Construction method of complex environment navigation strategy based on grid cells and place cells

Abstract

The present invention discloses a method for constructing a complex environment navigation strategy based on grid cells and place cells. The method includes the following steps: S1, constructing a scenario cognitive map based on the environmental cognition process of the hippocampus; S2, constructing a scenario cognitive model using an adaptive resonance theory network; S3, performing loop adjustment, topological connection, and time-dependent forgetting of the scenario cognitive process; S4, establishing a navigation model by combining the topological attributes in the scenario cognitive map with the metric attributes of vector calculation; S5, performing navigation using the navigation model; S6, performing segmented obstacle avoidance based on obstacle information; S7, performing obstacle avoidance based on a scenario playback mechanism; S8, using reward cells to define reward signals. The robot navigation model inspired by the hippocampus can generate a scenario cognitive map and use grid cell vector calculation, place cell topological connection, and boundary cell obstacle perception to navigate efficiently and reliably in a complex environment.

Description

Technical Field

[0001] The present invention relates to the technical field of robot navigation. Specifically, it relates to a method for constructing a complex environment navigation strategy based on grid cells and place cells. Background Art

[0002] When humans and animals move between two locations, they can intuitively find a shortcut. For this, scientists have discovered a cognitive map in the animal brain. This map can help animals establish an internal model in the brain according to the external environment they are in, so as to plan the optimal path. The brain is the most intelligent because it has the ability to recognize spatio-temporal sequences. This ability can recall the time sequence and spatial locations of past events through cognitive stimuli. This episodic cognitive memory can connect knowledge such as time, place, event, and behavior experienced in the past. It is a collection of individual experiences at specific times and places in the past.

[0003] Neuroscience research shows that the spatio-temporal mapping related to episodic cognition is rooted in a specific area of the brain - the hippocampus. The hippocampus is located below the cerebral cortex and plays a role in short-term and long-term episodic memory, as well as spatial orientation and navigation. The name of the hippocampus comes from the fact that the curved shape of this part resembles a seahorse. The hippocampus can form an "episodic cognitive map" through spatial and temporal exploration by the episodic cognitive system. It can be used for derivation and memory. This "internal map" allows organisms to simulate various possibilities through their brains and then make predictions.

[0004] Research has found that after damage to the hippocampus in the brain, rats show obstacles in finding the spatial location of the hidden platform in the water maze, and the volume of the hippocampus of taxi drivers is higher than the average value of ordinary people. When studying the hippocampus of rats, a special type of "place cell" was discovered. When a rat is in a specific position in the environment, this type of neuron becomes more active. In the entorhinal cortex adjacent to the hippocampus, researchers discovered another type of cell that responds to position. This type of cell can discharge at multiple positions. If the activity space is divided into countless equilateral triangles, then these cells always discharge when the rat walks to the vertices of the equilateral triangles. Therefore, researchers named them "grid cells". There are also some other cells that play important roles in positioning: for example, "reward" cells encode a reward mechanism related to position according to environmental information; there are also some cells that can track the forward speed and direction of the animal, just like a speedometer and a compass in the brain, and can calculate the complete movement process of the animal when exploring the environment; in addition, there is a type of cell that discharges when the animal approaches the edge of the map, which is called a "border cell".

[0005] In existing target navigation methods that imitate the hippocampal neural circuit, there are mainly two modes: topological navigation using place cells and vector navigation using grid cells. The place cells in the hippocampus are considered to form interconnected topological relationships through synapses, and the robot navigates to its target position by calculating the shortest path between the internal topological nodes of the environment. The grid cell neural network can support the vector navigation strategy. The grid cell phase moves with spatial movement. Since the grid metric property can be used as an expression of spatial coordinates, the robot can calculate the displacement vector using the grid firing modes between different positions.

[0006] Topological navigation and vector navigation each have their advantages. For example, Figure 2 For the typical task of an intelligent patrol robot, when the inspection robot starts from the stationing point and an emergency occurs during random exploration of the environment, the robot hopes to return to its own stationing point as soon as possible. Figure 3 The topological nature of the scenario cognition shown supports topological navigation based on empirical nodes, where the knowledge of the interconnectedness of place cells is used to reach the target, that is, topological navigation. Figure 4 The grid cell firing information of the scenario cognition may be able to calculate the distance and direction of the straight-line trajectory between any pair of previously visited positions, that is, vector navigation.

[0007] Both of these strategies have their advantages and disadvantages: Topological navigation is good at finding highly feasible paths in complex environments and can avoid obstacles. However, it requires sufficient prior exploration of the environment and the formation of a dense topological map. While vector navigation can directly navigate towards the target efficiently and can cross unknown environments. But vector navigation cannot anticipate obstacles and is easily trapped or spends a long path to bypass obstacles. Therefore, a single vector navigation strategy can only be used in open and obstacle-free environments.

[0008] In response to the problems in the related art, no effective solution has been proposed yet. Summary of the Invention

[0009] In response to the problems in the related art, the present invention proposes a method for constructing a complex environment navigation strategy based on grid cells and place cells to overcome the above-mentioned technical problems existing in the existing related art.

[0010] To this end, the specific technical solution adopted by the present invention is as follows:

[0011] A method for constructing a complex environment navigation strategy based on grid cells and place cells, the method comprising the following steps:

[0012] S1. Construct a scenario cognition map based on the environmental cognition process of the hippocampus;

[0013] S2. Construct a scenario cognition model using an adaptive resonance theory network;

[0014] S3. Perform loop adjustment, topological connection, and time-varying forgetting of the scenario recognition process;

[0015] S4. Establish a navigation model by combining the topological attributes in the scenario recognition map with the metric attributes of vector calculation;

[0016] S5. Use the navigation model for navigation;

[0017] S6. Perform segmented obstacle avoidance based on obstacle information;

[0018] S7. Perform obstacle avoidance based on the scenario playback mechanism;

[0019] S8. Use reward cells to define the reward signal.

[0020] Furthermore, the scenario recognition map is formed by connecting nodes composed of several scenario neurons, and each node maps specific perceptual information, where the perceptual information represents the neural firing information of the simulated cell model.

[0021] Furthermore, the simulated cell model includes an attractor model, a grid cell neural plate model, a place cell mathematical model, a boundary cell model, and a visual cell model;

[0022] Among them, the attractor model is used to simulate the encoding mechanism of head direction cells, and the calculation formula is:

[0023] s i (t) = v(t)·cos(θ i -θ t );

[0024] Among them, s i (t) represents the firing rate signal, θ t represents the head orientation of the robot at time t, v(t) represents the movement speed, and θ i represents the phase angle of the i-th head direction cell in the attractor model;

[0025] The place cell mathematical model is used to simulate the encoding mechanism of place cells, and the calculation formula is:

[0026]

[0027] Among them, represents the firing rate of place cell i at position r, r = [x, y] represents the position coordinates of the current robot in the environment, r i0 represents the position coordinates corresponding to the firing center of place cell i, and δ 2 represents the place cell firing adjustment coefficient;

[0028] The described boundary cell model is used to simulate the coding mechanism of boundary cells, and the calculation formula is:

[0029]

[0030] Among them, B i is the firing rate of boundary cell i, r represents the position of the current boundary cell, d i is the radius of boundary perception, θ represents the direction of the current boundary cell, is the main direction of boundary perception, σ rad , σ ang are the adjustment coefficients of the radius and angle responses respectively.

[0031] Furthermore, the grid cell neural plate model is used to simulate the coding mechanism of grid cells, including the following steps:

[0032] S11. Grid coding, the calculation formula is:

[0033]

[0034] Among them, represents the activation value of a single grid cell, κ represents the waveform adjustment coefficient, n max represents the peak value of the firing rate, represents the spatial displacement, represents the wave vector direction of each stripe;

[0035] S12. Grid decoding, the calculation formula is:

[0036]

[0037] Among them, represents the probability of the population firing rate vector composed of M cells at the spatial displacement , represents the firing rate of grid cell j, represents the response vector of the neural plate population, n j represents the firing rate of grid cell j at the current position;

[0038] S13. Multi-scale vector calculation, stacking the grid cell neural plates of each scale into a pyramid geometric model, mapping the position decoded from the grid firing information of the robot onto a virtual motion vector for expression, and then obtaining the direction vector towards the target position by taking the difference between the target motion vector and the current motion vector.

[0039] Furthermore, constructing the scenario cognitive model using the adaptive resonance theory network includes the following steps:

[0040] S21. The nerve cells encode the environmental perception information into a neuron activation vector as the input of the model;

[0041] S22. Apply the Adaptive Resonance Theory (ART) network for unsupervised self-learning of the information from the input layer to the event layer. The activation value of the event cells after learning gradually decays over time.

[0042] S23. Use the event activation sequence greater than the threshold as the input and apply the Adaptive Resonance Theory (ART) network for self-learning again. Combine the winner-takes-all rule to activate or newly generate a specific scenario cell, and use the activation state of this specific scenario cell as a node of a scenario cognitive map.

[0043] Further, the loop adjustment, topological connection, and time-varying forgetting in the scenario cognitive process include the following steps:

[0044] S31. If existing nerve cells are reactivated during the scenario cognitive process, it is determined that a loop of the motion trajectory is detected, and error correction is performed to prevent the accumulation of errors in path integration.

[0045] S32. Update the weights from the time cells to the scenario cells, reset the grid cell anti-electricity mode, and the reset amplitude is proportional to the event activation value.

[0046] S33. Create horizontal connections between the place cells to achieve the topological relationship on the scenario map.

[0047] S34. Update the horizontal topological connection of the topological layer by introducing synaptic connections between the scenario neurons.

[0048] Further, the navigation using the navigation model includes the following steps:

[0049] S51. After determining the final target scenario node, search for all possible paths according to the topological information of the scenario cognitive map.

[0050] S52. Endow the grid cells with high-frequency discharge cycles to drive the grid mode to traverse along the possible topological paths, and select the one with the shortest time consumption as the planned basic path to complete the simulation deduction.

[0051] S53. During the navigation process, before departing from the position of each scenario node, if the boundary cells of the current node are in the activated state, move towards the next topological node; if the boundary cells are not activated, obtain the vector navigation direction through the vector calculation of the grid cell mode, and the robot moves towards the final target direction.

[0052] S54. If a new obstacle is encountered during the vector navigation process, use the scenario playback mechanism to search for sub-goals in the order of the gradient descent of the activation value of the reward cells and calculate the vector direction until a vector direction of a sub-goal can take the robot away from the obstacle. After leaving the obstacle area, continue the vector navigation towards the final target.

[0053] Furthermore, the segmented obstacle avoidance based on obstacle information includes the following steps:

[0054] S61. Cluster and process the boundary recognition information, convert it into obstacle recognition, and use it as the basis for navigation decision-making;

[0055] S62. After obtaining the topological path through simulation and deduction, predict the existence of obstacles based on the activation information of boundary cells in the scenario nodes;

[0056] S63. Cut the planned topological path into three segments. The segment before the obstacle is the first segment, the segment near the obstacle is the second segment, and the segment after the obstacle is the third segment. Then, navigate according to the navigation model in sequence.

[0057] Furthermore, the obstacle avoidance based on the scenario replay mechanism includes the following steps:

[0058] S71. During vector navigation, when the normal vector of the obstacle surface faced by the robot forms an acute angle with the target vector, the robot adjusts its direction according to the components of the normal vector and continues to move forward;

[0059] S72. When the normal vector of the obstacle surface faced by the robot forms a right angle or an obtuse angle with the target vector, the deflection mechanism will not be able to find a feasible forward path. The robot searches for new topological nodes through scenario replay as sub-goals in the vector navigation process;

[0060] S73. When the target vector is blocked by an obstacle and cannot move forward, switch sub-goals step by step along the replay trajectory and combine the reward information.

[0061] Furthermore, the calculation formula for defining the reward signal using reward cells is:

[0062] R k =1 / (|V k |+1);

[0063] where R k represents the activation value of the reward cell of the k-th scenario node E k and V k represents the vector from E k to the target node.

[0064] The beneficial effects of the present invention are as follows: Based on the hippocampus-inspired robot navigation model, by generating a scenario cognitive map, using grid cell vector calculation, place cell topological connection, and boundary cell obstacle perception, it can navigate efficiently and reliably in complex environments; the proposed algorithm model can correspond to known anatomical and hippocampal structure mechanisms, and can generate different navigation behaviors through various target navigation strategies; thus, it successfully realizes the ability of efficient and robust navigation towards a target in complex environments through the combined strategy of vector navigation and topological navigation.

[0065] Specifically, in the process of imitating the environmental cognition of the hippocampus, a scenario cognitive map is constructed. This scenario cognitive map contains many interconnected scenario neurons. Each scenario neuron maps to a sequence of events, and each event node maps specific perceptual information, including the neural firing information of head direction cells, grid cells, place cells, visual cells, and boundary cells. Since grid cells have metric properties and place cells have topological properties, the scenario cognitive map can be used for both vector navigation and topological navigation. Therefore, it can greatly improve the navigation and obstacle avoidance accuracy of the robot in complex environments, significantly shorten the driving path, and improve efficiency. Compared with existing bionic navigation methods, the present invention constructs a reliable scenario cognitive map and is more applicable; in addition, it combines vector navigation, topological navigation, and segmented navigation strategies, is more flexible, and can continuously switch strategies during navigation to adapt to different scenarios and achieve obstacle avoidance. Brief Description of the Drawings

[0066] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0067] Figure 1 is a flowchart of a method for constructing a complex environment navigation strategy based on grid cells and place cells according to an embodiment of the present invention;

[0068] Figure 2 is the navigation task route of an existing intelligent patrol robot;

[0069] Figure 3 is the topological navigation route map in the hippocampal navigation mode of an existing intelligent patrol robot;

[0070] Figure 4 is the vector navigation route map in the hippocampal navigation mode of an existing intelligent patrol robot;

[0071] Figure 5It is a block diagram of a scenario recognition and navigation system in a method for constructing a complex environment navigation strategy based on grid cells and place cells according to an embodiment of the present invention;

[0072] Figure 6 It is a schematic diagram of a head direction cell model in a simulated cell model in a method for constructing a complex environment navigation strategy based on grid cells and place cells according to an embodiment of the present invention;

[0073] Figure 7 It is a schematic diagram of a grid cell model in a simulated cell model in a method for constructing a complex environment navigation strategy based on grid cells and place cells according to an embodiment of the present invention;

[0074] Figure 8 It is a schematic diagram of a mathematical model of a place cell in a simulated cell model in a method for constructing a complex environment navigation strategy based on grid cells and place cells according to an embodiment of the present invention;

[0075] Figure 9 It is a schematic diagram of the change of the firing mode of the grid cell neural plate from the starting point to the current point during the exploration and learning process of the robot in the calculation of the target navigation vector in a method for constructing a complex environment navigation strategy based on grid cells and place cells according to an embodiment of the present invention;

[0076] Figure 10 It is a schematic diagram of a vector calculation model for the robot to return from the current point to the starting point in the calculation of the target navigation vector in a method for constructing a complex environment navigation strategy based on grid cells and place cells according to an embodiment of the present invention;

[0077] Figure 11 It is a schematic diagram of the information structure of a scenario recognition map in a method for constructing a complex environment navigation strategy based on grid cells and place cells according to an embodiment of the present invention;

[0078] Figure 12 It is a schematic diagram of the structure of a scenario recognition model in a method for constructing a complex environment navigation strategy based on grid cells and place cells according to an embodiment of the present invention;

[0079] Figure 13 It is one of the schematic diagrams of obstacle anticipation-based navigation in a method for constructing a complex environment navigation strategy based on grid cells and place cells according to an embodiment of the present invention;

[0080] Figure 14 It is the second schematic diagram of obstacle anticipation-based navigation in a method for constructing a complex environment navigation strategy based on grid cells and place cells according to an embodiment of the present invention;

[0081] Figure 15 It is a schematic diagram of the motion calculation process of a scenario playback obstacle avoidance strategy in a method for constructing a complex environment navigation strategy based on grid cells and place cells according to an embodiment of the present invention;

[0082] Figure 16 One of the obstacle avoidance schematic diagrams of the scenario replay obstacle avoidance strategy in the complex environment navigation strategy construction method based on grid cells and place cells according to an embodiment of the present invention;

[0083] Figure 17 Another one of the obstacle avoidance schematic diagrams of the scenario replay obstacle avoidance strategy in the complex environment navigation strategy construction method based on grid cells and place cells according to an embodiment of the present invention;

[0084] Figure 18 The third one of the obstacle avoidance schematic diagrams of the scenario replay obstacle avoidance strategy in the complex environment navigation strategy construction method based on grid cells and place cells according to an embodiment of the present invention;

[0085] Figure 19 The fourth one of the obstacle avoidance schematic diagrams of the scenario replay obstacle avoidance strategy in the complex environment navigation strategy construction method based on grid cells and place cells according to an embodiment of the present invention;

[0086] Figure 20 Schematic diagram of the overview of experimental conditions in the complex environment navigation strategy construction method based on grid cells and place cells according to an embodiment of the present invention;

[0087] Figure 21 Schematic diagram of simply using vector navigation in the complex environment navigation strategy construction method based on grid cells and place cells according to an embodiment of the present invention;

[0088] Figure 22 Schematic diagram of using combined navigation in the complex environment navigation strategy construction method based on grid cells and place cells according to an embodiment of the present invention;

[0089] Figure 23 One of the comparison diagrams of the pure topological and combined vector position navigation effects in the complex environment navigation strategy construction method based on grid cells and place cells according to an embodiment of the present invention;

[0090] Figure 24 Another one of the comparison diagrams of the pure topological and combined vector position navigation effects in the complex environment navigation strategy construction method based on grid cells and place cells according to an embodiment of the present invention;

[0091] Figure 25 Schematic diagram of the proportion of the application of two navigation strategies under different exploration densities in the complex environment navigation strategy construction method based on grid cells and place cells according to an embodiment of the present invention;

[0092] Figure 26 Diagram for testing the ability to utilize maze shortcuts in the complex environment navigation strategy construction method based on grid cells and place cells according to an embodiment of the present invention;

[0093] Figure 27 It is one of the test diagrams for the maze environment change adaptation ability in the method for constructing a complex environment navigation strategy based on grid cells and place cells according to an embodiment of the present invention;

[0094] Figure 28 It is the second test diagram for the maze environment change adaptation ability in the method for constructing a complex environment navigation strategy based on grid cells and place cells according to an embodiment of the present invention;

[0095] Figure 29 It is the third test diagram for the maze environment change adaptation ability in the method for constructing a complex environment navigation strategy based on grid cells and place cells according to an embodiment of the present invention;

[0096] Figure 30 It is the first test diagram for finding a target in the open space of the maze in the method for constructing a complex environment navigation strategy based on grid cells and place cells according to an embodiment of the present invention;

[0097] Figure 31 It is the second test diagram for finding a target in the open space of the maze in the method for constructing a complex environment navigation strategy based on grid cells and place cells according to an embodiment of the present invention. Detailed implementation manners

[0098] According to an embodiment of the present invention, a method for constructing a complex environment navigation strategy based on grid cells and place cells is provided.

[0099] Now, the present invention will be further described in conjunction with the accompanying drawings and specific implementation manners. As Figure 1 shown, the method for constructing a complex environment navigation strategy based on grid cells and place cells according to an embodiment of the present invention includes the following steps:

[0100] S1. Based on the environmental cognition process of the hippocampus, construct a scenario cognitive map;

[0101] As Figure 5 shown, it is a structural block diagram of a hippocampal navigation model constructed based on the hippocampus. The model is composed of head direction cells encoding motion information, grid cells decoding target vectors, place cells for spatial position expression, boundary cells for local obstacle avoidance, and the constructed scenario cognitive map. During navigation, different types of obstacles bring different challenges. In the present invention, all spatial information and perception information are expressed based on neural firing.

[0102] The scenario cognitive map is formed by connecting nodes composed of several scenario neurons (a scenario neuron is a nerve cell that generates an activation response for a specific scenario). Each node maps specific perception information, and the perception information represents the neural firing information of the simulated cell model.

[0103] The described simulated cell model includes an attractor model, a grid cell neural plate model, a place cell mathematical model, a boundary cell model, and a visual cell model;

[0104] In the present invention, by introducing a rat as a simulated robot, its behavior is used to demonstrate the role of the hippocampus in the process of the rat's path and direction discrimination. As Figure 6 shown, among them, the head direction cell is a cell whose firing rate reaches a peak when the head direction of the rat approaches a specific direction. By constructing Figure 6 the ring attractor model shown to simulate the encoding mechanism of the head direction cells. Let the head direction of the rat at time t be θ t , the movement speed be v(t), and the phase angle of the i-th head direction cell in the attractor model be θ i . It is set that the movement speed of the rat is proportional to the firing rate of the corresponding head direction cell. Therefore, the following firing rate signals can be set for each head direction cell, and this firing rate signal contains the current head direction angle information and the movement speed information.

[0105] Among them, the attractor model is used to simulate the encoding mechanism of the head direction cells, and the calculation formula is:

[0106] s i (t) = v(t) · cos(θ i - θ t );

[0107] Among them, s i (t) represents the firing rate signal, θ t represents the head direction of the robot at time t, v(t) represents the movement speed, and θ i represents the phase angle of the i-th head direction cell in the attractor model;

[0108] The place cell mathematical model is used to simulate the encoding mechanism of the place cells. The place cells are a type of firing cells that are selective for spatial positions. Only when the rat is at a specific position in space, this cell will generate a firing activity, and no firing activity will occur at other positions in space. Therefore, the place cells establish a one-to-one correspondence between the neurons in the brain region and the real physical world. A place cell neural plate arranged on a two-dimensional plane is constructed, and the population firing of the place cells presents a unimodal mode, mapping the spatial position where the rat is located. The place cell mathematical model proposed by O'Keefe et al. is used to calculate the firing rate of each place cell, and the calculation formula is:

[0109]

[0110] Among them, represents the firing rate of the place cell i at the position r, r = [x, y] represents the position coordinates of the current robot in the environment, ri0 represents the position coordinates corresponding to the discharge center of place cell i, δ 2 represents the adjustment coefficient of place cell discharge;

[0111] The boundary cell model is used to simulate the coding mechanism of boundary cells. When the rat approaches obstacles in different directions, different boundary cells will be excited. The firing rate of boundary cells will increase as the distance of the rat from the obstacle decreases. In the constructed boundary cell model, the receptive field of each boundary cell is constructed by multiplying two Gaussian functions, one function represents the distance and the other function represents the non-egocentric direction. The calculation formula is:

[0112]

[0113] Among them, B i is the firing rate of the border cell i, r represents the current position of the border cell, d i is the radius of boundary perception, θ represents the direction of the current boundary cell, is the main direction of boundary perception, σ rad , σ ang are the adjustment factors for radius and angle response respectively.

[0114] The grid cell neural plate model is used to simulate the encoding mechanism of grid cells, and includes the following steps:

[0115] S11. Grid coding, for example, when a rat moves in a flat space, grid cells discharge repetitively and regularly at specific locations. Figure 7 The figure shows a grid cell neural plate model. The grid cells are evenly arranged in the neural plate. Each grid cell will produce periodic neural discharges as the spatial position changes. Since the periodic phases of different grid cells are uniformly offset, when the rat is in a certain spatial position, the grid cell group discharge will present a grid-like pattern. When the rat moves, the grid discharge mode also moves accordingly. The activation value of a single grid cell is calculated as follows:

[0116]

[0117] in, represents the activation value of a single grid cell, κ represents the waveform adjustment coefficient, and n max represents the peak discharge rate, represents the spatial displacement, represents the wave vector direction of each fringe;

[0118] S12. Grid decoding. For example, grid cells on each grid cell neural plate have the same scale λ, but different spatial phases. To decode the position of the rat through the firing activities of the grid cell population, record the firing rate n of grid cell i at the current position i , so that the response vector of the neural plate population When the rat is at The average value of the firing rate of each neuron Assume the firing rate of grid cell j Follows a Poisson distribution, and each neuron is statistically independent. Therefore, for the probability of the firing rate vector of the population composed of M cells at the phase The calculation formula of the probability is as follows:

[0119]

[0120] Among them, Represents the probability of the firing rate vector of the population composed of M cells at the spatial displacement , Represents the firing rate of grid cell j, Represents the response vector of the neural plate population, n j Represents the firing rate of grid cell j at the current position;

[0121] S13. Multi-scale vector calculation. During the vector navigation process, the next movement direction of the robot is mainly obtained by grid cell calculation. As Figure 9 The shown calculation model calculates the vector between two respective positions through the firing inputs of two independent grid cells. One represents the current position and the other represents the target position. According to the above formula, the larger the grid scale λ, the lower the accuracy of position estimation but the fewer the number of solutions, and the smaller the λ, the higher the accuracy of the position solution but the more the number of solutions. To obtain a displacement position solution with high accuracy, it is necessary to implement it through the method of multi-scale joint decoding. As Figure 10 Shown, stack the grid cell neural plates of each scale into a pyramid geometric model, map the position decoding obtained from the grid firing information of the robot onto a virtual motion vector for expression, and then obtain the direction vector towards the target position by taking the difference between the target motion vector and the current motion vector.

[0122] The vector to the target position can be calculated according to the following steps: (1) On the neural plate with the grid scale ratio λ 0 , through the probability distribution of Eq. (10), combine the maximum likelihood probability to solve for the roughest position estimate x 0 , and accordingly make a motion vector from the pole to this neural plate and extend it to each neural plate; (2) Solve for the position estimate x 1 On the neural plate with the grid scale ratio λ 1, there will be several solutions, that is, each hexagonal lattice has one solution. Find the position estimate closest to the motion vector and correct the direction of the motion vector to this closest position estimate point; (3) Repeat the method in the second step to correct the direction of the motion vector on the neural plate at scales λ 2 , λ 3 ... Finally, the obtained motion vector P 0 is projected onto the place cells as the peak firing position of the place cells; (4) Similarly, the motion vector P g of the target position can also be obtained through the target grid firing module. The direction vector T for target-oriented navigation is T = P g - P 0 .

[0123] In addition, each scenario node of the scenario cognitive map maps four types of information of cognitive memory: topological information (composed of the synaptic connection weight matrix of place cells), scale information (composed of the activation information of grid cells), perceptual information (mainly visual landmark information, boundary information), and reward information (mainly the reward cell information for reaching the target). The types and functions of various cells will be explained in detail below.

[0124] As Figure 11 shown (in the figure, the letters E, P, G, H, V, B, and R represent scenario cells, place cells, grid cells, head direction cells, visual cells, boundary cells, and reward cells respectively. The place cell layer is horizontally connected, representing the connectivity between scenario nodes), it is the hierarchical relationship of functional cells contained in scenario cognition. Each scenario cell E serves as the identifier of a specific scenario in the environment, without topological information or scale information, and its neural firing rate represents the activation intensity; each place cell P serves as the identifier of a specific spatial position, and the synaptic connections between cells represent the topological connection relationship between scenarios; the grid-like mode formed by the periodic firing of grid cells implies spatial scale information, which can integrate the movement path to obtain spatial cognition and can also calculate the vector direction and distance between any two spatial positions; the population firing of head direction cells represents the movement direction, that is, the pose information of the robot; visual landmark cells are the identifiers of specific visual targets, containing the spatial distribution information of the landmarks; boundary cells B serve as the perception of obstacle information, which are eight boolean quantities, providing necessary support for the robot's path planning and navigation;

[0125] The topological relationship between scenario nodes can be described by the adjacency matrix M, and each element M ij in the matrix represents the synaptic connection strength from scenario node i to j.

[0126] S2. Construct a scenario cognitive model using the adaptive resonance theory network;

[0127] Among them, as Figure 12As shown, the construction of the scenario cognitive model using the adaptive resonance theory network includes the following steps:

[0128] S21. The nerve cells encode the environmental perception information into a neuron activation vector as the input of the model.

[0129] S22. Apply the adaptive resonance theory (ART) network for unsupervised self-learning of the information from the input layer to the event layer. After learning, the activation value of the event cells gradually decays over time.

[0130] S23. Use the event activation sequence greater than the threshold as the input and apply the adaptive resonance theory (ART) network for self-learning again. Combining the winner-takes-all rule, activate or newly generate a specific scenario cell, and use the activation state of this specific scenario cell as a node of a scenario cognitive map.

[0131] In addition, to simplify the model, map the spatial position of the robot scenario node to the place cells. When the spatial position and perception information (such as vision) of the robot change to a certain extent during movement, new cognitive nodes will be created. If similar input information is encountered again, the previously generated scenario nodes can be activated again. Each scenario node remembers the firing information perceived by the head direction cells, grid cells, place cells, boundary information, and visual cells at that time, and this memory information can be replayed and retrieved in the future.

[0132] When the robot transfers from one scenario node to another, form a bidirectional connection between the scenario nodes, corresponding to a reinforcement learning of the synapses between the activated scenario cells. The obtained scenario cognitive map reflects the topological structure and spatial metric information of the environmental scenario. Based on this, the shortest topological path between any starting position unit and target position unit can be calculated. The robot can achieve the target navigation behavior by moving straight while always facing the adjacent scenario nodes located on this shortest topological path.

[0133] S3. Perform loop adjustment, topological connection, and time-varying forgetting of the scenario cognitive process.

[0134] Among them, the loop adjustment, topological connection, and time-varying forgetting of the scenario cognitive process include the following steps:

[0135] S31. If existing nerve cells are reactivated during the scenario cognitive process, it is determined that a loop of the movement trajectory is detected, and error correction is performed to prevent the accumulation of errors in path integration.

[0136] S32. Update the weights from the time cells to the scenario cells, reset the firing mode of the grid cells, and the reset amplitude is proportional to the event activation value.

[0137] S33. Achieve the topological relationship on the scenario map by creating lateral connections between place cells;

[0138] S34. Update the lateral topological connections of the topological layer by introducing synaptic connections between scenario neurons.

[0139] In addition, let the reset coefficient be μ ∈ (0, 1), and the event i activation vector y corresponding to the current scenario i , the phase deviation between the current scenario node and the scenario node in memory Then the grid phase reset vector Δ corresponding to each event neuron i ;

[0140]

[0141] Achieve the topological relationship on the scenario map by creating lateral connections between place cells. Each newly generated scenario cell p i is associated with a sequence of events, and each event includes a grid cell firing pattern, a place cell firing module, and a reward cell activation signal r i . The activation value y of each scenario cell is related to the length of time the robot has most recently activated this scenario. As long as the robot is currently in a certain scenario node, the corresponding event activation signal y remains at 1, otherwise it decays slowly, and the decay function is as follows:

[0142] y i (Δt) = e -t / τ ；

[0143] Each time the robot visits the scenario cell E i , the lateral connections of the place cells are strengthened through Hebbian update, and the weight matrix W u is updated using the following formula:

[0144] W u = W u ∪ H(q T - δ)*H(q - 1);

[0145] where q is the vector of cell activation values in the place cell neural plate under the current scenario, δ is the place cell activation threshold, H is the unit step function, and ∪ is the "or" operator for each element. The above formula updates the lateral topological connections of the topological layer by introducing or strengthening the synaptic connections between the most recently activated scenario cells. The threshold δ determines the critical value at which the scenario neurons are considered to be in an activated state.

[0146] S4. Establish a navigation model by combining the topological attributes in the scenario cognitive map with the metric attributes of vector calculation;

[0147] Among them, topological navigation does not require calculating the vector direction to the target location and always points to the next scenario node in the topological path during navigation. Topological navigation is reliable, but due to relying on experience, it cannot take some shortcuts during navigation. Vector navigation is the fastest, but often gets blocked by obstacles. However, combining the topological attributes of the scenario cognitive map and the metric attributes of vector calculation can produce a more powerful navigation function. In the scenario cognitive map, each scenario node is mapped to the grid cell firing mode in that scenario, so any topological node can be the starting point or ending point of vector navigation. Such a feature can help the robot overcome obstacles by selecting more appropriate scenario nodes as its sub-goals. It is reported that hippocampal replay occurs when rats stop at a path selection point or rest during a maze run. The replay phenomenon is characterized by rapid firing activity of hippocampal neurons, and the firing trajectory repeats the trajectory of early place cell activity. The hippocampal scenario replay mechanism can be used to sample possible sub-goals among scenario nodes to allow the robot to adjust its current target node.

[0148] S5. Navigate using the navigation model;

[0149] The navigation using the navigation model includes the following steps:

[0150] S51. After determining the final target scenario node, search for all possible paths according to the topological information of the scenario cognitive map;

[0151] S52. Assign a high-frequency firing period to the grid cells to drive the grid mode to traverse along the possible topological paths, and select the one with the shortest time consumption as the planned basic path to complete the simulation deduction;

[0152] S53. During the navigation process, before departing from the position of each scenario node, if the border cells of the current node are in the activated state, move towards the next topological node; if the border cells are not activated, obtain the vector navigation direction through the vector calculation of the grid cell mode, and the robot moves towards the final target direction;

[0153] S54. During the vector navigation process, if a new obstacle is encountered, use the scenario replay mechanism to find sub-goals in the order of the gradient descent of the reward cell activation value and calculate the vector direction until a vector direction of a sub-goal can lead the robot away from the obstacle. After leaving the obstacle area, continue the vector navigation towards the final target.

[0154] S6. Perform segmented obstacle avoidance based on obstacle information;

[0155] Among them, as Figure 13 - 14 shown, the segmented obstacle avoidance based on obstacle information includes the following steps:

[0156] S61. Cluster and process the boundary recognition information, convert it into obstacle recognition, and use it as the basis for navigation decision-making;

[0157] S62. After obtaining the topological path through simulation and deduction, predict the existence of obstacles based on the activation information of boundary cells in the scenario nodes;

[0158] S63. Cut the planned topological path into three segments. The segment before the obstacle is the first segment, the segment near the obstacle is the second segment, and the segment after the obstacle is the third segment. Then navigate according to the navigation model in sequence.

[0159] In addition, Figure 14 The navigation effect based on the navigation model without boundary information is shown. The robot first navigates according to vector navigation. When encountering an obstacle in the figure, it gradually bypasses the obstacle through the guidance of sub-goals and finally reaches the target position. Through comparison, it can be seen that the strategy based on obstacle prediction can reduce the navigation path. However, when the obstacle is relatively small, this method may increase the navigation path. Therefore, the use of this strategy should be selected according to the actual situation.

[0160] S7. Avoid obstacles based on the scenario replay mechanism;

[0161] Among them, as Figure 15 - 19 shown, the obstacle avoidance based on the scenario replay mechanism includes the following steps:

[0162] S71. During vector navigation, when the normal vector of the obstacle surface faced by the robot forms an acute angle with the target vector, the robot adjusts its direction according to the component of the normal vector and continues to move forward;

[0163] S72. When the normal vector of the obstacle surface faced by the robot forms a right angle or an obtuse angle with the target vector, the deflection mechanism will not be able to find a feasible forward path. The robot searches for a new topological node through scenario replay as a sub-goal during the vector navigation process;

[0164] S73. When the target vector is blocked by an obstacle and cannot move forward, then switch sub-goals step by step along the replay trajectory and combine the reward information.

[0165] In addition, the specific obstacle avoidance mechanism is as Figure 15 shown. Whenever a robot newly activates or generates a scenario node, the navigation controller calculates the vector to the target according to the vector algorithm, calculates the normal vector of the obstacle boundary according to the activation information of the boundary cells included in the current scenario node, and calculates the vector from the current position to the sub-goal. The target vector minus the boundary vector plus the sub-goal vector is used as the motion vector to control the movement of the robot's chassis. Initially, the robot is as Figure 16 shown and directly navigates towards the target position. When encountering an obstacle, calculate the vectors from each scenario to each sub-goal in sequence according to the gradient of the activation value of the reward cells, and according to Figure 15The computing mechanism in it calculates the motion vector, and when the included angle between the motion vector and the boundary vector is greater than the threshold, the robot executes the motion. Figure 16 、 17 In 18, the red circles are the sub-goals of each step of obstacle avoidance. Once the robot does not detect the activation information of the boundary cells (gets rid of the obstacle) during the obstacle avoidance process, it immediately returns to the vector navigation mode and directly moves towards the final goal, as Figure 19 shown.

[0166] S8. Use the reward cells to define the reward signal.

[0167] There are two important limitations in the robot's navigation method based on scenario awareness: (1) Due to the existence of cumulative errors, when the robot navigates to the recognized target position, it may still be a certain distance from the real target, forcing the robot to conduct random exploration to find the target; (2) When conducting topological path planning before navigation, in the face of numerous topological combinations, how to plan the fastest path; (3) When selecting sub-goals, when there are multiple optional scenario nodes, an evaluation mechanism is needed for decision-making.

[0168] One way to address these issues is to set each scenario node according to the reward signal gradient. The first problem can be solved by using the uphill method based on the reward information value for sub-goals. The second problem can be solved by the downhill method based on the reward information value or the optimal path. The third problem can be solved by finding the sub-goal closer to the final goal according to the magnitude of the reward signal. The prerequisite for achieving these is to let each scenario node have a corresponding reward signal, and this reward signal can reflect the distance to the final goal.

[0169] Among them, the calculation formula for using the reward cells to define the reward signal is:

[0170] R k = 1 / (|V k | + 1);

[0171] Among them, R k represents the activation value of the reward cells of the k-th scenario node E k , and V k represents the vector from E k to the target node.

[0172] The following shows the simulation results of the scenario navigation model of the present invention. First, it demonstrates successful navigation in a rescue scenario, then analyzes the characteristics of the environment suitable for the combined navigation mode, and finally demonstrates how this navigation mode is flexible enough to solve various navigation problems. It is assumed that the robot moves at a constant speed and it also has the ability to detect and avoid obstacles in the environment through ultrasonic ranging sensors. If the distance between the obstacle and the robot is less than 5 cm, it is regarded as an obstacle; if there is no obstacle within 5 cm, it is considered open. Throughout the simulation experiment, the speed of the virtual robot is constant at 50 mm / s. In addition, in the simulation experiment, the robot samples every 5 cm of movement, and after each sampling, scenario learning is carried out while calculating the target vector direction.

[0173] Example 1

[0174] Advantage test of combined navigation compared with vector navigation:

[0175] Figure 20 Shows the experimental conditions for testing the navigation ability. In a large open space, variously shaped obstacles are arranged around the location of the rescue target point. Each trial includes a training phase and a testing phase, and the robot is initially located at the rescue point. The robot first follows a given outbound path (the white curve in the figure) from this rescue point and continuously constructs new scenario nodes and forms a scenario cognitive map. Then, it moves along the edge arc (the gray dotted line) to a given starting position (the triangular position in the figure, a total of 12 starting positions are set), and then navigates back to the target rescue point based on the scenario cognitive map. To evaluate the robustness of the robot's navigation ability under different conditions, these starting positions are distributed along the perimeter.

[0176] Figure 21 Gives the simulation results of simply using the vector navigation mode. The tests show that the robot successfully reaches the rescue point from 6 of the tested starting positions, and the other 6 starting points are blocked by obstacles and fail.

[0177] Figure 22 Gives the test results of the combined navigation methods combining vector navigation, topological segmentation, scenario playback, etc. When the robot encounters a new obstacle, it starts scenario playback to find a new sub-goal and finally reaches the rescue position. The tests show that all robots starting from the starting points finally successfully navigate to the target.

[0178] Example 2

[0179] Advantages of combined navigation compared with topological navigation:

[0180] Dense exploration ( Figure 23 ) and sparse exploration ( Figure 24 ) are respectively used to test the advantages of combined navigation in the environment. Figure 25Describes a quantitative comparison of the navigation behavior of the robot in these configurations, showing the average length of the path required to return to the target position in 30 trials.

[0181] Dense exploration and sparse exploration refer to the degree of the robot's cognitive understanding of the environmental scenario before the navigation experiment. After dense exploration, a shorter topological path can be obtained only by relying on topological navigation without encountering obstacles ( Figure 24 ). However, when the environment is sparsely explored, obstacles can also be avoided based on the existing scarce environmental knowledge ( Figure 23 ), but it does not utilize any shortcuts, resulting in low efficiency (longer path). In a densely explored environment, the navigation performance of the combined vector robot is not much better than that of topological navigation, but it has obvious advantages in a sparsely explored scenario. The vector navigation function based on the scenario cognitive map enables the robot to quickly cross the initially unexplored open space. The quantitative comparison results ( Figure 25 ) show that the combined navigation is more efficient than topological navigation, especially in a sparsely explored environment.

[0182] Embodiment 3

[0183] Several typical maze experiments:

[0184] The first maze experiment demonstrates the ability of the navigation model to discover and utilize shortcuts never experienced in the environment. In these experiments, the environment is cognitively learned according to the Figure 26 trajectory each time, and then five tests are carried out in the maze. As Figure 26 shown, a specific shortcut is opened in the maze each time for testing whether the robot can utilize the shortcut.

[0185] The second maze is a communicating vessel. Different positions are blocked respectively as Figure 27 - 29 shown. The robot plans the topological path through re-scenario deduction, and guides the robot to bypass the obstacles to reach the target through the scenario playback mechanism.

[0186] The first two mazes are both navigations in narrow channels. The third maze is constructed as an Figure 30 - 31 open maze environment as shown. The black dotted line in the figure is the trajectory in the exploration stage, and the light-colored spheres are the topological paths planned by the robot. Figure 30 In [reference number], the exploration path only passed through 1 checkpoint, Figure 31 in [reference number], the exploration path passed through 2 checkpoints. The experimental results show that the hidden target (at the dark-colored sphere) was successfully found in both experiments. In addition, Figure 31 in the experiment of [reference number], due to more sufficient exploration, a shorter topological path was obtained during path deduction, and the finally navigated path was better than that of Figure 30 the experiment of [reference number].

[0187] In summary, by means of the above technical solutions of the present invention, the robot navigation model inspired by the hippocampus can generate a scenario cognitive map, and through grid cell vector calculation, place cell topological connection, and boundary cell obstacle perception, it can navigate efficiently and reliably in complex environments. The proposed algorithm model can correspond to the known anatomical and hippocampal structure mechanisms, and different navigation behaviors can be generated through various target navigation strategies. Thus, the combined strategy of vector navigation and topological navigation successfully realizes the ability to navigate efficiently and robustly towards the target in complex environments.

[0188] Specifically, in the process of imitating the environmental cognition of the hippocampus, a scenario cognitive map is constructed. This scenario cognitive map contains many interconnected scenario neurons. Each scenario neuron maps to a sequence of events, and each event node maps specific perceptual information, including the neural firing information of head direction cells, grid cells, place cells, visual cells, and boundary cells. Since grid cells have metric properties and place cells have topological properties, the scenario cognitive map can be used for both vector navigation and topological navigation. Therefore, it can greatly improve the navigation and obstacle avoidance accuracy of the robot in complex environments, greatly shorten the driving path, and improve efficiency. Compared with existing bionic navigation methods, the present invention constructs a reliable scenario cognitive map and is more applicable. In addition, it combines vector navigation, topological navigation, and segmented navigation strategies, is more flexible, and can continuously switch strategies during navigation to adapt to different scenarios and achieve obstacle avoidance.

[0189] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. Method for constructing complex environment navigation strategy based on grid cells and place cells, characterized in that, this method comprises the following steps: S1. Based on the environmental cognition process of the hippocampus, construct a scenario cognition map; S2. Use an adaptive resonance theory network to construct a scenario cognition model; S3. Perform loop adjustment, topological connection and time-varying forgetting of the scenario cognition process; S4. Combine the topological attributes in the scenario cognition map with the metric attributes of vector calculation to establish a navigation model; S5. Use the navigation model for navigation; S6. Perform segmented obstacle avoidance based on obstacle information; S7. Perform obstacle avoidance based on the scenario replay mechanism; S8. Use reward cells to define reward signals; The navigation using the navigation model comprises the following steps: S51. After determining the final target scenario node, search for all possible paths according to the topological information of the scenario cognition map; S52. Endow the grid cells with high-frequency firing cycles to drive the grid mode to traverse along the possible topological paths, and select the one with the shortest time consumption as the basic path for planning to complete the simulation deduction; S53. During the navigation process, before starting from the position of each scenario node, if the boundary cells of the current node are in the activated state, move towards the next topological node; if the boundary cells are not activated, obtain the vector navigation direction through the vector calculation of the grid cell mode, and the robot moves towards the final target direction; S54. If a new obstacle is encountered during the vector navigation process, use the scenario replay mechanism to find sub-goals in the order of the gradient descent of the activation value of the reward cells and calculate the vector direction until a vector direction of a sub-goal can take the robot away from the obstacle. After leaving the obstacle area, continue the vector navigation towards the final target; The segmented obstacle avoidance based on obstacle information comprises the following steps: S61. Cluster and process the boundary cognition information, convert it into obstacle cognition, and use it as the basis for navigation decision-making; S62. After the topological path obtained from the simulation deduction, predict the existence of obstacles according to the activation information of the boundary cells in the scenario nodes; S63. Cut the planned topological path into three segments, the front of the obstacle as the first segment, the vicinity of the obstacle as the second segment, and the back of the obstacle as the third segment, and then navigate according to the navigation model in turn; The obstacle avoidance based on the scenario replay mechanism comprises the following steps: S71. During vector navigation, when the normal vector of the obstacle surface faced by the robot forms an acute angle with the target vector, the robot adjusts its direction according to the component of the normal vector and continues to move forward; S72. When the normal vector of the obstacle surface faced by the robot forms a right angle or an obtuse angle with the target vector, the deflection mechanism will not be able to find a feasible forward path, and the robot searches for a new topological node as a sub-goal in the vector navigation process through scenario replay; S73. When the target vector is blocked by an obstacle and cannot move forward, switch sub-goals step by step along the replay trajectory and combine the reward information.

2. The method for constructing a complex environment navigation strategy based on grid cells and place cells according to claim 1, characterized in that, the scenario cognition map is formed by connecting nodes composed of a number of scenario neurons, each node maps specific perception information, and the perception information represents the neural discharge information of the simulated cell model.

3. The method for constructing a complex environment navigation strategy based on grid cells and place cells according to claim 2, characterized in that, the simulated cell model includes an attractor model, a grid cell neural plate model, a place cell mathematical model, a boundary cell model and a visual cell model; wherein, the attractor model is used to simulate the encoding mechanism of head direction cells, and the calculation formula is: si(t) = v(t)·cos(θi - θt); Among them, s i (t) represents the discharge rate signal, and θ t represents the heading of the robot at time t, v(t) represents the movement speed, and θ i represents the phase angle of the i-th heading cell in the attractor model; the place cell mathematical model is used to simulate the encoding mechanism of place cells, and the calculation formula is: Among them, represents the firing rate of place cell i at position r, where r = [x, y] represents the position coordinates of the current robot in the environment, and r i0 represents the position coordinates corresponding to the firing center of place cell i, and δ 2 represents the place cell firing adjustment coefficient; the boundary cell model is used to simulate the encoding mechanism of boundary cells, and the calculation formula is: Among them, B i is the firing rate of the border cell i, r represents the position of the current border cell, d i is the radius of border perception, θ represents the direction of the current border cell, is the main direction of border perception, σ rad and σ ang are the adjustment coefficients of the radius and angle responses respectively.

4. The method for constructing a complex environment navigation strategy based on grid cells and place cells according to claim 3, characterized in that, the grid cell neural plate model is used to simulate the encoding mechanism of grid cells, including the following steps: S11. Grid encoding, and the calculation formula is: Among them, represents the activation value of a single grid cell, κ represents the waveform adjustment coefficient, and n max represents the peak discharge rate, represents the spatial displacement, represents the wave vector direction of each stripe; S12. Grid decoding, and the calculation formula is: Among them, represents the probability of the population firing rate vector composed of M cells undergoing spatial displacement . represents the firing rate of grid cell j, represents the response vector of the neural plate population, n j represents the firing rate of grid cell j at the current position; S13. Multi-scale vector calculation, stacking the grid cell neural plates of each scale into a pyramid geometric model, mapping the position decoding obtained from the grid discharge information of the robot onto a virtual motion vector for expression, and then obtaining the direction vector towards the target position by taking the difference between the target motion vector and the current motion vector.

5. The method for constructing a complex environment navigation strategy based on grid cells and place cells according to claim 4, characterized in that, constructing a situation awareness model using an adaptive resonance theory network includes the following steps: S21. Nerve cells encode environmental perception information into a neuron activation vector as the input of the model; S22. Apply the adaptive resonance theory network for unsupervised self-learning of the information from the input layer to the event layer, and the activation value of the learned event cells gradually decays over time; S23. Use the event activation sequence greater than the threshold as the input and apply the adaptive resonance theory network for self-learning again. Combining the winner-takes-all rule, activate or newly generate a specific situation cell, and use the activation state of this specific situation cell as a node of a situation awareness map.

6. The method for constructing a complex environment navigation strategy based on grid cells and place cells according to claim 5, characterized in that, performing loop adjustment, topological connection and time-varying forgetting in the situation awareness process includes the following steps: S31. If existing nerve cells are reactivated during the situation awareness process, it is determined that a loop of the motion trajectory is detected, and error correction is performed to prevent the accumulation of errors in path integration; S32. Update the weights from time cells to situation cells, reset the grid cell anti-electricity mode, and the reset amplitude is proportional to the event activation value; S33. Create horizontal connections between place cells to achieve the topological relationship on the situation map; S34. Update the horizontal topological connection of the topological layer by introducing synaptic connections between situation neurons.

7. The method for constructing a complex environment navigation strategy based on grid cells and place cells according to claim 6, characterized in that, the calculation formula for defining the reward signal using reward cells is: R k = 1 / (|V k | + 1); Among them, R k represents the activation value of the reward cell of the k-th scenario node E k and V k represents the vector from E k to the target node.

Images