A mobile robot path planning method in a long-distance dense crowd scene

Abstract

The application discloses a kind of long-distance dense crowd scene under mobile robot path planning method, comprising: S1, the starting point and end point position of mobile robot on map are acquired, global path is generated using global planner;S2, waypoint and landmark are extracted on global path, and coordinate system transformation is carried out to obtain the waypoint and landmark under robot coordinate system;S3, the waypoint and landmark under robot coordinate system and the point cloud data collected by mobile robot are input into the decision strategy of deep learning neural network, and decision strategy learning is carried out to obtain optimal decision model;S4, the waypoint and landmark under robot coordinate system and the point cloud data collected by mobile robot are input into optimal decision model, and optimal local target point output according to fixed decision time is obtained, planning strategy learning is carried out to obtain optimal planning model, to obtain the path planning of mobile robot.The application can improve the safety and time efficiency of mobile robot in long-distance dense crowd environment.

Description

Technical Field

[0001] This invention relates to the field of mobile robot technology, and in particular to a method for mobile robot path planning in long-distance, densely populated crowd scenarios. Background Technology

[0002] With the rapid development of artificial intelligence and autonomous navigation technologies, and the industrial upgrading of manufacturing and service sectors under the "intelligent manufacturing" framework, the robotics industry has entered a period of rapid growth. Simultaneously, rapid economic development has led to higher demands for quality of life, prompting the introduction of intelligent, efficient, and convenient mobile robots into people's lives. In recent years, mobile robots have been widely applied in production and daily life scenarios such as warehousing and logistics, healthcare, e-commerce retail, and smart cities. The increasing complexity of these application scenarios has placed higher demands on the safety, flexibility, efficiency, and scalability of mobile robots.

[0003] Path planning algorithms are crucial for ensuring collision-free movement of robots from a starting point to a destination. In static scenarios, these algorithms determine the shortest path avoiding obstacles based on a scene map. In dynamic scenarios, timely avoidance of dynamic obstacles is also necessary. Long-distance, densely populated scenarios are common in real-world mobile robot applications. In such scenarios, traditional planning algorithms suffer from low success rates due to limited robot traversable areas and unpredictable pedestrian movements, resulting in low time efficiency for mobile robots in dense environments. To address this issue, numerous data-driven methods have been proposed. Among these, deep reinforcement learning-based methods have proven to be safer and more flexible in dynamic environments due to their advantages such as direct interaction with the environment, trial-and-error learning, model-free operation, and reward-driven approaches. However, agent-level deep reinforcement learning path planning methods rely on perfect perception, as the input consists of the agent's position and velocity information after target detection and tracking. Perception errors can affect the safety of robot planning. End-to-end deep reinforcement learning path planning methods often face convergence difficulties in long-distance scenarios due to "reward sparsity," making them difficult to directly apply to real-world, large-scale scenarios.

[0004] It should be noted that the information disclosed in the background section above is only for understanding the background of the present invention, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to solve the technical problems of low safety and time efficiency of mobile robots in long-distance dense crowd environments, and to propose a path planning method for mobile robots in long-distance dense crowd scenarios.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for path planning of mobile robots in long-distance, densely populated crowd scenarios includes the following steps:

[0008] S1. Obtain the starting and ending points of the mobile robot on the map, and use the global planner to generate a global path;

[0009] S2. Extract waypoints and landmarks on the global path and perform coordinate system transformation to obtain waypoints and landmarks in the robot coordinate system;

[0010] S3. Input the waypoints and landmarks in the robot coordinate system and the point cloud data collected by the mobile robot into the decision-making strategy of the deep learning neural network, and learn the decision-making strategy to obtain the optimal decision model.

[0011] S4. Input waypoints and landmarks in the robot coordinate system and point cloud data collected by the mobile robot into the optimal decision model to obtain the optimal local target point output at a fixed decision time. Then, learn the planning strategy to obtain the optimal planning model and thus obtain the path planning of the mobile robot.

[0012] In some embodiments, in step S1, the global planner employs the Dijkstra algorithm.

[0013] In some embodiments, step S2 specifically includes: extracting waypoints on the global path using a fixed-distance sampling method, and extracting landmarks using a global path curvature variation method.

[0014] In some embodiments, step S2, the coordinate system transformation includes: performing coordinate system transformation on waypoints and landmarks, and rotating and translating the waypoints to their coordinate positions (x, y, y) in the map coordinate system. wp ,y wp ) and the coordinates of the landmark in the map coordinate system (x lm ,y lm The formulas for transforming the coordinate systems of waypoints and landmarks from the map coordinate system to the robot coordinate system are as follows:

[0015]

[0016]

[0017] in, Indicates the angle between the map coordinate system and the robot coordinate system; x' wp This represents the x-coordinate of the waypoint in the robot's coordinate system; y' wp This represents the ordinate position of the waypoint in the robot's coordinate system; x' lm This represents the x-coordinate of the landmark in the robot's coordinate system; y' lm Indicates the vertical coordinate of the landmark in the robot's coordinate system; xbase_footprint This indicates the x-coordinate of the robot's coordinate system origin in the map coordinate system; y base_footprint This indicates the ordinate position of the robot's coordinate system origin in the map coordinate system.

[0018] In some embodiments, in step S3, the point cloud data is collected by the lidar of a mobile robot.

[0019] In some embodiments, the process of learning the decision strategy to obtain the optimal decision model specifically includes: selecting a local target point and using a local planner to plan and calculate the speed and angular velocity of the mobile robot; reselecting a new local target point and planning at fixed decision intervals; training a deep learning neural network to update the decision strategy and obtain the optimal decision model.

[0020] In some embodiments, the constraints of the decision-making strategy in step S3 are as follows:

[0021]

[0022] Where t represents time, s represents the state of the mobile robot in the environment, and π represents the strategy of the mobile robot; r represents the distance information measured by the mobile robot in the environment at time t. robot Indicates the radius of the mobile robot; Let d represent the positions of the moving robot and the destination at time t, respectively. goal This indicates the distance range to the destination, within which the mobile robot is considered to have reached the destination;

[0023] Reward function in the decision-making process The settings are as follows:

[0024]

[0025] The time reward is defined as follows:

[0026]

[0027] Where, ω t Let Δt be the coefficient for the time reward, and Δt be the decision time represented by each step update in the reinforcement learning decision. This is a small penalty imposed on the mobile robot for each decision, with the aim of reducing the decision-making time of the mobile robot throughout the decision-making and planning task, thereby allowing it to reach the destination faster.

[0028] The finish line reward is defined as follows:

[0029]

[0030] Where, r g This represents the reward the mobile robot receives when it reaches the finish line. and Let r represent the positions of the moving robot and the destination at time t, respectively. goal "Otherwise" indicates the range of the endpoint; when the mobile robot moves into this range, it is considered to have reached the endpoint.

[0031] The collision reward is defined as follows:

[0032]

[0033] Where, r collision It is the reward that the mobile robot receives when it collides with an obstacle. It is the set of lidar detection distance values ​​at time t, r robot It is the radius of the mobile robot;

[0034] The global feature point landmark reward is defined as follows:

[0035]

[0036] Where, ω l It is the coefficient of the landmark reward. and These represent the positions of the mobile robot at time t-1 and time t, respectively. and These represent the locations of the nearest landmarks to the mobile robot at time t-1 and time t, respectively.

[0037] The reward for changes in the angle of a local target point is defined as follows:

[0038]

[0039] Where, ω a It is the coefficient for the reward of changes in the angle of the local target point. and These represent the angles between the local target point at time t-1 and time t relative to the robot coordinate system.

[0040] In some embodiments, learning the planning strategy to obtain the optimal planning model specifically includes: inputting the optimal local target point, endpoint position, and point cloud data into the planning strategy of a deep learning neural network for planning, obtaining the speed and angular velocity of the mobile robot, training the deep learning neural network to update the planning strategy, and obtaining the optimal planning model; and a reward function during the planning process. The settings are as follows:

[0041]

[0042] The reward for a local target point is defined as follows:

[0043]

[0044] Where, ω sg The coefficient for the reward of the local target point. and These represent the positions of the mobile robot at time t-1 and time t, respectively. and These represent the positions of the local target point (subgoal) at time t-1 and time t, respectively.

[0045] The collision reward is defined as follows:

[0046]

[0047] Where, r collision The reward a mobile robot receives when it collides with an obstacle. Let r represent the set of detection range values ​​of the lidar at time t. robot Indicates the radius of the mobile robot; Otherwise indicates other cases;

[0048] The reward for static security zones is defined as follows:

[0049]

[0050] Where, ω sz The coefficient for the static security zone reward. Let r represent the set of detection range values ​​of the pedestrian layer lidar at time t. sz Indicates the extent of the static safety zone;

[0051] The speed reward is defined as follows:

[0052]

[0053] Where, ω vel w represents the coefficient for speed bonus. t Let t represent the robot's angular velocity at time t.

[0054] In some embodiments, the deep learning neural network employs CNN, LSTM, or RNN.

[0055] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in any of the preceding claims.

[0056] The beneficial effects of this invention compared to the prior art include:

[0057] The mobile robot path planning method provided by this invention for long-distance, densely populated scenarios obtains an optimal decision model through decision strategy learning, enabling the mobile robot to make decisions and avoid dense crowds, thus improving the time efficiency of the mobile robot in dense environments. Furthermore, by obtaining the optimal local target point output at a fixed decision time through the optimal decision model, and learning the planning strategy to obtain the optimal planning model, the method yields the mobile robot's path planning. This improves the mobile robot's obstacle avoidance ability against pedestrians whose speeds are unpredictable, thereby enhancing the safety of the mobile robot in pedestrian environments.

[0058] In some embodiments, a local target point is selected through the decision-making strategy of a deep learning neural network, and a local planner is used to plan and calculate the speed and angular velocity of the mobile robot. A new local target point is selected and planned again at fixed decision time intervals. The deep learning neural network is trained to update the decision-making strategy and obtain the optimal decision model, so that the mobile robot can make decisions in real time according to the dynamic changes of the environment and avoid dense crowds. This avoids the "robot freezing" problem caused by traditional planning algorithms and further improves the time efficiency of the mobile robot in dense environments.

[0059] In some embodiments, a planning strategy using a deep learning neural network is used to plan the optimal local target point, endpoint position, and point cloud data to obtain the speed and angular velocity of the mobile robot. The deep learning neural network is then trained to update the planning strategy and obtain the optimal planning model, thereby further improving the mobile robot's obstacle avoidance capability against pedestrians whose speeds are difficult to predict.

[0060] Other beneficial effects of the embodiments of the present invention will be further described below. Attached Figure Description

[0061] Figure 1 This is a flowchart of the path planning method for a mobile robot in an embodiment of the present invention.

[0062] Figure 2 This is a general flowchart of the path planning method for a mobile robot in an embodiment of the present invention.

[0063] Figure 3 This is an overall framework diagram of the simulation system in this embodiment of the invention.

[0064] Figures 4a to 4b This is a visualization of Waypoints and Landmarks in an embodiment of the present invention.

[0065] Figure 5 This is a schematic diagram of the coordinate system transformation between Waypoints and Landmarks in an embodiment of the present invention.

[0066] Figure 6 This is a flowchart of the intelligent decision-making module in an embodiment of the present invention.

[0067] Figure 7 This is a schematic diagram of the action space resolution calculation of the intelligent decision-making module in an embodiment of the present invention.

[0068] Figure 8 This is a schematic diagram of the action space of the intelligent decision-making module in an embodiment of the present invention.

[0069] Figure 9 This is a schematic diagram of the network structure in the intelligent decision-making stage of this invention.

[0070] Figure 10 This is a flowchart of the intelligent planning module in an embodiment of the present invention.

[0071] Figure 11 This is a schematic diagram of the network structure during the intelligent planning stage in an embodiment of the present invention. Detailed Implementation

[0072] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0073] In long-distance, densely populated scenarios, to reduce the complexity of dynamic scenes and address the convergence problem in such environments, two-stage path planning methods decouple decision-making and planning, using different strategies to determine where the mobile robot goes and how it gets there. Based on whether deep reinforcement learning is utilized, two-stage path planning methods can be categorized into three types: intelligent decision-making + traditional planning, rule-based decision-making + intelligent planning, and intelligent decision-making + intelligent planning. Intelligent decision-making + traditional planning is easy to deploy and offers a certain level of security, but its use of traditional planning algorithms makes it unsuitable for densely populated environments. Rule-based decision-making + intelligent planning is suitable for long-distance crowd navigation, but manually formulated rules cannot effectively handle unpredictable changes in pedestrian behavior. Intelligent decision-making + intelligent planning has proven to be efficient and safe in obstacle avoidance in multi-robot applications, but its application in indoor long-distance crowd navigation is currently lacking.

[0074] To address issues such as low success rate, low time efficiency, and low intelligence in mobile robot planning in real-world, large-scale dynamic scenarios, it is crucial to develop a safe and efficient two-stage path planning method that combines intelligent decision-making and intelligent planning for indoor, long-distance, densely populated crowd scenarios.

[0075] like Figure 1 As shown, this embodiment of the invention provides a method for mobile robot path planning in long-distance, densely populated crowd scenarios, including the following steps:

[0076] S1. Obtain the starting and ending points of the mobile robot on the map, and use the global planner to generate a global path.

[0077] The global planner uses Dijkstra's algorithm.

[0078] S2. Extract Waypoints and Landmarks on the global path and perform coordinate system transformation to obtain Waypoints and Landmarks in the robot coordinate system.

[0079] Specifically, Waypoints are extracted on the global path using a fixed-distance sampling method, and Landmarks are extracted based on the curvature variation of the global path.

[0080] Coordinate system transformation includes: transforming waypoints and landmarks into their coordinate systems, and then rotating and translating them to change the waypoints' coordinate positions (x, y, y) in the map coordinate system. wp ,y wp ) and the coordinates of the landmark in the map coordinate system (x lm ,y lm The formula for transforming the waypoint coordinates from the map coordinate system to the robot coordinate system is as follows:

[0081]

[0082] in, Indicates the angle between the map coordinate system and the robot coordinate system; x' wp This represents the x-coordinate of the waypoint in the robot's coordinate system; y' wp The x-coordinate represents the waypoint's position in the robot's coordinate system. base_footprint This indicates the x-coordinate of the robot's coordinate system origin in the map coordinate system; y base_footprint This indicates the ordinate position of the robot's coordinate system origin in the map coordinate system.

[0083] The formula for transforming the coordinate system of the landmark is as follows:

[0084]

[0085] in, Indicates the angle between the map coordinate system and the robot coordinate system; x' lm This represents the x-coordinate of the landmark in the robot's coordinate system; y' lmIndicates the vertical coordinate of the landmark in the robot's coordinate system; x base_footprint This indicates the x-coordinate of the robot's coordinate system origin in the map coordinate system; y base_footprint This indicates the ordinate position of the robot's coordinate system origin in the map coordinate system.

[0086] S3. Input the Waypoints and Landmarks in the robot coordinate system and the point cloud data collected by the mobile robot into the decision-making strategy of the deep learning neural network, and learn the decision-making strategy to obtain the optimal decision model.

[0087] Point cloud data is collected using the LiDAR of a mobile robot.

[0088] Specifically, the process of learning the decision-making strategy to obtain the optimal decision model includes: selecting a local target point and using a local planner to calculate the speed and angular velocity of the mobile robot; reselecting a new local target point and planning at fixed decision intervals; training a deep learning neural network to update the decision-making strategy and obtain the optimal decision model. The deep learning neural network can be CNN, LSTM, or RNN.

[0089] The constraints of the decision-making strategy are as follows:

[0090]

[0091] Where t represents time, s represents the state of the mobile robot in the environment, and π represents the strategy of the mobile robot; r represents the distance information measured by the mobile robot in the environment at time t. robot Indicates the radius of the mobile robot; Let d represent the positions of the moving robot and the destination at time t, respectively. goal This indicates the distance range to the destination, within which the mobile robot is considered to have reached the destination.

[0092] Reward function in the decision-making process The settings are as follows:

[0093]

[0094] The time reward is defined as follows:

[0095]

[0096] Where, ω t Let Δt be the coefficient for the time reward, and Δt be the decision time represented by each step update in the reinforcement learning decision. This is a small penalty imposed on the mobile robot for each decision, with the aim of reducing the decision-making time of the mobile robot throughout the decision-making and planning task, thereby allowing it to reach the destination faster.

[0097] The finish line reward is defined as follows:

[0098]

[0099] Where, r g This represents the reward the mobile robot receives when it reaches the finish line. and Let r represent the positions of the moving robot and the destination at time t, respectively. goal "Otherwise" indicates the range of the endpoint; when the mobile robot moves into this range, it is considered to have reached the endpoint.

[0100] The collision reward is defined as follows:

[0101]

[0102] Where, r collision It is the reward that the mobile robot receives when it collides with an obstacle. It is the set of lidar detection distance values ​​at time t, r robot It is the radius of the mobile robot;

[0103] The global feature point landmark reward is defined as follows:

[0104]

[0105] Where, ω l It is the coefficient of the landmark reward. and These represent the positions of the mobile robot at time t-1 and time t, respectively. and These represent the locations of the nearest landmarks to the mobile robot at time t-1 and time t, respectively.

[0106] The reward for changes in the angle of a local target point is defined as follows:

[0107]

[0108] Where, ω a It is the coefficient for the reward of changes in the angle of the local target point. and These represent the angles between the local target point at time t-1 and time t relative to the robot coordinate system.

[0109] S4. Input the Waypoints and Landmarks in the robot coordinate system and the point cloud data collected by the mobile robot into the optimal decision model to obtain the optimal local target point output at a fixed decision time. Then, learn the planning strategy to obtain the optimal planning model and thus obtain the path planning of the mobile robot.

[0110] Specifically, learning the planning strategy to obtain the optimal planning model includes: inputting the optimal local target point, endpoint position, and point cloud data into the planning strategy of a deep learning neural network for planning, obtaining the mobile robot's speed and angular velocity, training the deep learning neural network to update the planning strategy, and obtaining the optimal planning model. The deep learning neural network can be CNN, LSTM, or RNN.

[0111] Reward function during planning The settings are as follows:

[0112]

[0113] The reward for a local target point is defined as follows:

[0114]

[0115] Where, ω sg The coefficient for the reward of the local target point. and These represent the positions of the mobile robot at time t-1 and time t, respectively. and These represent the positions of the local target point (subgoal) at time t-1 and time t, respectively.

[0116] The collision reward is defined as follows:

[0117]

[0118] Where, r collision The reward a mobile robot receives when it collides with an obstacle. Let r represent the set of detection range values ​​of the lidar at time t. robot Indicates the radius of the mobile robot; Otherwise indicates other cases;

[0119] The reward for static security zones is defined as follows:

[0120]

[0121] Where, ω szThe coefficient for the static security zone reward. Let r represent the set of detection range values ​​of the pedestrian layer lidar at time t. sz Indicates the extent of the static safety zone;

[0122] The speed reward is defined as follows:

[0123]

[0124] Where, ω vel w represents the coefficient for speed bonus. t Let t represent the robot's angular velocity at time t.

[0125] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements any of the methods described above.

[0126] This invention proposes a two-stage deep reinforcement learning method for intelligent decision-making and intelligent planning in long-distance, densely populated scenarios. This method decouples decision-making and planning tasks, reducing the difficulty of learning dynamic scenarios and improving robot safety and runtime efficiency. Furthermore, it establishes a two-dimensional simulation environment based on a single-line LiDAR solution, compatible with ROS (Robot Operating System), facilitating the deployment of simulation training algorithms to real robots for joint testing. It has lower processor requirements and faster computation updates, effectively reducing the time and hardware costs of deep reinforcement learning training and updates. The simulation system uses a social force model to model pedestrian motion, presenting irregular pedestrian movements in real-world scenarios, thus more closely resembling real-world conditions and reducing the gap between simulation and reality, making it easier to deploy in real robots. Through intelligent decision-making strategies, this method enables mobile robots to make real-time decisions based on dynamic environmental changes and avoid dense crowds, avoiding the "robot freezing" problem caused by traditional planning algorithms, improving the time efficiency of mobile robots in dense environments. Additionally, the intelligent planning strategy enhances the mobile robot's obstacle avoidance ability against pedestrians with unpredictable speeds, improving the safety of mobile robots in pedestrian environments.

[0127] The following describes specific embodiments of mobile robot path planning methods in long-distance, densely populated crowd scenarios.

[0128] This embodiment proposes a two-stage deep reinforcement learning method combining intelligent decision-making and intelligent planning for long-distance, densely populated crowd scenarios. This method mainly consists of the following four steps:

[0129] The first step is to generate a safe global path that does not collide with static obstacles on the prior map using a global planner based on the Dijkstra algorithm, based on the starting and ending points.

[0130] The second step is to extract Waypoints by sampling at fixed distances along the global path and extract Landmarks by sampling based on the curvature variation of the global path.

[0131] The third step involves learning the intelligent decision-making strategy in the Phase 1 decision-making module. This involves observing the Waypoints and Landmarks generated in the second step, as well as the point cloud data collected by the mobile robot's own LiDAR, and then applying the decision-making strategy accordingly. Figure 7 Within the action space shown, a local target point is selected. The traditional optimization-based TEB algorithm is used as the local planner. At fixed decision intervals, a new decision point is selected based on observations. The network is then trained to update the decision strategy and the optimal decision model is saved. The TEB algorithm is an optimization-based algorithm with constraints including avoiding collisions with obstacles and minimizing the time to the target point. The optimal solution is obtained by optimizing based on these constraints. The optimal decision model can be the model with the highest reward function among every 200 tasks updated and saved during network training.

[0132] The fourth step is the learning of the planning strategy in the Phase 2 planning module. The Waypoints and Landmarks generated in the second step, along with data collected by the robot's own LiDAR, are input into the optimal decision model to obtain the optimal local target point output at a fixed decision time. The Phase 2 planning strategy is then used to replace the traditional TEB planning algorithm from the third step for planning, updating the planning strategy and saving the optimal planning model. The overall flow of the mobile robot path planning method in this embodiment is as follows: Figure 2 As shown.

[0133] This embodiment also establishes a simulation system based on the lightweight and convenient two-dimensional physics engine Flatland. Leveraging Flatland's compatibility with ROS, a deep reinforcement learning algorithm is integrated into the navigation framework NavigationStack as a plugin, facilitating the direct application of the simulation-learned algorithm to real robot planning tasks. This simulation environment offers a simple and convenient solution for path planning problems based on deep reinforcement learning algorithms in complex dynamic scenarios, boasts high scalability, and its ROS compatibility reduces the difficulty of transitioning simulation to real-world deployment.

[0134] The specific process of this embodiment is described as follows:

[0135] I. Building a simulation system for deep reinforcement learning training

[0136] In real-world scenarios, mobile robots face challenges in learning through interaction with their environment, such as low efficiency in collecting experience and the need for direct collisions with obstacles to collect failed experiences. Therefore, establishing an efficient and realistic offline simulation system is crucial for the learning planning strategy of deep reinforcement learning algorithms.

[0137] like Figure 3 As shown, this embodiment establishes a lightweight, simple, convenient, and realistic two-dimensional path planning simulation system for dynamic scenarios, facilitating the deployment of learning algorithms on real robots. The system utilizes the lightweight and efficient two-dimensional simulation simulator Flatland, the Pedsim plugin for driving pedestrian motion based on a social force model, the ROS navigation framework Navigation Stack, and integrates all components based on the ROS communication mechanism.

[0138] Flatland reduces the Z-axis information in 3D space, compressing the Gazebo environment into a plane. This reduction in information allows the simulator to run with fewer computational resources, making it more efficient. In Flatland, obstacle creation first requires creating YAML description files for various static obstacle and pedestrian models, then placing these files at designated locations on the map. Pedestrians are modeled as circles with a radius of 0.3m, similar to the radius of an adult in a real-world scene. The mobile robot is modeled as a circle with a radius of 0.3m, slightly larger than the radius of the Turtlebot2 chassis in a real-world scene, leaving sufficient safety thresholds. The robot configuration also utilizes Flatland's built-in differential motion and single-line LiDAR plugins to configure the mobile robot's kinematic model and sensor information. The LiDAR's detection range is set to [0, 12m], detection angle to [0, 360°], resolution to 1°, and update frequency to 10Hz. The LiDAR updates the obstacle situation in the environment every 0.1s and returns a sequence of 360 distance values.

[0139] The main process of pedestrian motion control is as follows: First, pedestrian motion parameters are configured in the Pedsim (pedestrian simulation) module. Then, the behavior of each pedestrian is calculated using a social force model, and real-time topics are published. Finally, the pedestrian model is controlled to move within the scene. In configuring the pedestrian motion parameters, the characteristics and speed of pedestrian movement in real-world scenarios are considered. The initial speed of the pedestrian is set to a Gaussian distribution v. pred Sampling is performed in ~N(0.6,0.2)m / s. The starting and ending positions of pedestrians are defined at random locations in the feasible space of the map. The pedestrian type is adults, and the number of pedestrians is randomly selected from 6 to 12.

[0140] The main process of motion control for the mobile robot is as follows: First, given the starting and ending points of the mobile robot on the map, a global planner is used to perform global path planning to obtain a safe, collision-free path that avoids static obstacles on the map. Then, according to the optimal decision model in the Stage 1 decision module, a local target point is selected in the action space. Guided by the local target point, local planning is performed to avoid dynamic obstacles. To facilitate the subsequent deployment of deep reinforcement learning-based path planning algorithms on real robots, the simulation system uses the ROS Navigation Stack (a collection of two-dimensional navigation function packages provided by ROS) framework for robot control. The ROS Navigation Stack provides interfaces to extend the path planning algorithm as a plugin. The local planner for the mobile robot's motion control inherits from the BaseLocalPlanner base class. The intelligent planning strategy in the Stage 2 planning module is customized as the rl_local_planner planner, which subscribes to the velocity output of the convolutional neural network and directly publishes the velocity to make the mobile robot move in the simulation environment.

[0141] II. Global Path Feature Point Extraction

[0142] Input: Prior map information, starting and ending points of the planned task.

[0143] Output: Global path points and landmarks

[0144] The "sparse reward" problem in end-to-end deep reinforcement learning causes mobile robots to become aimless and unable to reach the destination within the specified time in large-scale, long-distance scenarios when directly using deep reinforcement learning for path planning. The global path obtained through global planning, as prior information, can gradually guide the mobile robot towards the destination, enhancing the purposefulness of the robot's planning task. However, with too much information on long-distance global paths, not all points on the global path can enhance the robot's planning purposefulness; on the contrary, too much invalid information can negatively impact the robot's planning. Therefore, this embodiment proposes downsampling at fixed intervals of 3m along the global path to extract Waypoints to represent the approximate direction of the global path, and extracting Landmarks based on abrupt changes in the curvature of the global path to reflect changes in the surrounding environment. The visualization of Waypoints and Landmarks is shown below. Figures 4a to 4b As shown in the figure, circles represent Waypoints, and triangles represent Landmarks. It can be seen from the figure that Waypoints sparsify the global path while retaining the overall direction information of the global path, while Landmarks ensure that detailed changes in the global path are not lost.

[0145] The extracted Waypoints and Landmarks share the same global path coordinate system (map coordinate system). The point sets of Waypoints and Landmarks will serve as the observation space for the deep reinforcement learning algorithm, collecting the values ​​of observation data relative to the robot's coordinate system during real-time robot motion. Therefore, coordinate system transformation is required for the feature points. For example... Figure 5 As shown, the feature points of the global path, after rotation and translation, ultimately determine the coordinates (x, y) of the Waypoints in the map coordinate system. wp ,y wp ) and the coordinates of Landmarks in the map coordinate system (x lm ,y lm This involves transforming the coordinate system from the map coordinate system ( / map) to the robot coordinate system ( / base_footprint). The formula for coordinate system transformation in Waypoints is as follows:

[0146]

[0147] in, The meaning is the angle between the / map and / base_footprint coordinate systems; (x wp ,y wp The meaning of ) is the coordinate position of waypoints in the / map coordinate system; x' wp The meaning is the x-coordinate position of waypoints in the / base_footprint coordinate system; y' wp The meaning is the ordinate position of waypoints in the / base_footprint coordinate system; x base_footprint The meaning is the x-coordinate position of the origin of the / base_footprint coordinate system in the / map coordinate system; y base_footprint The meaning is the ordinate position of the origin of the / base_footprint coordinate system in the / map coordinate system.

[0148] The formula for coordinate system transformation in Landmarks is as follows:

[0149]

[0150] in, Indicates the angle between the map coordinate system and the robot coordinate system; x' lm This represents the x-coordinate of the landmark in the robot's coordinate system; y' lm Indicates the vertical coordinate of the landmark in the robot's coordinate system; x base_footprint This indicates the x-coordinate of the robot's coordinate system origin in the map coordinate system; y base_footprintThis indicates the ordinate position of the robot's coordinate system origin in the map coordinate system.

[0151] Phase 1: Intelligent Decision-Making Module

[0152] The core task of the intelligent decision-making module is to use deep reinforcement learning algorithms to learn a time-optimal decision-making strategy to avoid densely populated areas and dangerous zones. The overall process of this module is as follows: Figure 6 As shown, the observation information, consisting of the positions of the nearest feature points (Waypoints and Landmarks) and LiDAR data, is first input into the decision policy of the deep learning neural network, which outputs a decision point (local target point). Within a fixed decision period, the TEB planner calculates the motion speed to control the robot's movement. The decision policy gathers experience in a simulation environment and is gradually updated using deep reinforcement learning to finally obtain the optimal decision policy. The following sections detail the problem description, observation space, action space, reward function, and network structure construction for this stage.

[0153] Problem Description: Deep reinforcement learning models problems as Markov Decision Processes (MDPs) when the environment is fully perceived. However, since the unpredictable speed of pedestrians in dynamic scenes is unknown to mobile robots, path planning tasks involving pedestrians are often modeled as Partially Observable Markov Decision Processes (POMDPs). A Partially Observable Markov Process typically consists of a six-tuple (S, A, O, P, R, γ), where S is the state space, A is the action space, O is the observation space, P is the state transition probability, R is the reward function, and γ is the discount factor.

[0154] The goal of the decision-making module is to learn a decision-making strategy that minimizes the time it takes for the mobile robot to reach its destination, while satisfying constraints related to time, safety, and the destination. These constraints are summarized below:

[0155]

[0156] In the time constraint, t represents time, s represents the robot's state in the environment, and π represents the robot's strategy. The time constraint aims to get the robot to the destination in the shortest time under the guidance of the decision-making strategy; in the safety constraint... Let r be the distance information of the environment measured by the lidar at time t. robot The radius of the mobile robot is a safety constraint; the minimum distance measured by the lidar must be greater than the robot's radius to ensure the robot does not collide with obstacles. The endpoint constraint... Let d be the position of the moving robot at time t and the position of the destination, respectively. goal The distance range to the destination is considered as the distance within which the mobile robot has reached the destination.

[0157] Observation space: The observation space consists of LiDAR data and the nearest n coordinates in the robot coordinate system. wp Waypoints and n lm It consists of 10 Landmarks, i.e. Where, n wp The meaning is the number of waypoints observed by the mobile robot, n lm The meaning is the number of landmarks observed by the mobile robot. The meaning is the sequence data of the lidar at time t, R 8 The meaning is an 8-dimensional vector space, R 2 The meaning is a 2-dimensional vector space, R 360 The meaning is a 360-dimensional vector space.

[0158] Based on the configuration of the lidar in the simulation environment, it can be known that Based on the global feature point extraction method, it can be known that... It is a set consisting of the x and y coordinates of n points respectively ([x i ,y i [,i=0,1,…,n), in this embodiment, n is taken as wp =4,n lm =1, that is Therefore, the observation space O of the decision module t ∈R 370 .

[0159] Action space: The set of selectable local target points at each decision time (a t =[dist t ,angle t Drawing inspiration from the range limitations of automotive steering angles, the action space is divided into a forward zone and an adjustment zone. The forward zone is defined as dist. forward =2m, angle forward ∈[-π / 6,π / 6], the adjustment region is defined as ist adjustment =1m, angle adjustment ∈[-π / 2,-v / 6]∪[π / 6,π / 2]. Because an excessively large action space slows down the convergence speed of policy learning, and useless actions are not selected by the policy in the final learning process, the action space is discretized to ensure that action selection is as close as possible to the boundary of the action space. For example... Figure 7 As shown, the angle resolution in the motion space of the forward region is calculated as follows:

[0160]

[0161] in, The meaning is the resolution of the action space. The value represents the ratio of a pedestrian's radius to the distance in the movement space. To ensure that the number of movement spaces within the forward area is rounded up, a value was chosen.

[0162] The adjustment area helps the mobile robot escape from a predicament when there is no effective space to move forward. A schematic diagram of the intelligent decision-making module's motion space is shown below. Figure 8 As shown, when obstacles block the movement area, the robot is trapped in a corner, or a pedestrian enters a narrow passage and blocks the movement area, the robot has no choice of movement space. In this case, the robot's position can be adjusted by adjusting the area, so that the robot can get out of the trapped area.

[0163] Reward Function: The goal of the decision-making process is to minimize the time to reach the destination while ensuring safety and collision-free operation, enabling the mobile robot to respond quickly to dynamic scenarios while maintaining decision stability. Therefore, the reward function of this process is... The settings are as follows:

[0164]

[0165] The time reward is defined as follows:

[0166]

[0167] Where, ω t Let Δt be the coefficient for the time reward, and Δt be the decision time represented by each step update in the reinforcement learning decision. This is a small penalty for the mobile robot at each decision, with the aim of reducing the decision-making time throughout the decision-making and planning task so that it can reach the destination faster.

[0168] The finish line reward is defined as follows:

[0169]

[0170] Where, r g This represents the reward the mobile robot receives when it reaches the finish line. and Let r represent the positions of the moving robot and the destination at time t, respectively. goal "Otherwise" indicates the range of the endpoint. When the mobile robot moves into this range, it is considered to have reached the endpoint. "Otherwise" indicates other situations. A larger reward is given when the robot reaches the finish line, which enhances the importance of the finish line information and the purposefulness of the decision, ensuring that the robot can reach the finish line.

[0171] The collision reward is defined as follows:

[0172]

[0173] Where, r collision It is the reward that the mobile robot receives when it collides with an obstacle. It is the set of lidar detection distance values ​​at time t, r robot It is the radius of the mobile robot. A larger penalty is applied when the robot collides with an obstacle to encourage the robot to avoid obstacles and ensure safe movement.

[0174] The global feature point landmark reward is defined as follows:

[0175]

[0176] Where, ω l It is the coefficient of the landmark reward. and These represent the positions of the mobile robot at time t-1 and time t, respectively. and These represent the locations of the nearest landmarks to the mobile robot at time t-1 and time t, respectively. Encourage mobile robots to approach Landmarks, enabling them to reach their destination in long-distance tasks by gradually moving closer to the nearest global path feature points (Landmarks).

[0177] The reward for changes in the angle of a local target point is defined as follows:

[0178]

[0179] Where, ω a It is the coefficient for the reward of changes in the angle of the local target point. and These represent the angles between the local target point at time t-1 and time t relative to the robot coordinate system. The penalty is imposed when the angle of the local target point changes between the previous and current time steps. The greater the angle change, the greater the penalty and movement. This reward aims to ensure that the mobile robot selects local target points with continuity and stability.

[0180] Network structure: The network structure in the intelligent decision-making stage is as follows Figure 9 As shown, where Representing LiDAR data, Waypoints, and Landmarks observation data respectively, conv1d represents one-dimensional convolution, FC represents fully connected layer, ReLU represents non-linear activation function, CONCAT represents feature fusion, and A t Let V represent the action space, V represent the value of the estimated value function, and softmax represent the activation function that maps the distribution to probabilities. This embodiment uses three layers of one-dimensional convolutions to process the sequence data collected by the LiDAR. The number, size, and sampling interval of the convolutional kernels change sequentially from (32,7,4) to (64,5,2) to (64,3,2), and then a fully connected network is used to reduce the number of parameters. Simultaneously, Waypoints and Landmarks are processed through the fully connected network and combined with the extracted convolutional LiDAR features to form the observed environmental features. Because an actor-critic structure reinforcement learning algorithm is used, actor networks and critic networks need to be constructed separately, sharing the network structure before the final output layer. This shared structure improves the algorithm's learning ability in complex scenes. The final output layer, the actor network, outputs the probability distribution of the entire action space and samples actions based on a random strategy, while the critic network outputs the estimated value of the value function.

[0181] Phase 2: Intelligent Planning Module

[0182] The core task of the planning module is to use deep reinforcement learning algorithms to learn a local planning strategy with the local target point output by the intelligent decision module as the planning target. This strategy can comply with the social nature of pedestrian movement, has higher time efficiency, and can actively avoid obstacles.

[0183] The overall process of the intelligent planning module is as follows: Figure 10 As shown, the planning module first outputs an optimal local target point based on the intelligent decision-making module's decision strategy at each decision time, according to its observation space. Then, within each control cycle, this optimal local target point, the endpoint position, and LiDAR data are combined to form observation data, which is input into the planning strategy of the deep learning neural network, and the resulting velocity and angular velocity of the mobile robot are output to control the robot's movement. Compared to the decision-making module, the planning module's planning strategy replaces the traditional optimization-based TEB planning algorithm in the decision-making module, controlling the mobile robot's movement within multiple control cycles in each decision cycle. The planning strategy collects experience in the simulation environment and uses deep reinforcement learning to gradually update the strategy, ultimately obtaining a strategy that maximizes the reward. The following sections detail the problem description, observation space, action space, reward function, and network structure construction for this stage.

[0184] Problem Description: The goal of the planning module is to learn a safe and time-efficient planning strategy based on local target points of the intelligent decision-making module, while satisfying constraints of time, safety, and endpoint. This invention also models the intelligent planning module as a partially observable Markov decision process (POMDP).

[0185] The observation space consists of LiDAR data, the optimal local target point output by the decision module in the robot coordinate system, and the endpoint. Based on the configuration of the lidar, it can be known that It is the set ([x,y]) consisting of the x and y coordinates of the optimal local target point and the endpoint, i.e. Therefore, the observation space O of the decision module t ∈R 364 .in, The meaning is the data of the lidar at time t, R 2 The meaning is a 2-dimensional vector space, R 364 The meaning is a 364-dimensional vector space.

[0186] Action space: a set of optional control commands (a t =[v t ,w t ],v t ∈(0,v max ),w t ∈(-w max ,w max ), where a t The meaning is the action at time t, v t The meaning is the linear velocity at time t, w t The meaning is the angular velocity at time t, v max The meaning is the maximum linear velocity of the mobile robot, w max The meaning is the maximum angular velocity of the mobile robot.

[0187] Reward Function: The goal of the planning process is to respect the social nature of pedestrians and ensure the robot's safety and stability as much as possible during the planning process. Therefore, the reward function of this process is... The settings are as follows:

[0188]

[0189] The reward for a local target point is defined as follows:

[0190]

[0191] Where, ω sg The coefficient for the reward of the local target point. and These represent the positions of the mobile robot at time t-1 and time t, respectively. and These represent the positions of the local target point (subgoal) at time t-1 and time t, respectively. This is the reward a mobile robot receives when it gets closer to a local target point at the current moment compared to the previous moment. This reward encourages the mobile robot to move in the direction of getting closer to the local target point.

[0192] The collision reward is defined as follows:

[0193]

[0194] Where, r collision The reward a mobile robot receives when it collides with an obstacle. Let r represent the set of detection range values ​​of the lidar at time t. robot "Otherwise" indicates the radius of the mobile robot. A larger penalty is applied when a mobile robot collides with an obstacle to encourage the robot to avoid obstacles and ensure safe movement.

[0195] The reward for static security zones is defined as follows:

[0196]

[0197] Where, ω sz The coefficient for the static security zone reward. Let r represent the set of detection range values ​​of the pedestrian layer lidar at time t. sz This indicates the range of the static security zone. Mobile robots are encouraged to proactively respect the social attributes of pedestrians and plan "pedestrian-friendly" paths.

[0198] The speed reward is defined as follows:

[0199]

[0200] Where, ω vel w represents the coefficient for speed bonus. t Let t represent the robot's angular velocity at time t. To penalize excessive rotation caused by excessive angular velocity in mobile robots, this design encourages that the control angular velocity of mobile robots should not be too high, thereby ensuring that mobile robots operate more smoothly.

[0201] Network Construction: The network structure in the intelligent planning phase, such as... Figure 11As shown, where These represent observation data from LiDAR, local target points, and endpoints, respectively. `conv1d` represents a one-dimensional convolution, `FC` represents a fully connected layer, `relu` represents a non-linear activation function, and `CONCAT` represents the fusion of various feature connections. t Let V represent the action space, V represent the value of the estimated value function, and softmax represent the activation function that maps the distribution to probabilities. The LiDAR data is processed using three layers of one-dimensional convolutions to handle the sequence data collected by the LiDAR. The number, size, and sampling interval of the convolutional kernels change sequentially from (32,7,4) to (64,5,2) to (64,3,2), followed by a fully connected layer to reduce the number of parameters. Simultaneously, the optimal local target point and endpoint, along with the extracted convolutional LiDAR features, constitute the observed environmental features. The actor network and the critic network share the network structure before the final output layer. The final layer of the actor network outputs the probability of each action in the action space of the decision module and samples a specific control command. The final layer of the critic network outputs the estimated value of the value function.

[0202] V. Comparative Experiment of Two-Stage Deep Reinforcement Learning Path Planning Algorithms

[0203] This paper compares the proposed two-stage deep reinforcement learning algorithm with traditional path planning algorithms and other two-stage deep reinforcement learning algorithms. The DWA algorithm combines rule-based decision-making with DWA planning, the TEB algorithm combines rule-based decision-making with TEB planning, and Drl_Ros_Baseline and Arena-Rosnav combine rule-based decision-making with deep reinforcement learning planning. To demonstrate the average performance of each algorithm in long-distance dynamic indoor scenarios, the starting and ending points of the evaluation tasks are selected as random locations within the feasible space of the mobile robot on a static map. Evaluation tasks with a global path length greater than 15m are considered valid evaluation tasks. All evaluation experiments in this section select 200 valid evaluation tasks. The movement time of the mobile robot in the evaluation tasks is limited to within 100s. Valid evaluation tasks exceeding 100s without collisions are considered "timeout" and fail. The experimental results (performance of each algorithm) are shown in Table 1. Regarding the success rate of planning, the traditional path planning methods DWA and TEB achieved success rates of 75.5% and 85.5% respectively when there were 4 pedestrians, maintaining a high level. However, when the number of pedestrians increased to 8 and 12, the success rates of DWA and TEB decreased significantly due to the increased dynamic obstacles in the environment leading to denser conditions, reaching only 36.5% and 45% with 12 pedestrians. The deep reinforcement learning-based approach significantly outperformed the traditional algorithms in dense environments. With 12 pedestrians, all three learning methods maintained a success rate above 70%. This is because deep reinforcement learning does not require modeling and can directly interact with the environment to learn, better handling the unpredictable movement speeds of pedestrians. Furthermore, due to the differences in the decision-making modules among the three different deep reinforcement learning methods, it can be seen that TSDRL, which employs intelligent decision-making, outperforms Drl_Ros_Baseline (downsampling decision-making) by 7% and Arena-Rosnav (spatial distance decision-making) by 3.5% when there are 4 pedestrians. When there are 8 pedestrians, the results are 11.5% and 5.5% higher, respectively, and when there are 12 pedestrians, the results are 11% and 6% higher, respectively. Therefore, TSDRL outperforms other algorithms in terms of success rate, especially in densely populated crowds.

[0204] Table 1

[0205]

[0206] In terms of average running speed, deep reinforcement learning-based methods all outperformed traditional planning methods, with TSDRL achieving the highest speed among deep reinforcement learning-based methods. Regarding path ratios, TEB and Arena-Rosnav both employ spatial distance sampling decisions to keep the robot running within a fixed global path distance, resulting in shorter paths. DWA and Drl_Ros_Baseline have slightly longer paths, while TSDRL has the longest path.

[0207] In summary, the method proposed in this invention can improve the safety and time efficiency of mobile robots in long-distance, densely populated environments.

[0208] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, several equivalent substitutions or obvious modifications can be made without departing from the concept of the present invention, and all such modifications, achieving the same performance or purpose, should be considered within the scope of protection of the present invention.

Claims

1. A method for mobile robot path planning in long-distance, densely populated crowd scenarios, characterized in that, Includes the following steps: S1. Obtain the starting and ending points of the mobile robot on the map, and use the global planner to generate a global path; S2. Extract waypoints and landmarks on the global path and perform coordinate system transformation to obtain waypoints and landmarks in the robot coordinate system; S3. Input the waypoints and landmarks in the robot coordinate system and the point cloud data collected by the mobile robot into the decision-making strategy of the deep learning neural network, and learn the decision-making strategy to obtain the optimal decision model. S4. Input the waypoints and landmarks in the robot coordinate system and the point cloud data collected by the mobile robot into the optimal decision model to obtain the optimal local target point output at a fixed decision time. Perform planning strategy learning to obtain the optimal planning model, so as to obtain the path planning of the mobile robot; wherein, in step S3, the constraints of the decision strategy are as follows: in, Indicates time, s This indicates the state of the mobile robot in the environment. Indicates the strategy of the mobile robot; express t The robot constantly moves to measure distance information from the environment. Indicates the radius of the mobile robot; , They represent t The position of the robot and the destination are constantly being moved. This indicates the distance range to the destination, within which the mobile robot is considered to have reached the destination; Reward function in the decision-making process The settings are as follows: The time reward is defined as follows: in, The coefficient for time reward. To save time for each step of the reinforcement learning decision update. This is a small penalty imposed on the mobile robot for each decision, with the aim of reducing the decision-making time of the mobile robot throughout the decision-making and planning task, thereby allowing it to reach the destination faster. The finish line reward is defined as follows: in, This represents the reward the mobile robot receives when it reaches the finish line. They represent t The position of the robot and the destination are constantly being moved. "Otherwise" indicates the range of the endpoint; when the mobile robot moves into this range, it is considered to have reached the endpoint. The collision reward is defined as follows: in, It is the reward that the mobile robot receives when it collides with an obstacle. yes t The set of lidar detection distance values ​​at any given time. It is the radius of the mobile robot; The global feature point landmark reward is defined as follows: in, It is the coefficient of the landmark reward. They represent t-1 Time and t The robot's position is constantly being moved. They represent t-1 Time and t The location of the nearest landmark to the mobile robot at any given time; The reward for changes in the angle of a local target point is defined as follows: in, It is a change in the angle of a local target point. They represent t-1 Time and t The angle between the local target point and the robot coordinate system at any given time.

2. The mobile robot path planning method in long-distance, densely populated crowd scenarios according to claim 1, characterized in that, In step S1, the global planner uses the Dijkstra algorithm.

3. The mobile robot path planning method in long-distance, densely populated crowd scenarios according to claim 1, characterized in that, Step S2 specifically includes: extracting waypoints on the global path using a fixed-distance sampling method, and extracting landmarks using a global path curvature variation method.

4. The mobile robot path planning method in long-distance, densely populated crowd scenarios according to claim 1 or 3, characterized in that, In step S2, the coordinate system transformation includes: performing coordinate system transformation on waypoints and landmarks, and rotating and translating the waypoints to their coordinate positions in the map coordinate system. and the coordinates of landmarks in the map coordinate system The formulas for transforming the coordinate systems of waypoints and landmarks from the map coordinate system to the robot coordinate system are as follows: Indicates the angle between the map coordinate system and the robot coordinate system; express The vertical coordinate position in the robot coordinate system This indicates the vertical coordinate position of the landmark in the robot's coordinate system; This indicates the x-coordinate position of the robot's coordinate system origin in the map coordinate system. This indicates the ordinate position of the robot's coordinate system origin in the map coordinate system.

5. The mobile robot path planning method in long-distance, densely populated crowd scenarios according to claim 1, characterized in that, In step S3, the point cloud data is collected by the lidar of the mobile robot.

6. The mobile robot path planning method in long-distance, densely populated crowd scenarios according to claim 1, characterized in that, In step S3, the process of learning the decision strategy to obtain the optimal decision model specifically includes: selecting a local target point and using a local planner to plan and calculate the speed and angular velocity of the mobile robot; reselecting a new local target point and planning at fixed decision intervals; training a deep learning neural network to update the decision strategy and obtain the optimal decision model.

7. The mobile robot path planning method in long-distance, densely populated crowd scenarios according to claim 1, characterized in that, In step S4, the process of learning the planning strategy to obtain the optimal planning model specifically includes: inputting the optimal local target point, endpoint position, and point cloud data into the planning strategy of a deep learning neural network for planning, obtaining the speed and angular velocity of the mobile robot, training the deep learning neural network to update the planning strategy, and obtaining the optimal planning model; and the reward function during the planning process. The settings are as follows: The reward for a local target point is defined as follows: in, The coefficient for the reward of the local target point. They represent t-1 Time and t The robot's position is constantly being moved. They represent t-1 Time and t Local target point at time Location; The collision reward is defined as follows: in, The reward a mobile robot receives when it collides with an obstacle. express t Real-time LiDAR The set, Indicates the radius of the mobile robot; Otherwise indicates other cases; The reward for static security zones is defined as follows: in, The coefficient for the static security zone reward. express t Real-time pedestrian layer lidar The set, express Scope; The speed reward is defined as follows: in, The coefficient representing the speed bonus. express t The angular velocity of the robot at any given moment.

8. The mobile robot path planning method in long-distance dense crowd scenarios according to claim 6 or 7, characterized in that, The deep learning neural network used is CNN, LSTM, or RNN.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 8.

Images