Embodied intelligent navigation method based on collision risk calibration map attention mechanism

Abstract

The application provides a somatic intelligent navigation method based on a collision risk calibration graph attention mechanism. The method obtains the environment state of a robot, surrounding dynamic entities and static obstacles, constructs an interaction graph and a physical risk relationship matrix aligned with entity indexes; encodes the learning context relationship of node pairs, combines node features, physical risk relationships, relationship effectiveness or credibility and the difference between the two types of relationships, generates calibration coefficients for node pairs one by one, and fuses to obtain a calibration relationship matrix; the calibration relationship is mapped into a graph attention bias through monotonic mapping, and the information of safety-critical entities is aggregated; the dynamic environment representation, static obstacle representation, robot state and historical state are fused to generate a candidate navigation action, and safe actions can be selected according to short-term predicted risks. The application makes the attention distribution subject to an interpretable physical risk constraint, while retaining the context adaptability, improving the navigation safety, continuity and robustness in crowded and distributed changing environments.

Description

Technical Field

[0001] This invention relates to the fields of autonomous navigation for mobile robots, graph neural networks, environmental risk modeling, and reinforcement learning, and particularly to a method for reliably calibrating graph attention relationships using physical collision risks and generating robot navigation actions accordingly. Background Technology

[0002] In crowded environments with complex obstacle distribution, mobile robots need to generate navigation actions in real time based on a variable number of pedestrians and the state of obstacles. Graph neural networks can represent entities such as robots and pedestrians as nodes and use graph attention to learn the interaction relationships between different entities, thus they are used for robot navigation in crowded environments.

[0003] Existing graph attention mechanisms typically learn end-to-end from the geometric and motion features of nodes and task rewards. Collision penalties or discomfort costs only indicate whether the behavioral outcome is safe, but cannot directly identify which entity caused the risk. Attention weights learned from weak rewards represent statistical correlations and are not necessarily consistent with the urgency of physical collisions. When the number and density of pedestrians or the configuration of obstacles change, attention may be allocated to entities less relevant to the current safety decision, leading to a decline in navigation safety and generalization ability.

[0004] If fixed physical risk relationships are directly used to replace learning relationships, physical model errors may propagate to the policy network, and it will be difficult to express the contextual dependencies generated by scene structure and behavioral intentions. Therefore, a graph attention navigation method is needed that can both introduce entity-level physical risk priors and adaptively adjust the degree of influence of these priors according to the current interaction context. Summary of the Invention

[0005] The purpose of this invention is to provide an embodied intelligent navigation method based on a collision risk calibration map attention mechanism. By modeling the collision risk between the robot and dynamic subjects and static obstacles as a continuous, interpretable, fusionable risk field with action condition prediction capabilities, the method can identify and quantify potential collision trends in advance, thereby providing reliable risk input for robot safe navigation, reinforcement learning strategy optimization, or real-time obstacle avoidance control.

[0006] To address the issues of lack of entity-level physical risk constraints in learning attention and the difficulty of adapting fixed physical relationships to risk errors and complex contexts, this invention provides an embodied intelligent navigation method based on a collision risk calibration graph attention mechanism. Its core mechanism is as follows: for each directed node pair, a calibration coefficient is generated by combining physical risk relationships, learning context relationships, relationship differences, and risk effectiveness. The calibration relationship is used as a graph attention relationship bias, and the risk perception graph representation is further used for temporal navigation decision-making.

[0007] In a first aspect, this invention discloses an embodied intelligent navigation method based on a collision risk calibration map attention mechanism, the method comprising: Acquire the environmental state of the robot, several dynamic entities within the robot's perception range, and static obstacles; Based on the physical collision risk between the robot and each dynamic entity, obtain the physical risk relationship matrix corresponding to the entity index of the current interaction graph, and perform feature encoding on the node state in the interaction graph to obtain node features. For node pairs in the interaction graph, the learning context relationship is determined based on the corresponding node features to obtain the learning relationship matrix; Based on the node characteristics of the node pair, the physical risk relationship in the physical risk relationship matrix, the learning context relationship in the learning relationship matrix, and the difference in the relationship between the physical risk relationship and the learning context relationship, the calibration coefficient of the node pair is generated. Based on the calibration coefficients, a node-by-node pair fusion of physical risk relationships and learning context relationships is performed to obtain a calibration relationship matrix; By using the calibration relation matrix as the relation bias of the graph attention operation, and aggregating the messages of risk calibration of node features, a dynamic environment representation of risk perception is obtained. The system integrates the dynamic environmental representation of risk perception, the static obstacle representation obtained based on static obstacles, and the robot state representation, and outputs the robot navigation actions based on the fusion result.

[0008] Based on the above technical solution, the following improvements can be made: As a preferred approach, the physical risk relationship is determined based on the relative distance, relative speed, approach direction, closest approach time, closest approach distance, and combined safety radius between the robot and the dynamic entity, and is normalized to a preset relationship range; wherein, the physical risk relationship increases as the relative distance decreases, the approach speed increases, the closest approach time shortens, or the closest approach distance decreases.

[0009] As a preferred approach, the relationship difference is the absolute value of the difference between the physical risk relationship and the learning context relationship; the calibration coefficient is obtained by mapping the node characteristics, physical risk relationship, learning context relationship and relationship difference of the node pair to the calibration network.

[0010] As a preferred approach, node-by-node fusion includes: The processed calibration coefficient is multiplied by the physical risk relationship to obtain the first relationship component; Multiplying the calibration coefficient by the learning context relation yields the second relation component; The sum of the first relation component and the second relation component is determined as the calibration relation of the node pair.

[0011] As a preferred approach, multiple attention heads are used to perform graph attention operations, with each attention head sharing or using a calibration relation matrix separately. The outputs of each attention head are concatenated or averaged, and then updated node features are obtained sequentially through residual connections, layer normalization, and feedforward networks. Finally, the dynamic environment representation of risk perception is extracted from the robot node features of the last layer.

[0012] As a preferred approach, the static obstacle state includes a two-dimensional point cloud, a locally occupied grid, a set of boundary line segments, or a depth map; the static obstacle representation is obtained using a one-dimensional convolutional network, a two-dimensional convolutional network, or an ensemble encoding network; the risk-aware dynamic environment representation, the static obstacle representation, and the robot state representation are spliced, gated fusion, cross-attention fusion, or conditional modulation, and the fusion result, along with the temporal hidden state of the previous time step, is input into a gated recurrent unit, a long short-term memory network, a temporal convolutional network, or a temporal Transformer.

[0013] As a preferred approach, before executing a robot navigation action, the system determines whether the candidate navigation action meets safety conditions based on the short-term predicted trajectory, the latest physical risk relationship, and the static obstacle clearance corresponding to the candidate navigation action. If the highest priority candidate navigation action does not meet safety conditions, the system selects the next candidate navigation action that meets safety conditions, or outputs a preset safety action such as deceleration, braking, stopping, or steering to avoid obstacles.

[0014] Secondly, this invention discloses an embodied intelligent navigation system based on a collision risk calibration graph attention mechanism, comprising: an environment state acquisition module for acquiring robot state, dynamic entity state, and static obstacle state; a relationship acquisition module for generating a physical risk relationship matrix corresponding to the node index of the interaction graph and its validity information or credibility information, and acquiring a learning relationship matrix; an adaptive calibration module for generating calibration coefficients node-by-node and fusing them to obtain a calibration relationship matrix; a risk attention module for performing monotonic mapping on the calibration relationship matrix to form a relationship bias, and performing graph attention message aggregation based on the relationship bias to obtain a risk-aware dynamic environment representation; a multimodal temporal fusion module for fusing the risk-aware dynamic environment representation, static obstacle representation, robot state representation, and historical temporal state; and a policy output module for generating robot navigation actions based on the fusion results.

[0015] Thirdly, the present invention discloses an electronic device, including at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, which, when executed, cause the at least one processor to perform any of the methods described above.

[0016] Fourthly, the present invention discloses a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements any of the methods described above.

[0017] The present invention has at least the following beneficial effects: it directly injects entity-level physical risks into the messaging process, enabling safety-critical entities to obtain higher information aggregation weights; it retains the flexibility of learning relationships through per-interaction calibration, reducing the impact of fixed physical model errors; it encodes dynamic entities and static obstacles separately, adapting to variable numbers of entities and different obstacle densities; and it improves the continuity of navigation decisions by utilizing historical interaction trends through temporal state modeling. Attached Figure Description

[0018] Figure 1 A flowchart of the physical risk calibration map attention navigation method provided in this embodiment of the invention; Figure 2 This is a schematic diagram of the context-adaptive risk relationship calibration structure provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the network structure for a risk calibration navigation strategy provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the attention structure of the multi-head risk calibration map provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the training and online inference process provided in an embodiment of the present invention; Figure 6 A block diagram of a physical risk calibration map attention navigation system provided in an embodiment of the present invention. Detailed Implementation

[0019] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] Reference Figure 1 An embodiment of the present invention provides an embodied intelligent navigation method based on a collision risk calibration map attention mechanism, including steps S101 to S106.

[0021] Step S101: Obtain the environmental state of the robot, several dynamic entities within the robot's perception range, and static obstacles at the current moment. The robot state may include position, linear velocity, target position, orientation, and angular velocity; the dynamic entity state may include position and velocity relative to the robot; the static obstacle state may be represented by a two-dimensional laser point cloud, a local occupancy grid, boundary line segments, or a depth map.

[0022] Step S102: Construct the current interaction graph , where the set of nodes This includes robot nodes and currently observed dynamic entity nodes. A physical risk relationship matrix is ​​obtained based on the interaction between the robot and the dynamic entities. Physical risk relationships can be determined by distance, relative motion, predicted collision time, or a combination thereof; this embodiment does not limit the specific method of generating physical risk relationships.

[0023] Input the status of each node into the node encoder, for each node... , obtain node features , For each node pair The relation encoder calculates the learned context relations. .

[0024]

[0025] in, Indicates feature splicing, Represents a relational encoder. This represents a bounded activation function. Learning contextual relationships can simultaneously express the dependencies between robot-dynamic entities and between dynamic entities, arising from scene structure, motion patterns, and behavioral intentions.

[0026] Step S103, refer to Figure 2 According to physical risk relationships Learning contextual relationships And the differences between the two Generate node-by-node calibration coefficients .

[0027]

[0028]

[0029] The calibration coefficients are dynamically determined based on the node context of each interaction and the consistency between the two types of relationships, rather than using a uniform global fusion ratio for all node pairs.

[0030] Step S104: Integrate physical risk relationships and learning context relationships to obtain the calibration relationship matrix. .

[0031]

[0032] When the calibration coefficient is small, the calibration relationship retains more of the prior physical risk; when the calibration coefficient is large, the calibration relationship adopts more of the learning context relationship, thereby supplementing the interaction dependencies not expressed in the physical model.

[0033] Step S105: Use the calibration relation matrix as the relation bias for the graph attention Softmax operation. For the first... Layer node features Generate query matrices respectively Key matrix Sum matrix .

[0034]

[0035]

[0036]

[0037]

[0038]

[0039] in, Here is the attention weight matrix. To query the dimensions of the vector and key vector, The relation bias strength coefficient is calculated by Softmax along the neighbor dimension corresponding to each query node. The higher the calibration relation, the greater the relative weight of the corresponding node information aggregated to the query node. L layers of risk calibration graph attention layers can be stacked, and [the following text appears to be incomplete and requires further context: "from..."] Extracting robot node features as a dynamic environment representation for risk perception .

[0040] Step S106, refer to Figure 3 Static obstacle representations are obtained by encoding static obstacle point clouds using a one-dimensional convolutional network. ;Will and splicing together to form an environmental state characterization and the robot state representation The hidden state from the previous moment is input into the gating loop unit.

[0041]

[0042] in: These are the weights and bias parameters that need to be trained in the GRU.

[0043] The current temporal hidden state is input into the Actor-Critic policy network. The value head outputs the state value, and the policy head outputs the probability distribution of discrete navigation actions or continuous control variables. Discrete navigation actions can include combinations of linear and angular accelerations of different magnitudes.

[0044] During the training phase, the policy network can be trained using reinforcement learning based on task rewards. Task rewards may include rewards for reaching the target, collision penalties, rewards for progressing towards the target, and rewards for smooth motion. Physical risk relationships, as prior relationships within the graph attention mechanism, do not need to be directly added to the task rewards.

[0045] Implementation methods for environment state representation and interaction graph construction In some embodiments, the robot state includes the current position, linear velocity, angular velocity, target position, target orientation, robot radius, and the action executed in the previous control cycle. The robot state is first transformed to the robot body coordinate system, so that the angle between the target orientation and the robot's current orientation, the target distance, and the velocity components have translational and rotation-invariant expressions.

[0046] Dynamic entity states can include relative position, relative velocity, entity radius, entity category, orientation of motion, tracking confidence, and historical trajectory encoding. For entities whose velocity cannot be reliably estimated, adjacent observation difference, filtered state, or zero velocity initialization can be used, and the velocity confidence can be input into the node encoder.

[0047] Static obstacle states can be generated from 2D laser point clouds, locally occupied grids, boundary line segment sets, or depth maps. To reduce the impact of differences in the number of static obstacle points, sampling can be performed uniformly around the robot according to polar angles, or the nearest obstacle distance can be retained for each angular sector.

[0048] The node set of an interaction graph includes at least robot nodes and currently observed dynamic entity nodes. The edge set can be fully connected or constructed based on distance thresholds, field of view, visibility, and risk thresholds. For fully connected graphs, filler nodes can be masked using a validity mask; for sparse graphs, only the edge indices and attributes that satisfy the adjacency conditions can be stored.

[0049] When the number of dynamic entities exceeds the preset maximum number of nodes, entities with higher physical risk can be retained first based on their physical risk value. If the physical risk value has not yet been obtained, initial screening can be performed based on distance, estimated collision time, or observation confidence. Entities that are screened out can still provide information on the overall crowding level through aggregated nodes or local density features.

[0050] Node encoders can be multilayer perceptrons, shared convolutional networks, or encoding networks with entity type embeddings. Different types of nodes can share backbone parameters and be distinguished using type embeddings, or robot encoders and dynamic entity encoders can be set separately, and then the outputs can be mapped to the same feature dimension.

[0051] Implementation methods for obtaining physical risk relationships The robot-dynamic entity relationships in the physical risk relationship matrix can be determined by the current distance, combined safety radius, relative speed, closest approach time, closest approach distance, or a combination thereof. The physical risk value is preferably normalized to [value missing]. The range is defined by a larger value, which indicates a higher risk of collision.

[0052] For dynamic entity-dynamic entity node pairs, the physical risk relationship can be set to an empty relationship value, with the corresponding dependency supplemented by the learning context relationship; alternatively, the physical conflict relationship between the two dynamic entities can be calculated when their states are available. The diagonal elements of the matrix can be set to self-relation values ​​to preserve the node's own information.

[0053] The physical risk relationship matrix can be calculated in real time by the robot itself, or it can be generated by an independent safety module, edge server, or external perception system and then input into the navigation network. The navigation method only requires that the physical risk relationships are consistent with the entity indexes in the interaction graph, and does not limit the specific device that generates the risk matrix.

[0054] When the risk matrix and node states come from different timestamps, they are aligned according to the entity tracking identifier, and the risk value is updated to the index corresponding to the most recent node state. If a risk relationship exceeds a preset valid duration, it is marked as invalid, and expired risks are prevented from participating in the calibration through a relationship mask.

[0055] Learning Implementation Methods of Context Relation Encoding The relation encoder receives the two node features of a node pair, and can further receive relative position, relative velocity, node type combinations, and graph distance equilateral features. The concatenation order of the node features is used to distinguish directed relations, therefore... and They can be different to express information from nodes. n Aggregate to node m The directionality of time.

[0056] A relation encoder can consist of one or more fully connected networks, outputting relation values ​​between 0 and 1 mapped by a sigmoid function. Alternatively, it can employ a bilinear relation function, dot product function, cosine similarity, or a small Transformer encoder.

[0057] To reduce the scale shift caused by changes in the number of nodes, the input layer of the relation encoder can be normalized, and temperature scaling can be applied to the output. A smaller temperature parameter results in a sharper relation distribution, while a larger temperature parameter results in a smoother relation distribution. The temperature parameter can be used as a learnable parameter or set according to the number of nodes.

[0058] The learned relationship matrix is ​​used not only to represent physical collision risk, but also to express contextual relationships such as same-direction pedestrian groups, occupancy between pedestrians, passageway occupation, and potential collaborative behaviors. Therefore, the learned relationship and the physical risk relationship are complementary.

[0059] Implementation of per-interaction calibration coefficient generation The calibration network generates calibration coefficients individually for each node pair. Its inputs include at least the node characteristics of the two nodes, the physical risk relationship, the learning context relationship, and the absolute difference between the two types of relationships; it may also include the risk value change rate, observation confidence, entity category, and local node density.

[0060] When the physical risk is high and the physical relationship is consistent with the learned relationship, the calibration network can output a smaller calibration coefficient, so that the calibration relationship mainly retains the physical risk; when the physical relationship is missing, the confidence is low, or the scene context contains additional dependencies, the calibration network can increase the proportion of the learned relationship.

[0061] The calibration coefficients are constrained to between 0 and 1 by a sigmoid function. To avoid completely ignoring physical risks in the early stages of training, the output bias of the calibration network can be initialized to a negative value, making the initial calibration coefficients small; as training progresses, the network learns when to introduce more contextual relationships based on task data.

[0062] Regularization constraints can also be imposed on calibration coefficients. For example, calibration sparsity regularization can be used to reduce unnecessary relation switching, time smoothing regularization can be used to constrain the changes in calibration coefficients of the same entity relation at adjacent time points, or risk preservation regularization can be used to prevent high physical risk relations from being excessively weakened.

[0063] Risk calibration loss Defined as:

[0064] Among them, risk remains loss Defined as:

[0065] in, This is the set of currently valid node pairs; This is the set of node pairs with the same tracking identifier at adjacent time points; To avoid a denominator of 0 when there are no dynamic entities or no successfully matched entities. Used to constrain only high-physical-risk relationships; To allow for a calibration relationship that is less than the tolerance of a physical risk relationship; To allow for calibration relationships to be lower than the tolerance of physical relationships, and ; ; These are the weighting coefficients for each type of loss. ; In some embodiments, the calibration network outputs two or more relation gating weights, and the sum of the gating weights is made 1 using Softmax. This allows for three-way or multi-way fusion among physical risk relations, learning context relations, and historical relations.

[0066] Implementation of Risk Calibration Map Attention Relationship bias is adopted Logarithmic mappings preserve the discriminability of smaller relation values ​​while suppressing the over-dominance of compatible scores by maximal relation values. Linear mappings, learnable monotonic mappings, or piecewise mappings can also be used to generate relation biases.

[0067] Relationship strength coefficient You can set it to a fixed value, or you can set it to different attention layers according to the image. Alternatively, it can be dynamically generated based on robot speed, local environment density, and the number of nodes. When the environment is crowded or the robot speed is high, the relation bias strength can be increased to focus attention more on safety-critical entities.

[0068] Reference Figure 4 Graph attention can employ a multi-head structure. Each attention head uses independent query, key, and value projection matrices and can share or use calibration relation matrices separately. The outputs of different attention heads can be concatenated and linearly mapped, or averaged, to learn multiple interaction patterns simultaneously.

[0069] Each risk calibration graph attention layer can also include residual connections, layer normalization, a feedforward network, and random deactivation. Residual connections preserve the node's own representation, layer normalization stabilizes the feature scale under different numbers of nodes, and random deactivation is used to reduce training overfitting. The intermediate node feature matrix is ​​obtained after risk calibration attention aggregation. The definition is as follows:

[0070]

[0071] For invalid nodes or non-existent edges, a negative infinity mask is added to the compatibility score before Softmax, making the corresponding attention weights zero. For node pairs in the risk matrix that do not provide physical relationships, the relationship bias can be set to zero, and their attention weights can be determined by the learned compatibility score.

[0072] The updated representations of robot nodes are used to generate risk-aware dynamic environment representations. The updated representations of dynamic entity nodes can also be retained to the next layer, allowing information about high-risk entities to influence robot nodes after propagation through multiple layers; when only robot central aggregation is required, star graphs can also be used to reduce computational load.

[0073] Implementation methods for static obstacle coding and cross-modal fusion Static obstacle point clouds are arranged according to the robot's body coordinate system and then input into a one-dimensional convolutional network. The one-dimensional convolutional network can include several convolutional layers, nonlinear activation layers, and pooling layers, and obtains fixed-dimensional obstacle representations through global max pooling, global average pooling, or attention pooling.

[0074] When using an occupied grid or local map, a two-dimensional convolutional network can be used instead of a one-dimensional convolutional network; when using a set of boundary line segments, each line segment can be regarded as a static obstacle node, and a static obstacle representation can be obtained by using a shared encoder and ensemble pooling.

[0075] Risk perception can be achieved by directly stitching together dynamic environmental representations and static obstacle representations, or by gating fusion, cross-attention, or conditional modulation. Gating fusion determines the relative contribution of dynamic interactions and static constraints based on the robot's state, thereby adaptively switching the focus between crowded crowds and narrow obstacle environments.

[0076] The robot's state representation can be input into the GRU alone, or it can be concatenated with the environmental state representation first. The GRU hidden state is used to record historical pedestrian movement trends, recent robot actions, and local congestion evolution; in other embodiments, long short-term memory networks, temporal convolutional networks, or lightweight temporal Transformers can also be used.

[0077] Implementation methods for navigation actions and safe execution When the policy network outputs discrete actions, the action set can consist of multiple combinations of linear and angular accelerations; when outputting continuous actions, the policy head outputs parameters of linear velocity, angular velocity, or acceleration distribution. The state value head is used to estimate the expected cumulative reward of the current state.

[0078] Before an action is executed, a safety check unit can be set up to determine whether a candidate action meets a safety threshold based on the predicted short-term trajectory and the latest physical risk value. If the highest probability action does not meet the threshold, the next candidate action that meets the threshold is selected, or a preset safety action such as deceleration or stopping is executed.

[0079] The safety check unit and risk calibration attention are complementary structures: risk calibration attention is used to highlight hazardous entities during policy reasoning, while the safety check unit is used to prevent obviously unsafe actions at the execution level. Whether or not the safety check unit is enabled does not affect the core calibration relationship and attention bias mechanism of this invention.

[0080] The conditions for reaching the target can be that the distance between the robot and the target point is less than a preset threshold; the collision conditions can be determined by the intersection of the robot's outline with the outline of a dynamic entity or a static obstacle; and the timeout conditions can be determined by the number of navigation steps or the duration exceeding a preset upper limit.

[0081] Implementation of the training process Reference Figure 5 During the training phase, the robot interacts and rolls in a simulated or real environment, collecting data on its state, actions, rewards, next state, and termination flags. The collected data is organized into trajectories or experience batches to update the node encoder, relation encoder, calibration network, graph attention module, temporal network, and actor-critic network.

[0082] Task rewards can include positive rewards for reaching the objective. Collision negative reward Reward for the distance difference in progress towards the goal Speed ​​penalty and timeout penalty Physical risk does not need to be added as an additional scalar reward, but rather acts directly on the graph attention through the relation matrix.

[0083]

[0084] Policy updates can employ proximal policy optimization, dominant actor-critic, soft actor-critic, or other reinforcement learning algorithms suitable for discrete or continuous action spaces. Training loss includes policy loss. Value loss Entropy regularization And optional calibration loss .

[0085]

[0086] Among them: These are the weighting coefficients for each type of loss.

[0087] To improve robustness under varying distributions, the number of pedestrians, pedestrian speeds, number of obstacles, channel width, sensor noise, and robot dynamics parameters can be randomized during training. Furthermore, a learning process can be constructed that gradually transitions from the training distribution to high-density or high-constraint scenarios.

[0088] In the early stages of training, the physical risk generation module can be frozen and a larger relation strength can be used to enable the policy to acquire basic collision avoidance capabilities. Subsequently, the fixed bias limit can be gradually reduced and the calibration network can be trained so that the learned relations can supplement the contextual dependencies not covered by the physical model.

[0089] After the model is trained, the network parameters are fixed and deployed to the robot computing platform. Online inference only requires the current environment state, the hidden state of the previous time step, and the physical risk relationship matrix; it does not need access to the training labels or future trajectories.

[0090] Implementation of Online Reasoning and Exception Handling Online inference repeatedly executes state acquisition, graph construction, relation calibration, risk attention aggregation, temporal update, and action output according to the control cycle. The GRU hidden state generated at the current time step is cached and used as input for the next time step; when a new task starts, the robot is reset, or the state is interrupted for a long time, the hidden state is cleared or reinitialized.

[0091] When no dynamic entities are detected, the interaction graph may contain only robot nodes, and the dynamic environment representation for risk perception is generated by the robot nodes themselves. When static obstacle point clouds are unavailable, preset zero vectors and validity identifiers are used to replace obstacle representations.

[0092] When the physical risk relationship matrix is ​​temporarily unavailable, the validity of the physical relationship is input into the calibration network, and the calibration coefficients are biased towards the learning context relationship. When the node feature encoding is abnormal, the physical risk relationship can be preserved and a conservative deceleration based on the risk threshold can be performed.

[0093] To meet real-time requirements, the maximum number of nodes, the number of graph attention layers, and the number of attention heads can be limited, and sparse edge sets, low-precision quantization, or model distillation can be used. The risk matrix, as a structured prior, can provide safety-critical information even with relatively small network sizes.

[0094] Application Scenario Examples In the mall service robot embodiment, the robot node forms an interaction graph with surrounding pedestrian nodes. Pedestrians approaching or about to cross the street have a higher physical risk relationship, and the calibrated graph attention assigns more information weights to these pedestrians; pedestrians moving away in the same direction but related to the crowd structure can still participate in reasoning through learned relationships.

[0095] In the hospital delivery robot embodiment, static obstacle representation is used to describe narrow passages formed by hospital beds, walls, and equipment, while dynamic graph representation is used to describe medical staff, patients, and wheelchairs. Multimodal fusion and GRU work together to utilize passage constraints and historical motion trends to generate smooth navigation actions.

[0096] In this warehouse mobile robot implementation, dynamic entities include workers, forklifts, and other mobile robots. Physical risk relationships can use different safety radii according to entity categories; the relationship encoder learns contextual dependencies arising from vehicle queues, intersection passage, and multi-robot collaboration.

[0097] In the outdoor delivery embodiment, static obstacle states can be derived from a local map, while dynamic entities include pedestrians, bicycles, and vehicles. By adding prediction windows, node type embeddings, and category-related risk scales, the same risk calibration map attention framework can be applied to outdoor scenarios with a wider speed range. Example

[0098] Reference Figure 6 The environment state acquisition module is used to receive the robot state, dynamic entity state, and static obstacle state; the relationship acquisition module is used to generate a physical risk relationship matrix based on physical risks, and generate a learning relationship matrix through a relationship encoder.

[0099] The adaptive calibration module includes a relation difference calculation unit, a calibration coefficient generation unit, and a relation fusion unit. The relation difference calculation unit calculates the differences between the two types of relations on a node-by-node basis; the calibration coefficient generation unit generates calibration coefficients based on the node context and relation information; and the relation fusion unit outputs the calibration relation matrix.

[0100] The risk attention module includes a query projection unit, a key projection unit, a value projection unit, a relation bias unit, a softmax unit, and a message aggregation unit. The module can include a single-header or multi-header structure and can be stacked in multiple layers.

[0101] The multimodal fusion module includes a dynamic environment representation extraction unit, a static obstacle encoding unit, a robot state encoding unit, and a temporal update unit. The policy output module includes a value head, an action policy head, and an optional safety check unit.

[0102] The modules described above can be implemented in software on the same processor, or they can be deployed separately on the perception processor, navigation processor, and safety controller. Modules exchange node states, relationship matrices, environmental representations, and action commands via shared memory, message buses, or network interfaces.

[0103] The present invention also provides a corresponding robot navigation system, including an environment state acquisition module, a relationship acquisition module, an adaptive calibration module, a risk attention module, a multimodal fusion module, and a strategy output module. The specific processing procedures of each module can be referred to the above method embodiments.

[0104] Embodiments of the present invention also provide an electronic device and a computer-readable storage medium. The electronic device includes a processor and a memory, wherein a program stored in the memory, when executed by the processor, implements the methods described in any of the above embodiments.

[0105] The above description is merely a specific embodiment of the present invention. Equivalent substitutions or modifications made by those skilled in the art without departing from the concept of the present invention should all fall within the protection scope of the present invention.

Claims

1. An embodied intelligent navigation method based on a collision risk calibration map attention mechanism, characterized in that, The method includes: The environmental state of the robot, several dynamic entities within the robot's perception range, and static obstacles is acquired. Based on the physical collision risk between the robot and each of the dynamic entities, a physical risk relationship matrix corresponding to the entity index of the current interaction graph is obtained, and the node state in the interaction graph is feature-encoded to obtain node features. For the node pairs in the interaction graph, the learning context relationship is determined based on the corresponding node features to obtain the learning relationship matrix; Based on the node characteristics of the node pair, the physical risk relationship in the physical risk relationship matrix, the learning context relationship in the learning relationship matrix, and the difference in the relationship between the physical risk relationship and the learning context relationship, the calibration coefficient of the node pair is generated. Based on the calibration coefficients, the physical risk relationship and the learning context relationship are fused node-by-node to obtain the calibration relationship matrix; Using the calibration relation matrix as the relation bias for graph attention operation, the node features are aggregated for risk calibration to obtain a dynamic environment representation of risk perception. The dynamic environment representation of risk perception, the static obstacle representation obtained based on the static obstacles, and the robot state representation are fused together, and the robot navigation action is output according to the fusion result.

2. The method according to claim 1, characterized in that, The physical risk relationship is determined based on the relative distance, relative speed, approach direction, closest approach time, closest approach distance, and combined safety radius between the robot and the dynamic entity, and is normalized to a preset relationship range; wherein, the physical risk relationship increases as the relative distance decreases, the approach speed increases, the closest approach time shortens, or the closest approach distance decreases.

3. The method according to claim 1, characterized in that, The relationship difference is the absolute value of the difference between the physical risk relationship and the learning context relationship; the calibration coefficient is obtained by the calibration network mapping the node characteristics of the node pair, the physical risk relationship, the learning context relationship, and the relationship difference.

4. The method according to claim 1, characterized in that, The node-by-node fusion includes: The value obtained by processing the calibration coefficient is multiplied by the physical risk relationship to obtain the first relationship component; Multiplying the calibration coefficient by the learning context relation yields the second relation component; The sum of the first relation component and the second relation component is determined as the calibration relation of the node pair.

5. The method according to claim 1, characterized in that, Multiple attention heads are used to perform the graph attention operation, and each attention head shares or uses the calibration relation matrix separately. After concatenating or averaging the outputs of each attention head, the updated node features are obtained by sequentially passing through residual connections, layer normalization and feedforward networks, and the risk perception dynamic environment representation is extracted from the robot node features of the last layer.

6. The method according to claim 1, characterized in that, The static obstacle state includes a two-dimensional point cloud, a locally occupied grid, a set of boundary line segments, or a depth map; a static obstacle representation is obtained using a one-dimensional convolutional network, a two-dimensional convolutional network, or an ensemble coding network; the risk-aware dynamic environment representation, the static obstacle representation, and the robot state representation are spliced, gated fusion, cross-attention fusion, or conditional modulation, and the fusion result, along with the temporal hidden state of the previous time step, is input into a gated recurrent unit, a long short-term memory network, a temporal convolutional network, or a temporal Transformer.

7. The method according to claim 1, characterized in that, Before executing the robot navigation action, the safety conditions of the candidate navigation action are determined based on the short-term predicted trajectory, the latest physical risk relationship, and the static obstacle clearance corresponding to the candidate navigation action. If the highest priority candidate navigation action does not meet the safety conditions, the next candidate navigation action that meets the safety conditions is selected, or a preset safety action such as deceleration, braking, stopping, or steering to avoid a collision is output.

8. An embodied intelligent navigation system based on a collision risk calibration map attention mechanism, characterized in that, include: The environment status acquisition module is used to acquire the robot status, dynamic entity status, and static obstacle status. The relationship acquisition module is used to generate a physical risk relationship matrix corresponding to the node index of the interaction graph, along with its validity or credibility information, and to acquire the learning relationship matrix. The adaptive calibration module is used to generate calibration coefficients node by node and fuse them to obtain a calibration relationship matrix; The risk attention module is used to perform monotonic mapping on the calibration relation matrix to form relation biases, and to perform graph attention message aggregation based on the relation biases to obtain a dynamic environmental representation of risk perception. The multimodal temporal fusion module is used to fuse risk perception dynamic environment representation, static obstacle representation, robot state representation, and historical temporal state. The strategy output module is used to generate robot navigation actions based on the fusion results.

9. An electronic device, characterized in that, The method includes at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, which, when executed, cause the at least one processor to perform the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 7.

Images