Robot navigation method based on importance gaze representation and obstacle Gaussian constraint

Abstract

A robot navigation method based on importance gaze representation and obstacle Gaussian constraint collects the original point cloud data information of the robot and obtains the point cloud data at the current moment. t , the robot collects the historical point cloud data sequence O within the time T and retains the historical point cloud data sequence O within the time T; the obtained point cloud data sequence O is input into the importance gaze representation module, and the historical depth map sequence O is represented r , O r Input the decision network δ of reinforcement learning and output the navigation action a at the current moment t ; Input O into the Gaussian obstacle constraint module, and judge the safety of the current state through the value interval of the predefined safety set h(O); The present invention can solve the security problem of the end-to-end network, and ultimately realize the efficient and safe autonomous navigation of the mobile robot in unfamiliar dynamic scenes.

Description

Technical Field

[0001] The invention relates to a robot autonomous navigation method, in particular to a robot navigation method based on importance gaze representation and obstacle Gaussian constraint, belonging to the technical field of robot navigation and positioning. Background Art

[0002] In recent years, with the continuous promotion of related applications on mobile robots, many countries in the world, including developed countries and some emerging economies, have focused on planning and deploying robots as national plans. Robots are creating new industries and new formats, promoting the transformation of production and consumption to intelligence, and thus profoundly affecting human production and life. Robot navigation technology is one of the key basic supports in the field of robotics, which ensures that robots can have stable and safe mobility in mission scenarios. However, with the increasing demand for robots in the military, industrial, production and other fields, the application scenarios of mobile robots have gradually shifted from closed, static, simple single indoor scenes to open, dynamic, complex and unfamiliar outdoor scenes. However, in the face of the transformation of application scenarios and the need for security, traditional methods have been greatly challenged due to their relatively closed execution logic.

[0003] Traditional navigation algorithms are aimed at static navigation environments, such as the widely used dynamic window method, Dijkstra algorithm, A* algorithm, etc. In order to further solve the navigation function of robots in complex dynamic environments, algorithms based on real-time response are based on specific interaction rules to infer the next behavior of dynamic obstacle avoidance, such as obstacle avoidance algorithm and navigation dynamic avoidance algorithm. However, this kind of real-time response navigation algorithm usually causes the robot to be short-sighted and move unnaturally. In real scenes, due to the complexity and unpredictability of obstacle movement behavior, safety requirements cannot be met. As an effective technology for safety assurance, obstacle control function ensures navigation safety through precise dynamic modeling. However, the strong dependence of the above aspects on high-dimensional semantic information of the scene makes it difficult to quickly apply it in unfamiliar scenes.

[0004] Reinforcement learning provides a powerful tool for robots to navigate in dynamic scenes by encouraging agents to freely explore and spontaneously learn adaptive strategies in scenes. By introducing an attention mechanism to understand the complex interactions between dynamic obstacles, good navigation effects are achieved in highly complex and dynamic outdoor scenes. However, these methods rely on additional perception modules to provide high-dimensional semantic information of the scene. In recent years, some studies have built end-to-end navigation networks through cameras, but due to the influence of factors such as light, season, and weather, it is difficult to apply them in real scenes. As a commonly used sensor for robot navigation, 3D LiDAR obtains environmental information by detecting the distance between the body and obstacles through lasers. Its robust perception characteristics overcome the problems of SIM2REAL. However, the high dimensionality and disorder of 3D LiDAR data make it difficult to be directly learned by reinforcement learning networks. At the same time, due to its black box nature, reinforcement learning itself lacks security guarantees for practical applications. Summary of the invention

[0005] The purpose of the present invention is to provide a robot navigation method based on importance gaze representation and obstacle Gaussian constraints. By proposing an importance gaze representation for original three-dimensional lidar information, the dependence on external perception modules and the model migration problems of real scenes can be solved, and the feasibility problem in actual unfamiliar dynamic scenes can be overcome. An obstacle Gaussian constraint based on controlling obstacle functions and Gaussian processes is proposed to correct reinforcement learning navigation decisions in unsafe sets, which can solve the security problem of end-to-end networks and ultimately realize efficient and safe autonomous navigation of mobile robots in unfamiliar dynamic scenes.

[0006] In order to achieve the above object, the present invention provides a robot navigation method based on importance gaze representation and obstacle Gaussian constraint, comprising the following steps:

[0007] S1: Use the 3D laser radar carried by the robot to collect the original point cloud data information under the robot's first-person perspective observation, and obtain the point cloud data O at the current moment t , the robot collects time T and retains the historical point cloud data sequence O within time T: O = {O t ,O t-1 ,…,O t-T};

[0008] S2: The historical point cloud data sequence O is represented as a historical depth map sequence O of fixed size T×H×W through the important gaze representation module. r =θ(O),O r ∈R 3 ;

[0009] S3: The historical depth map sequence O processed by the importance gaze representation module in step S2 rInput the decision network δ of reinforcement learning, and after processing, output the navigation action a at the current moment t , that is, a t =δ(O r );

[0010] S4: The historical point cloud data sequence O is input into the Gaussian barrier constraint module, and the safety of the current state is judged by the value interval of the predefined safety set h(O). The barrier function serves as a trigger for safety correction to determine whether the current state requires safety correction of the navigation behavior and outputs the final navigation action. Right now

[0011] The importance gaze representation module of step S2 of the present invention includes three sub-modules executed sequentially, namely a two-dimensional compression representation module, a field of view enhancement representation module based on an adaptive learning factor, and a spatial distance enhancement representation module based on a self-learning inverse function.

[0012] The specific steps of step S2 are as follows:

[0013] S21: The two-dimensional compression representation module converts the high-dimensional, disordered historical point cloud data sequence O into a two-dimensional image. The pixel values ​​are used to retain the depth information in the three-dimensional space to construct a two-dimensional full-view depth map, which can solve the disorder in the original data, retain the three-dimensional space information and effectively reduce the data dimension;

[0014] S22: The field of view enhancement representation module based on the adaptive learning factor will analyze the movement trend of obstacles in the observation space through the robot's historical observation in the two-dimensional full-field depth map. Combined with the robot's own movement history and movement trajectory, it will enhance the spatial representation of the robot's future position and dynamic objects. By incorporating the prior knowledge of human experts to enhance the perception representation, the representation accuracy of important intervals will be further increased.

[0015] S23: The spatial distance reinforcement learning representation module based on the inverse function of the self-learning parameter further strengthens the resolution clarity of the representation information at close distances and weakens the spatial resolution at long distances. The introduction of self-learning parameters further optimizes the choice of clarity distance according to the navigation strategy in historical experience, so as to realize the scale of the robot's selection of spatial distance importance according to scene changes.

[0016] In step S3 of the present invention, the reward function used in reinforcement learning for robot navigation is improved, wherein a reward function R that can directly face the original point cloud data is innovatively constructed, specifically:

[0017] R=α+R l +R g +R d+R r +R o

[0018] Where: α is the time penalty factor;

[0019] R l Reward for open space;

[0020] R g For the end point reward;

[0021] R d Provide rewards for finishing line guidance;

[0022] R r Provides a bonus for yaw angle guidance;

[0023] R o Bonus for collision avoidance.

[0024] The Gaussian obstacle constraint module proposed in step S4 of the present invention includes an improved control obstacle function through Gaussian process regression, and Gaussian process regression strengthens the dynamic tolerance of the control obstacle function to high data noise in unfamiliar scenes; at the same time, an improved definition of a safe space set for original point cloud data information is used to determine the absolute safety of the scene, and a shielding function for original point cloud data information is designed as a trigger for safety constraints in the Gaussian obstacle constraint module. The safety constraint is solved by a quadratic programming, specifically:

[0025] S41: The dynamic model of the mobile robot is modeled as Define the safe set in the control barrier function as:

[0026]

[0027] C, Int(C) defines the entire set of safe sets, the edge of safe sets, and the interior of safe sets respectively;

[0028] Through Nance's theorem, we can deduce The above conditions can be used as the necessary and sufficient conditions for the forward invariance of the safety set for the system, and the forward invariance can be used as the necessary and sufficient conditions for determining the safety of the robot in the current scene;

[0029] S42: Train the Gaussian regression model as a determiner of unsafe edge points. First, define the definition of unsafe boundaries.

[0030] min j＝1,…,N |p oe,k (t)-x i (t)||>d sample

[0031] Where: p oe,k(t) is the obstacle position;

[0032] x i (t) is the current position of the robot;

[0033] d sample is a hyperparameter used to predefine the minimum obstacle distance. Note that d sample When applying, you should first select a sufficiently small value to avoid obstacles;

[0034] S43: The point cloud data O at the current time T t The barrier Gaussian constraints constructed by the control barrier equation and Gaussian process regression are input. Specifically, O t Enter the obstacle equation designed for the original point cloud:

[0035]

[0036] Where: d * (t) is the shortest distance between the robot and the obstacle in the current state

[0037] d safe It is a Gaussian regression kernel function trained under user-defined safety threshold conditions and a safety distance controlled by hyperparameters;

[0038] When the obstacle function is based on the point cloud data O at the current moment t When the current scene is safe, that is, h(x) ≥ 0, the output navigation action a of the reinforcement learning decision network is directly output t ;

[0039] When the obstacle function is based on the point cloud data O at the current moment t When the current scene is unsafe, the navigation strategy output by reinforcement learning is corrected through the safety control of obstacle Gaussian constraints. Specifically:

[0040]

[0041] Where: u(t) needs to satisfy L f h(x)+L g h(x)u+γ(h(x))≥0, where is the Lie derivative of h(x) along the vector field. By using the constraint conditions, the forward invariance of the safety set is guaranteed, thereby ensuring the safety of the constrained action u(t);

[0042] The obstacle Gaussian constraint not only ensures the security of the reinforcement learning method based on the black box principle, but also the constraints on the original point cloud information ensure the timeliness of the calculation and the frequency of the safety constraints, which is more conducive to high-frequency control on the real robot.

[0043] Compared with the prior art, the present invention directly targets the original three-dimensional laser radar data, efficiently reduces the dimension and transforms the point cloud information through the importance gaze representation, introduces the adaptive learning factor to strengthen the key scene information in the representation environment, and constructs the self-learning parameter inverse function to strengthen the spatial position relationship of the important distance interval in the rendering point cloud information; the represented point cloud information is sent to the reinforcement learning decision network to input the navigation decision, and the design of the reward function based on the original point cloud data is innovatively proposed in the reinforcement learning to further optimize the learning process and improve the network learning efficiency; in addition, the obstacle Gaussian constraint is constructed to correct the dangerous decisions that may be made by the reinforcement learning strategy network, and the original point cloud information is controlled by the obstacle function to detect whether the current scene meets the safety set constraint and needs to be corrected by the forced strategy, wherein the Gaussian process of non-parametric Bayesian learning is used to fit the shielding function. Finally, the mobile robot achieves safe navigation in dynamic unfamiliar scenes through the reinforcement learning navigation strategy of the obstacle Gaussian constraint. The present invention utilizes reinforcement learning to construct an end-to-end navigation method for three-dimensional lidar to ensure its dynamic adaptability in unfamiliar scenes. At the same time, it proposes an obstacle Gaussian constraint technology for original point clouds to enhance the safety of applications in real scenes, ultimately ensuring the efficiency and safety of the overall end-to-end navigation method of the robot in real scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is the overall framework diagram of the present invention;

[0045] Figure 2 A visualization diagram of the original point cloud data processed by the importance gaze characterization module of the present invention;

[0046] Figure 3 Network architecture diagram for reinforcement learning navigation decision network. DETAILED DESCRIPTION

[0047] The present invention will be further described below in conjunction with the accompanying drawings.

[0048] like Figure 1-Figure 3 As shown, a robot navigation method based on importance gaze representation and obstacle Gaussian constraint includes the following steps:

[0049] S1: Use the 3D laser radar carried by the robot to collect the original point cloud data information under the robot's first-person perspective observation, and obtain the point cloud data O at the current moment t , the robot collects time T and retains the historical point cloud data sequence O within time T: O = {O t ,O t-1 ,…,O t-T};

[0050] Specifically, the present invention relies on a three-dimensional laser radar (LiDAR) device carried by the robot to collect point cloud data of the robot's surrounding environment. t Specifically, the laser radar collects distance information within the 360-degree field of view every second and generates a point cloud map including all obstacles and structures in the environment. The robot retains the point cloud data of the most recent T moments to form a historical point cloud data sequence O = {O t ,O t-1 ,…,O t-T}, historical data helps the robot make predictions and decisions in dynamic environments during navigation.

[0051] The hardware requirements of the present invention are as follows: the laser radar should have a high resolution and refresh rate. The present invention recommends the use of a 64-line or higher LiDAR sensor with an acquisition frequency of 10 Hz or higher (the specific frequency is set according to the robot's own computing power) to ensure that the robot can obtain sufficient spatial information in a dynamic environment.

[0052] S2: The historical point cloud data sequence O is input into the importance gaze representation module, which performs a series of complex transformation operations to transform the original disordered point cloud information into a structured historical depth map sequence O r =θ(O),O r ∈R 3 , to improve the efficiency and accuracy of subsequent reinforcement learning decisions;

[0053] The importance gaze representation module of step S2 of the present invention includes three sub-modules executed sequentially, namely a two-dimensional compression representation module, a field of view enhancement representation module based on an adaptive learning factor, and a spatial distance enhancement representation module based on a self-learning inverse function.

[0054] The specific steps of step S2 are as follows:

[0055] S21: Input the historical data point cloud sequence O into the two-dimensional compression representation module. The core task of the two-dimensional compression representation module is to convert the high-dimensional, disordered point cloud data into a two-dimensional full-view depth map. The pixel value in each depth map represents the depth information of the corresponding spatial point. The converted depth map can effectively represent the three-dimensional spatial information in the environment. Use projection technology to project the three-dimensional point cloud data onto a two-dimensional plane. Commonly used projection methods include spherical projection or cylindrical projection, which maps the point cloud data onto a plane image to obtain the depth value of each pixel. In order to improve computational efficiency, after the depth map is generated, it can be downsampled and normalized to reduce the resolution and complexity of the image, and finally obtain a time series depth map O. d ;

[0056] S22: The time-series depth map O obtained in step S21 dThe data is sent to the field of view enhancement representation module based on the adaptive learning factor. The field of view enhancement representation module based on the adaptive learning factor analyzes the movement trend of obstacles in the current scene based on the robot's movement trajectory in historical time. Specifically, the adaptive learning factor learns the behavior pattern of dynamic obstacles through the robot's observation records and strengthens the prediction of future positions. Specifically, the long short-term memory network (LSTM) model is used to predict the movement trend of obstacles. LSTM can effectively analyze historical data and predict the future movement trend of obstacles, providing movement trend information for the adaptive learning factor constructed by the subsequent attention mechanism. In addition, the prior knowledge of human experts is further utilized. For example, in common indoor environments, robots give priority to avoiding static obstacles such as furniture and walls, and differential robots give priority to the space in the area where their movement is located. The initial value of the importance mechanism is given by integrating human experience to enhance the stability and convergence speed of learning. Finally, the representation of important spaces is strengthened through the attention mechanism, and the time-series depth map O of field of view enhancement is output. l ;

[0057] S23: The time-series depth map O outputted in step S22 with enhanced field of view l The spatial distance enhancement representation module based on the self-learning inverse function is sent to the spatial distance enhancement representation module based on the self-learning inverse function. The spatial distance enhancement representation module based on the self-learning inverse function automatically adjusts the representation clarity at close and long distances by introducing self-learning parameters. The close-range environmental information is given a higher resolution so that the robot can make more accurate decisions; while the long-range spatial information can be appropriately weakened to reduce excessive attention to distant obstacles. The specific setting of the self-learning inverse function is

[0058]

[0059] Where: y i Represents the time series depth map O l The specific value of each pixel in ;

[0060] β i is a hyperparameter that controls the range;

[0061] The self-learning inverse function learns the appropriate distance weight factor through the back propagation algorithm. The weight factor will be adjusted continuously as the robot navigates in the environment, thereby optimizing the clarity distribution of environmental perception, and finally outputting a feature history depth map sequence O processed by the complete importance gaze representation module. r , its visualization image is as follows Figure 2 shown.

[0062] S3: The historical depth map sequence O processed by the importance gaze representation module in step S2 r Input into the decision network δ of reinforcement learning, and after processing, output the navigation action a at the current momentt , that is, a t =δ(O r );

[0063] The decision network is designed based on the reinforcement learning framework. The decision network uses a three-layer convolutional neural network (CNN) with a residual structure to process the input depth map sequence. The CNN layer extracts the spatial features in the depth map, and then uses the LSTM network to further extract the temporal features. Finally, the robot's current navigation action a is generated through the fully connected layer. t , its detailed network structure is as follows Figure 3 As shown; in the reinforcement learning training stage, the present invention innovatively designs a reward function R based on point cloud data, which includes the following specific reward items:

[0064] R=α+R l +R g +R d +R r +R o

[0065] Where: α is the time penalty factor, which is the penalty that the robot will receive in each decision, used to guide the robot to take the shortest decision step navigation point and avoid the robot from being locked;

[0066] R l For the open space reward, the change in the number of close points in the robot's forward field of view between the current moment and the previous moment is counted to evaluate whether the robot is heading towards a more open scene, thereby encouraging the robot to avoid the locked area as much as possible during navigation;

[0067] R g The robot receives a large reward when it reaches the destination. This helps guide the robot to clarify the navigation purpose and follow the destination guidance through the reverse transfer of the Markov chain.

[0068] R d The reward for guiding the robot to the end point is to continuously give the robot intensive positive rewards by encouraging it to keep approaching the end point, thereby strengthening the robot's behavior of constantly approaching the end point. d The robot is not penalized for moving away from the end point, which encourages the robot to freely explore in complex scenes. In addition, when there are close obstacles in the robot's forward direction, the end point guidance reward will be blocked.

[0069] R r A yaw angle guidance reward ensures the efficiency of the robot during exploration by encouraging the robot to keep facing the end point.

[0070] R oFor collision avoidance rewards, the number of close points in the point cloud within the robot's forward range is detected to measure the degree of crowding in the obstacle area of ​​the robot, thereby guiding the robot to learn obstacle avoidance actions during navigation and avoid future crowded areas in open scenes.

[0071] S4: The historical point cloud data sequence O is input into the Gaussian obstacle constraint module. The purpose of the Gaussian obstacle constraint module is to ensure the safety of the robot in complex environments, especially when it may encounter unseen scenes during navigation. The Gaussian obstacle constraint module includes an improved control obstacle function through Gaussian process regression, which strengthens the dynamic tolerance of the control obstacle function to high data noise in unfamiliar scenes; at the same time, it improves the definition of the safe space set for the original point cloud data to determine the absolute safety of the scene, and designs a shielding function for the original point cloud data as a trigger for the safety constraint in the Gaussian obstacle constraint module. The safety constraint is solved by a quadratic programming.

[0072] The safety of the current state is determined by the value interval of the predefined safety set h(O), where the barrier function acts as a trigger for safety correction to determine whether the current state requires safety correction of the navigation behavior and outputs the final navigation action. ,Right now

[0073] Specific:

[0074] S41: The dynamic model of the mobile robot is modeled as Define the safe set in the control barrier function as:

[0075]

[0076] C, Int(C) defines the entire set of safe sets, the edge of safe sets, and the interior of safe sets respectively;

[0077] Through Nance's theorem, we can deduce

[0078] The above conditions can be used as the necessary and sufficient conditions for the forward invariance of the safety set for the system, and the forward invariance can be used as the necessary and sufficient conditions for determining the safety of the robot in the current scene;

[0079] S42: Training a Gaussian regression model as a determiner of unsafe edge points. First, the present invention defines the definition of an unsafe boundary:

[0080] min j＝1,…,N ||p oe,k (t)-x i (t)||>d sample

[0081] Where: p oe,k (t) is the obstacle position;

[0082] x i (t) is the current position of the robot;

[0083] d sample is a hyperparameter used to predefine the minimum obstacle distance. Note that d sample When applying, you should first select a sufficiently small value to avoid obstacles;

[0084] S43: The point cloud data O at the current time T t Input the obstacle Gaussian constraint constructed by the control obstacle equation and Gaussian process regression, specifically, the point cloud data O t Enter the obstacle equation designed for the original point cloud:

[0085]

[0086] Where: d * (t) is the shortest distance between the robot and the obstacle in the current state

[0087] d safe It is a Gaussian regression kernel function trained under user-defined safety threshold conditions and a safety distance controlled by hyperparameters;

[0088] When the obstacle function is based on the point cloud data O at the current moment t When the current scene is safe, that is, h(x) ≥ 0, the output navigation action a of the reinforcement learning decision network is directly output t ;

[0089] When the obstacle function is based on the point cloud data O at the current moment t When the current scene is unsafe, that is, h(x)<0, the navigation strategy output by reinforcement learning is corrected through the safety control of obstacle Gaussian constraints. Specifically:

[0090]

[0091] Where: u(t) needs to satisfy L f h(x)+L g h(x)u+γ(h(x))≥0

[0092] in is the Lie derivative of h(x) along the vector field. By using the constraint conditions, the forward invariance of the safety set is guaranteed, thereby ensuring the safety of the constrained action u(t);

[0093] The obstacle Gaussian constraint not only ensures the security of the reinforcement learning method based on the black box principle, but also the constraints on the original point cloud information ensure the timeliness of the calculation and the frequency of the safety constraints, which is more conducive to high-frequency control on the real robot.

Claims

1. A robot navigation method based on importance gaze representation and obstacle Gaussian constraint, characterized in that: The following steps are involved: S1: Use the 3D laser radar carried by the robot to collect the original point cloud data information under the robot's first-person perspective observation, and obtain the point cloud data O at the current moment t , the robot collects time T and retains the historical point cloud data sequence O within time T: O = {O t ,O t-1 ,…,O t-T }; S2: Input the point cloud data sequence O obtained in step S1 into the importance gaze representation module, and use the importance gaze representation module to represent the historical point cloud data sequence O as a historical depth map sequence O of a fixed size T×H×W. r =θ(O),O r ∈R 3 ; S3: The historical depth map sequence O processed by the importance gaze representation module in step S2 r Input the decision network δ of reinforcement learning, and after processing, output the navigation action a at the current moment t , that is, a t =δ(O r ); S4: The historical point cloud data sequence O is input into the Gaussian barrier constraint module, and the safety of the current state is judged by the value interval of the predefined safety set h(O). The barrier function serves as a trigger for safety correction to determine whether the current state requires safety correction of the navigation behavior and outputs the final navigation action. Right now The importance gaze representation module of step S2 includes three submodules executed sequentially, namely a two-dimensional compression representation module, a visual field enhancement representation module based on an adaptive learning factor, and a spatial distance enhancement representation module based on a self-learning inverse function; The specific steps of step S2 are as follows: S21: The two-dimensional compression representation module converts the high-dimensional, disordered historical point cloud data sequence I into a two-dimensional image, and the pixel values ​​are used to retain the depth information in the three-dimensional space to construct a two-dimensional full-view depth map; S22: The field of view enhancement representation module based on the adaptive learning factor will analyze the movement trend of obstacles in the observation space through the robot's historical observation in the two-dimensional full-field depth map. Combined with the robot's own movement history and movement trajectory, it will enhance the spatial representation of the robot's future position and dynamic objects. By incorporating the prior knowledge of human experts to enhance the perception representation, the representation accuracy of important intervals will be further increased. S23: The spatial distance reinforcement learning representation module based on the inverse function of the self-learning parameter further strengthens the resolution clarity of the representation information at close distances and weakens the spatial resolution at long distances. The introduction of self-learning parameters further optimizes the choice of clarity distance according to the navigation strategy in historical experience, so as to realize the scale of the robot's selection of spatial distance importance according to scene changes.

2. A robot navigation method based on importance gaze representation and obstacle Gaussian constraint according to claim 1, characterized in that: In step S3, the reward function used in reinforcement learning for robot navigation is improved, in which an innovative reward function R that can directly face the original point cloud data is constructed, specifically: R=α+R l +R g +R d +R r +R o Where: α is the time penalty factor; R l Reward for open space; R g For the end point reward; R d Provide rewards for finishing line guidance; R r Provides a bonus for yaw angle guidance; R o Bonus for collision avoidance.

3. A robot navigation method based on importance gaze representation and obstacle Gaussian constraint according to claim 1, characterized in that: The Gaussian obstacle constraint module proposed in step S4 includes an improved control obstacle function through Gaussian process regression. Gaussian process regression strengthens the dynamic tolerance of the control obstacle function for high data noise in unfamiliar scenes. At the same time, the safety space set for the original point cloud data information is improved to determine the absolute safety of the scene. A shielding function for the original point cloud data information is designed as a trigger for the safety constraint in the Gaussian obstacle constraint module. The safety constraint is solved by a quadratic programming. Specifically: S41: The dynamic model of the mobile robot is modeled as Define the safe set in the control barrier function as: C, Int(C) defines the entire set of safe sets, the edge of safe sets, and the interior of safe sets respectively; Through Nance's theorem, we can deduce The above conditions can be used as the necessary and sufficient conditions for the forward invariance of the safety set for the system, and the forward invariance can be used as the necessary and sufficient conditions for determining the safety of the robot in the current scene; S42: Train the Gaussian regression model as a determiner of unsafe edge points. First, define the definition of unsafe boundaries. min j＝1,…,N ||p oe,k (t)-x i (t)||>d sample Where: p oe,k (t) is the obstacle position; x i (t) is the current position of the robot; d sample It is a hyperparameter used to predefine the minimum obstacle distance; S43: The point cloud data O at the current time T t The barrier Gaussian constraints constructed by the control barrier equation and Gaussian process regression are input. Specifically, O t Enter the obstacle equation designed for the original point cloud: Where: d * (t) is the shortest distance between the robot and the obstacle in the current state d safe It is a Gaussian regression kernel function trained under user-defined safety threshold conditions and a safety distance controlled by hyperparameters; When the obstacle function is based on the point cloud data O at the current moment t When the current scene is safe, that is, h(x) ≥ 0, the output navigation action a of the reinforcement learning decision network is directly output t ; When the obstacle function is based on the point cloud data O at the current moment t When the current scene is unsafe, the navigation strategy output by reinforcement learning is corrected through the safety control of obstacle Gaussian constraints. Specifically: Where: u(t) needs to satisfy L f h(x)+L g h(x)u+γ(h(x))≥0, where is the Lie derivative of h(x) along the vector field. By using the constraints, the forward invariance of the safety set is guaranteed, thereby ensuring the safety of the constrained action u(t).

Images