Mobile robot robust path planning method and system based on space-time attention and adaptive entropy adjustment

Abstract

The invention discloses a mobile robot robust path planning method and system based on space-time attention and adaptive entropy adjustment, and belongs to the technical field of robot autonomous navigation. In order to solve the problem that deep reinforcement learning is easy to fall into local extremum and trajectory deadlock under local observation, a Recurrent-CBAM-PPO network model is constructed to extract environmental spatial and temporal characteristics; in the decision-making stage, an action filtering mechanism based on topological constraints is designed, and invalid cycle actions violating space connectivity are shielded from a physical rule by comparing historical tracks in real time; in the training stage, a stability-perceived adaptive entropy adjustment mechanism is constructed, an entropy coefficient in a PPO loss function is dynamically adjusted by monitoring trajectory repetition, and random exploratory performance is automatically enhanced to assist escape when a strategy stagnation trend is detected. According to the method, the decision robustness of the mobile robot in a complex dynamic environment can be remarkably improved, and the navigation fluency and success rate are guaranteed.

Description

Technical Field

[0001] This invention relates to the field of intelligent robot technology, and in particular to a robust path planning method and system for mobile robots based on spatiotemporal attention and adaptive entropy adjustment. Background Technology

[0002] Path planning is a key technology for achieving autonomous navigation in mobile robots. Traditional global planning algorithms (such as A) * Dijkstra's algorithm relies on a static global map, making it difficult to handle dynamic obstacles; while local planning methods based on deep reinforcement learning (DRL) possess dynamic obstacle avoidance capabilities, they face two major challenges in practical applications:

[0003] "State confusion" caused by local observation: Based solely on the laser or visual information of the current frame, the robot has difficulty distinguishing similar dead ends or narrow passages, and is prone to repeatedly switching between two state points, forming in-place oscillations or dead loops.

[0004] The failure of the balance between exploration and exploitation: When caught in oscillation, the traditional PPO algorithm often persists in executing actions that cause oscillation due to the high determinism of the policy network, and is unable to escape the local optimum through random exploration.

[0005] Existing technologies typically suppress oscillations through simple distance penalties, which have limited effectiveness. Summary of the Invention

[0006] The purpose of this invention is to provide a robust path planning method and system for mobile robots based on spatiotemporal attention and adaptive entropy adjustment, aiming to solve the technical problem that existing reinforcement learning path planning methods are prone to getting stuck in place and unable to effectively break out of dead loops.

[0007] The technical solution to achieve the objective of this invention is: a robust path planning method for mobile robots based on spatiotemporal attention and adaptive entropy adjustment, comprising the following steps:

[0008] Step 1: Obtain the local observation state of the mobile robot. The local observation state includes obstacle distribution information within the local field of view and path guidance information generated by the global planning algorithm.

[0009] Step 2: Construct a deep reinforcement learning policy network that includes a spatiotemporal feature extraction module, wherein the spatiotemporal feature extraction module is a convolutional neural network, an attention mechanism module, and a gated recurrent unit connected in series.

[0010] Step 3: Generate a topological constraint mask based on the robot's current position and previous position. The topological constraint mask is based on physical space connectivity and filters out candidate actions in the action space that could lead to deadlock or invalid loops.

[0011] Step 4: Input the local observation state into the network model, combine it with the topological constraint mask to output the action probability distribution at the current moment, sample and execute the action, and record the state transition data and the historical trajectory of the position.

[0012] Step 5: Calculate the trajectory repetition index based on the location history trajectory, and dynamically adjust the entropy regularization coefficient in the near-end policy optimization algorithm according to the index to balance the exploration and utilization capabilities of the policy network.

[0013] Step 6: Construct a loss function using the adjusted entropy regularization coefficient, and iteratively update the network model parameters until the model converges.

[0014] Furthermore, the specific process of obtaining the local observation state in step 1 includes:

[0015] Step 1-1, Global Guidance Generation: Generate a global path matrix from the starting point to the target point on a static global map using a global path planning algorithm;

[0016] Step 1-2, Local view capture: Centered on the robot's current position, capture a local area of ​​the corresponding size from the global path matrix as a path guidance layer, and at the same time obtain a local obstacle distribution layer of the same size;

[0017] Steps 1-3, Hybrid State Construction: Using frame stacking technology, the current path guidance layer, obstacle distribution layer, and historical observation frames are spliced ​​together in the channel dimension to form a high-dimensional state tensor, which is then combined with the hidden state vector from the previous time step as the final input state.

[0018] Furthermore, the network model in step 2 specifically includes:

[0019] Step 2-1, Feature Extraction Layer: A Convolutional Neural Network (CNN) is used to process the local observation state;

[0020] Step 2-2, Attention Enhancement Layer: Introduce a feature reweighting operator and embed a CBAM module after the convolutional layer. The key feature channels are weighted sequentially through the channel attention submodule, and the spatial attention submodule focuses on the spatial position of obstacles and paths.

[0021] Steps 2-3, Temporal Memory Layer: The GRU unit receives the flattened feature vector and the hidden state of the previous time step, outputs the feature code containing the temporal memory, and establishes the Markov dependency between local observations.

[0022] Furthermore, the specific logic for generating the topological constraint mask in step 3 is as follows:

[0023] Step 3-1: Calculate the number of non-obstacle neighbor nodes around the current position using the feasible region discrimination function, and initialize the mask vector to all 1s;

[0024] Step 3-2, Detect Action If the target location is a static obstacle or exceeds the map boundary, set the mask value. ;

[0025] Step 3-3: If the two sides overlap and there are other feasible areas around the current position, set the mask value corresponding to the "turn around" action to 0; if the current position is a dead end, keep the mask value of the action at 1 to allow backtracking.

[0026] Furthermore, step 4 also includes calculating the anti-oscillation reward function:

[0027] Step 4-1, if the current location is detected Triggering the "reciprocating oscillation mode" will grant a pre-set first negative reward. ;

[0028] Step 4-2, if the current location is detected Triggering the "loop oscillation mode" grants a pre-set second negative reward. ;

[0029] Step 4-3: The negative reward directly applies to the value evaluation function of the network model, enabling it to identify and avoid oscillating states at the value evaluation level.

[0030] Furthermore, the trajectory repeatability index calculation and adaptive entropy adjustment method in step 5 include:

[0031] Step 5-1: Maintain a fixed-length historical position queue. ;

[0032] Step 5-2, detect the current position Is it in The data presents a repeating pattern of ABA or ABCA, where A, B, and C represent different position nodes in the historical queue.

[0033] Step 5-3, Define the oscillation state identifier If a repeating pattern is detected, the system is determined to be in an oscillation state. The entropy regularization coefficient Adjust to high value This enhances the strategy's random exploration capabilities; no repeating patterns were detected, indicating a normal state. The entropy regularization coefficient Adjust to a low value This is to ensure the stability of strategy convergence.

[0034] Furthermore, the entropy regularization coefficient in step S6 Defined as:

[0035] in, As an indicator variable for oscillations, For a high exploration coefficient, It has a low convergence coefficient, and ;

[0036] Variable entropy objective function The structure is as follows:

[0037] ; in For time steps Expectations The objective function for shearing is... These are the loss weighting coefficients for the value function, balancing the strategy loss and the value loss. For the value function loss, This is the dynamic entropy regularization coefficient. The policy entropy represents the policy in the state. Randomness under [the following conditions].

[0038] This invention also provides a robust path planning simulation system for mobile robots based on spatiotemporal attention and adaptive entropy adjustment, comprising:

[0039] The perception and state construction module is used to perform local observation state acquisition tasks, acquire obstacle distribution information and path guidance information generated by the global planning algorithm within the local field of view of the mobile robot, and construct a high-dimensional state tensor containing spatiotemporal features.

[0040] The topology constraint generation module is used to monitor the robot's current position and previous position in real time, generate a topology constraint mask, and filter out candidate actions in the action space that may cause deadlock or invalid loops based on physical space connectivity.

[0041] The deep decision reasoning module is internally deployed with a Recurrent-CBAM-PPO policy network, which is used to receive the high-dimensional state tensor, extract features using a convolutional neural network, enhance the weight of key features through an attention mechanism module, process temporal memory using a gated recurrent unit, and output the action probability distribution at the current moment in combination with the topological constraint mask to control the robot's movement.

[0042] The adaptive optimization and training module is used to record historical location trajectories and calculate trajectory repetition index. Based on this index, the entropy regularization coefficient in the PPO algorithm is dynamically adjusted, thereby dynamically switching between policy exploration and policy exploitation. The adjusted entropy regularization coefficient is used to construct a loss function to iteratively update the network parameters.

[0043] The visualization simulation interaction module provides a graphical user interface and includes: a global situation rendering unit for real-time rendering of the global map background, the robot's current pose, and simulated LiDAR scanning points; and a decision logic monitoring unit for real-time display of the masking status of the topological constraint mask for actions in each direction, and plotting the waveform curve of the entropy regularization coefficient over time to intuitively demonstrate the system's decision security and strategy stability.

[0044] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the aforementioned robust path planning method for mobile robots based on spatiotemporal attention and adaptive entropy adjustment.

[0045] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned robust path planning method for mobile robots based on spatiotemporal attention and adaptive entropy adjustment.

[0046] Compared with the prior art, the significant advantages of the present invention are:

[0047] (1) Enhanced perception layer: The RCPPO architecture is adopted. The CBAM attention mechanism is introduced after the convolutional neural network. The feature map useful for obstacle avoidance is selected by channel attention, and the coordinates of the obstacle are located by spatial attention. At the same time, the GRU is used to remember the historical state information, so that the robot has short-term memory and can recognize its own movement trend.

[0048] (2) Decision-making constraints: Design a topological constraint mask. Before the action probability is output, not only static obstacles are masked, but the "turnaround" action is also dynamically masked in combination with the previous position. That is: if the robot has just moved from position A to position B, when making a decision at position B, it is forcibly prohibited to immediately return to position A (unless position B is a dead end). The simplest ABA oscillation path is cut off from the physical rules.

[0049] (3) Algorithm-level Adaptation: An adaptive entropy adjustment mechanism based on oscillation detection is proposed. In complex circular obstacles, the robot may form a large loop oscillation of A->B->C->A. Therefore, the position history queue is maintained in real time during the training cycle. Once the current position is detected to be repeated with the historical position, the entropy coefficient in the PPO loss function is immediately increased. The entropy coefficient determines the randomness of the strategy. Increasing the entropy coefficient forces it to try other actions with lower probability in this state, thereby breaking the equilibrium state of the dead loop. When the oscillation disappears, the entropy coefficient is automatically decayed, returning to a stable utilization stage. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the method of the present invention.

[0051] Figure 2 This is a flowchart of the overall algorithm of the method of the present invention.

[0052] Figure 3 The diagram shows the RCPPO network structure, illustrating the connection between CNN, CBAM, and GRU.

[0053] Figure 4 This is a schematic diagram of an adaptive entropy adjustment mechanism, illustrating oscillation detection and the entropy coefficient. The dynamic relationship of change.

[0054] Figure 5 This is a schematic diagram illustrating the logic for generating topological constraint masks.

[0055] Figure 6 This is a schematic diagram of the interactive interface of the mobile robot anti-vibration path planning simulation system constructed in this invention. Detailed Implementation

[0056] This invention proposes a robust path planning method and system for mobile robots based on spatiotemporal attention and adaptive entropy adjustment. Addressing the problem of deep reinforcement learning easily getting trapped in local optima and trajectory deadlock under local observation, this invention constructs a Recurrent-CBAM-PPO (RCPPO) network model to extract spatiotemporal features of the environment. In the decision-making stage, a topology-constrained action filtering mechanism is designed to physically filter invalid loop actions that violate spatial connectivity by comparing historical trajectories in real time. In the training stage, a stability-aware adaptive entropy adjustment mechanism is constructed to dynamically adjust the entropy coefficient in the PPO loss function by monitoring trajectory repetition, automatically enhancing random exploration to assist escape when a stagnant trend is detected. This invention can significantly improve the decision robustness of mobile robots in complex dynamic environments, ensuring navigation smoothness and success rate. Furthermore, this invention also constructs a supporting visualization simulation system, which realizes transparent monitoring and verification of the algorithm's internal logic by rendering the decision safety boundary and entropy adjustment curve in real time.

[0057] Combination Figure 1 The method includes the following steps:

[0058] The local observation state of the mobile robot is obtained, which includes obstacle distribution information within the local field of view and path guidance information generated by the global planning algorithm.

[0059] A deep reinforcement learning policy network containing a spatiotemporal feature extraction module is constructed, wherein the module is a convolutional neural network, an attention mechanism module and a gated recurrent unit connected in series;

[0060] Based on the robot's current position and previous position, a topological constraint mask is generated. The mask is based on physical space connectivity and filters out candidate actions in the action space that could lead to deadlock or invalid loops.

[0061] The local observation state is input into the network model, and the action probability distribution at the current moment is output by combining the topological constraint mask. The action is sampled and executed, and the state transition data and the historical trajectory of the position are recorded.

[0062] The trajectory repetition index is calculated based on the location history trajectory, and the entropy regularization coefficient in the near-end policy optimization (PPO) algorithm is dynamically adjusted according to the index to balance the exploration and utilization capabilities of the policy network.

[0063] The loss function is constructed using the adjusted entropy regularization coefficient, and the network model parameters are iteratively updated until the model converges.

[0064] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0065] like Figures 1-5 As shown, a robust path planning method for mobile robots based on spatiotemporal attention and adaptive entropy adjustment is proposed.

[0066] Step 1: Environment and State Definition. Construct a gridded environment and state space. It consists of three parts:

[0067]

[0068] in This is a local obstacle image. This is a global path guidance graph. This represents the hidden state of GRU at the previous moment.

[0069] Step 2: Topology constraint mask generation at each time step Obtain all candidate actions of the robot. ;

[0070] These correspond to movement up, down, left, and right respectively. Generate a mask vector. :initialization If the action This causes the person to hit the wall. .

[0071] If action The target position is equal to (Previous position), and current position There are other feasible solutions in the surrounding area (i.e.) ),but This step ensures that the robot will not frequently turn around due to minor fluctuations in the policy network.

[0072] Step 3: Network forward propagation to transfer state Input to the network, CNN extracts spatial features .

[0073] CBAM module processing: Enhance key features.

[0074] GRU processing: .

[0075] The Actor outputs the raw Logits, plus a mask penalty: .

[0076] Feature maps extracted by convolutional neural networks; This is the channel attention weight matrix; This is the spatial attention weight matrix; Let be the hidden state vector at time t; For GRU output features; This is the original action logits vector output by the Actor network; This is the logits vector after masking penalty correction;

[0077] Step 4: Adaptive Entropy Adjustment and Parameter Update While collecting data, maintain the queue. Record recent Step coordinates. Calculate the oscillation indicator: if If this occurs, an oscillation is determined. During the PPO update phase, an entropy coefficient is set. If the proportion of oscillating samples in the current batch exceeds the threshold, set... (High exploration); otherwise set (High utilization). By maximizing the objective function Update network weights.

[0078] in Indicates length is A historical position queue used to store recent The robot's coordinates for each step; Let be the robot's position coordinates at time t; This is the entropy regularization coefficient, which controls the intensity of policy randomness. The objective function for shearing; Indicates the policy in the state Randomness under [the following conditions].

[0079] To more accurately describe the technical solution of this invention, the core algorithm mathematical model proposed in this invention is defined in detail below. These models correspond to the network structure in the aforementioned steps, are used to realize the end-to-end mapping from local observation to action decision, and support the mathematical expression of the adaptive entropy adjustment mechanism.

[0080] (1) Memory encoding model with hybrid spatiotemporal features

[0081] Unlike traditional independent frame processing, this invention defines a temporal memory coding function. ,set up The local observation image at time is The memory state of the previous moment was First, through convolution operations... Basic features are extracted, and then a feature reweighting operator is introduced. :

[0082]

[0083] For CBAM attention mechanism operators, including channel and spatial attention; It is a multilayer perceptron; For feature map Global average pooling; For feature map Global max pooling; For activation functions; This is a convolution operation;

[0084] Although the above formula is based on an attention mechanism, this invention uses it as a state preprocessing operator. Subsequently, a gated loop unit is used. Update the path memory tensor at the current time step. :

[0085]

[0086] For flattening operation; Not only as the basis for generating the current action, but also as Prior knowledge of time is transmitted, thus establishing Markov dependencies between local observations at the mathematical level.

[0087] (2) Dynamic mask generation function based on topological constraints

[0088] This is the key difference between this invention and existing technologies. The robot's current position is defined as... The previous position is Define the feasible region discrimination function. Used to calculate position Number of non-obstacle neighboring nodes:

[0089]

[0090] in For action space; For action The corresponding displacement vector; This is an environment query function that returns the location. Map status; This is an indicator function that returns 1 if the condition is true, and 0 otherwise; this invention constructs an adaptive inverse suppression mask. Its definition is as follows:

[0091]

[0092] This formula precisely describes the anti-oscillation logic: the mask is only set to 0 when the system "hit a wall" or "attempts to turn around and is not currently in a dead end." The final policy output distribution... Revised to:

[0093]

[0094] in The raw logits values ​​output by the policy network; For a very small penalty constant (e.g.) ); This is a topology constraint mask; a value of 1 indicates that the action is allowed, and 0 indicates that it is prohibited.

[0095] (3) Oscillation-driven variable entropy objective function

[0096] To address loop deadlock, this invention proposes a loss function induced by oscillating potential energy. A historical trajectory set is defined. Define an oscillation indicator variable. :

[0097]

[0098] This is an indicator function that returns 1 if the condition is true, and 0 otherwise.

[0099] based on Construct the entropy regularization coefficient :

[0100]

[0101] in (For example The final policy optimization objective function. Defined as:

[0102]

[0103] Physical meaning: When the system detects trajectory overlap ( )hour, The jump to a higher position makes the objective function... Pay more attention to policy entropy By maximizing the gradient update direction, random noise is forcibly introduced to break the "overconfidence" of the policy network and achieve escape from local minima.

[0104] Based on the above method embodiments, this embodiment constructs a general-purpose intelligent planning simulation system for mobile robots. This system not only verifies the effectiveness of the aforementioned algorithm but also uses visualization to monitor the decision-making safety and strategy stability during the path planning process in real time, thereby verifying the effectiveness of the method described in claim 1 in complex dynamic environments.

[0105] The simulation system constructed in this invention runs on an electronic computing device. The system is logically divided into five core modules:

[0106] (1) Perception and State Construction Module: Responsible for map loading and tensor construction of local observation states;

[0107] (2) Topology constraint generation module: responsible for calculating topology constraint masks in real time to filter out deadlock actions;

[0108] (3) Deep decision reasoning module: responsible for running the Recurrent-CBAM-PPO network and outputting action probabilities;

[0109] (4) Adaptive optimization and training module: responsible for calculating trajectory repetition and dynamically adjusting the entropy coefficient;

[0110] (5) Visualization simulation interaction module: responsible for rendering the graphical user interface and human-computer interaction.

[0111] The system communicates with the underlying path planning algorithm library in real time through shared memory or data interface, and maps the state tensor, mask vector and entropy coefficient output by the algorithm to visual elements on the graphical user interface (GUI) in real time, realizing transparent mapping from the underlying logic to the top-level display.

[0112] like Figure 6 As shown, the simulation system's interactive interface is configured with three main functional areas, specifically including:

[0113] (1) Scene configuration and parameter control unit (corresponding to) Figure 6 The left-hand area is configured for system initialization settings, specifically including:

[0114] Map data interface: Configures the parsing of external input raster map files (such as .pgm or .yaml files built by SLAM) to construct the system's global static environment model;

[0115] Hyperparameter tuning interface: Configures deep reinforcement learning hyperparameters for receiving external input, including but not limited to learning rate, clipping range and entropy coefficient threshold, and transmits the parameters to the policy network in the background in real time to adjust the training or inference behavior of the network;

[0116] Layer Display Controller: Configures the overlay display status of the global cost map, local scrolling window, and LiDAR point cloud layers.

[0117] (2) Multi-source situational real-time rendering unit (corresponding to) Figure 6 The central area is the core display window of the system, configured to display the robot's motion state in a "virtual-real combination" manner.

[0118] Global Environment Layer: Renders a static distribution of obstacles against the background of the loaded raster map;

[0119] Subject pose identification: through specific graphic symbols (such as...) Figure 6 The arrow icon shown in the center maps the robot's global coordinates (x, y) and heading angle at the current time t in real time, displaying the robot's motion trajectory;

[0120] Local perception layer: A simulated LiDAR scan point or local obstacle distribution map is superimposed and displayed around the robot body. This layer corresponds to the "local field of view interception" process described in step 1 of the method of the present invention, and intuitively shows the perception range of the policy network.

[0121] (3) Algorithm internal state monitoring unit (corresponding to) Figure 6 The right-hand area (this unit) is configured to visualize the internal intermediate variables of deep reinforcement learning networks, specifically including:

[0122] Action Space Topology Constraint Monitor: A direction indicator control is located in the upper right corner of the interface. When the algorithm's topology constraint mechanism takes effect (i.e., actions that violate physical rules or cause deadlocks are filtered out), the corresponding indicator light will switch to the "blocked" state. This demonstrates that the system has the ability to perform logical safety checks before the neural network output, ensuring the physical feasibility of the planned actions.

[0123] Attention Mechanism Heatmap Subunit: Configured to map the spatial attention weight matrix in the Recurrent-CBAM module to a color heatmap. The variations in grid color intensity reflect the degree of attention the policy network pays to different obstacle regions in the local environment during the feature extraction stage.

[0124] Adaptive Entropy Coefficient Monitoring Subunit: Configured for plotting entropy regularization coefficients The waveform curve changes over time. When the system detects that the robot trajectory exhibits an ABA or loop oscillation mode according to step 5, the waveform curve will show the numerical jump of the entropy coefficient, indicating that the system is enhancing the stochastic exploration of the strategy to execute an escape action; when the oscillation is relieved, the curve falls back to the preset stable value.

[0125] After system startup, the environment configuration and access module loads the global map and calls the global path planning algorithm to generate a guiding path. During simulation, the deep decision-making inference engine receives local observation states and outputs action commands, while simultaneously transmitting its internal mask vector, attention weights, and entropy coefficients to the visualization and interactive rendering module in real time. If the algorithm state monitoring module detects an oscillation indicator, it triggers an alarm display in the adaptive entropy coefficient monitoring subunit until the robot leaves the oscillation area.

[0126] This invention avoids invalid reciprocating motions from a physical rule perspective through a topological constraint masking mechanism and solves the trajectory deadlock problem in complex environments at the algorithmic level using an adaptive entropy adjustment mechanism. In the model training and inference stages, the proposed robust path planning method and simulation system, which integrates spatiotemporal attention mechanisms, effectively extracts key environmental features and addresses the slow convergence speed of deep reinforcement learning under sparse rewards. Experimental results show that in highly dynamic environments with dense obstacles, this invention can significantly improve the navigation success rate and trajectory smoothness of mobile robots, greatly reducing uncertainty and redundant actions in the planning process, and has high engineering application value.

[0127] The above description is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. All components not explicitly stated in this embodiment can be implemented using existing technology.

Claims

1. A robust path planning method for mobile robots based on spatiotemporal attention and adaptive entropy adjustment, characterized in that, Includes the following steps: Step 1: Obtain the local observation state of the mobile robot. The local observation state includes obstacle distribution information within the local field of view and path guidance information generated by the global planning algorithm. Step 2: Construct a deep reinforcement learning policy network that includes a spatiotemporal feature extraction module, wherein the spatiotemporal feature extraction module is a convolutional neural network, an attention mechanism module, and a gated recurrent unit connected in series. Step 3: Generate a topological constraint mask based on the robot's current position and previous position. The topological constraint mask is based on physical space connectivity and filters out candidate actions in the action space that could lead to deadlock or invalid loops. Step 4: Input the local observation state into the network model, combine it with the topological constraint mask to output the action probability distribution at the current moment, sample and execute the action, and record the state transition data and the historical trajectory of the position. Step 5: Calculate the trajectory repetition index based on the location history trajectory, and dynamically adjust the entropy regularization coefficient in the near-end policy optimization algorithm according to the index to balance the exploration and utilization capabilities of the policy network. Step 6: Construct a loss function using the adjusted entropy regularization coefficient, and iteratively update the network model parameters until the model converges.

2. The method according to claim 1, characterized in that, The specific process of obtaining the local observation state in step 1 includes: Step 1-1, Global Guidance Generation: Generate a global path matrix from the starting point to the target point on a static global map using a global path planning algorithm; Step 1-2, Local view capture: Centered on the robot's current position, capture a local area of ​​the corresponding size from the global path matrix as a path guidance layer, and at the same time obtain a local obstacle distribution layer of the same size; Steps 1-3, Hybrid State Construction: Using frame stacking technology, the current path guidance layer, obstacle distribution layer, and historical observation frames are spliced ​​together in the channel dimension to form a high-dimensional state tensor, which is then combined with the hidden state vector from the previous time step as the final input state.

3. The method according to claim 1, characterized in that, The network model in step 2 specifically includes: Feature extraction layer: A convolutional neural network is used to process local observation states; Attention Enhancement Layer: Introduces a feature reweighting operator and embeds a CBAM module after the convolutional layer. Key feature channels are weighted sequentially through the channel attention submodule and the spatial attention submodule focuses on the spatial position of obstacles and paths. Temporal memory layer: Utilizes GRU units to receive flattened feature vectors and the hidden state of the previous time step, outputs feature codes containing temporal memory, and establishes Markov dependencies between local observations.

4. The method according to claim 1, characterized in that, The specific logic for generating the topological constraint mask in step 3 is as follows: The number of non-obstacle neighbor nodes around the current position is calculated using the feasible region discrimination function, and a mask vector of all 1s is initialized. Detection action If the target location is a static obstacle or exceeds the map boundary, set the mask value. ; If the two locations overlap and there are other feasible areas around the current location, the mask value corresponding to the "turn around" action is set to 0; if the current location is a dead end, the mask value of the action is kept at 1 to allow for backtracking.

5. The method according to claim 1, characterized in that, Step 4 also includes calculating the anti-oscillation reward function: If the current location is detected Triggering the "reciprocating oscillation mode" will grant the first preset negative reward. ; If the current location is detected Triggering the "loop oscillation mode" grants a pre-set second negative reward. ; The negative reward directly affects the value evaluation function of the network model, enabling it to identify and avoid oscillating states at the value evaluation level.

6. The method according to claim 1, characterized in that, Step 5 includes the calculation of trajectory repeatability index and adaptive entropy adjustment method, which includes: Step 5-1: Maintain a fixed-length historical position queue. ; Step 5-2, detect the current position Is it in The data presents a repeating pattern of ABA or ABCA, where A, B, and C represent different position nodes in the historical queue. Step 5-3, Define the oscillation state identifier If a repeating pattern is detected, the system is determined to be in an oscillation state. The entropy regularization coefficient Adjust to high value This enhances the strategy's random exploration capabilities; no repeating patterns were detected, indicating a normal state. The entropy regularization coefficient Adjust to a low value This is to ensure the stability of strategy convergence.

7. The method according to claim 1, characterized in that, Entropy regularization coefficient in step 6 Defined as: ； in, As an indicator variable for oscillations, For a high exploration coefficient, It has a low convergence coefficient, and ; Variable entropy objective function The structure is as follows: ； in For time steps Expectations The objective function for shearing is... These are the loss weighting coefficients for the value function, balancing the strategy loss and the value loss. For the value function loss, This is the dynamic entropy regularization coefficient. The policy entropy represents the policy in the state. Randomness under [the following conditions].

8. A robust path planning system for a mobile robot based on spatiotemporal attention and adaptive entropy adjustment, used to implement the method described in any one of claims 1 to 7, characterized in that the system... include: The perception and state construction module is used to perform local observation state acquisition tasks, acquire obstacle distribution information and path guidance information generated by the global planning algorithm within the local field of view of the mobile robot, and construct a high-dimensional state tensor containing spatiotemporal features. The topology constraint generation module is used to monitor the robot's current position and previous position in real time, generate a topology constraint mask, and filter out candidate actions in the action space that may cause deadlock or invalid loops based on physical space connectivity. The deep decision reasoning module is internally deployed with a Recurrent-CBAM-PPO policy network, which is used to receive the high-dimensional state tensor, extract features using a convolutional neural network, enhance the weight of key features through an attention mechanism module, process temporal memory using a gated recurrent unit, and output the action probability distribution at the current moment in combination with the topological constraint mask to control the robot's movement. The adaptive optimization and training module is used to record historical location trajectories and calculate trajectory repetition index. Based on this index, the entropy regularization coefficient in the PPO algorithm is dynamically adjusted, thereby dynamically switching between policy exploration and policy exploitation. The adjusted entropy regularization coefficient is used to construct a loss function to iteratively update the network parameters. The visualization simulation interaction module provides a graphical user interface and includes: a global situation rendering unit for real-time rendering of the global map background, the robot's current pose, and simulated LiDAR scanning points; and a decision logic monitoring unit for real-time display of the masking status of the topological constraint mask for actions in each direction, and plotting the waveform curve of the entropy regularization coefficient over time to intuitively demonstrate the system's decision security and strategy stability.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the robust path planning method for mobile robots based on spatiotemporal attention and adaptive entropy adjustment as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the robust path planning method for mobile robots based on spatiotemporal attention and adaptive entropy adjustment as described in any one of claims 1 to 7.

Images