Mobile robot autonomous navigation method with risk perception

Abstract

The invention provides an autonomous navigation method for a mobile robot with risk perception, which comprises the following steps: firstly, initializing the robot to obtain own state and radar information, and carrying out risk prediction to obtain the risk condition of the robot; and when the risk exceeds a threshold value, a low-risk area in the environment is found through risk measurement in combination with surrounding obstacles and target point information. Finally, whether the action output by the neural network is reasonable or not is evaluated for risk correction, so that the robot is guided to drive to a safer and more efficient route to complete navigation. Through the steps, the problem of map-free autonomous navigation of the robot in a complex dynamic environment is solved.

Description

Technical Field

[0001] This invention relates to the field of robot navigation technology, specifically to a method for autonomous navigation of a mobile robot with risk awareness. Background Technology

[0002] Autonomous robot navigation refers to a robot's ability to autonomously perceive, locate, plan paths, and successfully reach its target position within its environment. Environmental perception is the core foundation of autonomous navigation. To this end, robots rely on various sensors to collect information about their surroundings, including maps, obstacles, boundaries, and the positions of other objects. 2D radar, as a common environmental perception sensor, is widely used in robot navigation. It not only accurately measures the distance between the robot and surrounding objects but also has strong anti-interference capabilities, providing stable and reliable data in various complex environments. Most importantly, 2D radar is relatively inexpensive, making it an economical and efficient choice widely used in robot navigation.

[0003] In the early stages of autonomous robot navigation, robots are prone to collisions with surrounding obstacles, making the navigation strategy learning process both risky and extremely time-consuming. Furthermore, considering that robots typically operate in complex and dynamically changing environments, such as manufacturing workshops, smart factories, and hospitals, they must remain highly sensitive to their surroundings to ensure the successful completion of their tasks. Therefore, employing an effective risk perception method to assess the potential risks posed by pedestrians and obstacles in the environment is crucial. This helps robots identify safer and more efficient paths, improving the reliability of their autonomous navigation. Summary of the Invention

[0004] To address the problems of existing technologies, the purpose of this invention is to provide a risk-aware autonomous navigation method for mobile robots. This method aims to enhance the robot's ability to perceive environmental risks, thereby improving its navigation safety and efficiency in complex and dynamic environments. By using 2D radar data combined with risk assessment algorithms, the robot can perceive and predict surrounding obstacles, pedestrians, and other potential risk sources, and plan a safer and more efficient path based on this information.

[0005] To achieve the above-mentioned technical features, the objective of this invention is as follows: a risk-aware autonomous navigation method for a mobile robot, comprising the following steps:

[0006] S1: Initialize robot H, obtain its own status information and the latest radar data;

[0007] S2: Extract robot features and obstacle features from the current state of robot H and the k latest radar data to train a risk prediction network, and use the network to predict the risk situation of the robot at a future moment;

[0008] S3: When the predicted risk exceeds the threshold, combine information about surrounding obstacles and target points to find a relatively safe low-risk area in the environment through risk measurement.

[0009] S4: Input the state space into the neural network of the control strategy, train the network model using reinforcement learning, and output the predetermined action a to be executed by robot H. t ;

[0010] S5: Evaluate whether the action output by the deep neural network is reasonable. If reasonable, execute action a. t If robot H is placed in a greater risk, then risk correction will be performed, and an action will be executed. The robot H is guided to navigate a route that is low-risk and high-return.

[0011] S6: Determine whether robot H has reached the target point. If it has, navigation ends. If it has not reached the target point, continue to step S2. If robot H collides, continue to step S1 and repeat the above method until robot H reaches the target point.

[0012] Preferably, the robot H's own state information in step S1 includes linear velocity ν, angular velocity ω, and distance d from the target. targ The angle θ corresponding to the target targ Distance d from the obstacle obst and the angle θ of the corresponding obstacle obst At any given moment, the 2D lidar data consists of N angular distance pairs representing 360 degrees.

[0013] Preferably, the specific steps of step S2 are as follows:

[0014] S21: Transfer the six state information {ν,ω,d} of robot H goal ,d obst ,θ goal ,θ obst Input two fully connected layers to extract robot features;

[0015] S22: Raw radar data l =[l1,l2,…,l N After preprocessing, the original N dimensions were reduced to n. β Dimensions, processed radar data in α∈{1,2,···,n α}, β∈{1,2,···,n β}, l α,β Let l represent the length of the α-th laser in the β-th slice. max n represents the maximum sensing distance of the radar.α =N / n β ; Λ k consecutive latest radar data l The obstacle features are obtained by inputting two one-dimensional convolutional layers, then through an attention module, a pooling layer, and a flattening layer, and finally into a fully connected layer.

[0016] S23: Fuse robot features and obstacle features through a fully connected layer to generate risk predictions for future timeframes. The prediction network uses the loss function L risk Update model, L risk The definition is as follows:

[0017]

[0018] In the formula, s t+σ This represents the future state after σ time steps from the current state; in s t+σ Actual risk value of the state The evaluation is conducted using three indicators:

[0019]

[0020] In the formula, μ area The area of ​​the obstructed region within the radar range is calculated using mathematical approximation methods; k μ and k l The weighting coefficients; parameter n obst Related to the number of obstacles, it is defined as follows:

[0021] n obst =k num ·(num-1);

[0022] In the formula, num represents the number of obstacles, and k num These are the weighting coefficients;

[0023] Parameter l leng Related to the shortest laser length of radar, it is defined as follows:

[0024]

[0025] In the formula, min(l) is the shortest radar length, l crit As the risk length threshold, when l crit <min(l)≤l max This means that the obstacle is still some distance away from the robot and does not pose a risk to the robot at this time.

[0026] Preferably, the specific steps of step S3 are as follows:

[0027] S31: When the predicted risk value exceeds the threshold, it means that robot H is about to collide with an obstacle or pedestrian and needs to avoid obstacles. At this time, in order to find a navigation route with lower risk, the circular radar range is divided into layers and parts, a total of layer×part grid areas. By scoring the risk level of each grid, robot H can identify a safe and efficient movement route.

[0028] S32: The grid score of the i-th layer and j-th column in the radar range is denoted as score(i,j), and is defined as follows:

[0029] score(i,j)=Ψ lidar,i,j +Ψ area,i,j +Ψ angle,i,j ;

[0030] In the formula, the main indicator for measuring the risk of grid (i,j) is Ψ. area,i,j and Ψ lidar,i,j Ψ area,i,j The calculation calculates the area of ​​obstacles within the grid (i,j), while Ψ lidar,i,j Related to the length of the radar in the grid (i,j); Ψ lidar,i,j The definition is as follows:

[0031]

[0032] In the formula, n l The number of laser beams from the radar in each grid, n l =N / part;l i,j,h The length of the h-th laser beam in the grid (i,j) is represented by k. min These are the weighting coefficients;

[0033] Ψ angle,i,j Used to identify navigation routes with higher profitability in the surrounding area; Ψ angle,i,j The definition is as follows:

[0034] Ψ angle,i,j =|θ targ | / π;

[0035] In the formula, θ targ This represents the angle difference between the orientation angle of robot H relative to the grid (i,j) and its orientation angle relative to the target.

[0036] Preferably, the specific steps of step S4 are as follows:

[0037] S41: The observation space of robot H is composed of o l o t o v and o r Composition; radar data l=[l1,l2,…,l N ]; State o t = {ρ, φ}, where ρ represents the Euclidean distance between the robot and the target point, defined as follows:

[0038]

[0039] In the formula, (x c ,y c ) and (x t ,y t ) represent the robot's current position and the target position, respectively; φ represents the angle between the robot's position and the target position, calculated as follows:

[0040]

[0041] In the formula, and It is a vector representing the robot's direction of movement and its target position;

[0042] State o v ={ν,ω} includes the linear velocity ν and angular velocity ω output by robot H;

[0043] The last state o r ={risk} is the robot's risk value; when the accuracy of the prediction network training meets the requirement, o r It is the predicted risk. pred Otherwise, it is the actual calculated value of risk in the current state. real ;

[0044] At time t, the state space is described as follows:

[0045] s t =[l1,l2,···,l N ,ρ,φ,ν t ,ω t ,risk].

[0046] S42: The continuous motion space of robot H consists of two parts: linear velocity ν∈k max ·[0,1.0]m / s and angular velocity ω∈k max • [-1.0, 1.0] rad / s; where k max It is a scaling factor that adjusts the speed output range, and the robot H takes appropriate actions from the motion space according to the current state.

[0047] S43: A reward function is used to guide the robot's movement; the total reward is divided into sparse rewards r. sparse and dense reward r dense ;

[0048] The sparse reward takes into account the termination condition of robot H during navigation; if the distance between robot H and the obstacle or target point is less than a safety threshold d. end They will be subject to hefty penalties. collision Or reward r arrival Furthermore, if the risk predicted by the network is continuous for t... cons The time exceeds the risk threshold [k] risk Robot H will be subject to increased risk penalties. risk If the navigation time exceeds the scheduled time, a timeout penalty will also be imposed. overtime The definition of sparse reward is as follows:

[0049]

[0050] In the formula, d targ and d obst t represents the distance between robot H and the target point, and the distance between robot H and the obstacle, respectively; targ [t] represents the time taken to reach the goal and the time taken for the scheduled task; risk t+σ s t+σ Risk value under the current state;

[0051] Sparse rewards represent the basic task of navigation, but they are not effective in encouraging the robot to explore its surroundings. Therefore, dense rewards are designed, providing feedback at each time step before navigation is completed to encourage the robot H to move towards the target point. Dense rewards are defined as follows:

[0052] r dense =r t +r c +r v +r m ;

[0053] In the formula, r dense It contains four distinct terms, where r t It is a goal-oriented reward based on the distance and angle of robot H to the target; robot H receives a reward when it gets close to the target and is penalized when it deviates from the target direction; r t Defined as:

[0054]

[0055] In the formula, and represents the current distance and angle deviation between robot H and the target, while max(d) and max(θ) are the maximum distance and angle changes of robot H within a given moment; k d k a and k o These are the weighting coefficients;

[0056] r c The reward is related to obstacles; when the robot H is less than a predefined threshold away from an obstacle, it will be penalized. c Defined as:

[0057]

[0058] r v The reward is speed-related; to complete the navigation task quickly and smoothly, robot H should maintain a high linear velocity and a low angular velocity; r v The definition is as follows:

[0059]

[0060] In the formula, ν t and ω t These are the current linear velocity and angular velocity of robot H; ν max and ω max Indicates maximum linear velocity and maximum angular velocity; k ν and k ω These are the weighting coefficients;

[0061] Finally, r m The design aims to encourage the robot to actively explore and shorten its travel path; at each time step, the robot H will be subject to a slight penalty to accelerate the learning process and reduce the occurrence of timeouts.

[0062] Preferably, the specific steps of step S5 are as follows:

[0063] S51: Following step S3, robot H can identify the risk status of all grids within the radar range; it selects one grid from each layer, minimizing the sum of the risk scores of the selected layers, thus connecting the grids from different layers to form a reasonable navigation trajectory; the trajectory is represented by the number of columns of the selected grids in each layer:

[0064] column = [j layer＝1 ,j layer＝2 ,j layer＝3 ,j layer＝4 ,j layer＝5 ];

[0065] In the formula, j layer＝i Let represent the number of columns selected in the i-th layer; the optimization problem for selecting the grid is as follows:

[0066]

[0067] Optimal grid combination column * This is achieved through continuous sorting and comparison; an angle constraint is introduced in the selection of adjacent mesh layers, defined as follows:

[0068] st|θ up -θ down |≤[θ];

[0069] Parameter θ up and θ down An angular constraint is defined for mesh selection between two adjacent layers, where [θ] represents the angular difference between the meshes, and its expression is as follows:

[0070] [θ]=k θ ·θ(j layer ,j layer+1 );

[0071] In the formula, θ(j layer ,j layer+1 ) represents the angle between the bisectors of adjacent grids in different layers; k θ These are the weighting coefficients;

[0072] S52: Correct the actions output by the neural network to guide robot H to move along a reasonable navigation route; the correction method is as follows:

[0073]

[0074] In the formula, Used to calculate the angular velocity required for rotation between two adjacent grids at the current moment; j r This indicates that the initial orientation of robot H is in the column; the expression (j layer -j layer-1 This is used to calculate the difference in the number of selected grid columns between two adjacent layers;

[0075]

[0076] In the formula, Where ω t-1 Indicates the angular velocity output at the previous moment; hyperparameter k s and k p This is mainly to prevent the robot from moving too slowly when decelerating or not moving fast enough when accelerating;

[0077] S53: Has robot H learned a reasonable obstacle avoidance strategy π? * Determine using the following methods:

[0078]

[0079] In the formula, π′ represents that the robot has not yet fully mastered the obstacle avoidance strategy; the parameter k η The method for measuring the learning progress of the robot's policy is as follows:

[0080]

[0081] In the formula, k η k is only updated if the condition is met at consecutive time points; otherwise, it is not updated. η =0; when k η ≤[k η When [action a] is modified, the current action a is corrected. t = (ν,ω), robot H will perform safer, more rational, and more efficient actions. Complete navigation quickly and safely; if the adjusted values ​​exceed a reasonable range, then trim the actions.

[0082] The present invention has the following beneficial effects:

[0083] In the early stages of training for autonomous robot navigation, problems such as uneven trajectory smoothness and frequent collisions with obstacles often arise, making the navigation strategy learning process risky and time-consuming. Considering that robots typically operate in complex and dynamically changing environments, such as manufacturing workshops, smart factories, and hospitals, they must maintain a high degree of environmental sensitivity to ensure successful task completion. Therefore, this invention utilizes robot state and radar information, combined with a risk assessment algorithm, to perceive potential risks from surrounding obstacles and pedestrians, thereby planning safer and more efficient paths and improving navigation reliability. Attached Figure Description

[0084] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0085] Figure 1 This is a flowchart of the risk-aware autonomous navigation method for mobile robots according to the present invention.

[0086] Figure 2 This is a schematic diagram of robot radar grid division provided by the present invention.

[0087] Figure 3 This is a schematic diagram illustrating the robot's risk perception in the environment, as provided by the present invention.

[0088] Figure 4 This is a schematic diagram illustrating how the robot identified a safer navigation area, as provided by the present invention. Detailed Implementation

[0089] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0090] Example 1:

[0091] See Figure 1-4 A risk-aware autonomous navigation method for mobile robots includes the following steps:

[0092] S1: Initialize robot H, obtain its own status information and the latest radar data;

[0093] S2: Extract robot features and obstacle features from the current state of robot H and the k latest radar data to train a risk prediction network, and use the network to predict the risk situation of the robot at a future moment;

[0094] S3: When the predicted risk exceeds the threshold, combine information about surrounding obstacles and target points to find a relatively safe low-risk area in the environment through risk measurement.

[0095] S4: Input the state space into the neural network of the control strategy, train the network model using reinforcement learning, and output the predetermined action a to be executed by robot H. t ;

[0096] S5: Evaluate whether the action output by the deep neural network is reasonable. If reasonable, execute action a. t If robot H is placed in a greater risk, then risk correction will be performed, and an action will be executed. The robot H is guided to navigate a route that is low-risk and high-return.

[0097] S6: Determine whether robot H has reached the target point. If it has, navigation ends. If it has not reached the target point, continue to step S2. If robot H collides, continue to step S1 and repeat the above method until robot H reaches the target point.

[0098] Preferably, the robot H's own state information in step S1 includes linear velocity ν, angular velocity ω, and distance d from the target. targ The angle θ corresponding to the target targ Distance d from the obstacle obst and the angle θ of the corresponding obstacle obst At any given moment, the 2D lidar data consists of N angular distance pairs representing 360 degrees.

[0099] Preferably, the specific steps of step S2 are as follows:

[0100] S21: Transfer the six state information {ν,ω,d} of robot H goal ,d obst ,θ goal ,θ obst Input two fully connected layers to extract robot features;

[0101] S22: Raw radar data l =[l1,l2,…,lN After preprocessing, the original N dimensions were reduced to n. β Dimensions, processed radar data in α∈{1,2,···,n α}, β∈{1,2,···,n β}, l α,β Let l represent the length of the α-th laser in the β-th slice. max n represents the maximum sensing distance of the radar. α =N / n β ; Λ k consecutive latest radar data l The obstacle features are obtained by inputting two one-dimensional convolutional layers, then through an attention module, a pooling layer, and a flattening layer, and finally into a fully connected layer.

[0102] S23: Fuse robot features and obstacle features through a fully connected layer to generate risk predictions for future timeframes. The prediction network uses the loss function L risk Update model, L risk The definition is as follows:

[0103]

[0104] In the formula, s t+σ This represents the future state after σ time steps from the current state; in s t+σ Actual risk value of the state The evaluation is conducted using three indicators:

[0105]

[0106] In the formula, μ area The area of ​​the obstructed region within the radar range is calculated using mathematical approximation methods; k μ and k l The weighting coefficients; parameter n obst Related to the number of obstacles, it is defined as follows:

[0107] n obst =k num ·(num-1);

[0108] In the formula, num represents the number of obstacles; k num These are the weighting coefficients;

[0109] Parameter l leng Related to the shortest laser length of radar, it is defined as follows:

[0110]

[0111] In the formula, min(l) is the shortest radar length, l max For the radar's maximum sensing range, l crit As the risk length threshold, when l crit <min(l)≤l max This indicates that the obstacle is still some distance from the robot and does not pose a risk to the robot at this time.

[0112] Preferably, the specific steps of step S3 are as follows:

[0113] S31: When the predicted risk value exceeds the threshold, it means that robot H is about to collide with an obstacle or pedestrian and needs to avoid obstacles. At this time, in order to find a navigation route with lower risk, the circular radar range is divided into layers and parts, a total of layer×part grid areas. By scoring the risk level of each grid, robot H can identify a safe and efficient movement route.

[0114] S32: The grid score of the i-th layer and j-th column in the radar range is denoted as score(i,j), and is defined as follows:

[0115] score(i,j)=Ψ lidar,i,j +Ψ area,i,j +Ψ angle,i,j ;

[0116] In the formula, the main indicator for measuring the risk of grid (i,j) is Ψ. area,i,j and Ψ lidar,i,j Ψ area,i,j The calculation calculates the area of ​​obstacles within the grid (i,j), while Ψ lidar,i,j Related to the length of the radar in the grid (i,j); Ψ lidar,i,j The definition is as follows:

[0117]

[0118] In the formula, n l The number of laser beams from the radar in each grid, n l =N / part;l i,j,h The length of the h-th laser beam in the grid (i,j) is represented by k. min These are the weighting coefficients;

[0119] Ψ angle,i,j Used to identify navigation routes with higher profitability in the surrounding area; Ψ angle,i,j The definition is as follows:

[0120]

[0121] In the formula, θ targThis represents the angle difference between the orientation angle of robot H relative to the grid (i,j) and its orientation angle relative to the target.

[0122] Preferably, the specific steps of step S4 are as follows:

[0123] S41: The observation space of robot H is composed of o l o t o v and o r Composition; radar data l =[l1,l2,…,l N ]; State o t = {ρ, φ}, where ρ represents the Euclidean distance between the robot and the target point, defined as follows:

[0124]

[0125] In the formula, (x c ,y c ) and (x t ,y t ) represent the robot's current position and the target position, respectively; φ represents the angle between the robot's position and the target position, calculated as follows:

[0126]

[0127] In the formula, and It is a vector representing the robot's direction of movement and its target position;

[0128] State o v ={ν,ω} includes the linear velocity ν and angular velocity ω output by robot H;

[0129] The last state o r ={risk} is the robot's risk value; when the accuracy of the prediction network training meets the requirement, o r It is the predicted risk. pred Otherwise, it is the actual calculated value of risk in the current state. real ;

[0130] At time t, the state space is described as follows:

[0131] s t =[l1,l2,···,l N ,ρ,φ,ν t ,ω t [,risk];

[0132] S42: The continuous motion space of robot H consists of two parts: linear velocity ν∈k max·[0,1.0]m / s and angular velocity ω∈k max • [-1.0, 1.0] rad / s; where k max It is a scaling factor that adjusts the speed output range, allowing robot H to take appropriate actions from the motion space based on the current state.

[0133] S43: To ensure safe and efficient navigation, a reward function was designed to guide the robot's movement; the total reward is divided into sparse rewards r. sparse and dense reward r dense ;

[0134] The sparse reward takes into account the termination condition of robot H during navigation; if the distance between robot H and the obstacle or target point is less than a safety threshold d. end They will be subject to hefty penalties. collision Or reward r arrival Furthermore, if the risk predicted by the network is continuous for t... cons The time exceeds the risk threshold [k] risk Robot H will be subject to increased risk penalties. risk If the navigation time exceeds the scheduled time, a timeout penalty will also be imposed. overtime The definition of sparse reward is as follows:

[0135]

[0136] In the formula, d targ and d obst t represents the distance between robot H and the target point, and the distance between robot H and the obstacle, respectively; targ [t] represents the time taken to reach the goal and the time taken for the scheduled task; risk t+σ s t+σ Risk value under the current state;

[0137] Sparse rewards represent the basic task of navigation, but they are not effective in encouraging the robot to explore its surroundings. Therefore, dense rewards are designed, providing feedback at each time step before navigation is completed, encouraging the robot H to move towards the target point. Dense rewards not only speed up the training process but also make the trajectory smoother, and are defined as follows:

[0138] r dense =r t +r c +r v +r m ;

[0139] In the formula, r dense It contains four distinct terms, where r tIt is a goal-oriented reward based on the distance and angle of robot H to the target; robot H receives a reward when it gets close to the target and is penalized when it deviates from the target direction; r t Defined as:

[0140]

[0141] In the formula, and represents the current distance and angle deviation between robot H and the target, while max(d) and max(θ) are the maximum distance and angle that robot H can change within a given moment; k d k a and k o These are the weighting coefficients;

[0142] r c The reward is related to obstacles; when the robot H is less than a predefined threshold away from an obstacle, it will be penalized. c Defined as:

[0143]

[0144] r v The reward is speed-related; to complete the navigation task quickly and smoothly, robot H should maintain a high linear velocity and a low angular velocity; r v The definition is as follows:

[0145]

[0146] In the formula, ν t and ω t These are the current linear velocity and angular velocity of robot H; ν max and ω max Indicates maximum linear velocity and maximum angular velocity; k ν and k ω These are the weighting coefficients;

[0147] Finally, r m The design aims to encourage the robot to actively explore and shorten its travel path; at each time step, the robot H will be subject to a slight penalty to accelerate the learning process and reduce the occurrence of timeouts.

[0148] Preferably, the specific steps of step S5 are as follows:

[0149] S51: Following step S3, robot H can identify the risk status of all grids within the radar range. To ensure a safe and efficient navigation route, one grid is selected from each layer, minimizing the sum of the risk scores of the selected layers. Connecting the grids from different layers together will then form a reasonable navigation trajectory. The trajectory is represented by the number of columns of the selected grids in each layer.

[0150] column = [j layer＝1 ,j layer＝2 ,j layer＝3 ,j layer＝4 ,j layer＝5 ];

[0151] In the formula, j layer＝i Let represent the number of columns selected in the i-th layer; the optimization problem for selecting the grid is as follows:

[0152]

[0153] Optimal grid combination column * The results were obtained through continuous sorting and comparison. To ensure a smooth path and prevent robot H from taking dangerous sharp turns, an angle constraint was introduced in the selection of adjacent mesh layers, defined as follows:

[0154] st|θ up -θ down |≤[θ];

[0155] Parameter θ up and θ down An angular constraint is defined for mesh selection between two adjacent layers, where [θ] represents the angular difference between the meshes, and its expression is as follows:

[0156] [θ]=k θ ·θ(j layer ,j layer+1 );

[0157] In the formula, θ(j layer ,j layer+1 ) represents the angle between the bisectors of adjacent grids in different layers; k θ These are the weighting coefficients.

[0158] S52: To prevent robot H from colliding with obstacles before mastering the obstacle avoidance strategy, the action output by the neural network is corrected to guide robot H to move along a reasonable navigation route; the correction method is as follows:

[0159]

[0160] In the formula, Used to calculate the angular velocity required for rotation between two adjacent grids at the current moment; the initial orientation of robot H is in column 6, then j r =6; expression(j layer -j layer-1 This is used to calculate the difference in the number of selected grid columns between two adjacent layers;

[0161]

[0162] In the formula, Where ω t-1 Indicates the angular velocity output at the previous moment; hyperparameter k s and k p This is mainly to prevent the robot from moving too slowly when decelerating or not moving fast enough when accelerating;

[0163] S53: Has robot H learned a reasonable obstacle avoidance strategy π? * Determine using the following methods:

[0164]

[0165] In the formula, π′ represents that the robot has not yet fully mastered the obstacle avoidance strategy; the parameter k η The method for measuring the learning progress of the robot's policy is as follows:

[0166]

[0167] In the formula, k η k is only updated if the condition is met at consecutive time points; otherwise, it is not updated. η =0; when k η ≤[k η When [action a] is modified, the current action a is corrected. t = (ν,ω), robot H will perform safer, more rational, and more efficient actions. Complete navigation quickly and safely; if the adjusted values ​​exceed a reasonable range, then trim the actions.

[0168] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for autonomous navigation of a mobile robot with risk awareness, characterized in that, Includes the following steps: S1: Initialize robot H, obtain its own status information and the latest radar data; S2: Extract robot features and obstacle features from the current state of robot H and the k latest radar data to train a risk prediction network, and use the network to predict the risk situation of the robot at a future moment; S3: When the predicted risk exceeds the threshold, combine information about surrounding obstacles and target points to find a relatively safe low-risk area in the environment through risk measurement. S4: Input the state space into the neural network of the control strategy, train the network model using reinforcement learning, and output the predetermined action a to be executed by robot H. t ; S5: Evaluate whether the action output by the deep neural network is reasonable. If reasonable, execute action a. t If robot H is placed in a greater risk, then risk correction will be implemented, and an action will be taken. The robot H is guided to navigate a route that is low in risk and high in return. S6: Determine whether robot H has reached the target point. If it has, navigation ends. If it has not reached the target point, continue to step S2. If robot H collides, continue to step S1 and repeat the above method until robot H reaches the target point.

2. The autonomous navigation method for a mobile robot with risk perception according to claim 1, characterized in that, In step S1, the robot H's own state information includes linear velocity ν, angular velocity ω, and distance d from the target. targ The angle θ corresponding to the target targ Distance d from the obstacle obst and the angle θ of the corresponding obstacle obst ; At any given moment, the 2D LiDAR data consists of N angular distance pairs representing 360 degrees.

3. The autonomous navigation method for a mobile robot with risk perception according to claim 2, characterized in that, The specific steps of step S2 are as follows: S21: Transfer the six state information {ν,ω,d} of robot H goal ,d obst ,θ goal ,θ obst Input two fully connected layers to extract robot features; S22: Raw radar data l =[l1,l2,…,l N After preprocessing, the original N dimensions were reduced to n. β Dimensions, processed radar data in α∈{1,2,···,n α }, β∈{1,2,···,n β }, l α,β Let l represent the length of the α-th laser in the β-th slice. max n represents the maximum detection range of the radar. α =N / n β ; Λ k consecutive latest radar data l The input consists of two one-dimensional convolutional layers, followed by an attention module, a pooling layer, and a flattening layer, and finally input into a fully connected layer to obtain obstacle features.

4. The autonomous navigation method for a mobile robot with risk perception according to claim 3, characterized in that, Step S2 further includes: S23: Fuse robot features and obstacle features through a fully connected layer to generate risk predictions for future timeframes. The prediction network uses the loss function L risk Update model, L risk The definition is as follows: In the formula, s t+σ This represents the future state after σ time steps from the current state; in s t+σ Actual risk value of the state The evaluation is conducted using three indicators: In the formula, μ area The area of ​​the obstructed region within the radar range is calculated using mathematical approximation methods; k μ and k l The weighting coefficients; parameter n obst Related to the number of obstacles, it is defined as follows: n obst =k num ·(num-1); In the formula, num represents the number of obstacles, and k num These are the weighting coefficients; Parameter l leng Related to the shortest laser length of radar, it is defined as follows: In the formula, min(l) is the shortest radar length, l crit As the risk length threshold, when l crit <min(l)≤l max This means that the obstacle is still some distance away from the robot and does not pose a risk to the robot at this time.

5. The autonomous navigation method for a mobile robot with risk perception according to claim 4, characterized in that, The specific steps of step S3 are as follows: S31: When the predicted risk value exceeds the threshold, it means that robot H is about to collide with an obstacle or pedestrian and needs to avoid obstacles. At this time, in order to find a navigation route with lower risk, the circular radar range is divided into layers and parts, a total of layer×part grid areas. By scoring the risk level of each grid, robot H can identify a safe and efficient movement route. S32: The grid score of the i-th layer and j-th column in the radar range is denoted as score(i,j), and is defined as follows: score(i,j)=Ψ lidar,i,j +Ψ area,i,j +Ψ angle,i,j ; In the formula, the main indicator for measuring the risk of grid (i,j) is Ψ. area,i,j and Ψ lidar,i,j Ψ area,i,j The calculation calculates the area of ​​obstacles within the grid (i,j), while Ψ lidar,i,j Related to the length of the radar in the grid (i,j); Ψ lidar,i,j The definition is as follows: In the formula, n l The number of laser beams from the radar in each grid, n l =N / part;l i,j,h The length of the h-th laser beam in the grid (i,j) is represented by k. min These are the weighting coefficients; Ψ angle,i,j Used to identify navigation routes with higher profitability in the surrounding area; Ψ angle,i,j The definition is as follows: P angle,i,j =|θ targ | / π; In the formula, θ targ This represents the angle difference between the orientation angle of robot H relative to the grid (i,j) and its orientation angle relative to the target.

6. The autonomous navigation method for a mobile robot with risk perception according to claim 5, characterized in that, The specific steps of step S4 are as follows: S41: The observation space of robot H is composed of o l o t o v and o r Composition; radar data l =[l1,l2,…,l N ]; State o t = {ρ, φ}, where ρ represents the Euclidean distance between the robot and the target point, defined as follows: In the formula, (x c ,y c ) and (x t ,y t ) represent the robot's current position and the target position, respectively; φ represents the angle between the robot's position and the target position, calculated as follows: In the formula, and It is a vector representing the robot's direction of movement and its target position; State o v ={ν,ω} includes the linear velocity ν and angular velocity ω output by robot H; The last state o r ={risk} is the robot's risk value; when the accuracy of the prediction network training meets the requirement, o r It is the predicted risk. pred Otherwise, it is the actual calculated value of risk in the current state. real ; At time t, the state space is described as follows: s t =[l1,l2,···,l N ,r,f,n t ,oh t [risk].

7. The autonomous navigation method for a mobile robot with risk perception according to claim 6, characterized in that, Step S4 further includes the following steps: S42: The continuous motion space of robot H consists of two parts: linear velocity ν∈k max ·[0,1.0]m / s and angular velocity ω∈k max • [-1.0, 1.0] rad / s; where k max It is a scaling factor that adjusts the speed output range, allowing robot H to take appropriate actions from the motion space based on the current state.

8. The autonomous navigation method for a mobile robot with risk perception according to claim 6, characterized in that, Step S4 further includes the following steps: S43: A reward function is used to guide the robot's movement; the total reward is divided into sparse rewards r. sparse and dense reward r dense ; The sparse reward takes into account the termination condition of robot H during navigation; if the distance between robot H and the obstacle or target point is less than a safety threshold d. end They will be subject to hefty penalties. collision Or reward r arrival Furthermore, if the risk predicted by the network is continuous for t... cons The time exceeds the risk threshold [k] risk Robot H will be subject to increased risk penalties. risk If the navigation time exceeds the scheduled time, a timeout penalty will also be imposed. overtime The definition of sparse reward is as follows: In the formula, d targ and d obst t represents the distance between robot H and the target point, and the distance between robot H and the obstacle, respectively; targ [t] represents the time taken to reach the goal and the time taken for the scheduled task; risk t+σ s t+σ Risk value under the current state; Sparse rewards represent the basic task of navigation, but they are not effective in encouraging the robot to explore its surroundings. Therefore, dense rewards are designed, providing feedback at each time step before navigation is completed to encourage the robot H to move towards the target point. Dense rewards are defined as follows: r dense ＝r t +r c +r v +r m ； In the formula, r dense It contains four distinct terms, where r t It is a goal-oriented reward based on the distance and angle of robot H to the target; robot H receives a reward when it gets close to the target and is penalized when it deviates from the target direction; r t Defined as: In the formula, and represents the current distance and angle deviation between robot H and the target, while max(d) and max(θ) are the maximum distance and angle changes of robot H within a given moment; k d k a and k o These are the weighting coefficients; r c The reward is related to obstacles; when the robot H is less than a predefined threshold away from an obstacle, it will be penalized. c Defined as: r v The reward is speed-related; to complete the navigation task quickly and smoothly, robot H should maintain a high linear velocity and a low angular velocity; r v The definition is as follows: In the formula, ν t and ω t These are the current linear velocity and angular velocity of robot H; ν max and ω max Indicates the maximum linear velocity and maximum angular velocity; k ν and k ω These are the weighting coefficients; Finally, r m The design aims to encourage the robot to actively explore and shorten its travel path; at each time step, the robot H will be subject to a slight penalty to accelerate the learning process and reduce the occurrence of timeouts.

9. The autonomous navigation method for a mobile robot with risk perception according to claim 8, characterized in that, The specific steps of step S5 are as follows: S51: Following step S3, robot H can identify the risk status of all grids within the radar range; it selects one grid from each layer, minimizing the sum of the risk scores of the selected layers, thus connecting the grids from different layers to form a reasonable navigation trajectory; the trajectory is represented by the number of columns of the selected grids in each layer: column=[j layer＝1 ,j layer＝2 ,j layer＝3 ,j layer＝4 ,j layer＝5 ]; In the formula, j layer＝i Let represent the number of columns selected in the i-th layer; the optimization problem for selecting the grid is as follows: Optimal grid combination column * This is achieved through continuous sorting and comparison; an angle constraint is introduced in the selection of adjacent mesh layers, defined as follows: st|θ up -θ down |≤[θ]; Parameter θ up and θ down An angular constraint is defined for mesh selection between two adjacent layers, where [θ] represents the angular difference between the meshes, and its expression is as follows: [θ]=k θ ·θ(j layer ,j layer+1 ); In the formula, θ(j layer ,j layer+1 ) represents the angle between the bisectors of adjacent grids in different layers; k θ These are the weighting coefficients.

10. The autonomous navigation method for a mobile robot with risk perception according to claim 9, characterized in that, Step S5 also includes: S52: Correct the actions output by the neural network to guide robot H to move along a reasonable navigation route; the correction method is as follows: In the formula, Used to calculate the angular velocity required for rotation between two adjacent grids at the current moment; j r This indicates that the initial orientation of robot H is in the column; the expression (j layer -j layer-1 This is used to calculate the difference in the number of selected grid columns between two adjacent layers; In the formula, Where ω t-1 Indicates the angular velocity output at the previous moment; hyperparameter k s and k p This is mainly to prevent the robot from moving too slowly when decelerating or not moving fast enough when accelerating; S53: Has robot H learned a reasonable obstacle avoidance strategy π? * Determine using the following methods: In the formula, π′ represents that the robot has not yet fully mastered the obstacle avoidance strategy; the parameter k η The method for measuring the learning progress of the robot's policy is as follows: In the formula, k η k is only updated if the condition is met at consecutive time points; otherwise, it is not updated. η =0; when k η ≤[k η When [action a] is modified, the current action a is corrected. t = (ν,ω), robot H will perform safer, more rational, and more efficient actions. Complete navigation quickly and safely; if the adjusted values ​​exceed a reasonable range, then trim the actions.

Images