A virtual-real migration closed-loop feedback training method based on three-dimensional brain organoids

Abstract

The application discloses a kind of virtual-real migration closed-loop feedback training methods based on three-dimensional brain organoids, the method of the application adopts the technical route of "virtual-real migration", and three-dimensional brain organoids are trained by multimodal closed-loop feedback, i.e. Environmental feedback, punishment feedback and reward feedback effectively avoid the long time mechanical debugging and hardware wear problems caused by traditional physical environment training. In particular, in the initial training period, this method avoids the mechanical vibration caused by the violent collision of real mobile robots, prevents the damage of mechanical hardware and the damage of brain organoids activity, so as to maintain the activity of biological computing carrier and the stability of signal acquisition to the greatest extent while ensuring the training intensity.

Description

Technical Field

[0001] This invention relates to the field of brain organoid electrophysiology, and in particular to a virtual-real transfer closed-loop feedback training method based on three-dimensional brain organoids. Background Technology

[0002] The structural simplicity and functional limitations of biological computing carriers: Existing research on control based on biological neurons mostly focuses on two-dimensional (2D) neuron culture or invasive animal brain-computer interfaces. Two-dimensional culture lacks the complex three-dimensional neural network structure similar to the real brain, limiting its computational potential and cognitive functions; while live animal experiments are limited by ethical review, large individual differences, and difficulty in standardized long-term closed-loop control.

[0003] High risks associated with online training of physical robots: Existing methods typically involve directly connecting neurons to physical robots for training. Because living organisms react slowly and are prone to fatigue, and physical robots are slow to move and easily damaged, and because reset operations (such as manual intervention after a collision with a wall) are time-consuming, sample collection efficiency is very low, making high-intensity repeated training difficult. High-frequency failures in the early stages of training (high-frequency collisions in obstacle avoidance tasks) can also lead to mechanical damage. Summary of the Invention

[0004] The purpose of this invention is to address the technical deficiencies in the existing technology by providing a closed-loop feedback training method for virtual-real transfer based on three-dimensional brain organoids.

[0005] The technical solution adopted to achieve the purpose of this invention is: A closed-loop feedback training method for virtual-real transfer based on three-dimensional brain organoids includes the following steps: Step 1: Construct a virtual maze scene containing virtual mobile robots and static obstacles, and a real maze scene containing real mobile robots and static obstacles; Step 2: The real mobile robot detects the straight-line distance and angle of the nearest static obstacle in real time, converts it into an electrical stimulation signal, and uses it as an environmental perception stimulus to stimulate the three-dimensional brain organoid in real time. The data is recorded within a preset time. The three-dimensional brain organoid controls the real mobile robot to collide with each obstacle a number of times, which serves as the untrained baseline data. Step 3: Select stimulation sites of the three-dimensional brain organoid as input channels and select channels that elicit strong responses as output channels. The virtual mobile robot detects the straight-line distance and angle of the nearest static obstacle in real time and converts them into electrical stimulation signals, which are used as environmental perception stimuli to stimulate the three-dimensional brain organoid in real time. The three-dimensional brain organoid controls the virtual mobile robot to move in the virtual maze scene. When the virtual mobile robot collides, only the three-dimensional brain organoid is punished to suppress wrong decisions. When the virtual mobile robot successfully avoids obstacles, only the three-dimensional brain organoid is rewarded to reinforce the correct path, thereby training the three-dimensional brain organoid. Step 4: Use the trained 3D brain organoid to control the movement of a real mobile robot in a real maze scene, count the number of collisions per unit obstacle within the preset time, and compare it with the baseline data to verify the virtual-to-real transfer capability of the training effect.

[0006] In the above technical solution, in step 1, the kinematic parameters of the virtual mobile robot are proportional to those of the real mobile robot, and the kinematic parameters include the wheel diameter of the robot and the distance between the left and right wheel axles.

[0007] In the above technical solution, in steps 2 and 3, both the virtual mobile robot and the real mobile robot collect the straight-line distance and angle between themselves and the nearest static obstacle using a lidar sensor.

[0008] In the above technical solution, in steps 2 and 3, the straight-line distance and angle are normalized and then converted into electrical stimulation signals through electrical stimulation mapping. The normalization process uses a linear normalization algorithm to obtain the normalized distance. and normalization perspective The spatial data is expressed by the following formula:

[0009] in, d min As a safety threshold, d max Where d is the collision threshold and d is the straight-line distance. For angle; The Environmental Stimulation transforms spatial data into electrical stimulation signals, i.e., environmentally perceived stimuli, as shown in the following formula:

[0010] in, D To balance the intermediate metrics of normalized angle and distance information, an environmental (angle & distance) encoded metric is used to generate environmental perception stimuli. For environmental perception stimuli, The initial maximum stimulation threshold, This is the initial minimum threshold.

[0011] In the above technical solution, the three-dimensional brain organoid is connected to the electrode array, and the electrode array is connected to the electrophysiological acquisition system.

[0012] In the above technical solution, the selection of stimulation sites of three-dimensional brain organoids as input channels and the selection of channels that elicit strong responses as output channels are achieved through the following process: Step a: Using a bidirectional electrical pulse sequence as the test signal, the electrophysiological acquisition system applies stimulation to the three-dimensional brain organoid through all electrode sites of the electrode array, quantifies the neuronal response intensity of the three-dimensional brain organoid, and records the total number of electrode channels with effective response. Step b: Select two electrode sites that are physically far apart from the total number of electrode channels obtained in step a. Use one electrode site as the left wheel input channel of the virtual or real mobile robot and the other electrode site as the right wheel input channel of the virtual or real mobile robot to reduce signal crosstalk. Step c: Select the channels with the strongest neuron response from the remaining total number of electrode channels and divide them into two groups of channels. One group of channels serves as the output channel corresponding to the left wheel input channel, and the other group of channels serves as the output channel corresponding to the right wheel input channel.

[0013] In the above technical solution, in step 3, the electrophysiological acquisition system collects the discharge signal of the output channel in real time, extracts the discharge timing through Spike detection, and uses a decoding algorithm to convert the neural activity of the three-dimensional brain organoid into the left and right wheel speed commands of the virtual mobile robot, driving the virtual mobile robot to move in the virtual maze scene.

[0014] In the above technical solution, the frequency of the electrical stimulation signal is 4~40Hz.

[0015] In the above technical solution, the environmental perception stimulus is a low-amplitude 60~80uA pulse signal whose frequency changes linearly with the straight-line distance.

[0016] In the above technical solution, in step 3, the punishment stimulus includes a stimulation group and a non-stimulation group arranged at intervals. The total time for applying the punishment stimulus is divided into multiple time units. The time for applying the stimulus to each stimulation group includes at least two adjacent time units. Each time unit in the stimulation group is predefined as a stimulus event. The time for each non-stimulation group to stop applying the stimulus includes at least one time unit.

[0017] In the above technical solution, a control sequence is constructed to apply stimulation in the stimulation group and stop applying stimulation in the non-stimulation group. When the control sequence indicates an effective state, the corresponding stimulation event is output in the corresponding time unit; when the indication is an invalid state, the zero level is maintained to occupy the corresponding time length.

[0018] In the above technical solution, the stimulation pulse waveforms of stimulation events within all time units are consistent, which are biphasic stimulation pulses, and the amplitude and duration parameters of each phase remain fixed.

[0019] In the above technical solution, the reward stimulus is a high-frequency stimulation pulse signal, the amplitude of which is 80~100uA, the frequency is 80~120Hz, and the pulse width is 100~200us.

[0020] Compared with the prior art, the beneficial effects of the present invention are: 1. The method of this invention adopts a "virtual-real transfer" technical approach, using multimodal closed-loop feedback—namely, environmental feedback, penalty feedback, and reward feedback—to train three-dimensional brain organoids. This effectively avoids the problems of prolonged mechanical debugging and hardware wear and tear caused by traditional physical environment training. Especially in the early stages of training, this method avoids the mechanical vibrations caused by violent collisions with real mobile robots, preventing damage to mechanical hardware and impairment of brain organoid activity. Thus, while ensuring training intensity, it maximizes the activity of the biological computing carrier and the stability of signal acquisition.

[0021] 2. This invention provides a low-power computing solution for extreme environments. A well-trained brain organoid can serve as a milliwatt-level low-power computing core, integrated into real mobile robot systems. Compared to traditional silicon-based chips, this biological computing architecture (the three-dimensional brain organoid trained in this invention) has significant energy efficiency advantages when processing unstructured environmental information. This characteristic makes it particularly suitable for extreme application scenarios such as deep space exploration and disaster rescue, where energy supply, equipment heat dissipation, and endurance are strictly limited, thereby enhancing the robot's autonomous survivability in unknown environments.

[0022] 3. The training method of this invention is applicable to all three-dimensional brain organoids achieving a relatively consistent performance benchmark in physical control tasks. This state, solidified through a specific paradigm (stimuli throughout the training process), reduces errors caused by varying learning levels in traditional biological experiments. In drug screening, the three-dimensional brain organoids of this invention serve as drug delivery vehicles, and the drug efficacy is evaluated based on changes in obstacle avoidance performance. This stable benchmark allows any fluctuations in robot motion metrics (such as a decrease in obstacle avoidance rate) to be effectively correlated with the drug's interference with neural signal transmission, thus providing a quantitative detection method for evaluating neurotoxins or cognitive drugs. Attached Figure Description

[0023] Figure 1 This is a schematic flowchart of the method of the present invention.

[0024] Figure 2 To create a realistic maze scene and real mobile robots.

[0025] Figure 3 Schematic diagram of brain organoids, needle electrodes, and experimental platform.

[0026] Figure 4 This is a diagram illustrating obstacle avoidance training in a virtual environment.

[0027] Figure 5 This is a training and evaluation graph showing the number of obstacle collisions per 5 minutes, where test0 represents the 5 minutes of untrained data from step 2, and train... - 0 represents the first 5 minutes of training the brain organoids. - 1 represents 5 minutes after training the brain organoids, and test1 represents 5 minutes after completing the training in step 4. Detailed Implementation

[0028] The present invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0029] like Figure 1 As shown, a closed-loop feedback training method for virtual-real transfer based on three-dimensional brain organoids includes the following steps: Step 1: Construct a virtual maze scene containing virtual mobile robots and static obstacles, and a real maze scene containing real mobile robots and static obstacles (e.g., ...). Figure 2 As shown in the diagram, the kinematic parameters of the virtual mobile robot are proportional to those of the real mobile robot, including the wheel diameters and the distance between the left and right wheel axles. Preferably, the positions of static obstacles in the virtual maze scene differ from those in the real maze scene.

[0030] like Figure 3As shown, in step 2, a three-dimensional brain organoid is constructed, which includes a three-dimensional brain organoid, an electrode array connected to the three-dimensional brain organoid, and an Intan electrophysiological acquisition system connected to the electrode array. The real scene mode is activated, and the signal input source of the three-dimensional brain organoid is switched to the sensor interface of the real mobile robot. The real mobile robot detects the straight-line distance and angle of the nearest static obstacle in real time and converts it into an electrical stimulation signal (stimulation frequency of 4-40 Hz, stimulation current of 70 uA) as an environmental perception stimulus. The three-dimensional brain organoid is stimulated every 400 ms and recorded within a preset time (5 min). The three-dimensional brain organoid controls the real mobile robot to collide with a unit obstacle a number of times, which is used as the untrained baseline data. After resting for 5 min, the three-dimensional brain organoid is used as the baseline data.

[0031] Step 3 specifically includes the following steps: like Figure 4 As shown, step 3.1 involves activating the virtual scene mode, switching the signal input source of the 3D brain organoid to the virtual mobile robot sensor interface, selecting stimulation sites of the 3D brain organoid as input channels, and choosing channels that elicit strong responses as output channels; specifically, this includes the following steps: Step a: A bidirectional electrical pulse sequence (frequency 1Hz, number of pulses 10) is used as the test signal. The Intan electrophysiological acquisition system applies stimulation to the three-dimensional brain organoid through all electrode sites of the electrode array. The action potential spike count within 200ms after stimulation is counted to quantify the neuronal response intensity of the three-dimensional brain organoid. The higher the channel spike count, the stronger the neuron response to stimulation recorded in that channel. The channels with the top 32 spike counts are selected as the total number of electrode channels with effective response. Step b: Select two electrode sites that are physically far apart from the total number of electrode channels obtained in step a. Use one electrode site as the left wheel input channel of the virtual or real mobile robot and the other electrode site as the right wheel input channel of the virtual or real mobile robot to reduce signal crosstalk. Step c: Select the 12 channels with the strongest neuronal response from the remaining total number of electrode channels and divide them into two groups. Six groups of channels are used as the output channels corresponding to the left input channel, and the other six groups of channels are used as the output channels corresponding to the right input channel.

[0032] Step 3.2: The virtual mobile robot detects the straight-line distance and angle of the nearest static obstacle in real time, converts it into an electrical stimulation signal (70uA pulse, 4-40Hz), and uses it as an environmental perception stimulus (environmental feedback) to stimulate the three-dimensional brain organoid in real time.

[0033] Step 3.3: The electrophysiological acquisition system (RHS Stim / Recording System) acquires the discharge signals of the output channel in real time, extracts the discharge timing through Spike detection, and uses a decoding algorithm to convert the neural activity of the three-dimensional brain organoid into the left and right wheel speed commands of the virtual mobile robot, driving the virtual mobile robot to move in the virtual maze scene. When the virtual mobile robot collides, only the three-dimensional brain organoid is subjected to a punishment stimulus (punishment feedback) to suppress incorrect decisions. When the virtual mobile robot successfully avoids obstacles, only the three-dimensional brain organoid is subjected to a reward stimulus (reward feedback, the reward stimulus is a high-frequency pulse signal with an amplitude of 90uA, a frequency of 100Hz, and a pulse width of 200us) to reinforce the correct path. The three-dimensional brain organoid is trained in this way for 10 minutes and then rested for 5 minutes.

[0034] Step 4: Use the trained 3D brain organoid to control the real mobile robot to move in a real maze scene. Statistically count the number of collisions with a unit obstacle and the number of consecutive obstacle avoidances within the preset time (5 minutes), and compare it with the baseline data to verify the virtual-to-real transfer capability of the training effect.

[0035] Furthermore, both the virtual and real mobile robots acquire the straight-line distance and angle between themselves and the nearest static obstacle using lidar sensors. After normalization, the straight-line distance and angle are converted into electrical stimulation signals via electrical stimulation mapping. The normalization process employs a linear normalization algorithm to obtain the normalized distance. and normalization perspective The spatial data is expressed by the following formula:

[0036] in, d min As a safety threshold, d max Where d is the collision threshold and d is the straight-line distance. For angle.

[0037] The Environmental Stimulation transforms spatial data into electrical stimulation signals, i.e., environmentally perceived stimuli, as shown in the following formula:

[0038] .

[0039] in, D To balance the intermediate metrics of normalized angle and distance information, an environmental (angle & distance) encoded metric is used to generate optimized stimuli. For environmental perception stimuli, The initial maximum stimulation threshold, This is the initial minimum threshold.

[0040] Furthermore, the punishment stimulus includes stimulation groups and non-stimulation groups arranged at intervals. The total time for applying the punishment stimulus (4s) is divided into 20 time units (20ms). The application time of each stimulation group includes at least two adjacent time units. Within each time unit of the stimulation group, a predefined stimulation event (amplitude of 180µA, frequency of 5Hz, pulse width of 200µs) is defined. The time for stopping the application of the stimulus in each non-stimulation group includes at least one time unit. The stimulation pulse waveforms of the stimulation events in all time units are consistent, being biphasic stimulation pulses, and the amplitude and duration parameters of each phase remain fixed. The application time of the stimulation group accounts for 40-80% of the total time.

[0041] A control sequence is constructed to apply stimulation in the stimulation group and stop applying stimulation in the non-stimulation group. When the control sequence indicates an active state, the corresponding stimulation event is output in the corresponding time unit; when the indication is an inactive state, the zero level is maintained to occupy the corresponding time length.

[0042] like Figure 5 As shown in the figure, the number of collisions per unit obstacle can be evaluated, and it can be seen that, within the same time period, the number of collisions is significantly reduced when using a trained 3D brain organoid to control a real mobile robot.

[0043] Number of collisions per obstacle = Number of collisions of the mobile robot in the maze scene / Number of obstacles in the maze scene.

[0044] The above description is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A closed-loop feedback training method for virtual-real transfer based on three-dimensional brain organoids, characterized in that, Includes the following steps: Step 1: Construct a virtual maze scene containing virtual mobile robots and static obstacles, and a real maze scene containing real mobile robots and static obstacles; Step 2: The real mobile robot detects the straight-line distance and angle of the nearest static obstacle in real time, converts them into electrical stimulation signals, and uses them as environmental perception stimuli to stimulate the three-dimensional brain organoid in real time. The number of times the three-dimensional brain organoid controls the real mobile robot to collide with a single obstacle within a preset time is recorded as baseline data without training. Step 3: Select stimulation sites of the three-dimensional brain organoid as input channels and select channels that elicit strong responses as output channels. The virtual mobile robot detects the straight-line distance and angle of the nearest static obstacle in real time and converts them into electrical stimulation signals, which are used as environmental perception stimuli to stimulate the three-dimensional brain organoid in real time. The three-dimensional brain organoid controls the virtual mobile robot to move in the virtual maze scene. When the virtual mobile robot collides, only the three-dimensional brain organoid is subjected to a penalty stimulus. When the virtual mobile robot successfully avoids the obstacle, only the three-dimensional brain organoid is subjected to a reward stimulus, thereby training the three-dimensional brain organoid. Step 4: Use the trained 3D brain organoid to control the movement of a real mobile robot in a real maze scene, count the number of collisions per unit obstacle within the preset time, and compare it with the baseline data to verify the virtual-to-real transfer capability of the training effect.

2. The training method according to claim 1, characterized in that, In step 1, the kinematic parameters of the virtual mobile robot are proportional to those of the real mobile robot, and the kinematic parameters include the wheel diameter of the robot and the distance between the left and right wheel axles.

3. The training method according to claim 1, characterized in that, In steps 2 and 3, both the virtual mobile robot and the real mobile robot use lidar sensors to collect the straight-line distance and angle between themselves and the nearest static obstacle.

4. The training method according to claim 1, characterized in that, In steps 2 and 3, the straight-line distance and angle are normalized and then converted into electrical stimulation signals through electrical stimulation mapping. The normalization process uses a linear normalization algorithm to obtain the normalized distance. and normalization perspective The spatial data is expressed by the following formula: in, d min As a safety threshold, d max Where d is the collision threshold and d is the straight-line distance. For angle; The Environmental Stimulation transforms spatial data into electrical stimulation signals, i.e., environmentally perceived stimuli, as shown in the following formula: in, D To balance the intermediate indicators of normalized angle and distance information, For environmental perception stimuli, The initial maximum stimulation threshold, This is the initial minimum threshold.

5. The training method according to claim 1, characterized in that, The three-dimensional brain organoid is connected to an electrode array, which is in turn connected to an electrophysiological acquisition system.

6. The training method according to claim 1, characterized in that, The process of selecting stimulation sites of three-dimensional brain organoids as input channels and choosing channels that elicit strong responses as output channels is achieved through the following steps: Step a: Using a bidirectional electrical pulse sequence as the test signal, the electrophysiological acquisition system applies stimulation to the three-dimensional brain organoid through all electrode sites of the electrode array, quantifies the neuronal response intensity of the three-dimensional brain organoid, and records the total number of electrode channels with effective response. Step b: Select two electrode sites that are physically far apart from the total number of electrode channels obtained in step a. One electrode site is used as the left wheel input channel of the virtual mobile robot or the real mobile robot, and the other electrode site is used as the right wheel input channel of the virtual mobile robot or the real mobile robot. Step c: Select the channels with the strongest neuron response from the remaining total number of electrode channels and divide them into two groups of channels. One group of channels serves as the output channel corresponding to the left wheel input channel, and the other group of channels serves as the output channel corresponding to the right wheel input channel.

7. The training method according to claim 1, characterized in that, In step 3, the electrophysiological acquisition system collects the discharge signal of the output channel in real time, extracts the discharge timing through Spike detection, and uses a decoding algorithm to convert the neural activity of the three-dimensional brain organoid into the rotation speed commands of the left and right wheels of the virtual mobile robot, driving the virtual mobile robot to move in the virtual maze scene.

8. The training method according to claim 1, characterized in that, The frequency of the electrical stimulation signal is 4~40Hz.

9. The training method according to claim 1, characterized in that, The environmental perception stimulus is a pulse signal with an amplitude of 60~80uA, and the reward stimulus is a pulse signal with an amplitude of 80~100uA, a frequency of 80~120Hz, and a pulse width of 100~200us.

10. The training method according to claim 1, characterized in that, In step 3, the punishment stimulus includes a stimulation group and a non-stimulation group set at intervals. The total time for applying the punishment stimulus is divided into multiple time units. The time for applying the stimulus to each stimulation group includes at least two adjacent time units. Each time unit in the stimulation group is predefined as one stimulus event. The time for each non-stimulation group to stop applying the stimulus includes at least one time unit.

Images