A robotic local path planning method, system, device, and medium

Abstract

The application provides a robot local path planning method, system, device and medium, and belongs to the technical field of data processing. The method comprises a local path planning framework. Multi-source observation data of robots, pedestrians and obstacles are acquired through a perception layer and are mapped to a unified feature space. Then, a two-level SoftMoE mixed expert structure is adopted in an interaction modeling layer to hierarchically model the internal interaction relationship of the crowd and the robot-crowd interaction, respectively, to generate stable local state representations. Finally, a hysteresis routing mechanism based on a risk margin index is introduced in a decision layer. The mechanism adaptively switches between a rule control mode and a proximal policy optimization (PPO) mode according to the safety state of the environment, and a dynamic safety distance penalty and a speed-related collision penalty are introduced in a reward function to guide learning optimization. The method solves the problem that a single strategy cannot balance the traffic efficiency and safety in a complex dynamic crowd environment, and realizes robust and efficient adaptive navigation of the robot.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, specifically relating to a robot local path planning method, system, device, and medium. Background Technology

[0002] Local path planning is a core component of robot navigation and autonomous mobility systems. It aims to dynamically generate safe and feasible paths from the current location to local target points based on real-time perceived environmental information. Its core objective is to address uncertainties in dynamic environments, such as obstacle movement, sensor noise, and sudden environmental changes, while optimizing path smoothness, efficiency, and energy consumption.

[0003] Early advancements in computer vision and sensor technologies led to the mainstream adoption of geometric model-based methods for local path planning (such as artificial potential field methods and dynamic window methods). However, limited by their underlying assumptions and mathematical frameworks, these methods faced challenges in multi-objective optimization and insufficient adaptability to dynamic environments. With improved computing power and breakthroughs in machine learning theory, data-driven methods (such as reinforcement learning and deep learning) began to integrate into traditional planning frameworks, forming a hybrid architecture. Learning complex patterns and decision-making logic from data effectively compensates for the shortcomings of geometric model methods in local path planning. However, data-driven methods exhibit significant dynamism, randomness, and interactive uncertainty in densely populated scenarios. Pedestrian movement continuously alters the local feasible space, potentially causing short-term inaccessibility and channel contraction at the local scale, making it difficult to ensure that accessibility, continuity, and risk control are balanced in local path planning. Summary of the Invention

[0004] To address the shortcomings of existing technologies in local path planning under complex environments, this invention provides a robot local path planning method, system, device, and medium.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A robot local path planning method includes the following steps: The robot motion state sequence, pedestrian observation state sequence, and obstacle observation state sequence are acquired. The sequences are mapped to a unified feature space through a single-layer perception mechanism (MLP). The robot embedding features representing its own motion state and navigation target, the pedestrian embedding features representing dynamic attributes, and the obstacle embedding features representing distance distribution are obtained respectively. The pedestrian embedding features are processed using a SoftMoE network to obtain a crowd interaction representation; the crowd interaction representation and robot embedding features are input into a SoftMoE network to obtain a robot crowd interaction representation; the robot crowd interaction representation, robot embedding features, and obstacle embedding features are respectively input into an MLP for embedding processing; the output embedding features are concatenated and input into a GRU to obtain a local dynamic state representation. The risk margin index is calculated based on the pedestrian observation state sequence and the obstacle observation state sequence; the control mode is dynamically switched according to the preset control strategy switching rules based on the risk margin index; the local dynamic state characterization is input into the switched control mode to obtain the robot's control parameters.

[0006] Preferably, the control strategy switching rule is as follows: ; in, , and These are the complete safety threshold, the exit threshold, and the intervention threshold, respectively, which satisfy... ; For a safe environment, a fast straight-line rule mode is adopted; For low-risk environments, a slow-speed straight-line rule pattern is adopted; For medium- to high-risk environments, the PPO model is adopted. This is a risk margin indicator for the current moment. The control mode selected at time t.

[0007] Preferably, a hysteresis mechanism is used for mode switching in the preset control strategy switching rules, wherein the hysteresis mechanism specifically is: When the robot is in rule-based control mode, environmental changes affect the current risk margin index. If an imminent risk is detected, immediately switch to PPO mode; wherein, the rule control mode includes fast straight line rule mode and slow straight line rule mode; When the robot is in PPO mode, even the risk margin indicator rebounded to The above will not immediately switch back to rule-based control mode. And if the environment is deemed sufficiently open and safe after n consecutive time steps, switch back to rule control mode; In rule-based control mode, based on risk margin indicators The robot's desired speed is adjusted in real time, switching between a fast straight-line control mode and a slow straight-line control mode.

[0008] Preferably, the step of basing the risk margin index Adjusting the robot's desired speed in real time ; ; in, For a completely safe threshold, For the minimum speed ratio, The intervention threshold, The maximum speed ratio, These are the velocity interpolation coefficients obtained by normalizing based on the risk margin. This is the desired speed command.

[0009] Preferably, the step of using a SoftMoE network to process the pedestrian embedding features to obtain a crowd interaction representation specifically involves: introducing... A learnable Vectors, for pedestrian embedding features and The vectors are subjected to linear transformations and the scaling dot product affinity is calculated. Based on the scaling dot product affinity, the sequences are aggregated into slots. For each aggregated sequence... A two-layer MLP expert transformation is used to transform the processed data. Write back to using combined weights The space is processed, and residuals and layer normalization are performed to output a representation of crowd interaction.

[0010] Preferably, the calculation of the risk margin index specifically includes: Based on Euclidean distance and introducing a dynamic safe distance function Calculate the first The risk margin for each pedestrian is as follows: ; ; ; in, For robots and the first The relative velocity vector of each pedestrian Let be the unit vector of the line connecting the two positions. For robots and pedestrians Approximation velocity in the direction of the line connection. For the first Dynamic safe distance for pedestrians Based on the safe distance, To adjust the parameters, and Representing the robot and the first The position vector of a pedestrian in the world coordinate system. Denotes the Euclidean norm. and These represent the geometric radii of the robot and the pedestrian, respectively. Based on the set of static obstacle distance sequences, the obstacle risk margin is specifically as follows: ; in, This is a safety threshold; K represents the minimum distance measurement value; K is the index variable of the distance measurement sequence. Based on the principle of minimum risk, the minimum of the risk margins of visible pedestrians and obstacles is taken as the risk margin index at the current moment.

[0011] Preferably, the method further includes constructing a composite dense reward function comprising termination-type rewards and non-termination-type step rewards, used to guide the processing parameters of the updated control mode during the training phase; the control mode is specifically a rule controller or a PPO algorithm.

[0012] The present invention also provides a robot local path planning system, specifically comprising: The perception module is used to acquire robot motion state sequences, pedestrian observation state sequences, and obstacle observation state sequences. Through a single-layer perception mechanism (MLP), each sequence is mapped to a unified feature space to obtain robot embedding features that represent its own motion state and navigation target, pedestrian embedding features that represent dynamic attributes, and obstacle embedding features that represent distance distribution.

[0013] The interaction modeling module is used to process the pedestrian embedding features using the SoftMoE network to obtain the crowd interaction representation; input the crowd interaction representation and robot embedding features into the SoftMoE network to obtain the robot crowd interaction representation; input the robot crowd interaction representation, robot embedding features and obstacle embedding features into the MLP for embedding processing respectively; and concatenate the output embedding features and input them into the GRU to obtain the local dynamic state representation.

[0014] The navigation decision module calculates a risk margin index based on the pedestrian observation state sequence and the obstacle observation state sequence; dynamically switches the control mode according to the preset control strategy switching rules based on the risk margin index; and inputs the local dynamic state representation into the switched control mode to obtain the robot's control parameters.

[0015] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps described in the robot local path planning method.

[0016] The present invention also provides a computer-readable storage medium storing a computer program that, when loaded by a processor, can execute the steps described in the robot local path planning method.

[0017] The robot local path planning method provided by this invention has the following beneficial effects: This invention employs a two-level SoftMoE structure to replace traditional pairwise attention or graph neural networks. By using a fixed number of learnable slots to perform soft allocation and compression transformation on variable-length pedestrian sets, computational complexity is reduced, and representation oscillations caused by fluctuations in input size are avoided. This allows the robot to achieve smooth and stable interaction characteristics and more consistent obstacle avoidance behavior even in high-density scenarios with dynamically changing pedestrian numbers. A hysteresis-based routing design based on a risk margin index achieves adaptive fusion of rule-based control and the PPO algorithm. By introducing a dynamic safety distance driven by approximation velocity, the risk margin can more accurately reflect the real threat posed by rapidly approaching pedestrians, thus avoiding misjudgments caused by relying solely on Euclidean distance. This enables the robot to operate stably in open areas under an efficient and latency-free rule-based controller, while in high-risk scenarios, it can promptly switch to the PPO strategy to complete complex game-theoretic obstacle avoidance. Stable and efficient local navigation is achieved in complex crowd environments, providing a reliable motion execution foundation for the subsequent implementation of a high-level passage decision module. Attached Figure Description

[0018] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating a robot local path planning method according to an embodiment of the present invention.

[0020] Figure 2 This is a structural diagram of the sensing layer in an embodiment of the present invention.

[0021] Figure 3 This is a structural diagram of SoftMoE in an embodiment of the present invention.

[0022] Figure 4 This is a structural diagram of the interaction modeling layer in an embodiment of the present invention.

[0023] Figure 5 This is a schematic diagram of the dynamic safety distance in an embodiment of the present invention.

[0024] Figure 6 This is a diagram illustrating rule-based routing in an embodiment of the present invention.

[0025] Figure 7 This is a schematic diagram comparing the ablation experiment results in an embodiment of the present invention.

[0026] Figure 8 This invention compares the method of the present invention under different risk levels in various environments with two representative baseline methods. Figure 8 (a) shows the change in PPO usage rate with risk level; Figure 8 (b) represents the stability of the number of strategy switching. Detailed Implementation

[0027] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0028] Example This invention provides a robot local path planning method, such as... Figure 1 As shown, the specific steps include: S1: Construct a local path planning framework. Local path planning adopts a closed-loop decision-making process of "perception—interaction modeling—rule-based routing—action execution," and divides the whole into three parts: the perception layer, the interaction modeling layer, and the decision layer. In each control cycle... First, the robot's own state, the surrounding pedestrian state set, and obstacle observations are acquired at the perception layer and embedded into a unified feature representation. Then, the interaction modeling layer performs hierarchical modeling of the internal relationships of the crowd and the robot-crowd interaction, and extracts a dynamically stable local state representation by combining a temporal recursive structure. Finally, the decision layer selects a three-stage strategy branch based on the representation and generates control actions to complete the adaptive passage to the local dynamic environment.

[0029] (1) such as Figure 2 As shown, the goal of the perception layer is to formally represent the observations of robots, pedestrians, and obstacles and construct inputs that can be processed by the network. Discrete control cycles are employed. The closed-loop decision-making model, with a sampling period of 100 days. Within each cycle, the robot acquires its own motion state and local environmental observations, mapping them to a unified feature space for subsequent interaction modeling and policy generation. First, the robot is set at time... The status is:

[0030] (1); in, This represents the robot's current position in the environment. , These are the robot's linear velocity and angular velocity, respectively, and their ranges are determined by the chassis speed and safety constraints. , , For heading angle, These are the coordinates of the target point.

[0031] Set time The number of pedestrians visible within the robot's field of vision is It can be seen that the pedestrians gathered as For any pedestrian, the observed state is defined as:

[0032] (2); in, Indicates the first A pedestrian Coordinates of time, This represents the corresponding pedestrian velocity vector. Due to the number of pedestrians... Proximity is characterized by ranking pedestrians in ascending order of their relative distance from the robot over time.

[0033] Obstacle information is represented by a range sequence obtained from lidar: (3); in, This represents the number of laser beams. Ultimately, to unify the multi-source output, the perception layer employs a single-layer perception mechanism to map the original observations to a unified feature space:

[0034] (4); (5); (6); (7); (8).

[0035] (2) The interaction modeling layer adopts a series structure of "crowd relationship modeling → robot-crowd interaction fusion → temporal recursive integration" to form a representation that can be used for strategy selection. First, the interaction relationship of the pedestrian set is modeled through a SoftMoE-like aggregation module to obtain the crowd interaction representation. Then, through secondary interaction and fusion with the embedded robot, the robot-related user responses are obtained. The formula is as follows:

[0036] (9); (10); The original pedestrian data was then embedded and combined with the interaction representations of the pedestrian and the robot. A second embedding is performed with the obstacle embedding, and the results of the above embedding are concatenated as the input of the GRU, as shown in the following formula: (11); (12); (13); (14); (15); in, This is the output sequence of the GRU in this step. This is the hidden state of the GRU at this step. It is a time step mask to prevent the states of different episodes from being chained together.

[0037] (3) Decision-making level: Introduce a rule-based routing mechanism at the decision-making level in each control cycle. The system first calculates a scalar index to characterize the local safety margin. Subsequently based on A three-stage policy transition is implemented based on the policy state at the previous time step: (16); in, , and These are the complete safety threshold, the exit threshold, and the intervention threshold, respectively, which satisfy... The three types of strategies correspond to: Safe environment for rapid, straight-line passage; In low-risk environments, proceed slowly and in a straight line. The PPO algorithm is triggered in medium- to high-risk environments. A hysteresis mechanism is used for policy switching to avoid frequent switching near the threshold.

[0038] S2: To improve scalability and representation stability in dense scenarios, a SoftMoE-like interaction hybrid structure is introduced in the interaction modeling layer. This structure does not rely on explicitly constructing a fully connected interaction graph, but instead uses a fixed number of intermediate aggregation units ( The set is compressed, transformed, and then written back, transforming the main computation of interaction modeling from pairwise combination of subjects to subject-specific computation. The soft allocation allows for relatively stable computational overhead and representation quality even with a large number of subjects. Multi-expert parallel nonlinear transformations provide higher expressive capacity for different interaction patterns, enabling the network to simultaneously characterize multiple local interaction modes within a unified framework without relying on a single global aggregation mechanism. The SoftMoE structure is as follows: Figure 3 As shown.

[0039] SoftMoE's Compression and write-back mechanisms are naturally well-suited for handling variable pedestrian numbers. a fixed number It can be viewed as a low-dimensional prototype set of group interactions, such that the representation dimension and the main computational burden do not change with each other. The linear or quadratic growth and dramatic changes provide a smoother and more stable state representation for subsequent rule-based routing and PPO decisions. Meanwhile, due to different... For input The soft allocation weights are different, and A fixed mapping is used between experts, and different experts will face different statistical distributions and combinatorial structures during the training process. The input allows for a structurally compatible approach that focuses on different types of interaction information: for example, some experts tend to encode local geometric relationships related to nearest neighbor avoidance and collision risk, while others are more likely to capture complex interaction patterns among multiple pedestrians. This mechanism enables the model to accommodate multiple social interaction features within a unified framework without relying on a single global aggregation, providing a more discriminative state representation foundation for subsequent rule-based routing and PPO algorithms.

[0040] To address two types of information needs: intra-crowd interaction and robot-crowd interaction, such as Figure 4 As shown, a two-level SoftMoE structure was designed in the interaction modeling layer: the first level is used to extract the internal relationships (HH) of the crowd, and the second level is used to fuse the robot's interaction with the crowd and form a decision-oriented aggregate representation (RH). The two-level SoftMoE adopts a unified architecture, assuming the input of a certain layer... The sequence is:

[0041] (17); in, For batch size, for Quantity (length after padding) This is the latent space dimension. SoftMoE introduces this. A learnable vector:

[0042] (18); First of all and Perform linear projections separately and calculate the scaling dot product affinity (logits): (19); (20); (twenty one); in, For linear layer weights, They represent the first The first sample indivual The result of linear transformation of the feature and the m-th feature The result of a linear transformation of the parameter vector.

[0043] SoftMoE was subsequently in the same group. Construct two sets of softmax weights: Dispatch weight: For Dimensional normalization, each From all Collect information. (Set up / set) for Validity mask (1 represents valid, 0 represents padding), then:

[0044] (twenty two); Therefore, we obtain enter: (twenty three); combine weight pairs Dimensional normalization, each From all Recycling Information: (twenty four); For each A two-layer MLP expert system is used for transformation. Let the number of experts be... ,when Use round-robin mapping Will Assigned to experts:

[0045] (25); Each expert is a two-layer feedforward network, containing ReLU and Dropout: (26); After expert processing Write back to using combine weights space: (27); Finally, regarding padding The output is set to zero, and residual connections and LayerNorm are used to enhance stability. (28); (1) First layer SoftMoE: Pedestrian internal interaction (HH): Time The observable pedestrian embeddings are stacked according to their relative distances. sequence: (29); in The maximum number of pedestrians after padding, mask. Mark valid pedestrians within the field of vision After a series of interactive information extractions by SoftMoE, the content relationships of the pedestrians are obtained. express :

[0046] (30); The fused population representation is denoted as a matrix, representing the interactions of the entire population set, rather than a single vector: (31); (2) Second layer SoftMoE: Robot-pedestrian interaction fusion (RH): This layer features robots. Interaction with the crowd on the first floor A joint sequence is formed, and the 0th element in the joint sequence is taken after mixing with SoftMoE. (robot The output of () is used as a pooled representation, thus obtaining a contextual representation that fuses robot and pedestrian information. The robot's embedded representation is obtained from equation (4). embedding robots with pedestrians Mapped to the RH mixing space, then concatenated to obtain the RH input. sequence:

[0047] (32); (33); (34); The corresponding RH validity mask is defined as follows: (35); in For pedestrians in the HH phase The validity mask, the robot itself Always valid, its mask is set to a constant value of 1.

[0048] The obtained spliced ​​sequence With validity mask The input to the SoftMoE interactive layer yields the output sequence: (36); Finally, extract the 0th element of the output sequence. As a pooled representation, it is used to characterize the context in which the robot incorporates pedestrian interaction information: (37).

[0049] S3: In real-world, densely populated environments, robots inevitably face highly dynamic and uncertain local interaction problems. Especially in areas with dense pedestrian traffic or complex movement patterns, a single strategy is insufficient to simultaneously ensure both traffic efficiency and movement safety.

[0050] The decision-making level introduces a rule-based routing mechanism, specifically a hysteresis-based rule-based routing mechanism based on risk margin. This mechanism constructs a unified environmental risk index to quantitatively assess the safety of the current scenario and dynamically switches between multiple control strategies accordingly, thereby improving overall traffic efficiency while ensuring safety. Hysteresis-based strategy switching is also introduced to avoid frequent policy switching near thresholds, enhancing the system's stability and robustness.

[0051] (1) Risk margin indicators The core design goal is to quantify the safety status between the robot and surrounding obstacles, and its value directly reflects the risk level of the local environment. A smaller value indicates a lower safety margin for the robot and obstacles, and a higher environmental risk. Conversely, a larger value indicates a lower risk, allowing for more efficient straight-line strategies. This metric considers not only the physical distance to static obstacles but also incorporates a pedestrian collision risk assessment based on dynamic comfort distance.

[0052] Considering the movement characteristics of pedestrians, using Euclidean distance alone is insufficient to characterize risk. For example... Figure 5 As shown, a dynamic safety distance function is introduced. The function varies with the approach speed of the robot and the pedestrian. Linear growth. (The...) The risk margin for an individual pedestrian is calculated as follows:

[0053] (38); (39); (40); in, For robots and the first The relative velocity vector of each pedestrian Let be the unit vector of the line connecting the two positions. For robots and pedestrians Approximation speed in the direction of the line. Based on the safe distance, To adjust the parameters. and Representing the robot and the first The position vector of a pedestrian in the world coordinate system. This represents the Euclidean norm. and These represent the geometric radii of the robot and the pedestrian, respectively. The physical meaning is: the remaining safety range after deducting the dynamic safety buffer that takes into account relative velocity.

[0054] For the set of static obstacle distance sequences detected by lidar Its risk margin is defined as the minimum ranging value minus the safety threshold: (41); Considering that the robot is constrained by both pedestrians and obstacles during actual operation, the principle of minimum risk is adopted, unifying the two types of risk into a single indicator: (42).

[0055] (2) Hysteresis-based policy switching: The core challenge of policy switching is to avoid jitter near the threshold, as frequent policy switching can lead to unstable robot motion. Therefore, a policy switching rule with hysteresis characteristics is designed, based on calculated indicators. The routing layer needs to switch between rule-based mode and PPO mode. A demonstration of rule-based routing is shown below. Figure 6 As shown.

[0056] Set two key thresholds: intervention threshold With exit threshold And satisfy The strategy switching logic is as follows:

[0057] (1) Non-PPO mode PPO mode (rapid response): When the robot is in rule-based control mode, once environmental changes cause... The system determines that a risk is imminent and immediately switches to the PPO strategy. PPO networks excel at handling complex group games and can execute precise obstacle avoidance actions.

[0058] (2) PPO model Non-PPO mode (conservative exit): When the robot is in PPO mode, even if rebounded to The system will not immediately switch back to rule-based mode. Only when... Furthermore, the system only determines that the environment is sufficiently open and safe after maintaining this state for five consecutive time steps, thus switching back to the rule controller. This hysteresis design, which makes it easy to enter but difficult to exit, ensures the integrity of the obstacle avoidance process.

[0059] (3) Adaptive speed adjustment in rule mode: In rule control mode (non-PPO), in order to further improve smoothness, a constant speed is not used, but rather the speed is adjusted according to the rules. Real-time switching between fast straight-ahead and slow straight-ahead driving. Desired speed. and The relationship is shown in the following formula:

[0060] (43); (44); in, For a completely safe threshold, For the minimum speed ratio, These are the velocity interpolation coefficients obtained by normalizing based on the risk margin, and their range is determined by the value of the clamp. The range is limited to [0,1]. This mechanism ensures that when the risk increases slightly but the PPO has not yet been triggered, the robot can decelerate smoothly, rather than brake or steer abruptly.

[0061] Through the above design, the routing layer achieves efficient, linear navigation with no computational delay in open areas using rule controllers; it achieves safe interactive obstacle avoidance in congested areas using PPO; and it smooths the transition between modes through hysteresis logic and speed adjustment, meeting the navigation needs in complex scenarios.

[0062] S4: In the deep reinforcement learning framework, the reward function is the core factor guiding the agent to learn the optimal strategy. To address the problems of sparse rewards, high collision risk, and uneven motion in dense crowd navigation scenarios, a composite dense reward function integrating task completion, dynamic safety constraints, and kinematic constraints is constructed. Overall reward... It consists of two parts: termination reward and non-termination progress reward, and its mathematical form is defined as:

[0063] (45); in, Used to describe the task result at the end of the round. Guided rewards are continuously applied at each time step to mitigate the sparse feedback problem and constrain the robot's motion behavior.

[0064] 1. Termination-type reward functions: These functions are triggered at the end of an episode to explicitly inform the agent of the final result of the task. They mainly include three cases: reaching the target, collision, and timeout.

[0065] Target reward: When the distance between the robot and the target point is less than the threshold radius. When the time is up, the task is considered successful and a substantial positive reward is given. This is to encourage intelligent agents to complete navigation tasks.

[0066] Velocity-based collision reward: Traditional collision penalties are usually set to a fixed constant, which ignores the differences in collision severity at different relative velocities. A dynamic collision penalty mechanism based on normal approximation velocity is designed. Under this mechanism, the penalty generated by a collision is no longer a single constant, but is composed of a basic penalty term and a dynamic velocity term, with the penalty intensity increasing as the normal approximation velocity increases at the instant of collision.

[0067] This design significantly enhances the model's ability to identify different collision risks, enabling it to effectively distinguish the fundamental difference between "low-speed contact" and "high-speed impact." This mechanism guides the agent to learn two key advanced safety strategies: defensive deceleration, which involves actively reducing speed to minimize the penalty when a collision is deemed unavoidable; and reverse avoidance, which involves the agent learning to take reverse acceleration or retreat actions when faced with pedestrians who suddenly turn back or approach rapidly, thereby mitigating high-risk conflicts by actively reducing the relative approach rate.

[0068] (46); in: The basic collision penalty constant; This refers to the speed penalty weighting coefficient; This is the speed scaling factor, which determines the sensitivity of the penalty function to changes in speed; It is a power of the nonlinear penalty, used to nonlinearly amplify the penalty intensity of high-speed collisions; The truncation function is defined as follows: ,in This is a velocity threshold to prevent gradient explosion or excessive penalties caused by extremely high velocities. The core variables in the formula represent... Normal approach velocity is the relative approach speed of the two objects in the direction normal to the contact surface at the instant of collision. Its calculation varies depending on the objects colliding:

[0069] (1) When colliding with a pedestrian, the velocity vectors are determined by the projection of the relative velocities of the two parties onto the line connecting their centers. Let the velocity vectors of the robot and the pedestrian be respectively... and Unit direction vector To point to a pedestrian, the formula is as follows:

[0070] (47); (2) When colliding with a static obstacle, since the obstacle is stationary, to conservatively estimate the risk, the robot's velocity modulus is directly taken: (48); Timeout handling: If the robot fails to reach the target and avoid collision within the specified time step, the round ends forcibly without any additional reward or penalty, prompting the agent to find a better time-efficient strategy.

[0071] 2. Non-terminating reward function: Non-terminating rewards are calculated at every time step and are designed to provide continuous dense feedback to the agent. They mainly include potential-based path guidance, dynamic safe distance penalty, and time step penalty.

[0072] (1) Potential Energy-Based Guided Reward: To guide the robot to continuously approach the target, a potential energy-based method is adopted. Define the potential energy function. This represents the negative value of the Euclidean distance between the robot's current position and the target point. It also represents the guidance reward for each step. Equal to the change in potential energy:

[0073] (49); in For the robot's position, For the target location, This is the scaling factor. Simultaneously, to avoid extreme step sizes leading to abnormal returns, the potential energy return is clipped to... This reward system encourages the robot to move closer to the goal with each step, effectively mitigating the reward sparsity problem in navigation tasks.

[0074] (2) Static-friendly dynamic safety distance penalty: In order to ensure safety while avoiding the robot being too conservative towards static pedestrians and thus causing the robot to freeze, and to more accurately characterize social distance, the reward function adopts a nonlinear dynamic distance different from Equation 3-39. When pedestrians quickly approach the robot, The increased size requires the robot to maintain a greater distance. When pedestrians are stationary or moving away, Retraction allows robots to pass close together in crowded environments. Dynamic safe distance is defined as... The function is given by the following formula:

[0075] (50); in: Basic safe distance; Let be the normal phase approximation velocity between the robot and the nearest pedestrian, as defined in Equation 3.4.1; and These are the contraction coefficient and expansion coefficient, used to adjust the safety distance; To adjust parameters and control the sensitivity of contraction speed.

[0076] If the actual clearance distance between the robot and the pedestrian Less than Intrusion penalty will then be imposed: (51); This mechanism enables robots to flexibly distinguish between mobile threats and static obstacles, exhibiting obstacle avoidance behavior that is more in line with human social norms.

[0077] (3) Time step penalty: A small constant reward is deducted after each time step. This enables robots to complete tasks in the shortest possible time.

[0078] The effectiveness of the proposed rule-based router-based adaptive navigation method is verified through system simulation experiments. First, the experimental platform, environment configuration, and evaluation index system are introduced. Then, the contribution of each improved module is analyzed through ablation experiments. Finally, the generalization ability and robustness of the method are verified through comparative experiments under different risk levels.

[0079] To verify the effectiveness of the improved modules of the local path planning framework, an ablation study was designed to compare the navigation performance under four different configurations. E1-Base is the basic PPO method without any improvements; E2-Reward improves the reward function based on E1; E3-SoftMoE adds a SoftMoE hybrid expert architecture to E1; and E4-Full is the complete method. All configurations were trained and tested under the same random number seed conditions, and comparative experiments were conducted in environments with a dynamic pedestrian scale of 2–10 people. The policy training for each method used the same training duration to ensure consistency in experimental settings and fairness in comparison. Experimental results are as follows: Figure 7 As shown, a comprehensive evaluation was conducted from multiple dimensions, including task completion, navigation efficiency, and safety.

[0080] Ablation experiments revealed that the E1-Based PPO method performed poorly in complex crowd environments. From a task completion perspective, the success rate was only 2.0%, meaning that only about 10 out of 500 test episodes were successfully completed. A timeout rate as high as 80.0% indicates that the robot failed to reach the target within the time limit in most cases. Navigation efficiency was also poor, with an average navigation time of 29.19 seconds, almost reaching the time limit, and an average path length of 19.75 meters, far exceeding the straight-line distance from the starting point to the target point. This reflects that even in the few successful cases, the robot employed numerous detours and waiting strategies. Regarding safety, the collision rate was 18.0%, with human collisions accounting for 15.0%, obstacle collisions for 3.0%, and personal space intrusion for 4.15%. These results expose a core problem with traditional reinforcement learning methods: in complex scenarios with insufficient training data coverage, the learning strategy is prone to getting trapped in local optima, leading to frequent timeouts or overly conservative strategies.

[0081] Improvements to the reward function significantly enhanced navigation performance. The success rate of E2-Reward increased from 2.0% to 75.0%, meaning that navigation tasks could be successfully completed in the vast majority of test scenarios. Furthermore, the timeout rate plummeted from 80.0% to 1.0%, indicating that the robot can now efficiently complete tasks within the time limit. Thanks to the velocity-related collision penalty and dynamic safety distance mechanism proposed in Section 3.4, the average navigation time was reduced to 18.25 seconds, and the path length was optimized to 10.45 meters, the latter being only about half the path length of E1-Base. In terms of safety, the obstacle collision rate decreased from 3.0% to 2.0%, the human collision rate increased from 15.0% to 22.0%, and the personal space intrusion rate also increased from 4.15% to 12.3%. However, considering the extremely low success rate and high timeout rate of E1-Base, this comparison is not practically meaningful. In fact, the safety performance exhibited by E2-Reward has reached an acceptable level, and its breakthrough in success rate fully demonstrates the core role of reward function design in reinforcement learning.

[0082] The introduction of the SoftMoE hybrid expert architecture enabled multi-dimensional performance optimizations in E3-SoftMoE. In terms of task completion performance, E3-SoftMoE demonstrated excellent performance comparable to E2-Reward, achieving a success rate of 76.0%, a timeout rate of 1.0%, and a collision rate of 23.0%, all significantly better than the base method E1-Base. Navigation efficiency was further improved, with the average navigation time reduced to 17.64 seconds and the path length optimized to 10.19 meters, not only far superior to E1-Base but also showing a slight improvement over E2-Reward. Safety results showed a different optimization direction compared to E2-Reward; the human collision rate decreased to 21.0%, lower than E2-Reward's 22.0%, but the personal space intrusion rate increased to 17.87%. This phenomenon reflects the characteristics of the SoftMoE architecture, which more accurately predicts pedestrian behavior through a hybrid expert mechanism, enabling the robot to adopt a more aggressive navigation strategy while ensuring collision avoidance. Therefore, although the frequency of entering personal space increased, the actual collision rate decreased.

[0083] The complete E4-Full approach integrates the reward function improvements of E2-Reward and the hybrid expert architecture of E3-SoftMoE, achieving optimal overall performance through the combination of these two effective improvements. The success rate is increased to 85.0%, the timeout rate is reduced to 0%, and the robot no longer fails to reach the target within the time limit. The collision rate is controlled at 15.0%, the lowest among all configurations. Navigation efficiency also reaches an optimal level, with an average navigation time of 16.57 seconds and a path length of 9.94 meters, both the best among the four configurations, demonstrating high path planning quality. Safety features are comprehensively optimized, with the human collision rate significantly reduced to 13.0%, both lower than E2-Reward's 22.0% and E3-SoftMoE's 21.0%. The personal space intrusion rate is 13.25%, falling between E2-Reward and E3-SoftMoE, achieving a good balance. The combination of these two improvements gives the system both excellent reward guidance and powerful environmental awareness and interaction modeling capabilities, thus achieving optimal performance in three dimensions: task completion rate, navigation efficiency, and safety, fully demonstrating the advantages of multi-module collaborative optimization.

[0084] In order to evaluate the performance advantages, comparative experiments were conducted under different risk levels, comparing the method of this invention with two representative baseline methods: the DWA traditional local path planning method and the HEIGHT navigation strategy network architecture.

[0085] The tests were conducted in three different population densities: low-risk (2-6 people), medium-risk (6-10 people), and high-risk (10-14 people). 500 episodes were tested at each risk level to comprehensively evaluate the performance of each method under different environmental complexities. The results are shown in Table 1.

[0086] Table 1. Detailed performance comparison under different risk levels As shown in Table 1, the three methods exhibit significant differences in task completion performance. The method of this invention achieves the highest success rate and the lowest collision rate across all risk levels, fully validating its superiority. In low-risk environments, the success rate and collision rate of this invention are 94.6% and 5.4%, respectively, both superior to the HEIGHT method's 86.6% and 13.4%, and DWA's 44.2% and 39.4%. This indicates that the proposed method can handle most low-density population situations. As the environmental risk level increases, the advantages of this invention's method gradually expand, fully demonstrating its adaptability in complex scenarios. In medium-risk environments, the success rate of this invention's method is 85.4%, an improvement of 12.8 percentage points compared to RH-HH, while the collision rate is reduced by 12.6 percentage points. In high-risk environments, the success rate of this invention's method is 80.4%, an improvement of 18 percentage points compared to the HEIGHT method, while the collision rate is reduced by 17 percentage points. The success rate advantage of the method in this invention compared to the HEIGHT method increases from 8 percentage points in low-risk environments to 18 percentage points in high-risk environments. This demonstrates that the SoftMoE architecture significantly improves the ability to handle complex interactions in high-density crowds by allocating dedicated expert networks for different interaction modes. In contrast, the performance of the DWA method drops sharply with increasing environmental complexity, with the success rate decreasing from 44.2% to 10.0% and the collision rate soaring from 39.4% to 78.4%. This clearly exposes the lack of predictive ability for dynamic crowd behavior in traditional local path planning methods, making effective navigation difficult in complex crowd environments.

[0087] In terms of navigation efficiency, the method of this invention achieves the shortest navigation time under all risk levels. In low, medium, and high risk environments, the navigation times of the method are 14.40 seconds, 15.43 seconds, and 15.78 seconds, respectively, which are 21.3%, 21.8%, and 23.0% shorter than the HEIGHT method's 18.29 seconds, 19.73 seconds, and 20.48 seconds. This efficiency advantage is mainly attributed to the synergistic effect of the dynamic safety distance mechanism and the rule-based router. The dynamic safety distance mechanism allows the robot to take a more direct path while remaining safe, reducing unnecessary detours compared to the RH-HH method's use of a fixed safety distance. The rule-based router, through risk perception, uses rule-based strategies to quickly traverse low-risk areas, avoiding redundant calculations of the learned strategy in simple scenarios and further improving navigation efficiency.

[0088] A comprehensive analysis of safety and comfort indicators reveals that the method of this invention effectively maintains the safety boundary while ensuring high performance. Regarding the minimum distance, the method of this invention achieves 0.595 meters in low-risk environments, superior to HEIGHT's 0.551 meters and DWA's 0.409 meters. In medium- and high-risk environments, the method of this invention maintains a comparable level to HEIGHT, but significantly outperforms DWA, indicating that the method of this invention can still maintain a reasonable safety boundary in high-density crowds. Regarding personal space intrusion rate, the method of this invention performs excellently in medium- and high-risk environments, with rates of 12.92% and 12.90% respectively, significantly lower than RH-HH's 15.38% and 19.73%. Although the personal space intrusion rate of the method of this invention is 12.97% in low-risk environments, slightly higher than HEIGHT's 9.46%, this is because the rule-based router uses rule policies more frequently in low-risk areas for faster passage, optimizing navigation efficiency while ensuring overall safety. In terms of comfort indicators, the average acceleration of the method of this invention is 0.063 m / s², 0.065 m / s², and 0.059 m / s² in the three environments, all of which are better than RH-HH's 0.070 m / s², 0.078 m / s², and 0.081 m / s², and significantly better than DWA's 0.095 m / s², 0.101 m / s², and 0.105 m / s². The lower acceleration indicates that the motion trajectory of the method of this invention is smoother, without abrupt acceleration and deceleration, which is beneficial to improving travel comfort and system stability.

[0089] The adaptive characteristics of rule-based routers have been fully demonstrated in environments with different risk levels. For example... Figure 8As shown, the PPO usage rate increases significantly with the increase of environmental risk level, rising from 67.91% in low-risk environments to 80.69% in medium-risk environments, and then to 91.04% in high-risk environments. This trend reflects the characteristic of rule-based routers to adaptively adjust policy usage according to environmental risk: using rule-based policies for faster passage in low-risk areas, and using PPO policies for more precise obstacle avoidance in high-risk areas, achieving an optimal balance between efficiency and security. Regarding the number of policy switching times, the average switching times were 0.93, 0.96, and 0.88 in low, medium, and high-risk environments, respectively, with the average switching times remaining close to one and stable. This indicates that the rule-based routers did not exhibit excessively frequent switching issues under different complexity environments, fully demonstrating the effectiveness of the hysteresis mechanism.

[0090] The present invention also provides a robot local path planning system, comprising: The perception module is used to acquire robot motion state sequences, pedestrian observation state sequences, and obstacle observation state sequences. Through a single-layer perception mechanism (MLP), each sequence is mapped to a unified feature space to obtain robot embedding features that represent its own motion state and navigation target, pedestrian embedding features that represent dynamic attributes, and obstacle embedding features that represent distance distribution.

[0091] The interaction modeling module is used to process pedestrian embedding features using the SoftMoE network to obtain crowd interaction representation; the crowd interaction representation and robot embedding features are input into the SoftMoE network to obtain robot crowd interaction representation; the robot crowd interaction representation, robot embedding features and obstacle embedding features are respectively input into the MLP for embedding processing, and the output embedding features are concatenated and input into the GRU to obtain local dynamic state representation.

[0092] The navigation decision module calculates a risk margin index based on pedestrian and obstacle observation state sequences; dynamically switches control modes according to preset control strategy switching rules based on the risk margin index; and inputs the local dynamic state representation into the switched control mode to obtain the robot's control parameters. Each module in the above-described robot local path planning system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, allowing the processor to call and execute the corresponding operations of each module.

[0093] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps in an embodiment of a robot local path planning method. Specific implementation methods can be found in the method embodiments, and will not be repeated here.

[0094] Furthermore, the present invention also provides a non-transitory computer-readable storage medium containing instructions on which a computer program is stored. For example, a memory containing instructions that can be executed by a processor of a computer device to perform the above-described method. For example, the non-transitory computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc. When the computer program is executed by the processor, it can implement the steps in an embodiment of a robot local path planning method. Specific implementation methods can be found in the method embodiments, which will not be repeated here.

[0095] Those skilled in the art will understand that embodiments of the present invention can provide methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0096] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0097] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0098] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1The steps of the function specified in one or more boxes.

[0099] It should be noted that the specific embodiments described above enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although the present invention has been described in detail in this specification and embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention; and all technical solutions and improvements that do not depart from the spirit and scope of the present invention are covered within the protection scope of the present invention patent. No reference numerals in the claims should be construed as limiting the scope of the claims. Any simple variations or equivalent substitutions of technical solutions that can be readily obtained by those skilled in the art within the scope of the technology disclosed in the present invention are within the protection scope of the present invention.

Claims

1. A robot local path planning method, characterized in that, Includes the following steps: The robot motion state sequence, pedestrian observation state sequence, and obstacle observation state sequence are acquired. The sequences are mapped to a unified feature space through a single-layer perception mechanism (MLP). The robot embedding features representing its own motion state and navigation target, the pedestrian embedding features representing dynamic attributes, and the obstacle embedding features representing distance distribution are obtained respectively. The pedestrian embedding features are processed using a SoftMoE network to obtain a crowd interaction representation; the crowd interaction representation and robot embedding features are input into a SoftMoE network to obtain a robot crowd interaction representation; the robot crowd interaction representation, robot embedding features, and obstacle embedding features are respectively input into an MLP for embedding processing; the output embedding features are concatenated and input into a GRU to obtain a local dynamic state representation. The risk margin index is calculated based on the pedestrian observation state sequence and the obstacle observation state sequence; the control mode is dynamically switched according to the preset control strategy switching rules based on the risk margin index; the local dynamic state characterization is input into the switched control mode to obtain the robot's control parameters.

2. The robot local path planning method according to claim 1, characterized in that, The control strategy switching rules are as follows: ； in, , and These are the complete safety threshold, the exit threshold, and the intervention threshold, respectively, which satisfy... ; For a safe environment, a fast straight-line rule mode is adopted; For low-risk environments, a slow-speed straight-line rule pattern is adopted; For medium- to high-risk environments, the PPO model is adopted. This is a risk margin indicator for the current moment. The control mode selected at time t.

3. The robot local path planning method according to claim 2, characterized in that, The preset control strategy switching rules employ a hysteresis mechanism for mode switching. Specifically, the hysteresis mechanism is as follows: When the robot is in rule-based control mode, environmental changes affect the current risk margin index. If an imminent risk is detected, immediately switch to PPO mode; wherein, the rule control mode includes fast straight line rule mode and slow straight line rule mode; When the robot is in PPO mode, even the risk margin indicator rebounded to The above will not immediately switch back to rule-based control mode. And if the environment is deemed sufficiently open and safe after n consecutive time steps, switch back to rule control mode; In rule-based control mode, based on risk margin indicators The robot's desired speed is adjusted in real time, switching between a fast straight-line control mode and a slow straight-line control mode.

4. The robot local path planning method according to claim 3, characterized in that, According to the risk margin index Adjusting the robot's desired speed in real time ； ； in, For a completely safe threshold, For the minimum speed ratio, The intervention threshold, The maximum speed ratio, These are the velocity interpolation coefficients obtained by normalizing based on the risk margin. This is the desired speed command.

5. The robot local path planning method according to claim 1, characterized in that, The process of using the SoftMoE network to process the pedestrian embedding features to obtain a crowd interaction representation specifically involves: introducing... A learnable Vectors, for pedestrian embedding features and The vectors are subjected to linear transformations and the scaling dot product affinity is calculated. Based on the scaling dot product affinity, the sequences are aggregated into slots. For each aggregated sequence... A two-layer MLP expert transformation is used to transform the processed data. Write back to using combined weights The space is processed, and residuals and layer normalization are performed to output a representation of crowd interaction.

6. The robot local path planning method according to claim 1, characterized in that, The specific indicators for calculating the risk margin are: Based on Euclidean distance and introducing a dynamic safe distance function Calculate the first The risk margin for each pedestrian is as follows: ； ； ； in, For robots and the first The relative velocity vector of each pedestrian Let be the unit vector of the line connecting the two positions. For robots and pedestrians Approximation velocity in the direction of the line connection. For the first Dynamic safe distance for pedestrians Based on the safe distance, To adjust the parameters, and Representing the robot and the first The position vector of a pedestrian in the world coordinate system. Denotes the Euclidean norm. and These represent the geometric radii of the robot and the pedestrian, respectively. Based on the set of static obstacle distance sequences, the obstacle risk margin is specifically as follows: ； in, This is a safety threshold; K represents the minimum distance measurement value; K is the index variable of the distance measurement sequence. Based on the principle of minimum risk, the minimum of the risk margins for visible pedestrians and the risk margins for obstacles is used as the risk margin indicator at the current moment.

7. The robot local path planning method according to claim 1, characterized in that, It also includes constructing a composite dense reward function that includes termination-type rewards and non-termination-type step rewards, used to guide the processing parameters of the updated control mode during the training phase; the control mode is specifically a rule controller or a PPO algorithm.

8. A robot local path planning system, characterized in that, include: The perception module is used to acquire robot motion state sequences, pedestrian observation state sequences, and obstacle observation state sequences. Through a single-layer perception mechanism (MLP), each sequence is mapped to a unified feature space to obtain robot embedding features that represent its own motion state and navigation target, pedestrian embedding features that represent dynamic attributes, and obstacle embedding features that represent distance distribution. The interaction modeling module is used to process the pedestrian embedding features using the SoftMoE network to obtain the crowd interaction representation; input the crowd interaction representation and robot embedding features into the SoftMoE network to obtain the robot crowd interaction representation; input the robot crowd interaction representation, robot embedding features and obstacle embedding features into the MLP for embedding processing respectively; and concatenate the output embedding features and input them into the GRU to obtain the local dynamic state representation. The navigation decision module calculates a risk margin index based on the pedestrian observation state sequence and the obstacle observation state sequence; dynamically switches the control mode according to the preset control strategy switching rules based on the risk margin index; and inputs the local dynamic state representation into the switched control mode to obtain the robot's control parameters.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is loaded by the processor, it is able to perform the steps of the method according to any one of claims 1 to 7.

Images