Robot navigation method and system in dynamic environment, robot and storage medium
By generating and filtering local candidate paths in a dynamic environment and using a probabilistic ensemble neural network model to assess risks, the robot is controlled to maintain a safe distance in the dynamic environment. This solves the navigation failure problem of supervised learning methods in unseen scenarios and achieves effective and safe navigation.
Patent Information
- Application Number
- CN202511399026.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-28
- Publication Date
- 2026-01-23
AI Technical Summary
Existing technologies are prone to navigation failure when robots navigate in dynamic environments and encounter unfamiliar scenarios. The diversity and coverage dependence of supervised learning methods lead to policy failure.
By controlling the robot to walk along a global reference path, continuously monitoring whether there are pedestrians within the perception range, generating multiple local candidate paths, and filtering out the target local path, controlling the robot to move along the target local path and maintain a preset distance, updating the global reference path when the destination is reached, and using a probabilistic ensemble neural network model for risk assessment and path selection.
Effectively address dynamic scenarios, ensure the effectiveness and safety of navigation, maintain a preset distance from pedestrians, adapt to complex environments, and enhance the robustness and social adaptability of navigation.
Smart Images
Figure CN121384013A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of robot navigation, and particularly relates to a robot navigation method and system in a dynamic environment, a robot and a storage medium. BACKGROUND
[0002] In the process of performing a task, a robot needs to effectively avoid obstacles. With the development of deep learning and reinforcement learning, learning methods are gradually introduced into robot group navigation. Supervised learning methods usually use expert demonstration data (such as rule-based planners, human teleoperation data or real human trajectories) to train navigation strategies, so that the robot can imitate expert behavior to complete obstacle avoidance and navigation in known scenes. However, the supervised learning method is highly dependent on the diversity and coverage of the training data. When facing dynamic scenes that have not been seen during training, the strategy learned by imitation learning is likely to fail, resulting in navigation failure.
[0003] Therefore, the prior art has defects and needs to be improved and developed. SUMMARY
[0004] The technical problem to be solved by the present application is to provide a robot navigation method and system in a dynamic environment, aiming to solve the problem of navigation failure in the prior art when facing dynamic scenes that have not been seen during training.
[0005] The technical solution adopted by the present application to solve the technical problem is as follows: In a first aspect, the present application provides a robot navigation method in a dynamic environment, comprising: controlling the robot to walk along a global reference path generated by a path planning algorithm, and continuously monitoring whether there are pedestrians in the perception range of the robot; if there are pedestrians in the perception range of the robot, generating a plurality of local candidate paths, and selecting a target local path from all the local candidate paths; controlling the robot to move along the target local path, and constraining the robot to maintain a preset distance from the pedestrians during movement; when the robot reaches the end point of the target local path, updating the global reference path by using the path planning algorithm, and controlling the robot to move along the updated global reference path.
[0006] In one embodiment, if there are pedestrians in the perception range of the robot, a plurality of local candidate paths are generated, comprising: if there are pedestrians in the perception range of the robot, a circle is generated with the current position of the robot as the center and a preset distance as the radius; uniformly sampling a plurality of candidate target points on the circle; Generate, in sequence, paths from the current position of the robot as the starting point and each candidate target point as the ending point as local candidate paths.
[0007] In an embodiment, screening the target local path from all the local candidate paths comprises: Inputting all the local candidate paths into the trained probabilistic ensemble neural network model for risk prediction to obtain the risk probability distribution of each local candidate path output by each sub-model in the probabilistic ensemble neural network model; Screening based on the risk probability distribution of all the local candidate paths output by all the sub-models to obtain the target local path.
[0008] In an embodiment, the training process of the probabilistic ensemble neural network model comprises: Constructing a training scene containing pedestrians and road segment features of different densities, and generating a training circle with a preset radius centered on the current training position of the robot, and uniformly sampling a preset number of candidate training target points on the training circle; Generating, in sequence, paths from the current training position of the robot as the starting point and each candidate training target point as the ending point as local training candidate paths; For each candidate training path, evaluating according to a preset risk indicator to obtain the risk label of each candidate training path; Initializing an initial probabilistic ensemble neural network model to be trained; Inputting the candidate training paths with risk probability distribution into the initial probabilistic ensemble neural network model for training until the end condition is reached to obtain the trained probabilistic ensemble neural network model.
[0009] In an embodiment, screening based on the risk probability distribution of all the local candidate paths output by all the sub-models to obtain the target local path comprises: Sequentially performing cognitive uncertainty filtering and data uncertainty filtering on the risk probability distribution of all the local candidate paths output by all the sub-models; Among the filtered local candidate paths, the local candidate path with the lowest mean of risk probability distribution is taken as the target local path.
[0010] In an embodiment, sequentially performing cognitive uncertainty filtering and data uncertainty filtering on the risk probability distribution of all the local candidate paths output by all the sub-models comprises: Calculating the divergence between the risk probability distributions output by any two sub-models for the same local candidate path; If any of the local candidate paths has a divergence less than a preset threshold, the risk probability distribution of the local candidate path is filtered out; If the divergence is greater than a preset cognitive uncertainty threshold, the local candidate path is eliminated; For the local candidate path filtered by the cognitive uncertainty, the risk probability distribution of all sub-models of the local candidate path is aggregated, and a conditional risk value of a preset confidence level is calculated; If the conditional risk value is greater than a preset risk tolerance threshold, the local candidate path is eliminated.
[0011] In an embodiment, after the global reference path is updated by the path planning algorithm and the robot is controlled to move along the updated global reference path, the method further comprises: The process of monitoring whether there is a pedestrian within the perception range of the robot, screening a target local path, controlling the robot to move along the target local path, updating the global reference path, and controlling the robot to move along the updated global reference path is continuously performed until the end point is reached.
[0012] In a second aspect, an embodiment of the present application further provides a robot navigation system in a dynamic environment, which comprises: A first control module configured to control the robot to move along a global reference path generated by a path planning algorithm, and continuously monitor whether there is a pedestrian within the perception range of the robot; A target local path generation module configured to generate a plurality of local candidate paths if there is a pedestrian within the perception range of the robot, and screen a target local path from all the local candidate paths; A local obstacle avoidance module configured to control the robot to move along the target local path, and constrain the robot to maintain a preset distance from the pedestrian during the movement; A second control module configured to update the global reference path by using the path planning algorithm when the robot reaches the end point of the target local path, and control the robot to move along the updated global reference path.
[0013] In a third aspect, an embodiment of the present application further provides a robot, which comprises a memory, a processor, and a robot navigation system in a dynamic environment stored on the memory and executable on the processor, wherein the robot navigation system in a dynamic environment, when executed by the processor, implements the steps of the robot navigation method in a dynamic environment as described above.
[0014] In a fourth aspect, an embodiment of the present application further provides a computer readable storage medium, which stores a robot navigation system in a dynamic environment, wherein the robot navigation system in a dynamic environment can be executed to implement the steps of the robot navigation method in a dynamic environment as described above.
[0015] The robot is controlled to walk along the global reference path, and whether there is a pedestrian in the perception range of the robot is continuously monitored; if there is a pedestrian in the perception range of the robot, a plurality of local candidate paths are generated, and a target local path is selected from all the local candidate paths; the robot is controlled to move along the target local path, and the robot is constrained to maintain a preset distance from the pedestrian during the movement; when the robot reaches the end point of the target local path, the global reference path is updated, and the robot is controlled to move along the updated global reference path. The robot generates a target local path to move after encountering a pedestrian in the perception range, and the preset distance from the pedestrian is constrained during the movement, which can effectively deal with dynamic scenes and ensure the effectiveness of navigation. BRIEF DESCRIPTION OF DRAWINGS
[0016] Figure 1 is a flowchart of a preferred embodiment of the robot navigation method in a dynamic environment in the present application.
[0017] Figure 2 is a schematic diagram of a target local path in the present application.
[0018] Figure 3 is a structural schematic diagram of a preferred embodiment of the robot navigation system in a dynamic environment in the present application.
[0019] Figure 4 is a principle block diagram of the robot in the present application. DETAILED DESCRIPTION
[0020] In order to make the purpose, technical scheme and advantages of the present application more clear and definite, the present application will be further described in detail below with reference to the drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.
[0021] In the process of performing tasks, the robot needs to effectively avoid obstacles. With the development of deep learning and reinforcement learning, learning methods are gradually introduced into robot group navigation. Supervised learning methods usually use expert demonstration data (such as rule-based planners, artificial teleoperation data or real human trajectories) to train navigation strategies, so that the robot can imitate expert behavior to complete obstacle avoidance and navigation in known scenes. However, the supervised learning method is highly dependent on the diversity and coverage of the training data. When facing dynamic scenes that have not been seen during training, the strategy learned by imitation learning is likely to fail, resulting in navigation failure.
[0022] In view of the above defects of the prior art, the present application provides a robot navigation method, system, robot and medium in a dynamic environment, the method comprising: controlling the robot to walk along a global reference path, and continuously monitoring whether there are pedestrians in the perception range of the robot; if there are pedestrians in the perception range of the robot, generating a plurality of local candidate paths, and screening a target local path from all the local candidate paths; controlling the robot to move along the target local path, and constraining the robot to maintain a preset distance from the pedestrians during the movement; and when the robot reaches the end point of the target local path, updating the global reference path, and controlling the robot to move along the updated global reference path. The present application can effectively deal with dynamic scenes and ensure the effectiveness of navigation by generating a target local path for movement after encountering pedestrians in the perception range, and constraining the preset distance from the pedestrians during the movement.
[0023] Referring to Figure 1 The robot navigation method in a dynamic environment according to the embodiments of the present application comprises the following steps: Step S100, controlling the robot to walk along a global reference path generated by a path planning algorithm, and continuously monitoring whether there are pedestrians in the perception range of the robot.
[0024] Specifically, the path planning algorithm can be any one of Multi-RRT algorithm, A* algorithm, and D* Lite algorithm. Whether there are pedestrians in the perception range is monitored in real time by using sensors equipped on the robot, including but not limited to laser radar, depth camera, millimeter wave radar, etc.
[0025] In one implementation, the global reference path is generated by using Multi-RRT algorithm.
[0026] Specifically, compared with the traditional single RRT algorithm, the path can be quickly generated by using Multi-RRT algorithm, effectively improving the efficiency of global planning.
[0027] Referring to Figure 1 The robot navigation method in a dynamic environment according to the embodiments of the present application further comprises the following steps: Step S200, if there are pedestrians in the perception range of the robot, generating a plurality of local candidate paths, and screening a target local path from all the local candidate paths.
[0028] Specifically, if there are pedestrians in the perception range of the robot, a circle is generated with the current position of the robot as the center and a preset distance as the radius; a plurality of candidate target points are uniformly sampled on the circle; and a path is generated in turn with the current position of the robot as the starting point and each candidate target point as the end point as a local candidate path. The range of the preset number is 6-12. For example, if the preset number is 8, eight candidate target points are uniformly sampled on the circle, and eight local candidate paths are generated in turn. Figure 2As shown, if the preset number is 12, 12 candidate target points are uniformly sampled on the circumference, and then 12 local candidate paths are generated. Subsequently, a target local path is selected from the 12 local candidate paths, and the robot is controlled to move along the target local path. In this way, the pedestrian can be effectively avoided.
[0029] In an implementation manner, the target local path is selected from all the local candidate paths, comprising: all the local candidate paths are input into the trained probabilistic ensemble neural network model for risk prediction, to obtain a risk probability distribution of each local candidate path output by each sub-model in the probabilistic ensemble neural network model; all the local candidate paths output by all the sub-models are filtered based on the risk probability distribution, to obtain the target local path.
[0030] Specifically, the probabilistic ensemble neural network (PENN) comprises a plurality of sub-models with the same structure and different parameters. When all the local candidate paths are input into the probabilistic ensemble neural network model, a risk probability distribution of the local candidate path output by each sub-model can be obtained. Then, the risk probability of all the local candidate paths output by all the sub-models is filtered in terms of cognitive uncertainty and data uncertainty; among the filtered local candidate paths, the local candidate path with the lowest mean value of the risk probability distribution is selected as the target local path. Through the display of the risk quantification, the target local path can be generated by comparing different local candidate paths in real time during the movement of the robot, and then the robot moves according to the target local path. In this way, the dynamic environment can be flexibly coped with. In addition, through the multi-level fine evaluation of the local path, a more robust decision can be made.
[0031] In an implementation manner, the training process of the probabilistic ensemble neural network model comprises: a training scene comprising pedestrians and road features with different densities is constructed, and a training circumference is generated with a preset radius centering on the current training position of the robot, and a plurality of candidate training target points are uniformly sampled on the training circumference; paths from the current training position of the robot as a starting point to each candidate training target point as an ending point are sequentially generated as local training candidate paths; for each candidate training path, a risk label of each candidate training path is obtained by evaluating the candidate training path according to a preset risk index; an initial probabilistic ensemble neural network model to be trained is initialized; the candidate training paths with the risk labels are input into the initial probabilistic ensemble neural network model for training until an end condition is reached, to obtain the trained probabilistic ensemble neural network model.
[0032] Specifically, in the training process, the risk of the candidate training path is evaluated by a preset risk indicator. The risk indicator includes path distance, direction deviation, relative speed of the robot and the pedestrian, distance to the pedestrian, and direction of the pedestrian, etc. After evaluation, a candidate training path with risk annotation can be generated, and then the initial probability integrated neural network model is trained using the candidate training path with risk annotation to obtain a trained probability integrated neural network model. In the training stage, the features of human space science are introduced, including the distance to the pedestrian, the relative speed, the direction of the pedestrian and other parameters, so that the robot can actively avoid entering the personal space of the pedestrian in the subsequent path selection process, effectively maintaining the preset distance, thereby realizing the navigation behavior more in line with social norms. The present application proposes a path risk modeling method for generating risk annotation data through heuristic model predictive control (HR-MPC) and training using a probability integrated neural network, realizing explicit risk prediction of the candidate path and ensuring safety and sociality in the navigation process. By replacing the artificial heuristic rule with a data-driven method, the risk prediction can have stronger generalization ability.
[0033] In an implementation manner, the risk probability distribution of all the local candidate paths output by all the sub-models is sequentially subjected to cognitive uncertainty filtering and data uncertainty filtering, including: calculating the divergence between the risk probability distributions output by any two sub-models for the same local candidate path; if any divergence of any local candidate path is greater than a preset cognitive uncertainty threshold, the local candidate path is removed; for the local candidate path subjected to cognitive uncertainty filtering, aggregating the risk probability distributions of all the sub-models of the local candidate path to calculate the conditional value at risk of the preset confidence level; if the conditional value at risk is greater than a preset risk tolerance threshold, the local candidate path is removed.
[0034] Specifically, the prediction results are screened by divergence and conditional value at risk (CVaR) in sequence, effectively removing low-confidence or high-risk paths, which can improve the robustness and reliability of the decision.
[0035] In addition to the above manner, the application also supports filtering cognitive uncertainty, data uncertainty and environment perception uncertainty in sequence on the risk probability distribution of all the local candidate paths output by all the sub-models; and taking the local candidate path with the lowest mean of the risk probability distribution as the target local path in the filtered local candidate paths. Specifically, the cognitive uncertainty filtering and the data uncertainty filtering of this process are the same as the above-mentioned manner, and will not be described here. The process of the environment perception uncertainty filtering is as follows: for the local candidate paths filtered by the data uncertainty, the perception uncertainty score is calculated by using a preset perception uncertainty evaluation index. The perception uncertainty evaluation index can include sensor noise, perception delay time, etc. Then, the perception uncertainty score of each candidate local path is compared with a preset perception uncertainty threshold, and the local candidate path higher than the perception uncertainty threshold is removed. Then, the local candidate path with the lowest mean of the risk probability distribution is taken as the target local path in the filtered local candidate paths. In this way, the robustness of the system in the real complex environment can be effectively enhanced.
[0036] Referring to Figure 1 The robot navigation method in a dynamic environment according to the embodiment of the application further includes the following steps: Step S300, controlling the robot to move along the target local path and constraining the robot to keep a preset distance from the pedestrian during the movement.
[0037] Specifically, after the target local path is generated, it will be transmitted to the model predictive controller (MPC) of the robot. The model predictive controller will introduce a control barrier function (CBF) constraint to ensure that the robot keeps a preset distance from the pedestrian during the movement. This constraint converts the safety requirement of keeping a preset distance from the pedestrian into a specification that must be followed by the controller, so that the motion control amount (such as speed and angular velocity) of the robot is dynamically adjusted in each step of rolling optimization. By combining the optimization ability of the model predictive controller with the strict safety guarantee of the control barrier function, the application can output smooth and stable control instructions in a complex dynamic environment, effectively avoiding the robot from experiencing sharp acceleration or deceleration or emergency stop, and significantly improving the fluency, safety and user experience of the navigation process.
[0038] Referring to Figure 1 The robot navigation method in a dynamic environment according to the embodiment of the application further includes the following steps: Step S400, when the robot reaches the end point of the target local path, updating the global reference path by using a path planning algorithm, and controlling the robot to move along the updated global reference path.
[0039] Specifically, the end point of the target local path is the termination point of the local obstacle avoidance behavior, usually located behind or safely aside of the pedestrian motion trajectory, indicating that the robot has successfully avoided the current pedestrian interference. Upon reaching this point, the local obstacle avoidance task is completed, and the robot needs to return to its global navigation task. At this time, the robot current position is taken as the new starting point, and the original task target point is taken as the end point, and the path planning algorithm is called to update the global reference path, and then the robot is controlled to move along the updated global reference path. This step ensures that the local obstacle avoidance behavior will not cause the robot to deviate from the final target, and realizes the unification of short-term obstacle avoidance safety and long-term task efficiency.
[0040] In an implementation manner, after the global reference path is updated by using the path planning algorithm and the robot is controlled to move along the updated global reference path, the method further comprises: The process of monitoring whether there is a pedestrian in the perception range of the robot, screening a target local path, controlling the robot to move along the target local path, updating the global reference path and controlling the robot to move along the updated global reference path is continuously performed until the end point is reached.
[0041] Specifically, pedestrians in a dynamic environment often appear continuously and multiple times. A single local obstacle avoidance behavior is not enough to complete the entire navigation task. The method can continuously monitor and timely respond to the presence of pedestrians in the perception range, so that the robot has the ability to continuously respond to multiple and continuous dynamic interference. Whenever a new pedestrian enters the perception range, the system can quickly start a new round of local obstacle avoidance and global planning process, thereby maintaining high environmental adaptability and navigation robustness throughout the task period. Although local obstacle avoidance ensures short-term safety, it may cause the robot to deviate from the original optimal path. By updating the global reference path after each local obstacle avoidance, it is ensured that the motion direction of the robot always points to the final target point.
[0042] In complex environments such as shopping malls, hospitals, stations and airports, robots not only need to complete efficient path planning and obstacle avoidance tasks, but also must consider human social behavior norms, i.e. should maintain a predetermined distance from humans. Psychological research shows that humans have clear "space needs" in interaction, i.e. personal space, social space, etc. If the robot frequently enters the personal space of humans during navigation, it is easy to cause discomfort, anxiety and even aversion, thereby reducing the acceptability in human-robot coexistence environment. Therefore, it is an industry focus to realize robot navigation in dense crowd environment while considering safety, efficiency and social perception.
[0043] Currently, in addition to supervised learning methods, reinforcement learning methods can learn complex obstacle avoidance and cooperation behaviors in diverse training scenarios by interacting with simulated environments and constantly optimizing strategies. Although reinforcement learning methods can learn complex behaviors in simulated environments, they lack explicit risk assessment mechanisms and are insufficient in handling uncertainty, which may result in unsafe or unpredictable behaviors in real-world applications. At the same time, reinforcement learning has low sample efficiency and high training costs, making it difficult to be directly transferred to real-world environments. Furthermore, in the field of robot swarm navigation, existing methods can be divided into reactive methods and predictive methods. The typical representatives of reactive methods are Optimal Reciprocal Collision Avoidance (ORCA) and Social Force (SF) models. ORCA is based on the idea of velocity obstacle and assumes that robots and humans share the responsibility of obstacle avoidance. SF model simulates human group dynamics by constructing an artificial force field. These methods can quickly generate collision-free trajectories in low-density scenarios. In contrast, predictive methods such as Dynamic Window Approach (DWA), TimedElastic Band (TEB), and Multi-Risk-RRT method generate smooth and dynamically feasible paths by predicting environmental changes within a short time range and optimizing trajectories. Reactive methods rely heavily on strong assumptions (such as humans actively cooperating to avoid obstacles) and are prone to failure in crowded environments. They are also sensitive to parameter settings, often resulting in overly conservative paths or getting stuck in local difficulties. Moreover, they lack modeling of social distance, which can lead to unnatural behaviors such as invading human personal space. Although predictive methods can generate relatively smooth paths, they may not be reliable when faced with complex human behaviors, lack consideration of social factors, and may still collide or excessively avoid in dense dynamic crowds, resulting in insufficient robustness.
[0044] To improve the social adaptability of robots in group environments, some studies have introduced social perception models. Related methods are usually based on proxemics and add constraints of human personal space or group dynamics in planning. For example, some methods model the degree of invasion of human space by robots through inadmissibility functions or CCR maps; some studies use Gaussian functions or Gaussian mixture models to represent human social space and interaction areas; in addition, attention mechanisms and spatiotemporal interaction graphs are also used to capture potential social signals between humans and robots, so as to achieve navigation behaviors that conform to social norms. Most social perception methods rely on hand-designed spatial models or fixed heuristic rules, lack flexibility and adaptive ability, and have poor generalization performance in different cultural habits or different density of human group environments. In addition, such methods usually fail to explicitly model and handle uncertainty, which limits their effectiveness and promotion in practical applications.
[0045] The method proposed in the present application continuously monitors whether there are pedestrians in the perception range of the robot, determines a target local path in real time for pedestrian avoidance whenever pedestrians appear, and introduces distance constraints in the process of avoidance to ensure that the robot actively respects the personal and social space of humans when performing local obstacle avoidance. This approach effectively overcomes the problems of insufficient consideration of social factors and overly strong assumptions in reactive and predictive methods. In the process of determining the target local path, a double-layer evaluation framework of cognitive uncertainty and data uncertainty is constructed by fusing the probability ensemble neural network (PENN) and divergence, conditional value at risk (CVaR) filtering, which explicitly handles prediction risk and model disagreement, so as to achieve safe, reliable and socially acceptable navigation decisions in complex dynamic environments. This approach effectively overcomes the lack of explicit risk assessment mechanism and insufficient handling of uncertainty in reinforcement learning methods, and also overcomes the problem of lack of flexibility and adaptive ability in social perception methods that rely on hand-designed spatial models or fixed heuristic rules, and poor generalization performance in different cultural habits or different density of human group environments.
[0046] The scheme of the present application is not only applicable to robots (such as delivery robots, security patrol robots, medical auxiliary robots, etc.), but also applicable to autonomous driving technology.
[0047] In one embodiment, as shown in Figure 3 Based on the above robot navigation method in dynamic environments, the present application also correspondingly provides a robot navigation system in dynamic environments, which comprises: a first control module 100 for controlling the robot to walk along a global reference path generated by a path planning algorithm and continuously monitoring whether there are pedestrians in the perception range of the robot; The target local path generation module 200 is configured to generate a plurality of local candidate paths if there is a pedestrian within the perception range of the robot, and to select a target local path from all the local candidate paths. The local obstacle avoidance module 300 is configured to control the robot to move along the target local path and to constrain the robot to maintain a preset distance from the pedestrian during movement. The second control module 400 is configured to update the global reference path using a path planning algorithm when the robot reaches the end point of the target local path, and to control the robot to move along the updated global reference path.
[0048] In an embodiment, the system further comprises: The circumcircle generation unit is configured to generate a circumcircle with the current position of the robot as the center and a preset distance as the radius if there is a pedestrian within the perception range of the robot. The sampling unit is configured to uniformly sample a preset number of candidate target points on the circumcircle. The local candidate path generation unit is configured to generate, in sequence, paths with the current position of the robot as the starting point and each candidate target point as the end point as local candidate paths.
[0049] In an embodiment, the system further comprises: The risk probability distribution generation unit is configured to input all the local candidate paths into a trained probabilistic ensemble neural network model for risk prediction to obtain a risk probability distribution of each local candidate path output by each sub-model in the probabilistic ensemble neural network model. The screening unit is configured to screen based on the risk probability distributions of all the local candidate paths output by all the sub-models to obtain the target local path.
[0050] In an embodiment, the system further comprises: The construction unit is configured to construct a training scene containing pedestrians and road segment features of different densities, and to generate a training circumcircle with a preset radius centered on the current training position of the robot, and to uniformly sample a preset number of candidate training target points on the training circumcircle. The training data generation unit is configured to generate, in sequence, paths with the current training position of the robot as the starting point and each candidate training target point as the end point as local training candidate paths. The labeling unit is configured to evaluate each candidate training path according to a preset risk indicator to obtain a risk label for each candidate training path. The initialization unit is configured to initialize an initial probabilistic ensemble neural network model to be trained. The training unit is configured to input the candidate training path with the risk probability distribution into the initial probabilistic ensemble neural network model for training until a termination condition is reached, so as to obtain a trained probabilistic ensemble neural network model.
[0051] In an embodiment, the system further comprises: The filtering unit is configured to sequentially perform cognitive uncertainty filtering and data uncertainty filtering on the risk probability distribution of all the local candidate paths output by all the sub-models. The target local path generation unit is configured to select, from the filtered local candidate paths, a local candidate path with the lowest mean value of the risk probability distribution as the target local path.
[0052] In an embodiment, the system further comprises: The first calculation unit is configured to calculate the divergence between the risk probability distributions output by any two sub-models for the same local candidate path. The divergence; The first elimination unit is configured to eliminate a local candidate path if the divergence of the local candidate path is greater than a preset cognitive uncertainty threshold. The divergence; The second calculation unit is configured to calculate the conditional risk value at a preset confidence level of the risk probability distribution of all the sub-models of a local candidate path that has passed the cognitive uncertainty filtering. The second elimination unit is configured to eliminate the local candidate path if the conditional risk value is greater than a preset risk tolerance threshold.
[0053] In an embodiment, the system further comprises: The iteration unit is configured to continuously perform the process of monitoring whether there is a pedestrian within the perception range of the robot, screening out the target local path, controlling the robot to move along the target local path, updating the global reference path, and controlling the robot to move along the updated global reference path until the end point is reached.
[0054] Based on the above embodiments, the application further provides a robot, a structure diagram of which can be as shown in Figure 4The robot shown in the figure includes a processor, a memory, a network interface and a display screen connected through a system bus. The processor of the robot is configured to provide computing and control capabilities. The memory of the robot includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a robot navigation system in a dynamic environment. The internal memory provides an environment for the operating system and the robot navigation system in a dynamic environment in the non-volatile storage medium. The network interface of the robot is configured to communicate with an external robot through a network connection. The robot navigation system in a dynamic environment is executed by the processor to implement the steps of any of the robot navigation methods in a dynamic environment. The display screen of the robot can be a liquid crystal display screen or an electronic ink display screen.
[0055] Those skilled in the art can understand that, Figure 4 The structural diagram shown in the figure is only a schematic diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the robot to which the scheme of the present application is applied. The specific robot can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.
[0056] In one embodiment, a robot is provided. The robot includes a memory, a processor, and a robot navigation program in a dynamic environment stored in the memory and executable on the processor. The robot navigation program in a dynamic environment, when executed by the processor, implements the steps of any of the robot navigation methods in a dynamic environment provided by the embodiments of the present application.
[0057] The embodiments of the present application also provide a computer readable storage medium. The computer readable storage medium stores a robot navigation program in a dynamic environment. The robot navigation program in a dynamic environment, when executed by the processor, implements the steps of any of the robot navigation methods in a dynamic environment provided by the embodiments of the present application.
[0058] It should be understood that the sequence numbers of the steps in the above embodiments do not mean the order of execution. The execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.
[0059] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the above-mentioned division of each functional unit and module is exemplified, and in actual application, the above-mentioned functions can be completed by different functional units and modules according to needs, that is, the internal structure of the above-mentioned system is divided into different functional units or modules to complete all or part of the functions described above. Each functional unit and module in the embodiment can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be realized in the form of hardware or software. In addition, the specific names of each functional unit and module are only for easy distinction, and do not limit the protection scope of the present application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the foregoing method embodiments, which will not be described here.
[0060] In the above embodiments, the description of each embodiment has its own emphasis, and the parts not described or recorded in detail in a certain embodiment can be referred to the related description of other embodiments.
[0061] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. A person skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.
[0062] In the embodiments provided by the present application, it should be understood that the disclosed system / robot device and method can be implemented in other ways. For example, the system / robot device embodiments described above are only schematic, for example, the division of the above-mentioned modules or units is only a logical function division, and actual implementation can be in another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed.
[0063] The above-described embodiments are only used to illustrate the technical solutions of the present application, but not limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand; it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not deviate from the spirit and scope of the corresponding technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.
Claims
1. A robot navigation method in a dynamic environment, characterized in that, The method includes: The robot is controlled to walk along a global reference path generated by a path planning algorithm, and the presence of pedestrians within the robot's perception range is continuously monitored. If there are pedestrians within the robot's perception range, multiple local candidate paths are generated, and the target local path is selected from all the local candidate paths. Control the robot to move along the target local path, and constrain the robot to maintain a preset distance from pedestrians during the movement; Once the robot reaches the end of the target local path, the global reference path is updated using a path planning algorithm, and the robot is controlled to move along the updated global reference path.
2. The robot navigation method in a dynamic environment according to claim 1, characterized in that, If a pedestrian is present within the robot's perception range, multiple local candidate paths are generated, including: If there are pedestrians within the robot's perception range, a circle with the robot's current position as the center and a preset distance as the radius is generated; A predetermined number of candidate target points are uniformly sampled on the circumference; Local candidate paths are generated sequentially, starting from the robot's current position and ending at each candidate target point.
3. The robot navigation method in a dynamic environment according to claim 1, characterized in that, Selecting the target local path from all the local candidate paths includes: All the local candidate paths are input into the trained probabilistic ensemble neural network model for risk prediction, and the risk probability distribution of each local candidate path output by each sub-model in the probabilistic ensemble neural network model is obtained. The target local path is obtained by filtering based on the risk probability distribution of all local candidate paths output by all sub-models.
4. The robot navigation method in a dynamic environment according to claim 3, characterized in that, The training process of the probabilistic ensemble neural network model includes: Construct a training scene containing pedestrian and road segment features of different densities, and generate a training circle with a preset radius centered on the robot's current training position. On this training circle, uniformly sample a preset number of multiple candidate training target points. The robot generates local training candidate paths sequentially, starting from its current training position and ending at each candidate training target point. For each candidate training path, an evaluation is conducted according to a preset risk index to obtain a risk label for each candidate training path; Initialize the initial probabilistic ensemble neural network model to be trained; The candidate training paths with risk probability distributions are input into the initial probabilistic ensemble neural network model for training until the termination condition is met, resulting in the trained probabilistic ensemble neural network model.
5. The robot navigation method in a dynamic environment according to claim 3, characterized in that, The target local path is obtained by filtering based on the risk probability distribution of all local candidate paths output by all sub-models, including: For the risk probability distribution of all local candidate paths output by all sub-models, cognitive uncertainty filtering and data uncertainty filtering are performed sequentially. Among the filtered local candidate paths, the local candidate path with the lowest mean of risk probability distribution is selected as the target local path.
6. The robot navigation method in a dynamic environment according to claim 5, characterized in that, For the risk probability distribution of all local candidate paths output by all sub-models, cognitive uncertainty filtering and data uncertainty filtering are performed sequentially, including: Calculate the risk probability distributions output by any two sub-models for the same local candidate path. Divergence; If any local candidate path is any If the divergence is greater than the preset cognitive uncertainty threshold, the local candidate path is eliminated. For local candidate paths filtered by cognitive uncertainty, aggregate the risk probability distributions of all sub-models of the local candidate path and calculate the conditional value of risk at a pre-set confidence level. If the conditional risk value is greater than the preset risk tolerance threshold, then the local candidate path is eliminated.
7. The robot navigation method in a dynamic environment according to claim 1, characterized in that, After updating the global reference path using a path planning algorithm and controlling the robot to move along the updated global reference path, the method further includes: The process of continuously monitoring whether there are pedestrians within the robot's perception range, filtering out the target local path, controlling the robot to move along the target local path, updating the global reference path, and controlling the robot to move along the updated global reference path continues until the destination is reached.
8. A robot navigation system in a dynamic environment, characterized in that, include: The first control module is used to control the robot to walk along the global reference path generated by the path planning algorithm and to continuously monitor whether there are pedestrians within the robot's perception range; The target local path generation module is used to generate multiple local candidate paths if there are pedestrians within the robot's perception range, and to select the target local path from all the local candidate paths. The local obstacle avoidance module is used to control the robot to move along the target local path and constrain the robot to maintain a preset distance from pedestrians during the movement. The second control module is used to update the global reference path using a path planning algorithm after the robot reaches the end of the target local path, and then control the robot to move along the updated global reference path.
9. A robot, characterized in that, The robot includes: a memory, a processor, and a dynamic environment robot navigation program stored in the memory and executable on the processor. When the dynamic environment robot navigation program is executed by the processor, it implements the steps of the dynamic environment robot navigation method as described in any one of claims 1-7.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a robot navigation program for a dynamic environment. When the robot navigation program for a dynamic environment is executed by a processor, it implements the steps of the robot navigation method for a dynamic environment as described in any one of claims 1-7.