Risk awareness path planning method and system based on time sequence prediction

Abstract

The invention relates to the technical field of path planning, in particular to a risk awareness path planning method and system based on time sequence prediction, and the method comprises the steps: predicting the passability probability of a dynamic node at a future moment, calculating the predictive risk cost of the node, and adding the predictive risk cost into a cost function of an A * algorithm, thereby enabling a robot to actively avoid a high-risk region, and improving the path planning efficiency. According to the method, the path planning is prospective, the path which is failed during arrival is prevented from being planned, the path planning can be quantitatively balanced between the path length and the blocking risk, the optimal path selection is realized, and the efficiency and the quality of the path planning are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of path planning, in particular to a risk perception path planning method and system based on time sequence prediction. BACKGROUND

[0002] With the wide application of mobile robots in service, logistics, manufacturing and other fields, their autonomous navigation ability in dynamic environments has become a core technical challenge. Obstacles in dynamic environments, such as automatic doors, elevators, and moving people, have time-varying characteristics, and their state changes directly affect the path passability of the robot.

[0003] Existing path planning methods mainly include the following three types:

[0004] 1. Static path planning method based on A* algorithm:

[0005] A* algorithm is a classical heuristic search algorithm, which finds the optimal path in static environment through cost function . Wherein represents the actual cost from the starting point to node , represents the heuristic estimated cost from node to the end point. This method performs excellently in static environment, but in dynamic environment, due to the inability to consider the state change of obstacles at future time, the robot often finds that the path is blocked when it reaches the dynamic node.

[0006] 2. Responsive re-planning method based on D*Lite:

[0007] D*Lite algorithm is an improvement of D* algorithm, which can efficiently repair the path after detecting obstacles. Its core idea is: when the robot finds obstacles while executing the path, immediately recalculate the optimal path from the current position to the target. Although this method improves the robustness in dynamic environment, it is essentially a passive response: only when the obstacle actually appears, the adjustment is made, which cannot avoid risks in advance.

[0008] 3. Deterministic planning method based on space-time A*:

[0009] Spatio-Temporal A algorithm incorporates time dimension into search space, searching for path in three-dimensional space . This method assumes that the future state of dynamic obstacles is known, and can find the theoretically optimal space-time path. However, this method has two key limitations: (1) it needs to know the complete schedule of all dynamic obstacles in advance, which is difficult to obtain in real environment; (2) it regards dynamic obstacles as deterministic entities, ignoring the uncertainty and prediction error in actual environment.

[0010] In general, the existing path planning methods have the following problems:

[0011] 1. Lack of foresight: the traditional A* algorithm only considers the current environmental state, and cannot predict the passability of dynamic nodes at future time, leading to the planned path being invalid when reaching the dynamic node.

[0012] 2. Passive response mechanism: D*Lite and other responsive algorithms can only be adjusted passively after the obstacle appears, and cannot actively avoid potential risks, leading to frequent waiting or inefficient detouring of the robot; and frequent path failure leads to a large amount of computing resources wasted on re-planning, resulting in increased cost of path planning.

[0013] 3. Deterministic assumption is too idealistic: the space-time A* algorithm assumes that the future state is completely known, but the state of dynamic obstacles in the real environment has uncertainty, leading to reduced quality of path planning.

[0014] 4. Lack of risk quantification mechanism: existing methods cannot quantify the blocking risk of the path at future time, and cannot optimize between path length and risk. SUMMARY

[0015] To this end, the technical problem to be solved by the present application is to overcome the problems of short-sightedness, passive response, and lack of risk quantification in the prior art.

[0016] To solve the above technical problems, the present application provides a risk perception path planning method based on time sequence prediction, comprising: using a large visual-language model to convert the image frames of dynamic nodes in the environment into time sequence state sequences; constructing a state prediction model for each dynamic node, which is used to obtain the passable probability of the dynamic node at the predicted time based on the historical time sequence state sequence of the dynamic node; and constructing a time sequence prediction topology graph at the predicted time based on the passable probability of all dynamic nodes at the predicted time; calculating the predicted time of the robot reaching the candidate node; if the candidate node is a dynamic node, obtaining the passable probability of the dynamic node at the predicted time, and calculating the predictive risk cost of the dynamic node at the predicted time based on the passable probability; if the candidate node is a static node, obtaining the predictive risk cost of the dynamic nodes in the neighborhood of the static node at the predicted time based on the time sequence prediction topology graph at the predicted time, and propagating it to the static node to obtain the predictive risk cost of the static node at the predicted time; using a risk perception A* algorithm for path planning, including: in the process of node expansion, constructing an expansion cost function based on the predictive risk cost of the candidate node at the predicted time; selecting a target node in the candidate node with the expansion cost function to achieve path planning.

[0017] Preferably, the image frames of dynamic nodes in the environment are converted into a temporal state sequence using a large-scale vision-language model, the steps of which include: The image frames and structured text prompts collected at each time step of the dynamic node are input into a large-scale vision-language model to obtain the semantic state vector of the dynamic node at each time step. The semantic state vector of a dynamic node at each time step is mapped to a binary state vector, which is used to characterize whether the dynamic node is passable. The binary state vector of the time step and the dynamic node at that time step are combined into sample pairs, and the temporal state sequence is constructed according to the time sequence.

[0018] Preferably, the state prediction model employs a long short-term memory network.

[0019] Preferably, the predictive risk cost of the dynamic node at the prediction time is calculated based on the passability probability, using the following method: The blocking risk and prediction uncertainty of a dynamic node at the prediction time are calculated based on the passability probability of the dynamic node at the prediction time. Based on the blocking risk and prediction uncertainty of the dynamic node at the prediction time, the predictive risk cost of the dynamic node at the prediction time is calculated using the following formula: ; in, For dynamic nodes At the predicted time The cost of predictive risk The predicted time for the robot to reach the dynamic node. For dynamic nodes, For dynamic nodes At the predicted time The risk of congestion For indexes of dynamic nodes. For dynamic nodes At the predicted time The uncertainty of prediction This represents the uncertainty penalty coefficient.

[0020] Preferably, the congestion risk of a dynamic node at the predicted time is calculated based on the passability probability of the dynamic node at the predicted time, and the calculation formula is as follows: ; in, For dynamic nodes At the predicted time The risk of congestion For indexes of dynamic nodes. The predicted time for the robot to reach the dynamic node. For dynamic nodes The weighting coefficients, For dynamic nodes The probability of passage at the predicted time; The prediction uncertainty of a dynamic node at the prediction time is calculated based on the walkability probability of the dynamic node at the prediction time. The calculation formula is as follows: ; in, For dynamic nodes At the predicted time The uncertainty of prediction For dynamic nodes The probability of passage at the predicted time.

[0021] Preferably, the predictive risk cost of dynamic nodes within the neighborhood of a static node at the prediction time is obtained, and then propagated to the static node to obtain the predictive risk cost of that static node at the prediction time. The calculation formula is as follows: ; in, Static node At the predicted time The cost of predictive risk For static nodes, Static node Dynamic nodes within the neighborhood, For dynamic nodes At the predicted time The cost of predictive risk Static node To dynamic nodes The Euclidean distance between them This is the spatial attenuation coefficient.

[0022] Preferably, an extended cost function is constructed based on the predictive risk cost of a node at the prediction time, expressed as follows: ; in, Let n be the expansion cost function for node n. Let n be the path length from the starting point to node n. For heuristic estimation from node n to the endpoint, For node n at the prediction time The cost of predictive risk This is the risk weighting coefficient.

[0023] Preferably, during the node expansion process, before selecting the target node from the candidate nodes using the expansion cost function, if the candidate node is a dynamic node, the process further includes: If the passable probability of the dynamic node at the predicted time is less than the confidence threshold, the dynamic node is removed from the candidate node queue.

[0024] Preferably, further comprising: if it is detected that the current node is impassable during execution of the path, triggering a re-planning decision mechanism, the steps comprising: Based on the state prediction model, the waiting time required for the current node to change to a passable state is obtained; and based on the waiting time, the waiting cost is calculated; Based on the calculation overhead of the path planning, the cost difference between the new path cost of the re-planned path from the current node and the remaining path cost of the original path, the re-planning cost is calculated; If the waiting cost is less than the re-planning cost, the robot waits until the current node changes to a passable state, and continues to execute the path according to the original path; If the waiting cost is greater than or equal to the re-planning cost, the risk-aware A* algorithm is used to re-plan the path with the current node as the starting point.

[0025] The application also provides a risk-aware path planning system based on time sequence prediction, comprising: A semantic perception module is used to convert the image frames of the dynamic nodes in the environment into time sequence state sequences using a large visual-language model; A time sequence prediction module is used to construct a state prediction model for each dynamic node, which is used to obtain the passable probability of the dynamic node at the predicted time based on the historical time sequence state sequence of the dynamic node; and based on the passable probability of all dynamic nodes at the predicted time, a time sequence prediction topology graph at the predicted time is constructed; A risk cost calculation module is used to calculate the predicted time when the robot reaches the candidate node; if the candidate node is a dynamic node, the passable probability of the dynamic node at the predicted time is obtained, and the predictive risk cost of the dynamic node at the predicted time is calculated based on the passable probability; if the candidate node is a static node, the predictive risk cost of the dynamic nodes in the neighborhood of the static node at the predicted time is obtained based on the time sequence prediction topology graph at the predicted time, and is propagated to the static node to obtain the predictive risk cost of the static node at the predicted time; A path planning module is used to plan the path using the risk-aware A* algorithm, comprising: during node expansion, an expansion cost function is constructed based on the predictive risk cost of the candidate node at the predicted time; and a target node is selected from the candidate nodes using the expansion cost function to realize path planning.

[0026] The above technical solutions of the application have the following beneficial effects compared with the prior art:

[0027] The present invention discloses a risk-aware path planning method based on time-series prediction. By predicting the accessibility probability of dynamic nodes at future times, calculating the predictive risk cost of nodes, and incorporating it into the cost function of the A* algorithm, the robot can actively avoid high-risk areas. This not only makes path planning more forward-looking, avoiding the planning of paths that are already invalid upon arrival, but also enables path planning to quantitatively weigh path length and congestion risk, achieving better path selection and improving the efficiency and quality of path planning.

[0028] Furthermore, this invention introduces prediction uncertainty into the calculation of predictive risk cost, quantifies prediction confidence through entropy value, and penalizes low-confidence predictions to improve the robustness of path planning.

[0029] Furthermore, this invention performs node pruning based on the prediction confidence of dynamic nodes. Before calculating the next node, nodes with low predicted passability are removed from the candidate nodes to avoid searching high-risk path branches and improve search efficiency.

[0030] Furthermore, based on the probability of the traversability of dynamic nodes in the future, this invention proposes an adaptive path replanning decision-making mechanism to calculate and measure waiting costs and replanning costs, thereby achieving intelligent decision-making, avoiding blind waiting or unnecessary replanning by the robot, reducing the computational overhead of path replanning, and improving the efficiency of path planning. Attached Figure Description

[0031] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This is a flowchart of a risk perception path planning method based on time-series prediction according to the present invention; Figure 2 This is an overall framework diagram of the method of the present invention; Figure 3 This is a schematic diagram of the construction of a time series prediction topology graph, in which... Figure 3 (a) in the image is a static topological map. Figure 3 (b) in the figure is the topological map after probabilistic embedding. Figure 3 (c) in the diagram represents the time-series prediction topology. Figure 4 It is a visual breakdown of the costs of predictive risk; Figure 5 This is a comparison diagram of path planning between the method of this invention and the traditional A* algorithm; Figure 6 This is a comparison chart of key performance indicators (KPIs) for different path planning methods. Figure 6 (a) in the figure is a comparison chart of the number of times the target is reached by different path planning methods. Figure 6(b) in the figure is a comparison of the total time taken by different path planning methods. Figure 6 (c) in the figure is a comparison chart of the total path length of different path planning methods. Figure 6 (d) in the figure is a comparison of the average time taken by different path planning methods. Figure 6 (e) in the figure is a comparison chart of the average path length of different path planning methods. Figure 6 (f) in the figure is a comparison of the number of replanning steps for different path planning methods. Detailed Implementation

[0032] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0033] Reference Figure 1 and Figure 2 As shown, this invention provides a risk perception path planning method based on time-series prediction, comprising: S1: Use a large-scale vision-language model to replace image frames of dynamic nodes in the environment with temporal state sequences; S2: Construct a state prediction model for each dynamic node to obtain the passability probability of the dynamic node at the prediction time based on the historical time-series state sequence of the dynamic node; construct the time-series prediction topology map at the prediction time based on the passability probabilities of all dynamic nodes at the prediction time. S3: Calculate the predicted time when the robot will arrive at the candidate node; if the candidate node is a dynamic node, obtain the passability probability of the dynamic node at the predicted time, and calculate the predictive risk cost of the dynamic node at the predicted time based on the passability probability; if the candidate node is a static node, obtain the predictive risk cost of the dynamic nodes in the neighborhood of the static node at the predicted time based on the temporal prediction topology graph at the predicted time, and propagate it to the static node to obtain the predictive risk cost of the static node at the predicted time. S4: Path planning is performed using the risk-aware A* algorithm, including: during the node expansion process, constructing an extended cost function based on the predictive risk cost of candidate nodes at the prediction time; and selecting the target node from the candidate nodes using the extended cost function to achieve path planning.

[0034] Since traditional sensors (such as LiDAR and depth cameras) can only provide geometric obstacle information and cannot understand the semantic state of dynamic obstacles (such as the opening and closing state of a door), this invention S1 uses a large-scale vision-language model LLaVA as the perception front end to solve the problem of insufficient semantic understanding.

[0035] The robot's operating environment includes dynamic nodes and static nodes. Dynamic nodes correspond to physical locations with passage status that changes over time (such as doors and elevator entrances), while static nodes represent ordinary passable spatial points.

[0036] Specifically, the steps of S1 include: S11: Input the image frames and structured text prompts collected at each time step of the dynamic node into the large-scale vision-language model to obtain the semantic state vector of the dynamic node at each time step.

[0037] For time Image frames of captured dynamic nodes This embodiment features structured text prompts. This is used to guide large-scale vision-language models in recognizing and reporting the states of dynamic entities, obtaining the semantic state vector of the dynamic node at each time step. This process can be formalized as: ; in, It is a semantic state vector. For large-scale vision-language models, For image frames, This is a structured text prompt, where t is the time index.

[0038] For example, for a gate node, the hint is... The design is as follows: "Analyze the gate status in the image and return JSON format: {"status":"open / closed / half_open", "confidence":0-100}.

[0039] LLaVA's response parser converts natural language output into structured semantic state vectors. For a gate node, the semantic state vector includes: classification state. and model confidence For example, the output of the parsed model is This indicates that the door is in a half-open state, with a confidence level of 90%.

[0040] S12: Map the semantic state vector of the dynamic node at each time step to a binary state vector to characterize whether the state node is passable.

[0041] Specifically, mapping semantic states to binary décor labels includes: for states ,like Then the binary state vector (Passable); If ,but (Not passable).

[0042] S13: Combine the binary state vector of the time step and the dynamic node at that time step into a sample pair, and construct a temporal state sequence according to the time order, specifically represented as follows: ; in, It is a time-series state sequence. For a moment The semantic state vector, where n is the length of the temporal state sequence.

[0043] For each dynamic node, a historical time-series state sequence of that dynamic node is formed by extracting a subsequence from its corresponding time-series state sequence.

[0044] This invention uses the LLaVA model to overcome the limitations of semantic understanding in traditional sensors, converting unstructured visual data into machine-readable structured time-series data, providing high-quality input for subsequent time-series prediction.

[0045] Preferably, in S2, the state prediction model uses a Long Short-Term Memory (LSTM) network.

[0046] For dynamic nodes Its complete historical sequence is ,in Indicates the dynamic node at time... The passable status. , For dynamic nodes The length of the historical sequence.

[0047] This embodiment uses the sliding window method to construct the historical time-series state sequence: according to the length of the historical time-series state sequence. (e.g., 50 time steps) The complete historical sequence is segmented by sliding the window to form training sample pairs. ;in, The historical time series state sequence at time t is the input to the state prediction model; For a moment The true state, i.e., the true label. The state prediction model is used to predict the time of input from a historical time-series state sequence at time t. state, For the prediction time window.

[0048] Specifically, the present invention provides for each dynamic node. Train an independent LSTM state prediction model The network structure of the state prediction model includes: Input layer: receiving The training sample sequence consists of states at each time step, with an input dimension of 1 (binary state). Two stacked LSTM layers: each layer contains 64 hidden units, and learns long and short-term temporal dependencies through a gating mechanism; Fully connected layer: maps the hidden state of the last layer of the LSTM to a single output; Sigmoid activation layer: Output time probability of passage .

[0049] The core update process of LSTM is described by the following set of equations: ; in, For a moment Input status, , , and Both are weight matrices. , , and Both are bias vectors. It is the Sigmoid activation function. This represents element-wise multiplication. , and These are the forget gate, input gate, and output gate, respectively. and Let be the hidden states at time t and time t-1, respectively. and These represent the cell states at time t and time t-1, respectively. This represents candidate information for time t.

[0050] The loss function used during training of the state prediction model is binary cross-entropy (BCE), and the formula is as follows: ; in, The number of training samples, For real labels, This represents the predicted probability output by the model.

[0051] The Adam optimizer was used during the training phase, with a learning rate of 0.001, a batch size of 64, and 50 training epochs. The data was divided into two parts: the first 80% of the data was used as the training set, and the last 20% was used as the test set.

[0052] Once the state prediction model is trained, it can predict the probability of passage at several future time steps based on the historical time-series state sequence at the current moment. For future moments... Dynamic nodes In the future Probability of passage: ; in, and For the weights and biases of the output layer, This represents the final hidden state of the LSTM.

[0053] Integrate the predicted mobility probabilities into a dynamic temporal predictive topological map (TPTM). Among them It is a set of nodes, including dynamic nodes and static nodes; Let be the set of edges, representing the walkable connections between nodes. When a path passes through a dynamic node, the validity and cost of the edge are determined by the walkable probability of the dynamic node at a future time, as output by the state prediction model, thereby realizing risk-aware path planning based on time-series prediction. For each node at a future time The set of predicted probabilities. Time series prediction topology graph. This enables traditional static topology maps to have time-series prediction capabilities, providing forward-looking information for path planning. Figure 3 This is a schematic diagram of the construction of a time series prediction topology graph, in which... Figure 3 (a) in the image is a static topological map. Figure 3 (b) in the figure is the topological map after probabilistic embedding. Figure 3 (c) in the diagram represents the time-series prediction topology.

[0054] This step extends the traditional static topology map into a predictive temporal topology map. By using an LSTM network to learn the inherent temporal patterns of each dynamic node, it achieves probabilistic predictions of future drivability, providing a forward-looking risk assessment basis for path planning. The temporal predictive topology map serves as the core input data for risk perception calculations in subsequent path planning. Specifically, it is used to calculate the predicted time when the robot will arrive at candidate dynamic nodes. Then, the planner obtains the predicted passability probability of the corresponding dynamic nodes at future times by querying the time-series prediction topology graph. And calculate the predictive risk cost accordingly. This risk cost is further used in the A* extended cost function and a confidence-based node pruning strategy to achieve forward-looking path planning. Furthermore, during the execution phase, when predictions differ from reality, the time-series forecast topology graph provides future traversable trends for comparing waiting costs and replanning costs, enabling intelligent decision-making.

[0055] S3, based on the passability probability obtained from S2, transforms this prediction result into a quantifiable risk cost, which is then incorporated into the cost function of path planning. Specific steps include:

[0056] S31: Calculate the predicted time when the robot will arrive at the candidate node.

[0057] For candidate nodes in path planning First, calculate the predicted time when the robot is expected to arrive at the node. .

[0058] Assuming the robot starts from the starting point Departure, current time is The robot's speed is Then the robot reaches the candidate node. Predicted time for: ; in, Indicates from node arrive Path length, From the starting point to the candidate node The number of path segments.

[0059] S32: If the candidate node is a dynamic node, obtain the passability probability of the dynamic node at the prediction time.

[0060] If candidate nodes For dynamic nodes Then query the time-series prediction topology graph to obtain the dynamic node. At any moment probability of passage .

[0061] S33: Calculate the blocking risk and prediction uncertainty of a dynamic node at the prediction time based on the passability probability of the dynamic node at the prediction time.

[0062] The congestion risk of a dynamic node at the predicted time is calculated based on its passability probability at the predicted time. The calculation formula is as follows: ; in, For dynamic nodes At the predicted time The risk of congestion For indexes of dynamic nodes. The predicted time for the robot to reach the dynamic node. For dynamic nodes The weight coefficient reflects the importance of the dynamic node (e.g., the gate node of a critical channel has a larger weight). For dynamic nodes The probability of passage at the predicted time.

[0063] To address the uncertainty in prediction, this invention introduces a prediction uncertainty term. The prediction confidence is quantified using entropy values. The prediction uncertainty of a dynamic node at the prediction time is calculated based on its walkability probability at that time, using the following formula: ; in, For dynamic nodes At the predicted time The uncertainty of prediction For dynamic nodes The probability of passage at the predicted time. The predicted time for the robot to reach the dynamic node. It is a dynamic node.

[0064] The formula for predicting uncertainty is used to calculate the entropy value of the predicted probability distribution; a higher entropy value indicates higher prediction uncertainty. When At that time, uncertainty is greatest; when The uncertainty is minimized when the value is close to 0 or 1.

[0065] S34: Reference Figure 4 As shown, based on the blocking risk and prediction uncertainty of the dynamic node at the prediction time, the predictive risk cost of the dynamic node at the prediction time is calculated by the following formula: ;

[0066] The formula expands to: ; in, For dynamic nodes At the predicted time The cost of predictive risk The predicted time for the robot to reach the dynamic node. For dynamic nodes, For dynamic nodes At the predicted time The risk of congestion For indexes of dynamic nodes. For dynamic nodes At the predicted time The uncertainty of prediction This is the uncertainty penalty coefficient, used to penalize low-confidence predictions.

[0067] In this embodiment, the uncertainty penalty coefficient is calibrated through Bayesian optimization so that one unit of uncertainty penalty is equivalent to the cost of a path length of 0.15 meters.

[0068] S35: If the candidate node is a static node, then based on the time-series prediction topology graph at the prediction time, obtain the predictive risk cost of the dynamic nodes in the neighborhood of the static node at the prediction time, propagate it to the static node, and obtain the predictive risk cost of the static node at the prediction time.

[0069] The risk cost of dynamic nodes can be propagated to static nodes in local regions through kernel functions, thereby providing risk information for planning within the local neighborhood of the time-series prediction topology graph.

[0070] Obtain the predictive risk cost of dynamic nodes within the neighborhood of a static node at the prediction time, propagate it to the static node, and obtain the predictive risk cost of that static node at the prediction time. The calculation formula is as follows: ; in, Static node At the predicted time The cost of predictive risk For static nodes, Static node Dynamic nodes within the neighborhood, For dynamic nodes At the predicted time The cost of predictive risk Static node To dynamic nodes The Euclidean distance between them This is the spatial attenuation coefficient, used to control the range of risk impact.

[0071] This step transforms the predicted probability into a quantifiable risk cost, while simultaneously taking into account the uncertainty of the prediction, enabling path planning to make a quantitative trade-off between path length and future congestion risk, thus achieving risk-aware path optimization.

[0072] In S4, the risk-aware A* algorithm is used for path planning. The specific steps include:

[0073] S41: During the node expansion process, an expansion cost function is constructed based on the predictive risk cost of the node at the prediction time.

[0074] Based on the classic A* algorithm, this invention incorporates predictive risk costs into the cost function to form a risk-aware A* algorithm, achieving an improvement from pure spatial search to a unified spatiotemporal framework.

[0075] For any node The extended cost function is expressed as follows: ; in, Let n be the expansion cost function for node n. Let n be the path length from the starting point to node n. This is a heuristic estimate of the distance from node n to the endpoint (usually using Euclidean distance). For node n at the prediction time The cost of predictive risk This is the risk weighting coefficient, used to balance path length and risk.

[0076] Preferably, the risk weighting coefficient The method for determining this is as follows: different [values] are evaluated on the validation set using Bayesian optimization or grid search. Given the path planning performance under the given value, choose the one that minimizes the total navigation time. Value. Typical value range is... .

[0077] S42: During the node expansion process, the expansion cost function is used in candidate nodes. Before selecting a target node, if the candidate node is a dynamic node, the process also includes: executing a node pruning strategy based on confidence level.

[0078] The confidence-based node pruning strategy is as follows: if the probability of a dynamic node being passable at the prediction time is less than the confidence threshold, then the dynamic node is removed from the candidate node queue.

[0079] The formula for the confidence-based node pruning strategy is expressed as follows: ; in, For dynamic nodes At the predicted time The probability of passage. This is the confidence threshold.

[0080] like If the node is not selected, it will be pruned, meaning the dynamic node will be removed from the candidate node queue and not added to the search queue to avoid searching for high-risk path branches.

[0081] If the candidate node is a static node, then the expansion cost of that node is calculated directly. .

[0082] S43: Select the target node from the candidate nodes using an extended cost function to achieve path planning.

[0083] Store all candidate nodes and their extended cost functions in a priority queue, and select the node with the extended cost function from the priority queue as the target node.

[0084] The complete risk perception A* algorithm flowchart is shown in Table 1. This algorithm introduces time-dimensional state updates and TPTM-based risk cost calculation within the standard A* framework.

[0085] Table 1. Risk Perception A* Algorithm Flow

[0086] in, As a cost of predictive risk, For nodes To the node Euclidean distance, Indicates the distance from the starting point to the node. The cumulative actual physical cost (path length). Indicates candidate nodes In the case of the next node, from the starting point to the node Path length, To start from candidate nodes Heuristic estimation of the endpoint.

[0087] This step directly uses the predicted probability for node pruning, enabling forward-looking obstacle avoidance and avoiding the search of high-risk path branches; at the same time, it incorporates the risk costs of the spatiotemporal dimensions into the cost function, thus upgrading path planning from pure spatial search to spatiotemporally unified risk perception optimization.

[0088] Even with predictive foresight, offline calculated paths can still fail due to prediction errors or unmodeled environmental events. Therefore, this invention also designs an adaptive execution and intelligent replanning mechanism, enabling the robot to intelligently decide whether to wait or replan during path execution.

[0089] During the execution of the path, at any time... Reaching dynamic nodes The system checks the consistency between the actual node state and the predicted state. If the actual node state is consistent with the predicted state, meaning the node is passable, the robot continues to execute the original path; if the actual node state is inconsistent with the predicted state, meaning the node is detected as impassable, a replanning decision mechanism is triggered.

[0090] The specific implementation steps of the replanning decision-making mechanism are as follows:

[0091] S51: Calculate the waiting cost.

[0092] Based on the state prediction model, obtain the waiting time required for the current node to transition to a passable state. The waiting time is defined as the time since the dynamic node started from the current moment. The probability of passage first exceeds the waiting confidence threshold The moment, that is: ; in, For the waiting time, This represents the number of future time steps that need to be waited from the current moment. Expressing expectations, This is the confidence threshold for waiting.

[0093] The waiting cost is calculated based on the aforementioned waiting time, using the following formula: ; in, These are time-weighting coefficients used to convert time into cost. For waiting costs.

[0094] If within the maximum tolerable waiting time If the passable time cannot be predicted, then it is considered... .

[0095] S52: Calculate the cost of replanning.

[0096] The cost of replanning is calculated based on the computational overhead of path planning, the cost difference between the cost of the new path replanned from the current node and the remaining cost of the original path, and is expressed by the following formula: ; in, To prioritize planning costs, The computational cost of path planning (usually set to a fixed value, such as the cost corresponding to 0.1 seconds). To replan a new path from the current node, This represents the remaining path of the original path. The cost of the path includes path length and risk cost.

[0097] S53: Make intelligent decisions based on waiting costs and replanning costs.

[0098] Instruct the robot to choose the action that minimizes expected future costs: ; in, For the robot's actions, For Action The cost, Act towards the goal.

[0099] If the waiting cost is less than the replanning cost, the robot waits until the current node becomes passable, and then continues to execute the path according to the original path. If the waiting cost is greater than or equal to the replanning cost, then the risk-aware A* algorithm is used to replan the path starting from the current node.

[0100] The complete adaptive execution and replanning decision-making mechanism process is shown in Table 2.

[0101] Table 2. Adaptive Execution and Replanning Decision-Making Mechanism

[0102] in, This indicates the new path after replanning. This is the next path point of the current path point.

[0103] The adaptive execution and replanning decision-making mechanism in this step quantitatively calculates waiting costs and replanning costs based on a predictive state model, enabling intelligent decision-making and avoiding blind waiting or unnecessary replanning; it formalizes the decision problem into an online cost minimization problem, transforming the robot from a passive responder to an active decision-maker.

[0104] Preferably, the key system parameters of the present invention and their typical value ranges are shown in Table 3.

[0105] Table 3. Key parameters of the present invention

[0106] The computational complexity analysis of this invention is as follows:

[0107] In S1, LLaVA inference time is approximately 0.1-0.5 seconds per frame (depending on GPU performance), which can be optimized through batch processing.

[0108] In S2, the training complexity of LSTM is: ,in The number of training samples, The number of hidden units is 64, and it can be used offline after training once.

[0109] LSTM prediction complexity: The prediction time for a single prediction is approximately 1-5 milliseconds, which meets the real-time requirements.

[0110] In S4, the path planning module: the time complexity of the risk perception A* algorithm is... Where b is the branching factor and d is the path depth. Due to the node pruning strategy, the actual search space is usually smaller than that of the traditional A* algorithm, and the search efficiency is improved by about 30-50%.

[0111] In S5, the decision calculation is as follows: It does not increase the additional computational burden.

[0112] To ensure the robustness of the system in real-world environments, the present invention may employ the following mechanism: Prediction error handling: When prediction uncertainty At that time, the risk weight is automatically increased to make the planning more conservative; This represents the uncertainty threshold. Non-stationary adaptation: For non-stationary environments, a piecewise stationary assumption is adopted, within a time window... The LSTM model is retrained internally to adapt to environmental changes; Anomaly detection and recovery: When multiple consecutive predictions are inconsistent with the actual state, the state prediction model is retrained or the prediction confidence weight is reduced. Multimodal fusion: In the semantic perception stage, information from multiple sensors (such as vision + LiDAR) can be fused to improve perception reliability.

[0113] To verify the effectiveness of this invention, this embodiment constructs a two-dimensional grid map containing multiple dynamic doors in a Python simulation environment. The map size is 50×50, containing 4 dynamic door nodes (Door 1-4) and several static obstacles. The state of each dynamic door changes periodically according to a preset timing pattern (e.g., opening and closing once every 10 seconds).

[0114] The specific implementation steps of this embodiment include:

[0115] S101. Construct a time series prediction topology graph.

[0116] First, the state sequence of each dynamic gate node over the past 1000 time steps is collected as training data. The first 80% of the data (800 time steps) is used as the training set, and the last 20% of the data (200 time steps) is used as the test set.

[0117] Independent LSTM models were trained for each dynamic gate. The Adam optimizer was used during training, with a learning rate of 0.001, a batch size of 64, and a training duration of 50 epochs. After training, the model achieved a prediction accuracy of 96.5% on the test set, indicating that the model can effectively learn the temporal variation patterns of the dynamic gates.

[0118] S102, Risk Perception Path Planning.

[0119] Given a starting point (5,5) and a target point (45,45), use the risk-aware A* algorithm to perform path search.

[0120] During the search process, for each candidate node n: Calculate the estimated arrival time of the robot from the starting point to node n. ; If node n is located near the dynamic gate (distance less than the threshold), then query the LSTM model of the corresponding gate to obtain... The probability of being passable at any given time; Calculate risk cost And add an extended cost function ; If the predicted passability probability is lower than the threshold If so, then prune that node.

[0121] S103, Path Execution and Adaptive Replanning.

[0122] The robot follows the planned path. When the robot reaches a dynamic gate node: if the actual state of the gate matches the prediction, the robot passes through directly; if the gate is actually closed but the prediction shows that it will be passable in the future, the robot calculates the waiting cost and the replanning cost, and selects the strategy with the lower cost.

[0123] To verify the beneficial effects of the present invention, a comparative experiment was also included in this embodiment. The method for conducting the comparison includes: Baseline 1: Traditional A* algorithm (considering only static environment); Baseline 2: D*Lite algorithm (responsive replanning); Baseline 3: Spatiotemporal A* algorithm (assuming perfect prediction information); The method of this invention: Risk perception A* algorithm.

[0124] Figure 5 The diagram comparing the path planning of the method of this invention and the traditional A* algorithm clearly demonstrates the forward-looking advantages of this invention. The traditional A* algorithm selects the shortest path, but this path passes through a door that is about to close, forcing the robot to wait or replan upon arrival. In contrast, the method of this invention identifies high-risk areas in advance through a predictive model and proactively selects a slightly longer but more reliable path, avoiding waiting and replanning, ultimately resulting in a shorter total time.

[0125] Comparative experimental results are shown in Table 4 and Figure 6 As shown, where Figure 6 (a) in the figure is a comparison chart of the number of times the target is reached by different path planning methods. Figure 6 (b) in the figure is a comparison of the total time taken by different path planning methods. Figure 6 (c) in the figure is a comparison chart of the total path length of different path planning methods. Figure 6(d) in the figure is a comparison of the average time taken by different path planning methods. Figure 6 (e) in the figure is a comparison chart of the average path length of different path planning methods. Figure 6 Figure 6 (f) in the figure is a comparison of the number of replanning steps for different path planning methods.

[0126] Table 4. Path planning performance comparison (average of 50 tasks)

[0127] As can be seen from Table 4: (1) The present invention is optimal in terms of navigation time: it reduces navigation time by 6.3% compared to D*Lite and by 27.6% compared to traditional A*; (2) The present invention is optimal in terms of path length: it reduces the path length by 13.0% compared to the traditional A* and by 10.8% compared to D* Lite; (3) The invention has the fewest replanning times: 61.1% less than D*Lite and 50.5% less than traditional A*; (4) All methods achieved a 100% success rate: This indicates that the present invention significantly improves navigation efficiency while maintaining a high success rate.

[0128] Based on experiments, this embodiment demonstrates that, compared to existing technologies, the present invention has the following beneficial effects:

[0129] 1. Solving the problem of short-sighted planning: By using the LSTM prediction model to predict the future accessibility of dynamic nodes in advance, the path planning is forward-looking and avoids planning paths that are no longer viable by the time of arrival.

[0130] 2. Significantly reduces the number of replanning operations: Experimental data shows that compared with the traditional D* Lite algorithm, the number of replanning operations in this invention is reduced by 61.1%, which greatly reduces the computational cost.

[0131] 3. Improve navigation efficiency: Through the risk perception cost function, the robot actively avoids high-risk areas, reducing the average navigation time by 6.3%-27.6% and shortening the path length by 7.2%-13.0%.

[0132] 4. Quantitative Risk Assessment: Transforming predicted probabilities into risk costs enables path planning to make quantitative trade-offs between path length and congestion risk, achieving better path selection.

[0133] 5. Intelligent decision-making mechanism: The adaptive replanning mechanism calculates waiting costs and replanning costs quantitatively based on the predictive model, thereby achieving intelligent decision-making and avoiding blind waiting or unnecessary replanning.

[0134] 6. Improved navigation success rate: The navigation success rate is significantly improved in dynamic environments. Experiments show that all test tasks were completed successfully, while traditional methods have a certain failure rate.

[0135] This embodiment fully verifies the effectiveness and superiority of the present invention. By incorporating LSTM temporal prediction into path planning, the present invention achieves a paradigm shift from passive response to active prediction, effectively solving the problem of short-sighted planning and significantly improving navigation efficiency in dynamic environments.

[0136] In summary, the risk-aware path planning method based on time-series prediction described in this invention predicts the drivability probability of dynamic nodes at future times, calculates the predictive risk cost of nodes, and incorporates it into the cost function of the A* algorithm. This enables the robot to actively avoid high-risk areas, not only making path planning more forward-looking and avoiding planning paths that are already invalid upon arrival, but also allowing path planning to quantitatively weigh path length and congestion risk, achieving better path selection and improving the efficiency and quality of path planning.

[0137] Furthermore, this invention introduces prediction uncertainty into the calculation of predictive risk cost, quantifies prediction confidence through entropy value, and penalizes low-confidence predictions to improve the robustness of path planning.

[0138] Furthermore, this invention performs node pruning based on the prediction confidence of dynamic nodes. Before calculating the next node, nodes with low predicted passability are removed from the candidate nodes to avoid searching high-risk path branches and improve search efficiency.

[0139] Furthermore, based on the probability of the traversability of dynamic nodes in the future, this invention proposes an adaptive path replanning decision-making mechanism to calculate and measure waiting costs and replanning costs, thereby achieving intelligent decision-making, avoiding blind waiting or unnecessary replanning by the robot, reducing the computational overhead of path replanning, and improving the efficiency of path planning.

[0140] Based on the aforementioned risk perception path planning method based on time-series prediction, this invention also provides a risk perception path planning system based on time-series prediction, comprising: The semantic awareness module is used to convert image frames of dynamic nodes in the environment into temporal state sequences using a large vision-language model; The temporal prediction module is used to build a state prediction model for each dynamic node, and to obtain the passability probability of the dynamic node at the prediction time based on the historical temporal state sequence of the dynamic node; and to construct the temporal prediction topology map at the prediction time based on the passability probabilities of all dynamic nodes at the prediction time. The risk cost calculation module is used to calculate the predicted time when the robot arrives at the candidate node; if the candidate node is a dynamic node, the module obtains the passability probability of the dynamic node at the predicted time and calculates the predictive risk cost of the dynamic node at the predicted time based on the passability probability; if the candidate node is a static node, the module obtains the predictive risk cost of the dynamic nodes in the neighborhood of the static node at the predicted time based on the temporal prediction topology graph at the predicted time, and propagates it to the static node to obtain the predictive risk cost of the static node at the predicted time. The path planning module is used to perform path planning using the risk-aware A* algorithm. It includes: constructing an extended cost function based on the predictive risk cost of candidate nodes at the prediction time during the node expansion process; and selecting the target node from the candidate nodes using the extended cost function to achieve path planning.

[0141] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A risk perception path planning method based on time-series prediction, characterized in that, include: Large-scale vision-language models are used to replace image frames of dynamic nodes in the environment with temporal state sequences; A state prediction model is constructed for each dynamic node to obtain the passability probability of the dynamic node at the prediction time based on the historical time-series state sequence of the dynamic node; a time-series prediction topology map at the prediction time is constructed based on the passability probabilities of all dynamic nodes at the prediction time. Calculate the predicted time when the robot will arrive at the candidate node; if the candidate node is a dynamic node, obtain the passability probability of the dynamic node at the predicted time, and calculate the predictive risk cost of the dynamic node at the predicted time based on the passability probability. If the candidate node is a static node, then based on the time-series prediction topology graph at the prediction time, the predictive risk cost of the dynamic nodes in the neighborhood of the static node at the prediction time is obtained, and it is propagated to the static node to obtain the predictive risk cost of the static node at the prediction time. Path planning using the risk-aware A* algorithm includes: during node expansion, constructing an expansion cost function based on the predictive risk cost of candidate nodes at the prediction time; The target node is selected from the candidate nodes using an extended cost function to achieve path planning.

2. The risk perception path planning method based on time-series prediction according to claim 1, characterized in that, The steps for converting image frames of dynamic nodes in an environment into temporal state sequences using a large-scale vision-language model include: The image frames and structured text prompts collected at each time step of the dynamic node are input into a large-scale vision-language model to obtain the semantic state vector of the dynamic node at each time step. The semantic state vector of a dynamic node at each time step is mapped to a binary state vector, which is used to characterize whether the dynamic node is passable. The binary state vector of the time step and the dynamic node at that time step are combined into sample pairs, and the temporal state sequence is constructed according to the time sequence.

3. The risk perception path planning method based on time-series prediction according to claim 1, characterized in that, The state prediction model employs a long short-term memory network.

4. The risk perception path planning method based on time-series prediction according to claim 1, characterized in that, The predictive risk cost of the dynamic node at the predicted time is calculated based on the aforementioned passability probability, using the following method: The blocking risk and prediction uncertainty of a dynamic node at the prediction time are calculated based on the passability probability of the dynamic node at the prediction time. Based on the blocking risk and prediction uncertainty of the dynamic node at the prediction time, the predictive risk cost of the dynamic node at the prediction time is calculated using the following formula: ； in, For dynamic nodes At the predicted time The cost of predictive risk The predicted time for the robot to reach the dynamic node. For dynamic nodes, For dynamic nodes At the predicted time The risk of congestion For indexes of dynamic nodes. For dynamic nodes At the predicted time Predictive uncertainty, This represents the uncertainty penalty coefficient.

5. The risk perception path planning method based on time-series prediction according to claim 4, characterized in that, The congestion risk of a dynamic node at the predicted time is calculated based on its passability probability at the predicted time. The calculation formula is as follows: ； in, For dynamic nodes At the predicted time The risk of congestion For indexes of dynamic nodes. The predicted time for the robot to reach the dynamic node. For dynamic nodes The weighting coefficients, For dynamic nodes The probability of passage at the predicted time; The prediction uncertainty of a dynamic node at the prediction time is calculated based on the walkability probability of the dynamic node at the prediction time. The calculation formula is as follows: ； in, For dynamic nodes At the predicted time Predictive uncertainty, For dynamic nodes The probability of passage at the predicted time.

6. The risk perception path planning method based on time-series prediction according to claim 1, characterized in that, Obtain the predictive risk cost of dynamic nodes within the neighborhood of a static node at the prediction time, propagate it to the static node, and obtain the predictive risk cost of that static node at the prediction time. The calculation formula is as follows: ； in, Static node At the predicted time The cost of predictive risk For static nodes, Static node Dynamic nodes within the neighborhood, For dynamic nodes At the predicted time The cost of predictive risk Static node To dynamic nodes The Euclidean distance between them This is the spatial attenuation coefficient.

7. The risk perception path planning method based on time-series prediction according to claim 1, characterized in that, An extended cost function is constructed based on the predictive risk cost of a node at the prediction time, and the formula is expressed as: ； in, Let n be the expansion cost function for node n. Let n be the path length from the starting point to node n. For the heuristic estimation from node n to the endpoint, For node n at the prediction time The cost of predictive risk This is the risk weighting coefficient.

8. The risk perception path planning method based on time-series prediction according to claim 1, characterized in that, During node expansion, before selecting the target node from the candidate nodes using the expansion cost function, if the candidate nodes are dynamic nodes, the process also includes: If the probability of a dynamic node being passable at the predicted time is less than the confidence threshold, then the dynamic node is removed from the candidate node queue.

9. The risk perception path planning method based on time-series prediction according to claim 1, characterized in that, Also includes: If the robot detects that the current node is impassable during path execution, it triggers a replanning decision-making mechanism, which includes the following steps: Based on the state prediction model, the waiting time required for the current node to become passable is obtained; the waiting cost is calculated based on the waiting time. The cost of replanning is calculated based on the computational overhead of path planning and the cost difference between the cost of the new path replanned from the current node and the remaining path cost of the original path. If the waiting cost is less than the replanning cost, the robot waits until the current node becomes passable, and then continues to execute the path according to the original path. If the waiting cost is greater than or equal to the replanning cost, then the risk-aware A* algorithm is used to replan the path starting from the current node.

10. A risk perception path planning system based on time-series prediction, characterized in that, include: The semantic awareness module is used to convert image frames of dynamic nodes in the environment into temporal state sequences using a large vision-language model; The temporal prediction module is used to build a state prediction model for each dynamic node, and to obtain the passability probability of the dynamic node at the prediction time based on the historical temporal state sequence of the dynamic node; and to construct the temporal prediction topology map at the prediction time based on the passability probabilities of all dynamic nodes at the prediction time. The risk cost calculation module is used to calculate the predicted time when the robot will arrive at the candidate node; If the candidate node is a dynamic node, obtain the passability probability of the dynamic node at the prediction time, and calculate the predictive risk cost of the dynamic node at the prediction time based on the passability probability. If the candidate node is a static node, then based on the time-series prediction topology graph at the prediction time, the predictive risk cost of the dynamic nodes in the neighborhood of the static node at the prediction time is obtained, and it is propagated to the static node to obtain the predictive risk cost of the static node at the prediction time. The path planning module is used to perform path planning using the risk perception A* algorithm, including: during the node expansion process, constructing an expansion cost function based on the predictive risk cost of candidate nodes at the prediction time; The target node is selected from the candidate nodes using an extended cost function to achieve path planning.

Images