Visual clue-free brain-like SLAM (Simultaneous Localization and Mapping) method based on grid cells

Abstract

The invention relates to a non-visual clue brain-like SLAM (Simultaneous Localization and Mapping) method based on grid cells, which refers to a neural mechanism of the grid cells in a mammalian brain olfactory cortex, tracks and encodes the position of a robot only by utilizing self-motion information of the robot, and constructs a cognitive map of the robot and an interaction environment of the robot. The method comprises the following steps: firstly, constructing a multi-level grid cell continuous attractor network module, and encoding self-motion information of a mobile robot into cognitive representation of a grid cell module through a cognitive space transformation module; and then constructing a cognitive map in a navigation environment without visual clues by utilizing the cognitive representation of the grid cell modules. The invention provides a robust and efficient brain-like SLAM navigation method without visual clues, which can be used for intelligent autonomous positioning and navigation of the mobile robot in satellite denial and visual clue-free environments, and has important significance for expanding the application scene of the mobile robot.

Description

Technical Field

[0001] The invention relates to a visual clue-free brain-like SLAM method based on grid cells, belongs to the field of navigation positioning and artificial intelligence, and is a method for navigation and positioning of a robot in an environment without visual clues. Background Art

[0002] When a robot moves in an unknown environment to perform tasks, it is crucial to estimate its position in the environment. Simultaneous localization and mapping (SLAM) is one of the most promising and practical strategies to solve this problem. SLAM is a process in which a robot uses sensor information to model the environment, draw an environmental map, and simultaneously estimate its position in the environmental map while moving in an unknown environment. Although current visual SLAM methods have achieved some results, these methods still perform poorly in dynamic and unstructured environments, especially in environments without visual cues. Visual SLAM cannot effectively extract image features, resulting in navigation and mapping failure. Summary of the invention

[0003] In response to the above problems, the present invention proposes a visual clue-free brain-like SLAM method based on grid cells. This method draws on the neural dynamics mechanism of grid cells in the entorhinal cortex of the mammalian brain to encode the robot's self-motion information and solve the robot's navigation vector. First, a multi-level grid cell continuous attractor network module is constructed to encode the mobile robot's self-motion information into the cognitive representation of the grid cell module; then, the cognitive representation of these grid cell modules is used to construct a cognitive map in the navigation environment without visual clues.

[0004] The technical solution adopted by the present invention to achieve the above-mentioned purpose is:

[0005] A visual clue-free brain-like SLAM method based on grid cells comprises the following steps:

[0006] 1) Construct a multi-level grid cell module network and initialize it;

[0007] 2) Receive the robot’s velocity information through the robot’s sensor and decompose it at multiple scales;

[0008] 3) Input the decomposed speed into the continuous attractor network of the corresponding level according to the scale;

[0009] 4) The continuous attractor network at each level moves dynamically according to speed information and time;

[0010] 5) When the state of each level in the grid cell module network reaches a steady state, determine whether there is a step situation in the attractor network of each level. If so, perform a level step transfer on the entire grid cell module network. Otherwise, execute step 6);

[0011] 6) Perform multi-scale decoding on the entire grid cell module network to obtain the robot position encoded by it;

[0012] 7) Repeat steps 2) to 6) until the robot stops moving and obtains the complete movement trajectory of the robot.

[0013] The multi-level grid cell module network is formed by coupling multiple single-level two-dimensional continuous attractor networks, and the dynamic equation of each single-level two-dimensional continuous attractor network is characterized as follows:

[0014]

[0015]

[0016]

[0017] Among them, u is the firing rate of the neuron in the network, θ is the preferred position of the neuron, t is time, e is the external input stimulus of the neuron, w is the weight between the neuron at the preferred position θ and the neuron at the preferred position θ′ in the network, r is the normalized neuron firing rate, τ is the time constant of the neuron response, ρ is the degree to which the neuron is affected by other neurons in the network, a is a hyperparameter that controls the range of interaction between neurons, w0 is the scale factor, and k is a hyperparameter.

[0018] The step 2) is specifically as follows:

[0019] The robot's speed information is decomposed according to the number of layers of the multi-level grid cell module network, namely:

[0020]

[0021] Among them, v t represents the speed received by the robot at time t, Indicates speed v t The i-th component of , M is the number of levels of velocity decomposition, i.e., the number of levels of grid modules, and e represents the difference ratio between adjacent velocity components. The difference ratio between different modules is determined by the power of e i Indicates that b is a positive integer used to characterize the range of speed levels.

[0022] The step 4) comprises the following steps:

[0023] 4.1) For the i-th level of the grid cell module network, convert the speed received by this level into the grid cell cognitive space;

[0024] 4.2) Decomposing the velocity converted to the grid cell cognitive space in the x-direction and the y-direction;

[0025] 4.3) Dynamically move in the x direction of the grid cell cognitive space according to the velocity component in the x direction;

[0026] 4.4) Dynamically move in the y direction of the grid cell cognitive space according to the velocity component in the y direction.

[0027] The step 4.1) is specifically as follows:

[0028]

[0029] Among them, v i is the decomposed velocity component, It is the speed information after being converted into the grid cell cognitive space. is the change matrix, expressed as:

[0030]

[0031] Among them, s i is the step size of the grid cells in the i-th layer, θ i Indicates the orientation of the grid cells.

[0032] The step 4.3) is specifically as follows:

[0033] The neuron state update in the network can be expressed as:

[0034]

[0035] in is the neuron state in the updated network, is the normalized neural state of the network at time t, n is the number of neurons in the network along the x-direction or y-direction, and the weight matrix W′ is:

[0036]

[0037] It is described as:

[0038]

[0039] B is the basic circulant matrix:

[0040]

[0041] Δt is the interval time, is the number of neurons in the x direction,

[0042]

[0043] s i is the step size of the grid cells at this level;

[0044] Update the normalized network neural state, that is:

[0045]

[0046] Among them, μ is a hyperparameter, indicating the degree of normalization, is a diagonal matrix constructed from neuron states, which can be described as:

[0047]

[0048] The step 5) is specifically as follows:

[0049] When a continuous attractor network at a certain level receives speed information that causes its network activity to reach its network epidemic boundary, it is considered that there is a step in the continuous attractor network at the current level. The activity position of the continuous attractor network at this level is reset to 0, and its accumulated activity is transferred to the next level, and the network at the previous level is moved by 1 position according to the direction of movement.

[0050] In step 6), the position P of the robot after decoding t for:

[0051]

[0052] in, is the position encoded by the i-th module at time t, and M is the number of grid cell modules.

[0053] A visual clue-free brain-like SLAM system based on grid cells comprises a memory and a processor; the memory is used to store a computer program; the processor is used to implement the visual clue-free brain-like SLAM method based on grid cells when executing the computer program.

[0054] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the grid cell-based visual cue-free brain-like SLAM method is implemented.

[0055] The present invention has the following beneficial effects and advantages:

[0056] 1. The method of the present invention draws on the spatial navigation mechanism of mammalian grid cells and adopts a neurodynamic approach to model grid cell modules, which has a high biological similarity and is conducive to promoting the development of brain-like intelligence.

[0057] 2. The present invention couples multiple single grid cell continuous attractor network modules to form a multi-scale encoded grid cell module. Through dynamic transfer of unified levels and information transmission between levels, it can accurately characterize the movement trajectory of the robot and reduce the cumulative error.

[0058] 3. The present invention draws on the characteristic of grid cells that can be released in dark environments. By embedding the grid cell module into an actual robot platform, compared with the classic RatSLAM algorithm, it can achieve intelligent autonomous positioning and mapping in satellite-denied and visually clue-free environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 The framework diagram of the brain-like SLAM system based on grid cells proposed in the present invention;

[0060] Figure 2 The multi-level grid cell module structure diagram constructed by the present invention;

[0061] Figure 3 Schematic diagram of the dynamic movement of a single-layer grid cell module over time;

[0062] Figure 4 Comparison diagram of cognitive map constructed by the proposed method on Kitti dataset and its ground truth;

[0063] Figure 5 Schematic diagram of the cumulative error of the proposed method on the Kitti dataset;

[0064] Figure 6 Comparison of cognitive maps constructed by two methods in the Maze-1 scenario;

[0065] Figure 7 Comparison of cognitive maps constructed by two methods in the Maze-2 scenario. DETAILED DESCRIPTION

[0066] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0067] like Figure 1As shown, the present invention discloses a brain-like SLAM method without visual cues based on grid cells, which is used for intelligent autonomous positioning and mapping in satellite denial and visual clue-free environments, and belongs to the field of navigation positioning and artificial intelligence. Drawing on the spatial cognitive neural mechanism of grid cells in scenes without visual cues, the grid cells are modeled in a neurodynamic way, the self-motion information of the robot is received, and a situational memory cognitive map that can characterize the interaction between the robot and the environment is output. First, a continuous attractor network model of a single grid cell module based on spatial transformation is constructed, and the self-motion information of the robot is spatially transformed into the grid cell cognitive space for encoding; secondly, by coupling multiple single grid cell modules, a multi-scale grid cell module is formed, and the dynamic transfer of the unified level and the information transmission between levels can accurately characterize the movement trajectory of the robot; finally, the present invention draws on the characteristics of grid cells that can be released in an environment without visual cues, decodes the spatial representation of the grid cell module, and outputs the robot's cognitive map in the environment.

[0068] The specific implementation of the present invention is as follows:

[0069] 1. Construct a multi-level grid cell module network, each level is a two-dimensional continuous attractor network, and initialize the entire grid cell module network;

[0070] The dynamic equation of the two-dimensional continuous attractor network at each level is characterized as:

[0071]

[0072] Where u is the firing rate of the neuron in the network, θ is the preferred position of the neuron, t is time, e is the external input stimulus of the neuron, w is the weight between the neuron at the preferred position θ and the neuron at the preferred position θ′ in the network, r is the normalized neuron firing rate, τ is the time constant of the neuron response, and the ρ parameter describes the extent to which the neuron is affected by other neurons in the network.

[0073] (1.1) w follows a Gaussian distribution, which can be expressed as:

[0074]

[0075] Among them, a is a hyperparameter that controls the range of interaction between neurons, and w0 is the scale factor.

[0076] (1.2)γ is used to maintain the stability of the network and prevent the energy in the network from growing infinitely, which can be expressed as:

[0077]

[0078] Here, k is a hyperparameter used to adjust the strength of normalization.

[0079] After constructing a single-level grid cell continuous attractor network, multiple levels of grid cell continuous attractor networks are coupled to form a multi-level grid cell module network, such as Figure 2 shown.

[0080] 2. The robot's sensor receives speed and angle information and decomposes the received speed information at multiple scales;

[0081] The robot's speed information can be decomposed into the following according to the number of grid module levels in step 1:

[0082]

[0083] Among them, v t represents the speed received by the robot at time t, Indicates speed v t The i-th component of , N is the level of velocity decomposition, e represents the difference ratio between adjacent velocity components, and the difference ratio between different modules is determined by the power of e i To represent, b is a positive integer used to characterize the range of speed levels.

[0084] 3. The decomposed speed in step 2 is input into the continuous attractor network at different levels according to different scales;

[0085] 4. After receiving the speed, each level of the continuous attractor network moves dynamically according to the speed information and time;

[0086] The dynamic movement process includes the following steps:

[0087] (4.1) The velocity received by the hierarchical continuous attractor network from the physical space is converted into the grid cell cognitive space. Specifically,

[0088]

[0089] in, is the speed information collected by the robot in the physical space. It is the speed information after being converted into the grid cell cognitive space. is the phase of the grid cell. R pg can be expressed as:

[0090]

[0091] Among them, s i is the step size of the grid cell, θ i Indicates the orientation of the grid cells.

[0092] (4.2) The speed converted to the grid cell cognitive space is decomposed in two directions x and y, namely The two components of represent its velocity components in the two directions x and y respectively;

[0093] (4.3) Dynamically moving in the x direction of the grid cell cognitive space according to the speed, the neuron state in the network can be expressed as:

[0094] U t+1 =ρW′R t #(7)

[0095] Among them, R t is the normalized network neural state at time t, and W′ can be expressed as:

[0096]

[0097] Among them, B is the basic circulant matrix, which can be expressed as:

[0098]

[0099] And Δt is the interval time, and

[0100]

[0101] At the same time, the normalized network neural state is updated, which can be expressed as:

[0102]

[0103] Among them, μ is a hyperparameter, which indicates the degree of normalization.

[0104] (4.4) Similarly, dynamic movement is performed in the y direction of the grid cell cognitive space according to the speed;

[0105] like Figure 3 The figure shows the dynamic movement process of a single-layer grid cell module over time after receiving the velocity component of this layer. The figure shows the firing state of the continuous attractor network neurons of the robot in three different motion tests. The position of the robot in the cognitive space is successfully encoded in the network in a group coding manner.

[0106] 5. After the network module is stable, determine whether there is a step situation in each level of the attractor network. If so, perform a hierarchical step transfer on the entire grid cell module network;

[0107] When a continuous attractor network at a certain level receives speed information that causes its network activity to reach its network epidemic boundary, the activity position of the continuous attractor network at this level is reset to 0, and its accumulated activity is transferred to the next level, and the network at the previous level is moved by 1 position according to the direction of movement.

[0108] 6. Perform multi-scale decoding on the entire grid cell module network to obtain the robot position encoded by it;

[0109] The entire grid cell module is decoded, and the decoded animal position can be described as:

[0110]

[0111] in, is the position encoded by the i-th module at time t, and M is the number of grid cell modules.

[0112] 7. Repeat steps 2-6 until the robot stops moving;

[0113] like Figure 4 The cognitive map constructed by the proposed method on the Kitti dataset is shown. In the initial stage, the proposed method is very close to the theoretical calculation results, which shows that the proposed method can perform accurate path integration to estimate the robot's trajectory. As the motion process increases, the path integration results obtained gradually differ from the theoretical calculation results, but the error does not increase all the time, and the accumulated error will be reduced in some positions. Figure 5 This result is further confirmed, which shows that the proposed method has the ability to reduce the cumulative error due to the dynamic stability property of the attractor network.

[0114] like Figure 6 (a) and Figure 7 (a) shows two maze scenes of robot movement without visual cues. In this scene, due to the consistency of texture between the maze wall and the ground, the robot cannot collect effective visual features in this scene. Figure 6 (b), (c) and Figure 7 (b) and (c) show the cognitive maps constructed by the classic RatSLAM and this method in Maze 1 and Maze 2, respectively. Figure 6 and Figure 7 The comparison of cognitive map construction with the classic brain-like SLAM algorithm RatSLAM in the Maze-1 and Maze-2 scenes is shown respectively. Both methods can receive image information and self-motion information from the car, and can generate corresponding cognitive maps in the initial stage. However, as the car moves, the proposed method can still form an accurate cognitive representation of the space, while the cognitive map generated by the brain-like navigation method based on place cells gradually becomes chaotic. This is because the structure of the maze environment makes the robot's visual input not only unable to provide more information for spatial cognition, but will introduce cognitive ambiguity and difficulties in scene recognition. Since the response of position cells is directly related to the visual input, when external visual stimulation cannot provide cognitive information or even causes cognitive ambiguity, the brain-like navigation method based on place cells will diverge or fail to establish a cognitive map.

Claims

1. A visual clue-free brain-like SLAM method based on grid cells, characterized in that: The following steps are involved: 1) Construct a multi-level grid cell module network and initialize it; 2) Receive the robot’s velocity information through the robot’s sensor and decompose it at multiple scales; 3) Input the decomposed speed into the continuous attractor network of the corresponding level according to the scale; 4) The continuous attractor network at each level moves dynamically according to speed information and time; 5) When the state of each level in the grid cell module network reaches a steady state, determine whether there is a step situation in the attractor network of each level. If so, perform a level step transfer on the entire grid cell module network. Otherwise, execute step 6); 6) Perform multi-scale decoding on the entire grid cell module network to obtain the robot position encoded by it; 7) Repeat steps 2) to 6) until the robot stops moving and obtains the complete movement trajectory of the robot.

2. According to claim 1, a grid cell-based brain-like SLAM method without visual clues is characterized in that: The multi-level grid cell module network is formed by coupling multiple single-level two-dimensional continuous attractor networks, and the dynamic equation of each single-level two-dimensional continuous attractor network is characterized as follows: Where u is the firing rate of the neuron in the network, θ is the preferred position of the neuron, t is time, e is the external input stimulus to the neuron, and w is the number of neurons in the network from the neuron at the preferred position θ to the neuron at the preferred position θ. ′ is the weight between neurons, r is the normalized neuron firing rate, τ is the time constant of neuron response, ρ is the degree to which a neuron is affected by other neurons in the network, a is a hyperparameter that controls the range of interaction between neurons, w0 is the scale factor, and k is a hyperparameter.

3. The visual clue-free brain-like SLAM method based on grid cells according to claim 1, characterized in that: The step 2) is specifically as follows: The robot's speed information is decomposed according to the number of layers of the multi-level grid cell module network, namely: Among them, v t represents the speed received by the robot at time t, Indicates speed v t The i-th component of , M is the number of levels of velocity decomposition, i.e., the number of levels of grid modules, and e represents the difference ratio between adjacent velocity components. The difference ratio between different modules is determined by the power of e i Indicates that b is a positive integer used to characterize the range of speed levels.

4. The visual clue-free brain-like SLAM method based on grid cells according to claim 1, characterized in that: The step 4) comprises the following steps: 4.1) For the i-th level of the grid cell module network, convert the speed received by this level into the grid cell cognitive space; 4.2) Decomposing the velocity converted to the grid cell cognitive space in the x-direction and the y-direction; 4.3) Dynamically move in the x direction of the grid cell cognitive space according to the velocity component in the x direction; 4.4) Dynamically move in the y direction of the grid cell cognitive space according to the velocity component in the y direction.

5. The visual clue-free brain-like SLAM method based on grid cells according to claim 4, characterized in that: The step 4.1) is specifically as follows: Among them, v i is the decomposed velocity component, It is the speed information after being converted into the grid cell cognitive space. is the change matrix, expressed as: Among them, s i is the step size of the grid cells in the i-th layer, θ i Indicates the orientation of the grid cells.

6. The visual clue-free brain-like SLAM method based on grid cells according to claim 4, characterized in that: The step 4.3) is specifically as follows: The neuron state update in the network can be expressed as: in is the neuron state in the updated network, is the normalized neural state of the network at time t, n is the number of neurons in the network along the x-direction or y-direction, and the weight matrix W ′ for: It is described as: B is the basic circulant matrix: Δt is the interval time, is the number of neurons in the x direction, s i is the step size of the grid cells at this level; Update the normalized network neural state, that is: Among them, μ is a hyperparameter, indicating the degree of normalization, is a diagonal matrix constructed from neuron states, which can be described as:

7. The visual clue-free brain-like SLAM method based on grid cells according to claim 1, characterized in that: The step 5) is specifically as follows: When a continuous attractor network at a certain level receives speed information that causes its network activity to reach its network epidemic boundary, it is considered that there is a step in the continuous attractor network at the current level. The activity position of the continuous attractor network at this level is reset to 0, and its accumulated activity is transferred to the next level, and the network at the previous level is moved by 1 position according to the direction of movement.

8. The visual clue-free brain-like SLAM method based on grid cells according to claim 1, characterized in that: In step 6), the position P of the robot after decoding t for: in, is the position encoded by the i-th module at time t, and M is the number of grid cell modules.

9. A visual clue-free brain-like SLAM system based on grid cells, characterized in that: It comprises a memory and a processor; the memory is used to store a computer program; the processor is used to implement a visual clue-free brain SLAM method based on grid cells as described in any one of claims 1-8 when executing the computer program.

10. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, a visual cue-free brain-like SLAM method based on grid cells is implemented as described in any one of claims 1 to 8.

Images