library for storing basic actions that the robot can execute and their parameter interfaces; an example library for storing the correspondence between verified and correct commands and sub-action sequences; and a similar example retrieval unit for matching historical similar examples for the current command through text vectorization and vector database retrieval.

3. The intelligent decision-making system for community service robots based on large models and behavior trees according to claim 1, characterized in that, The behavior tree construction module uses a recursive descent parsing algorithm to convert JSON-formatted sub-action sequences into behavior trees, and uses the "behavior tree blackboard" as a globally shared data area to realize parameter passing and state sharing between nodes.

4. The intelligent decision-making system for community service robots based on large models and behavior trees according to claim 1, characterized in that, The behavior execution module uses a multi-process architecture to run the behavior tree executor and includes execution status monitoring and node retry mechanisms.

5. The intelligent decision-making system for community service robots based on large models and behavior trees according to claim 1, characterized in that, The autonomous learning module includes: a behavior recorder, used to record user commands, generated action sequences and execution results; a sample diversity assessment agent, which judges the uniqueness of new samples based on vector similarity; and a sample correctness assessment agent, which calls multiple large language models to evaluate the action sequence logic of new samples.

6. The intelligent decision-making system for community service robots based on large models and behavior trees according to claim 1, characterized in that, The knowledge base module stores information including the location of key points on the community indoor map, access control device IDs, channel attributes, and performance parameters of robot motion and interaction devices.

Images