Closed-loop training device and method for controlling dolly to avoid obstacles by using in-vitro neural chip

Abstract

The invention discloses a closed-loop training device and a closed-loop training method for an in-vitro neural chip to control obstacle avoidance of a trolley, and belongs to the field of neurobiology and in-vitro brain-computer interfaces, and the closed-loop training device shapes a biological neural network structure and regulates and controls a discharge mode in a closed-loop training mode, so that the device makes a correct decision in different obstacle situations to control obstacle avoidance of the trolley. According to the device, by means of sensory electrical stimulation coding, the neuron culture can'sense 'the obstacle condition faced by the trolley in the environment, and then motion control over the external trolley is achieved by collecting and decoding the discharge signals of the neuron. According to the method, training electrical stimulation is adopted to provide feedback for a neural network, a system applies reward or punishment electrical stimulation to the neural network according to whether a trolley successfully avoids an obstacle or not, a synaptic structure and a discharge mode of the neural network are adaptively shaped and optimized, and the safety of the trolley is improved. And a network structure capable of autonomously controlling the trolley to avoid obstacles in different obstacle environments is gradually formed.

Description

Technical Field

[0001] The present disclosure relates to the fields of neurobiology and in vitro brain-computer interface, and in particular to a closed-loop training device and method for controlling a car to avoid obstacles using an in vitro neural chip. Background Art

[0002] In recent years, artificial intelligence technology has made significant progress in the past few years, but it still faces some bottlenecks and challenges. In terms of data quality and quantity, the simulated neural network in artificial intelligence requires a large amount of data set for pre-training, but in reality it is often difficult to obtain sufficient quantity and high-quality data. In addition, due to the imbalance and bias of the data, the trained model may have certain errors; in terms of algorithm efficiency and energy consumption, artificial intelligence algorithms require a lot of computing resources for training and reasoning, which not only increases the cost of computing, but also may lead to energy waste.

[0003] At present, with the rapid development of brain-computer interface technology, the neural network prototype in artificial intelligence - the neural network in the biological system has gradually become a research hotspot. The biological neural network is composed of neurons and synapses, has the biological characteristics of real neurons, has a high degree of complexity and plasticity, and can realize complex functions such as perception, processing, and decision-making. Compared with artificial intelligence, the biological neural network has more network nodes and synaptic connections, and its complex network structure makes its information processing ability more advantageous. The biological neural network adopts a distributed information processing method, which greatly improves the efficiency of task processing. The frequency of discharge activities between neurons is lower, which makes it have lower energy consumption. In addition, the biological neural network has the ability of self-organization and evolution, which enables it to process tasks more efficiently.

[0004] Microelectrode array technology provides a method for efficiently recording and stimulating biological neuronal activity and is widely used to study the function and behavior of neural networks. Although biological neuronal cultures and MEA technology have made significant progress in research, there are still many challenges in applying them to actual intelligent systems. Most of the current intelligent systems rely on silicon-based neural networks or traditional computer algorithms, which have limitations in dealing with real-time decision-making and adaptive learning in complex environments. In order to achieve true biological intelligence, it is necessary to effectively combine biological neuronal cultures with devices in dynamic environments (such as mobile robots) to establish mechanisms for real-time interaction and adaptive learning. Such a system will not only promote the study of biological neuroscience, but also provide a new path for the development of intelligent robots and adaptive systems.

[0005] Based on the closed-loop training method of synthetic biological intelligence, some studies have successfully enabled biological neural networks to learn to play the classic arcade game "Pong" in a simulated environment. This study used "unpredictable stimulation" and "predictable stimulation" based on the free energy principle (FEP) as feedback methods. The feedback training mode adopted by this method is relatively simple. When facing complex tasks that require more targeted shaping effects on neural networks, training based solely on the principle of free energy minimization is not an efficient method. In addition, the neuronal discharge decoding algorithm used in this study decodes by counting and comparing the number of neuronal discharges in different partitions, but this decoding method has significant limitations when the distribution of neuronal cultures in microelectrode arrays (MEAs) is uneven, thereby limiting the self-organizing ability of the neuronal network. In response to the problems existing in the above studies, the present invention optimizes the design and improvement of the system. The system introduces the function of customizing electrical stimulation sequences, which can flexibly set the frequency, amplitude and location of stimulation, providing convenience for exploring and verifying more accurate "reward" and "punishment" feedback methods. In addition, the system uses the activity center method for decoding, and takes the difference between the CA vector in the training process and the initial state as the decoding basis, which effectively solves the problem of poor training effect caused by the uneven distribution of neurons in MEA.

[0006] Based on the above research, the present invention aims to build a closed-loop training system that shapes biological neural networks to control the car to perform corresponding obstacle avoidance actions in different obstacle environments. The system integrates a complete experimental paradigm, and the adaptive training method and integration of the functions of each module significantly improve the efficiency of training, making the development of synthetic biological intelligence systems more systematic, and laying the foundation for future research and application of biological intelligence. Summary of the invention

[0007] The problem to be solved by the present invention is to build a complete biological neural network closed-loop training system based on the goal of controlling the car to avoid obstacles using a biological neural network. The system includes a complete experimental paradigm, process, and specific relevant parameters in the experiment, and links the biological system with the artificial components, making the training process highly automated and systematic.

[0008] The technical solution adopted by the present invention is:

[0009] A closed-loop training device for controlling a car to avoid obstacles using an in vitro neural chip, including a biological neural network, a computer, an intelligent car, an obstacle simulation device, and an electric stimulator;

[0010] The biological neuron network comprises neuron cells and a microelectrode array, wherein the neuron cells are seeded on the microelectrode array;

[0011] The computer has a built-in main control program, which executes:

[0012] (1) transmitting obstacle position information parameters to the obstacle simulation device, thereby setting obstacles in the simulation environment;

[0013] (2) Acquiring and filtering the neuronal electrical signals of the microelectrode array, displaying the processed waveform, and extracting the number of spikes recorded within the detection time from the signal;

[0014] (3) Process the signal and send motion control instructions to the smart car;

[0015] (4) Send feedback stimulation instructions to the electrical stimulator according to the movement of the smart car;

[0016] The obstacle simulation device includes obstacles arranged around the smart car, and the obstacles are divided into two states: open and closed;

[0017] The smart car is provided with an infrared sensor and a Bluetooth module, wherein the infrared sensor is used to detect the position information of obstacles in the environment, and the Bluetooth module is used to transmit the position information of obstacles to the main control program of the computer;

[0018] The electrical stimulator applies sensory electrical stimulation to the biological neuron network according to the feedback stimulation instruction.

[0019] In the above technical solution, the microelectrode array is prepared by a micro-electromechanical system process, the number of effective sites of the microelectrode array ranges from 64 to 288 channels, and the sites are modified by nanomaterials.

[0020] In the above technical solution, obstacles are set in the left, front and right directions of the smart car.

[0021] The closed-loop training method for controlling a car to avoid obstacles using an in vitro neural chip includes the following steps:

[0022] Step 1, environment initialization: set the obstacle positions in the environment where the smart car is located, and obtain the target motion instructions of the smart car in this round of training;

[0023] Step 2, environmental information encoding: The location information of obstacles in the environment is encoded through sensory electrical stimulation and transmitted to neuronal cells;

[0024] Step 3, motion decision: by processing and decoding the collected data, establish the mapping relationship between electrophysiological activity and actual motion of the smart car;

[0025] Step 4, shaping the neural network: according to the movement state of the smart car in the environment, punishment / reward electrical stimulation is applied to the neuron cells to shape the synaptic structure between the neuron cells and the neuron discharge pattern in different regions.

[0026] In the above technical solution, in step 1, obstacles are placed at two positions in the left, front and right directions of the smart car. The smart car uses infrared sensors to detect obstacles and obtain target motion instructions in the scene.

[0027] In the above technical scheme, in step 2, the obstacle position information obtained in step 1 is converted into a sensory electrical stimulation instruction. There are three cases of obstacle position information, namely: obstacles on the left and right sides; obstacles on the right and front; obstacles on the left and front. Three sensory areas are divided on the neuronal cells, namely site area 1, site area 2 and site area 3. When the obstacle position information is the first case, the sensory electrical stimulation is applied to the stimulation site of site area 1. When the obstacle position information is the second case, the sensory electrical stimulation is applied to the stimulation site of site area 2. When the obstacle position information is the third case, the sensory electrical stimulation is applied to the stimulation site of site area 3.

[0028] In the above technical solution, the sensory electrical stimulation method is to use a negative leading two-phase rectangular voltage pulse sequence with an amplitude of 300mV, a pulse length of 200μs and a frequency of 1Hz.

[0029] In the above technical solution, in step 3, after the obstacle position information is transmitted to the neuron culture through sensory stimulation, the discharge pattern of the neuron cluster needs to be detected. Using a multi-threaded approach: Thread 1 obtains the neuron discharge data of the effective channel collected in the acquisition card; Thread 2 visualizes the acquired data and displays the discharge waveform of each channel in the user interface; Thread 3 uses a threshold method to detect the number of Spikes in the data acquired by each channel during the detection period for calculating the firing rate. Based on the position coordinates and firing rate of the effective channel, the two-dimensional vector of the activity center (CA) of the entire neuron network is calculated:

[0030] ,

[0031] Assume that the multi-channel microelectrode array has 8×8 sites, CA represents the two-dimensional vector of the activity center of the whole neuronal network, FR(k) represents the discharge rate at electrode k during the recording phase, Col(k) represents the column coordinate of electrode k, Row(k) represents the row coordinate of electrode k, and the value range of Col(k) and Row(k) is 1-8. The difference between the CA vector in the training state and the initial spontaneous discharge state is calculated to obtain the difference CA. The sensory area pointed by the vector determines the actual motion command of the car. The mapping relationship between electrophysiological activity and actual motion of the car is established according to the following rules to determine the Car_Real_Motion of the smart car:

[0032] ,

[0033] In the above technical solution, in step 4, by comparing the target motion instruction of the car obtained in step 2 and the actual motion instruction obtained in step 3, the training method to be adopted for the neural network in this round of training can be obtained. If the target motion instruction is not equal to the actual motion instruction, it is necessary to weaken the neuronal synaptic connection between the two areas involved, that is, to apply penalty electrical stimulation to the central sites of the two areas. For example: if the target motion instruction is n (n=1, 2, 3), the actual motion instruction is m (m=1, 2, 3), and m≠n, then penalty electrical stimulation is applied to areas n and m. The method of penalty electrical stimulation is: applying low-frequency alternating electrical stimulation with a pulse interval of 300ms to the central sites of the two areas; if the target motion instruction is equal to the actual motion instruction, then applying reward electrical stimulation to the site of the decision area to stabilize the neuronal synaptic connection in the area. The method of reward electrical stimulation is: electrically stimulating the electrodes in the decision area at a high frequency of 100Hz at the same time.

[0034] Beneficial effects:

[0035] First, the present invention realizes the real-time interaction between biological neuron cultures and artificial cars in a dynamic environment, and cultivates the neural network in vitro, making the electrical stimulation regulation of the neural network more direct and effective. This method allows the neural network to perform adaptive learning in a constantly changing external environment, improving the flexibility and practicality of the system; secondly, the system adopts a "reward" and "punishment" stimulation mechanism, which can apply real-time feedback to the neural network, thereby adjusting its network structure and discharge pattern. This mechanism makes full use of the adaptability of biological neural networks in network structure and synaptic connection, and enhances the training effect of the system; in response to the shortcomings of existing research, the present invention introduces the function of customizing electrical stimulation sequences, allowing the frequency, amplitude and site parameters of the stimulation to be flexibly set, which provides convenience for exploring and verifying more accurate and effective "reward" and "punishment" feedback methods. In addition, the system uses the activity center method for decoding, and uses the difference between the CA vector in the training process and the initial state as the basis for decoding. This method is more in line with the memory and learning mechanism of in vitro neurons, and quantifies the learning behavior of the biological neural network through the change of the network synaptic weight, thereby mapping the movement direction of the car, and effectively solving the problem of poor training effect caused by the uneven distribution of neurons in the MEA; the present invention adopts a modular design, covering a neural network module, a Spike signal detection and signal display module, a discharge signal processing module, an intelligent car module, and an electrical stimulation module. In order to improve training efficiency, the present invention also designs a training system in a simulation environment without a hardware car, and modularizes the scenes in the simulation environment. The modular design facilitates the expansion and upgrade of the system, allowing users to independently optimize and improve each module of the system according to their needs.

[0036] The present invention aims to enable the biological neural network to learn to make three corresponding correct decisions according to the three obstacles in the external environment through multi-objective training, so as to effectively control the obstacle avoidance action of the car. This training method provides an important research means for synthetic biological intelligence to realize complex functions. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The above contents and other objects, features and advantages of the present disclosure will become more apparent through the following description of the embodiments of the present disclosure with reference to the accompanying drawings, in which:

[0038] Figure 1 This is the module operation flow chart of the closed-loop training system for controlling the car's obstacle avoidance with the neural chip;

[0039] Figure 2 A diagram of the maturation of biological neuronal networks cultured on MEA;

[0040] Figure 3 Functional validation results for training electrical stimulation;

[0041] Figure 4 This is the k-means clustering result diagram based on the effective site location distribution;

[0042] Figure 5 It is the Spike detection and signal display diagram;

[0043] Figure 6 is the center of activity (CA) vector pointing diagram;

[0044] Figure 7 Schematic diagram of two training modes of the closed-loop training system;

[0045] Figure 8 This is the effect diagram of the virtual car successfully avoiding obstacles controlled by neuron discharge signals in the simulation scene;

[0046] Fig. 9 This is a picture showing the failure of the virtual car to avoid obstacles in the simulation scene due to the neuron discharge signal controlling it.

[0047] Fig.10 It is the operation flow chart of the closed-loop training system in the simulation scenario;

[0048] Fig.11 This is a diagram showing the effect of using neuron discharge signals to control the motion of a smart car in a real scene;

[0049] Fig.12 This is a flowchart of the operation of the closed-loop training system in a real scenario. DETAILED DESCRIPTION

[0050] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. However, the following embodiments are limited to explaining the present invention, and the protection scope of the present invention should include the entire contents of the claims, and through the description of the following embodiments, those skilled in the art can fully implement the entire contents of the claims of the present invention.

[0051] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present disclosure. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure. However, it is apparent that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of known structures and technologies are omitted to avoid unnecessary confusion of the concepts of the present disclosure.

[0052] The terms used herein are only for describing specific embodiments and are not intended to limit the present invention. The terms "comprise", "include", etc. used herein indicate the existence of the features, steps, operations and / or components, but do not exclude the existence or addition of one or more other features, steps, operations or components.

[0053] The present invention proposes a closed-loop training device and method for controlling a car to avoid obstacles using an in vitro neural chip. The invention relies on the detection and electrical stimulation functions of MEA on neurons to achieve interaction and coordination between biological neural networks and external devices, and realizes motion control of a physical intelligent car in a real environment or a virtual car in a simulated environment by processing and decoding the collected electrophysiological signals. Different types of stimulation and feedback are applied to the biological neural network according to the position information of the car in the obstacle environment and the actual motion pattern of the car to establish a closed-loop training system. The system uses the adaptability of neuronal cultures to shape the network structure and discharge pattern, aiming to train the neural network to accurately control the car to successfully avoid obstacles in the environment.

[0054] like Figure 1 As shown in the figure, the normal operation of the system relies on the mutual coordination and interconnection of five main modules. The main functions of the five modules are:

[0055] (1) Biological neuron network module: Biological neuron culture, as the decision-making unit of synthetic biological intelligence system, needs to be cultured in an in vitro microelectrode array and produce a complex network structure (such as Figure 2 as shown).

[0056] (2) Spike detection and signal display module: The module acquires and filters the neuronal electrical signals collected by the acquisition card, displays the processed waveforms of each channel in the interface, and extracts the number of spikes recorded by each channel during the detection time from the signal (e.g. Figure 5 as shown).

[0057] (3) Discharge signal processing module: In each round of closed-loop training, according to the activity center (CA) motion mapping method (such as Figure 6 As shown in the figure), the command for controlling the movement of the car is calculated by using the spike firing rates and coordinates of the sites on the three network areas recorded in real time during a specified time period.

[0058] (4) Smart car module: When controlling a virtual car in a simulation environment, the system automatically generates different obstacle scenes in the simulation environment and obtains obstacle location information. When controlling a physical hardware car in a real environment, in each round of training, the smart car uses infrared sensors to detect obstacle location information in the environment and transmits it to the main control program on the PC via Bluetooth. The main control program transmits control instructions to the car based on the discharge patterns of different areas of the neuron culture to control the movement of the car in the real environment or virtual simulation environment.

[0059] (5) Electrical stimulation module: Based on the obstacle position information obtained by the main control program on the PC, the system applies sensory electrical stimulation to the biological neural network; the system will judge the feedback to the neural network based on whether the car successfully avoids the obstacle, and apply reward / punishment electrical stimulation to the biological neural network. The verification results of the training electrical stimulation function are as follows: Figure 3 As shown, it can be found that the network similarity after punishment electrical stimulation is significantly lower than the network similarity after reward electrical stimulation, that is, punishment electrical stimulation can disrupt the synaptic connections of the existing neuronal network; while reward electrical stimulation can consolidate the synaptic connections of the existing neuronal network.

[0060] This system aims to use the decision-making of neurons to control the car to avoid obstacles in the actual obstacle environment. It shapes the neural network on the microelectrode array through closed-loop training and forms three neuron functional partitions corresponding to different movements of the car. The system first divides the effective sites on the microelectrode array that can record the spontaneous discharge signals of neurons, and then divides the effective sites into regions based on the location information of the effective sites through the k-means clustering algorithm (such as Figure 4 In the subsequent closed-loop training process, the present invention designs two sets of closed-loop control methods for the simulation scenario and the real scenario respectively.

[0061] Example 1: Controlling a virtual car in a simulation scenario

[0062] according to Figure 7 In the training mode shown, the obstacle position information in training is set in advance in the main control program, and the target motion instructions of the car in each round of training are determined. The simulation module obtains the obstacle position information in the main control program through the interface and generates a simulation scene. After applying specific sensory stimulation to the neural network according to the obstacle position, the system's discharge signal processing module processes the neuronal discharge signals recorded by the effective sites in the three areas of the microelectrode array and decodes the actual motion instructions of the car. After the simulation module obtains the actual motion instruction, it displays the movement of the virtual car in the simulation scene according to the instruction on the screen. Figure 8 , Fig. 9The system's user interface and the use of neuron discharge signals to control the virtual car's successful and failed obstacle avoidance in an obstacle environment are demonstrated.

[0063] Relying on the functions of the simulation module and the adaptive closed-loop training method, this system interacts the obstacle environment with the biological system. According to the correctness of the decision-making of the neural network on the movement of the car, the feedback of different training stimuli is used to shape the network structure and regulate the network discharge mode. The main steps and methods of system operation include:

[0064] Step 1, environment initialization: set the obstacle positions in the environment where the car is located, and obtain the label for this round of training (the target movement direction of the car);

[0065] Step 2, environmental information encoding: the location information of obstacles in the environment is transmitted to the biological neuron culture through sensory electrical stimulation;

[0066] Step 3, motion decision: through data analysis and decoding, establish the mapping relationship between electrophysiological activity and actual movement of the car.

[0067] Step 4, shaping the neuronal network: applying punishment / reward electrical stimulation to the culture according to the movement state of the car in the environment to shape the synaptic structure and firing pattern of biological neurons.

[0068] Specific user operation steps are as follows: Fig.10 shown.

[0069] Example 2: Controlling the smart car in a real scenario

[0070] The hardware structure of the smart car includes three modules: sensor module, actuator module and Bluetooth communication module. The sensor module of the car contains three HC-06 infrared distance sensors, which are placed at the front, left and right ends of the car. This module can be used to detect obstacles in real scenes, so as to obtain the target motion instructions of the car. The smart car interacts with the main control system on the PC through its Bluetooth communication module and passes the target motion instructions to the main control program. The main control program then applies sensory electrical stimulation to the neural network according to the target motion instructions. The discharge signal processing module of the system processes the neuronal discharge signals recorded by the effective sites in the three areas of the microelectrode array to obtain the actual motion instructions of the car. After the smart car obtains the actual motion instructions obtained from the main control program through the Bluetooth module, it controls the specific movement of the car according to the motion instructions by controlling the motor module, such as Fig.11 After execution, the robot is controlled to stop moving and wait for the instruction to use the sensor module to detect the position information of external obstacles again.

[0071] Relying on the functions of the above hardware modules and based on the adaptive closed-loop training method, this system interacts the obstacle environment with the biological system. According to the correctness of the decision-making of the neural network on the movement of the car, the feedback of different training stimuli is used to shape the network structure and regulate the network discharge mode. The main steps and methods of system operation include:

[0072] Step 1, environment initialization: set the obstacle positions in the environment where the car is located, and obtain the label for this round of training (the target movement direction of the car);

[0073] Step 2, environmental information encoding: the location information of obstacles in the environment is transmitted to the biological neuron culture through sensory electrical stimulation;

[0074] Step 3, motion decision: through data analysis and decoding, establish the mapping relationship between electrophysiological activity and actual movement of the car.

[0075] Step 4, shaping the neuronal network: applying punishment / reward electrical stimulation to the culture according to the movement state of the car in the environment to shape the synaptic structure and firing pattern of biological neurons.

[0076] Specific user operation steps are as follows: Fig.12 shown.

[0077] The above description is only a specific implementation of the present application, so that those skilled in the art can understand or implement the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest range consistent with the principles and novel features applied for herein.

Claims

1. A closed-loop training device for controlling a car to avoid obstacles using an in vitro neural chip, characterized in that: It includes biological neural network, computer, intelligent car, obstacle simulation device and electric stimulator; The biological neuron network comprises neuron cells and a microelectrode array, wherein the neuron cells are seeded on the microelectrode array; The computer has a built-in main control program, which executes: (1) transmitting obstacle position information parameters to the obstacle simulation device, thereby setting obstacles in the simulation environment; (2) Acquiring and filtering the neuronal electrical signals of the microelectrode array, displaying the processed waveform, and extracting the number of spikes recorded within the detection time from the signal; (3) Process the signal and send motion control instructions to the smart car; (4) Send feedback stimulation instructions to the electrical stimulator according to the movement of the smart car; The obstacle simulation device includes obstacles arranged around the smart car, and the obstacles are divided into two states: open and closed; The smart car is provided with an infrared sensor and a Bluetooth module, wherein the infrared sensor is used to detect the position information of obstacles in the environment, and the Bluetooth module is used to transmit the position information of obstacles to the main control program of the computer; The electrical stimulator applies sensory electrical stimulation to the biological neuron network according to the feedback stimulation instruction.

2. The closed-loop training device for controlling a car to avoid obstacles using an in vitro neural chip according to claim 1 is characterized in that: The microelectrode array is prepared by a micro-electromechanical system process, the number of effective sites of the microelectrode array ranges from 64 to 288 channels, and the sites are modified by nanomaterials.

3. The closed-loop training device for controlling a car to avoid obstacles using an in vitro neural chip according to claim 2 is characterized in that: The obstacles are arranged in the left, front and right directions of the smart car.

4. A closed-loop training method for controlling a car to avoid obstacles using an in vitro neural chip, characterized in that: The steps include: Step 1, environment initialization: set the obstacle positions in the environment where the smart car is located, and obtain the target motion instructions of the smart car in this round of training; Step 2, environmental information encoding: The location information of obstacles in the environment is encoded through sensory electrical stimulation and transmitted to neuronal cells; Step 3, motion decision: by processing and decoding the collected data, establish the mapping relationship between electrophysiological activity and actual motion of the smart car; Step 4, shaping the neural network: according to the movement state of the smart car in the environment, punishment or reward electrical stimulation is applied to the neuron cells to shape the synaptic structure between the neuron cells and the neuron discharge pattern in different areas.

5. The closed-loop training method for controlling a car to avoid obstacles using an in vitro neural chip according to claim 4 is characterized in that: In step 1, obstacles are placed at two positions in the left, front and right directions of the smart car. The smart car uses infrared sensors to detect obstacles and obtain target motion instructions in the scene.

6. The closed-loop training method for controlling a car to avoid obstacles using an in vitro neural chip according to claim 4, characterized in that: In step 2, the obstacle position information obtained in step 1 is converted into a sensory electrical stimulation instruction. There are three types of obstacle position information, namely: obstacles on the left and right sides; obstacles on the right and front; obstacles on the left and front. Three sensory areas are divided on the neuron cells, namely site area 1, site area 2 and site area 3. When the obstacle position information is the first case, the sensory electrical stimulation is applied to the stimulation site of site area 1. When the obstacle position information is the second case, the sensory electrical stimulation is applied to the stimulation site of site area 2. When the obstacle position information is the third case, the sensory electrical stimulation is applied to the stimulation site of site area 3.

7. The closed-loop training method for controlling a car to avoid obstacles using an in vitro neural chip according to claim 6, characterized in that: The sensory electrical stimulation method is to use a negative leading two-phase rectangular voltage pulse sequence with an amplitude of 300mV, a pulse length of 200μs and a frequency of 1Hz.

8. The closed-loop training method for controlling a car to avoid obstacles using an in vitro neural chip according to claim 4, characterized in that: In step 3, after the obstacle position information is transmitted to the neuron culture through sensory stimulation, the discharge pattern of the neuron cluster is detected, using a multi-threaded approach: Thread 1, obtains the neuron discharge data of the valid channels collected in the acquisition card; Thread 2, visualizes the acquired data and displays the discharge waveform of each channel in the user interface; Thread 3 uses a threshold method to detect the number of spikes in the data obtained from each channel during the detection period, which is used to calculate the firing rate; based on the position coordinates and firing rate of the effective channel, the two-dimensional vector of the activity center CA of the entire neural network is calculated.

9. The closed-loop training method for controlling a car to avoid obstacles using an in vitro neural chip according to claim 4, characterized in that: In step 4, by comparing the target motion instruction of the car obtained in step 2 and the actual motion instruction obtained in step 3, the training method for the neural network to be adopted in this round of training is obtained. If the target motion instruction is not equal to the actual motion instruction, the synaptic connection between the two areas involved is weakened, that is, a penalty electrical stimulus is applied to the central site of the two areas; if the target motion instruction is equal to the actual motion instruction, a reward electrical stimulus is applied to the site of the decision area to stabilize the synaptic connection of the neurons in the area.

Images