A control system and method for a fruit fly neural bionic decision-making robot

Abstract

This invention relates to the field of neuro-bionic algorithm technology, specifically disclosing a lightweight bionic intelligent decision-making system and method based on the neural replication of a fruit fly brain. The system includes a fruit fly brain bionic decision-making module, a large-model inference unit based on DeepSeek, an instruction parsing module, and a hardware communication interface compatible with Moying mobile collaborative robots. The fruit fly brain bionic decision-making module simulates the neural behavioral logic of a fruit fly, the large-model inference unit outputs real-time decision instructions, and the instruction parsing module converts the AI ​​output into executable actions for the robot. This invention constructs a bionic AI decision-making kernel that simulates the biological neural logic of a fruit fly, performs lightweight fine-tuning and optimization based on a DeepSeek large model, and achieves low-latency, highly adaptive autonomous control of the robot. It can directly interface with the hardware execution units of mobile collaborative robots from brands such as Moying, filling the technological gap in seamless integration of large models and industrial-grade mobile collaborative robots.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of bionic neural networks, artificial intelligence, and robot control, specifically to a control system and method for a fruit fly-based neuronic decision-making robot. Background Technology

[0002] The following pain points are prevalent in the current fields of artificial intelligence and robot control: 1. Large models have a large number of parameters, high computing power requirements, and high inference latency, making them difficult to run in real time on edge embedded devices; 2. Traditional bionic neural networks mostly focus on the motor control of the mammalian brain or cerebellum, which are complex in structure, have high training costs, and are difficult to implement; 3. Industrial mobile robots, AGVs, and collaborative robots lack the ability to autonomously avoid obstacles, make task priority decisions, and adapt to the environment in unstructured environments, relying on preset paths and strong human intervention.

[0003] As a classic model organism, the fruit fly possesses a small number of neurons in its brain, a clear structure, and complete behavioral logic. It exhibits strong environmental perception, risk aversion, autonomous navigation, and decision-making abilities, and is naturally lightweight and low-power. Currently, there is no mature solution that can completely replicate the core neural pathways of the fruit fly brain and directly integrate it with industrial-grade mobile collaborative robot hardware to achieve real-time edge-side decision-making, indicating a significant technological gap. Summary of the Invention

[0004] The purpose of this invention is to propose a lightweight biomimetic intelligent decision-making system based on the neural replication of the fruit fly brain. By highly replicating the complete biological neural pathway of the fruit fly brain—perception, processing, decision-making, and execution—it achieves autonomous intelligent decision-making with low computing power, low latency, and high robustness at the edge, solving the problems of bulky, high-latency, and poor environmental adaptability of existing robot decision-making systems.

[0005] The objective of this invention can be achieved through the following technical solutions: A lightweight biomimetic intelligent decision-making system based on the replication of the fruit fly brain's neural pathways includes: The Drosophila neural bionic core module replicates the core neural structure of the Drosophila brain, including olfactory sensory neurons, projection neurons, Kenyon cells, mushroom body output neurons, and descending neurons, forming a closed-loop decision-making link; it uses a spiking neural network (SNN) to achieve neuron-level signal transmission and plasticity regulation; The multimodal lightweight sensing access module accesses data from sensors such as vision, infrared, ultrasound, and odometer, performs low-computation feature extraction, and directly maps it into stimulus signals that can be recognized by the fruit fly's neural kernel. The behavior decision-making and task scheduling module, based on the biological behavior logic of fruit flies—seeking benefits and avoiding harm, being goal-oriented, and prioritizing decisions—outputs motion control commands, task switching commands, and obstacle avoidance commands. The robot hardware adapter interface module can be directly connected to mobile collaborative robots, AGVs, and robotic arm actuators, and supports industrial communication protocols such as CAN, RS485, and Ethernet to achieve real-time control at the end.

[0006] Preferably, the large model adopts LoRA lightweight fine-tuning and 4-bit quantization deployment, which can complete real-time inference in an edge computing unit or free GPU environment.

[0007] Preferably, the hardware communication interface is compatible with the MCR mobile collaborative robot standard control protocol and can be directly connected to the Moying robot chassis and robotic arm execution unit.

[0008] This application also proposes a lightweight biomimetic intelligent decision-making method based on the neural replication of a fruit fly brain, comprising the following steps: collecting data on the robot's surrounding environment and task → inputting the data into the fruit fly brain-inspired Prompt engine → completing inference and decision-making through the DeepSeek large model → parsing and verifying the decision results → issuing control commands to the Moying robot hardware for execution → collecting the execution status in real time and feeding it back to the AI ​​model for iterative optimization.

[0009] The beneficial effects of this invention compared to the prior art are as follows: - Extremely lightweight: No high-end GPU required, can run in real time on the embedded MCU / ARM side; - Ultra-low latency: Decision cycle is much shorter than traditional large models and deep neural networks; - Strong environmental adaptability: Inheriting the fruit fly's natural ability to seek advantages and avoid disadvantages, dynamically avoid obstacles, and navigate autonomously; - Easy to industrialize: It can be directly connected to existing industrial robot, AGV, and service robot hardware platforms; - Iterative upgradeable: It supports continuous fine-tuning and optimization while preserving the basic neural structure, forming a technological barrier. Detailed Implementation

[0010] The technical solutions in the embodiments of the present invention have been clearly and completely described. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0011] A lightweight biomimetic intelligent decision-making system based on the replication of the fruit fly brain's neural pathways includes: 1) The Drosophila neuromorphic core module replicates the core neural structure of the Drosophila brain: olfactory sensory neurons, projection neurons, Kenyon cells, mushroom body output neurons, and descending neurons, forming a closed-loop decision-making link; a spiking neural network (SNN) is used to realize neuron-level signal transmission and plasticity regulation.

[0012] 2) Multimodal lightweight sensing access module By accessing data from sensors such as vision, infrared, ultrasound, and odometer, low-computing-power feature extraction is performed and directly mapped to stimulus signals that can be recognized by the fruit fly's neural core.

[0013] 3) Behavioral Decision-Making and Task Scheduling Module Based on the biological behavioral logic of fruit flies—which seeks advantage and avoids harm, is goal-oriented, and prioritizes decision-making—it outputs motion control commands, task switching commands, and obstacle avoidance commands.

[0014] 4) Robot hardware adapter interface module It can be directly connected to mobile collaborative robots, AGVs, and robotic arm actuators, and supports industrial communication protocols such as CAN, RS485, and Ethernet to achieve real-time control at the end.

[0015] A fruit fly neural biomimetic decision-making model was constructed to fully simulate the biological neural pathways of fruit flies, including perception, pursuit of benefits, obstacle avoidance, decision-making, and execution, thereby enabling robot environmental adaptive behavior planning and task priority judgment. The DeepSeek large model is fine-tuned using LoRA low-rank adaptation and 4-bit quantization technology, enabling real-time inference on free GPUs and edge computing units with end-to-end control latency ≤300ms. Design a large-scale model and a standardized instruction interface for robot hardware to automatically convert AI decision outputs into executable control protocols for the robot chassis, robotic arm, and sensors; A fully closed-loop real-time control system was built, encompassing "sensing and data acquisition - AI inference - command issuance - hardware execution - status feedback - iterative optimization," to improve the stability of the robot's autonomous operation.

[0016] System overall structure; Perception layer: Equipped with robot vision cameras, LiDAR, and environmental sensors, it completes real-time data collection of the surrounding environment and tasks; The fruit fly brain AI layer integrates the DeepSeek lightweight fine-tuning model with the fruit fly neural bionic decision-making logic, serving as the core decision-making unit; Instruction translation layer: Enables the parsing, translation, and compliance verification of AI decision outputs and robot control protocols; Execution layer: Interacts with the Moying mobile collaborative robot chassis, robotic arm, and drive execution unit to implement commands; Feedback layer: Collects the robot's execution status in real time and sends it back to the AI ​​layer for secondary iterative decision-making.

[0017] Example 1: Autonomous Obstacle Avoidance and Navigation of Mobile Collaborative Robots The system acquires obstacle information through visual and ultrasonic sensors, processes it through a fruit fly neural bionic kernel, and outputs real-time obstacle avoidance and path adjustment commands, enabling the robot to autonomously navigate to target points in an unstructured environment.

[0018] Example 2: Multi-task priority decision The system receives multi-task instructions, and automatically sorts task priorities according to preset goals and environmental constraints, following the fruit fly's behavioral decision-making logic, to achieve autonomous task scheduling and switching.

[0019] The operation of a lightweight biomimetic intelligent decision-making system V1.0 based on the replication of the fruit fly brain's neural network is presented as follows: 1. Basic Software Information - Full software name: Fruit Fly Brain AI Robot Decision Control System V1.0 - Development language: Python - Operating environment: Linux / Windows operating system, CUDA acceleration environment, Google Colab free GPU, Python 3.8 or above. - Core Application: Neuro-bionic autonomous decision-making and real-time closed-loop control for mobile collaborative robots 2. Software Core Architecture and Innovation Description This system uses the fruit fly biological neural decision-making link as a biomimetic prototype and combines DeepSeek large model lightweight inference technology to build a fully autonomous control architecture of "environmental perception - biomimetic decision-making - instruction conversion - hardware execution - state feedback". It is different from traditional fixed logic robot control systems and has original technical features such as strong environmental adaptation, low latency inference and lightweight deployment on the edge. The overall technical solution and functional implementation are independently designed and do not copy the architecture of existing similar systems.

[0020] 3. Operating Environment Requirements 3.1 Hardware Environment Basic operation: General-purpose computer CPUs are sufficient for startup and debugging; High-performance mode: Supports NVIDIA discrete graphics cards (≥8GB VRAM) with CUDA acceleration enabled; Deployment environment: Compatible with edge computing units and Google Colab free GPU cloud computing power.

[0021] 3.2 Software Environment Operating System: Windows 10 / 11, Ubuntu 20.04 and above; Runtime Dependencies: Python 3.8+, PyTorch 2.0+, quantization inference related libraries; Communication Compatibility: Supports standard communication protocol integration with mobile collaborative robot hardware.

[0022] 4. Installation and Deployment Process 4.1. Basic Environment Configuration: Install the specified version of Python environment, complete the installation of third-party libraries through a dedicated dependency configuration file. All installation scripts and configuration items are independently written, without using any existing project's complete configuration copy.

[0023] 4.2. Model Loading and Initialization: The inference model is loaded using 4-bit quantization and LoRA lightweight loading. The model call interface and parameter configuration are self-encapsulated. Only the public API is used for application layer development. The model body and core training code are not copied.

[0024] 4.3. Hardware Communication Configuration: The robot's communication address, port number, and protocol format are set independently in the system configuration file to achieve point-to-point connection with the execution hardware. The communication parsing logic is an original design.

[0025] 4.4. System Startup Verification: Run the main program entry file, complete environment self-check, model loading status verification, and hardware connection status verification, and enter the running state after ensuring that the entire link is normal.

[0026] 5. Function Operation Instructions 5.1 Operation of the Bionic Decision Engine After the system starts up, it automatically enters the fruit fly neural bionic decision-making mode, receives sensor and environmental data in real time, and completes task priority judgment, obstacle avoidance planning and path decision-making according to independently designed bionic behavior rules. All decision-making logic and rules are original settings.

[0027] 5.2 Instruction Parsing and Execution The AI ​​inference output is processed by the original instruction conversion module, automatically adapted to the robot hardware execution protocol, and sent to execution units such as the chassis and robotic arm to achieve precise implementation of decisions into actions.

[0028] 5.3 Real-time closed-loop feedback The system continuously collects data on hardware execution status and environmental changes, and sends it back to the decision-making level for dynamic iterative optimization, forming an autonomous closed-loop control. The feedback mechanism and data processing logic are designed independently.

[0029] 5.4 Customizing runtime parameters It supports users to independently configure core parameters such as decision sensitivity, execution speed, and safety distance. The parameter interaction interface and configuration logic are all original implementations.

[0030] 6. Troubleshooting and Maintenance - Model loading error: Check the model path configuration and the integrity of the dependency environment, and reinitialize as prompted by the system. - Hardware communication error: Check communication configuration parameters and restart the communication service and hardware connection. - Slow performance: Release system resources and switch to lightweight inference mode. All exception detection rules and handling solutions are independently designed and do not copy the processes of similar software.

[0031] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A lightweight bionic intelligent decision-making system based on fruit fly brain neural replication, comprising: a fruit fly neural bionic kernel module replicating the core neural structure of the fruit fly brain, including olfactory sensory neurons, projection neurons, Kenyon cells, mushroom body output neurons, and descending neurons, forming a closed-loop decision-making link; using a spiking neural network (SNN) to realize signal transmission and plasticity adjustment at the neuron level; a multi-modal lightweight perception access module that accesses visual, infrared, ultrasonic, odometer, and other sensor data, performs low-power feature extraction, and directly maps to stimulus signals recognizable by the fruit fly neural kernel; a behavior decision and task scheduling module based on the biological behavior logic of fruit fly seeking benefits and avoiding harm, goal-oriented, and priority decision-making, outputting motion control instructions, task switching instructions, and obstacle avoidance instructions; a robot hardware adaptation interface module that directly connects with mobile collaborative robots, AGVs, and mechanical arm actuators, supports CAN, RS485, Ethernet, and other industrial communication protocols, and realizes real-time control on the end side.

2. The lightweight bionic intelligent decision-making system based on fly brain neural reenactment according to claim 1, characterized in that, The large model uses LoRA lightweight fine-tuning and 4-bit quantization deployment, which can complete real-time inference in edge computing units or free GPU environments.

3. The lightweight bionic intelligent decision-making system based on fly brain neural reenactment according to claim 2, characterized in that, The hardware communication interface adapts to the MCR mobile collaborative robot standard control protocol, which can directly interface with the Moxi robot chassis and mechanical arm execution unit.

4. A lightweight bionic intelligent decision-making method based on neural reconstruction of a fruit fly brain, characterized in that: The steps include: Collecting robot peripheral environment and task data→ inputting data into the fruit fly brain bionic Prompt engine→ completing inference decision-making through the DeepSeek large model→ analyzing and verifying the compliance of the decision-making results→ issuing control instructions to the Moxi robot hardware for execution→ real-time collection of execution status and feedback to the AI model for iterative optimization.